The role of language and pragmatics in conceptual abstraction: Interactive experiments and emergent communication models using reference games
Abstract
Choosing the right word to convey a particular concept is one of the central decisions in communication, as natural languages offer multiple ways to communicate the same object (e.g., dalmatian, dog, animal). Such terms reflect different conceptual perspectives, and they can be situated within a conceptual hierarchy, ranging from subordinate (dalmatian) to superordinate (animal) terms (E. V. Clark, 1997; Rosch et al., 1976). The primary objective of this dissertation is to investigate conceptual abstraction, that is, the abstraction process involved in forming and communicating concepts that lie on different levels of a conceptual hierarchy. To this end, I present four case studies that collectively provide insights into answering the following overarching research question: “What is the role of language and pragmatics in conceptual abstraction?” We employ an interactive, reference-game-based experimental paradigm in case study 1 and a computational modeling approach in case studies 2-4 to address this question. The computational approach models language evolution and emergence through repeated interactions between Artificial Neural Network agents, where we manipulate the communicative need via conceptual and contextual factors. In case study 1, we find that cognitive and pragmatic factors play a role in the communication of concepts and that speakers, but not listeners, benefit from the production of cognitively economical terms. Case study 2 demonstrates that the availability of context during communication shapes an emerging language to the extent that it becomes more efficient and less overinformative. Case study 3 extends on these insights and generates three main research findings. First, context-based pragmatics leads to the emergence of very efficient languages. Second, utility-based pragmatics improves efficiency further, but only if languages evolved under the condition of context-based pragmatics. Third, the structure of categories and linguistic systems can both be explained by efficiency, operationalized as a tradeoff between the speaker’s need for simplicity and the listener’s need for informativeness during communicative interaction. Finally, case study 4 probes which linguistic strategies artificial agents spontaneously use when they need to communicate concepts at a higher or lower level of the conceptual hierarchy than those conventionalized in their languages. This study shows that generalization to lower levels of the conceptual hierarchy is based on a compositional strategy, whereas abstraction to higher levels of the conceptual hierarchy benefits from a meaning extension approach. Based on these findings, I argue that language reflects the structure of conceptual hierarchies due to their shared origin in interaction. Pragmatics shapes these linguistic interactions during language use and during language emergence, thereby influencing conceptual abstraction. Moreover, novel communicative needs can be met by using creative linguistic strategies and utility-based pragmatic reasoning, preserving the flexibility of language. The here presented work has implications for current and future research in Cognitive Science, Linguistics, and Pragmatics by demonstrating that interaction and efficiency critically determine conceptual abstraction, by providing first evidence for a joint evolution of efficient language and category systems through communicative interaction between speakers and listeners, and by offering ways forward for the research on the role of pragmatics in conceptualization and lexical choice.
Full text
!
Disse a ion
an de Uni e si ä Osnab ück
zu E langung des Dok o g ades D . e . na .
Thema
The ole o language and p agma ics in
concep ual abs ac ion: In e ac i e expe imen s
and eme gen communica ion models using
e e ence games
K is ina Kob ock
Die!Disse a ion!wu de!be eu !du ch
P o .!D .!Nicole!Go zne !
P o .!D .!Elia!B uni!
Uni e si ä Osnab ück
Fachbe eich Humanwissenscha en
Kogni ionswissenscha
Osnab ück 2025
blue
Agen s a e embedded in an en i onmen cons uc ed by he esea che . I is
his en i onmen ha de e mines he social in e ac ions o agen s and also
wha he agen s will communica e abou .
— Ch is iansen & Ki by (2003, p. 304)
This documen was ypese wi h he help o KOMASc ip and LaTeX using
he kaobook class.
The ole o language and p agma ics
in concep ual abs ac ion
In e ac i e expe imen s and eme gen communica ion models
using e e ence games
K is ina Kob ock
Acknowledgemen s
In e ac ion is no only one o he co e opics in his disse a ion, bu i also played a signi ican ole
in i s de elopmen p ocess. Du ing he ime I was wo king on his disse a ion, I had he pleasu e
o in e ac ing wi h many g ea scien is s, amazing pe sonali ies, and wonde ul colleagues.
Fi s and o emos , I would like o wholehea edly hank my supe iso s, Nicole Go zne and Elia
B uni, o emba king on his jou ney wi h me and o hei con inuous suppo o e he pas h ee
yea s. I am g a e ul o you o being open o and encou aging he esea ch p ojec , o helping me
g ow pe sonally and academically, and o always asking me he igh challenging ques ions and
gi ing me c i ical eedback when I needed i . You suppo has been in aluable o me on my way o
becoming an independen esea che !
Nicole, you ha e in eg a ed me in o he SPA lab and shown me how suppo i e and welcoming
academia can be. No only a e you an excep ional academic supe iso , bu I am also e y g a e ul o
you o belie ing in me, encou aging me, and celeb a ing wi h me. Thank you o gi ing me secu i y
amid he p eca i y o academic posi ions and o e ing me he oppo uni y o con inue ou joy ul
and ui ul collabo a ion! Many hanks also o all membe s o he SPA lab who ha e accompanied
me, cu en and o me : Cha lo e Uhlemann, Elli Tou ou i, Mo wenna Hoeks, Radim Lacina, Ka ja
Ruge, Julia Reu e , and S a oula Alexand opoulou. Especially o Cha lo e, who accompanied me
h ough he challenges o de eloping and implemen ing a no el expe imen al pa adigm, who was
he i s s uden I e e supe ised, and who has become my iend, o ice ma e, co-au ho , and oad
ip buddy.
Special hanks also belong o Xenia Ohme , who was like an academic big sis e o me. Thank you
o aking he ime o ou egula mee ings and discussions. I lea ned a lo om you, and I could
always coun on you hough ul ad ice, imely eedback, keen eye o small mis akes, and c ea i e
ideas!
I would also like o exp ess my g a i ude o he people a he Ins i u e o Cogni i e Science who
con ibu e o his being such an ideal place o his academic jou ney. In pa icula , I would like o
hank all he membe s o he Compu a ional Cogni ion esea ch aining g oup o he s imula ing
scien i ic discussions, he help ul ca ee ad ice, and o making me eel ha I am no alone wi h my
doub s and p oblems. Ano he g ea sou ce o inspi a ion and joy du ing his jou ney has been
wo king wi h he amazing Coxi s uden s. Thank you, Il a Ho emann, Anna B io o, Eosand a
G und, Vanessa Ve dugo, Muhip Tezcan, Vanessa Obi, Felix Jap ok, Isabella del Pozo, and Ma iia
Gudko a, o suppo ing his esea ch as a s uden assis an o while w i ing you hesis wi h me. I
could always lea n some hing new while wo king wi h you, and I wish you all he bes o you
u u e! My own jou ney as a Coxi was made special by my dea iends Julian, Johanna, Jana, Me i,
Daniel, and Kla a (no all o whom a e Coxis, bu hey all we e o a leas one day).
The e a e many mo e people wi h whom I ha e enjoyed deligh ul imes a con e ences, elaxing
and s imula ing lunch and co ee b eaks, and un DnD e enings a he ins i u e. Thank you all e y
much!
Whene e I needed a b eak, a hug, o a lis ening ea , my amily and iends we e he e o me.
I wan o hank my pa en s and my sis e . Fo as long as I can emembe , you ha e looked a e
me, ad ised and mo i a ed me, suppo ed me, and lo ed me. Thank you o my pa ne Vincen ,
who dis ac ed me om wo k when I couldn’ s op, helped me elax when I was s essed, and o
you uncondi ional lo e and suppo . Anne, hank you o being my bes iend and my cons an
con e sa ion pa ne since i s g ade.
Las bu no leas , I owe a huge hank you o my p e-PhD collabo a o s and men o s: Timo Rö ge ,
Michael F anke, I is an Rooij, Go don Pipa, and Pe e König. Wi hou you, none o his would
ha e been possible.
Abs ac
Choosing he igh wo d o con ey a pa icula concep is one o he cen al decisions in communi-
ca ion, as na u al languages o e mul iple ways o communica e he same objec (e.g., dalma ian,
dog,animal). Such e ms e lec di e en concep ual pe spec i es, and hey can be si ua ed wi hin a
concep ual hie a chy, anging om subo dina e (dalma ian) o supe o dina e (animal) e ms (E. V.
Cla k, 1997; Rosch e al., 1976). The p ima y objec i e o his disse a ion is o in es iga e concep ual
abs ac ion, ha is, he abs ac ion p ocess in ol ed in o ming and communica ing concep s ha
lie on di e en le els o a concep ual hie a chy.
To his end, I p esen ou case s udies ha collec i ely p o ide insigh s in o answe ing he ollow-
ing o e a ching esea ch ques ion: “Wha is he ole o language and p agma ics in concep ual
abs ac ion?” We employ an in e ac i e, e e ence-game-based expe imen al pa adigm in case
s udy 1 and a compu a ional modeling app oach in case s udies 2-4 o add ess his ques ion. The
compu a ional app oach models language e olu ion and eme gence h ough epea ed in e ac ions
be ween A i icial Neu al Ne wo k agen s, whe e we manipula e he communica i e need ia
concep ual and con ex ual ac o s. In case s udy 1, we ind ha cogni i e and p agma ic ac o s
play a ole in he communica ion o concep s and ha speake s, bu no lis ene s, bene i om he
p oduc ion o cogni i ely economical e ms. Case s udy 2 demons a es ha he a ailabili y o
con ex du ing communica ion shapes an eme ging language o he ex en ha i becomes mo e
e icien and less o e in o ma i e. Case s udy 3 ex ends on hese insigh s and gene a es h ee
main esea ch indings. Fi s , con ex -based p agma ics leads o he eme gence o e y e icien
languages. Second, u ili y-based p agma ics imp o es e iciency u he , bu only i languages
e ol ed unde he condi ion o con ex -based p agma ics. Thi d, he s uc u e o ca ego ies and
linguis ic sys ems can bo h be explained by e iciency, ope a ionalized as a adeo be ween he
speake ’s need o simplici y and he lis ene ’s need o in o ma i eness du ing communica i e
in e ac ion. Finally, case s udy 4 p obes which linguis ic s a egies a i icial agen s spon aneously
use when hey need o communica e concep s a a highe o lowe le el o he concep ual hie a chy
han hose con en ionalized in hei languages. This s udy shows ha gene aliza ion o lowe le els
o he concep ual hie a chy is based on a composi ional s a egy, whe eas abs ac ion o highe
le els o he concep ual hie a chy bene i s om a meaning ex ension app oach.
Based on hese indings, I a gue ha language e lec s he s uc u e o concep ual hie a chies due o
hei sha ed o igin in in e ac ion. P agma ics shapes hese linguis ic in e ac ions du ing language
use and du ing language eme gence, he eby in luencing concep ual abs ac ion. Mo eo e , no el
communica i e needs can be me by using c ea i e linguis ic s a egies and u ili y-based p agma ic
easoning, p ese ing he lexibili y o language. The he e p esen ed wo k has implica ions o
cu en and u u e esea ch in Cogni i e Science, Linguis ics, and P agma ics by demons a ing ha
in e ac ion and e iciency c i ically de e mine concep ual abs ac ion, by p o iding i s e idence
o a join e olu ion o e icien language and ca ego y sys ems h ough communica i e in e ac ion
be ween speake s and lis ene s, and by o e ing ways o wa d o he esea ch on he ole o
p agma ics in concep ualiza ion and lexical choice.
Zusammen assung
Die Wahl des passenden Wo es zu Übe mi lung eines bes imm en Konzep s s ell einen de
zen alen En scheidungsp ozesse in de Kommunika ion da , denn na ü liche Sp achen bie en
meh e e Möglichkei en, dasselbe Objek zu kommunizie en (z. B. Dalma ine ,Hund,Tie ). Die
Beg i s a ian en spiegeln un e schiedliche konzep uelle Pe spek i en wide und können inne halb
eine Beg i shie a chie angesiedel we den, die on un e geo dne en Beg i en (Dalma ine ) zu
übe geo dne en Beg i en (Tie ) eich (E. V. Cla k, 1997; Rosch e al., 1976). Das Haup ziel diese
Disse a ion is es, die konzep uelle Abs ak ion, d. h. den Abs ak ionsp ozess, de bei de Bildung
und Kommunika ion on Konzep en au e schiedenen Ebenen eine Beg i shie a chie s a inde ,
zu un e suchen.
Zu diesem Zweck geben ie Falls udien gemeinsam Einblicke in die Bean wo ung de ol-
genden übe geo dne en Fo schungs age: „Welche Rolle spielen Sp ache und P agma ik in de
konzep uellen Abs ak ion?“ Zu Bean wo ung diese F age inden ein in e ak i es, au Re e en-
zspielen basie endes Expe imen in Falls udie 1 und ein compu e ges ü z e Modellie ungsansa z
in den Falls udien 2-4 Anwendung. De compu e ges ü z e Ansa z modellie Sp achen wicklung
und -en s ehung du ch wiede hol e In e ak ionen zwischen als küns liche neu onale Ne ze imple-
men ie en Agen en, wobei die Kommunika ionsbedü nisse übe konzep ionelle und kon ex uelle
Fak o en manipulie we den. In Falls udie 1 s ell sich die E kenn nis he aus, dass kogni i e
und p agma ische Fak o en bei de Kommunika ion on Konzep en eine Rolle spielen und dass
Sp eche *innen da on p o i ie en, kogni i ökonomische Beg i e zu wählen, jedoch nich ih e
Zuhö e *innen. Falls udie 2 zeig , dass die Ve ügba kei on Kon ex wäh end de Kommunika ion
eine en s ehende Sp ache insowei p äg , dass sie e izien e und wenige übe in o ma i ges al e
wi d. Falls udie 3 bau au diesen E kenn nissen au und lie e d ei wesen liche Fo schungs esul a e.
E s ens üh kon ex basie e P agma ik zu He ausbildung seh e izien e Sp achen. Zwei ens
e besse nü zlichkei sbasie e P agma ik die E izienz wei e , jedoch nu , wenn sich Sp achen
un e den Bedingungen kon ex basie e P agma ik en wickel haben. D i ens lassen sich sowohl
die S uk u on Ka ego ien als auch die S uk u on Sp achsys emen du ch E izienz e klä en, die
als Komp omiss zwischen dem Bedü nis de Sp eche *innen nach Ein achhei und dem Bedü nis
de Zuhö e *innen nach In o ma i i ä wäh end de kommunika i en In e ak ion ope a ional-
isie we den kann. Schließlich e mi el Falls udie 4, welche sp achlichen S a egien küns liche
Agen en spon an anwenden, wenn sie Konzep e au eine höhe en ode nied ige en Ebene de
Beg i shie a chie kommunizie en müssen als diejenigen, die in ih en Sp achen kon en ionalisie
sind. Diese S udie läss die Schluss olge ung zu, dass die Ve allgemeine ung au nied ige e Ebenen
de Beg i shie a chie au eine komposi ionalen S a egie basie , wäh end die Abs ak ion au
höhe e Ebenen de Beg i shie a chie on einem Ansa z de Bedeu ungse wei e ung p o i ie .
Diese E kenn nisse üh en zu de A gumen a ion, dass sich die S uk u on Beg i shie a chien
in de Sp ache inde , da diese ih en gemeinsamen U sp ung in de In e ak ion haben und dem
E izienzd uck un e liegen, de sich aus dem Bedü nis de Sp eche *innen nach ein ache Sp ache
und dem Bedü nis de Zuhö e *innen nach in o ma i en Bo scha en e gib . Die P agma ik p äg
6.9
Con ex -awa e: Mean en opy sco es ac oss all da ase s o di e en con ex condi ions
indica ed by he numbe o sha ed a ibu es ........................ 63
6.10
Con ex -unawa e: Mos e o s occu on he diagonal om op le o bo om igh , i.e. in
he ines possible con ex condi ions ............................. 67
6.11
Con ex -awa e: Mos e o s occu on he diagonal om op le o bo om igh , i.e. in he
ines possible con ex condi ions ............................... 68
6.12
Con ex -unawa e: Mos alse nega i e e o s occu in he op le , i.e. in condi ions whe e
gene ic concep s need o be disc imina ed in ine con ex s ................. 69
6.13
Con ex -awa e: Mos alse nega i e e o s occu in he op le , i.e. in condi ions whe e
gene ic concep s need o be disc imina ed in ine con ex s ................. 69
6.14
Con ex -unawa e: Mos alse posi i e e o s occu in he bo om igh , i.e. in condi ions
whe e speci ic concep s ha e o be disc imina ed in a ine con ex . . . . . . . . . . . . 70
6.15
Con ex -awa e: Mos alse posi i e e o s occu in he bo om igh , i.e. in condi ions
whe e speci ic concep s ha e o be disc imina ed in a ine con ex . . . . . . . . . . . . 71
6.16
Con ex -unawa e: Mean NMI sco es ac oss all da ase s o di e en concep (# ixed
a ibu es) and con ex condi ions (# sha ed a ibu es) .................. 73
6.17
Con ex -awa e: Mean NMI sco es ac oss all da ase s o di e en concep (# ixed
a ibu es) and con ex condi ions (# sha ed a ibu es) .................. 73
7.1 Model a chi ec u e ....................................... 83
7.2 Game scena ios in Expe imen 1 ............................... 86
7.3 Game scena ios added in Expe imen 2 ........................... 87
7.4 Message leng hs pe concep hie a chy le el ........................ 90
7.5 NMI, consis ency and e ec i eness sco es o each le el o he concep ual hie a chy . . 91
7.6 Dis ibu ion o message leng hs o di e en le els o he concep ual hie a chy . . . . . 95
7.7 T adeo be ween lexicon size and in o ma i eness ..................... 96
7.8 Zip ’s law like dis ibu ion o message equency plo ed o da ase D(4,4) ....... 97
7.9 Zip ’s law like dis ibu ion o message leng h plo ed o da ase D(4,4) ......... 97
7.10 Da ase examples ........................................ 109
7.11 In o ma ion- heo e ic sco es and ambigui y in language .................. 112
7.12
T aining ajec o y plo s wi h aining and alida ion accu acies o one example da ase
D(4,4) and bo h condi ions. .................................. 117
7.13 Zip ’s law like dis ibu ion o message equency plo ed o each da ase . ....... 119
7.14 Zip ’s law like dis ibu ion o message leng h plo ed o each da ase . . . . . . . . . . 120
8.1
Examples o speake inpu s o aining and es ing in he wo ze o-sho es condi ions
“ o speci ic” and “ o gene ic” ................................. 128
8.2 A chi ec u e ........................................... 129
8.3
To speci ic: Topog aphic simila i y sco es calcula ed on messages om he ain and es
spli s ............................................... 134
8.4
To gene ic: Topog aphic simila i y sco es calcula ed on messages om he ain and es
spli s ............................................... 135
Lis o ables
6.1 Da ase s wi h 𝑛a ibu es and 𝑘 alues, labeled as 𝐷(𝑛, 𝑘).. . . . . . . . . . . . . . . . 57
6.2
Con ex -unawa e: Unique messages used o e e o a andomly picked speci ic concep
in he D(4,4) da ase o e di e en con ex condi ions ................... 58
6.3
Con ex -awa e: Unique messages used o e e o a andomly picked speci ic concep in
he D(4,4) da ase o e di e en con ex condi ions ..................... 59
6.4
Accu acy means o agen s ained in he con ex -unawa e and con ex -awa e se ing
a e aged o e i e uns wi h s anda d de ia ions. ..................... 66
6.5
Con ex -unawa e: Unique messages used o e e o a andomly picked speci ic concep
in he D(3,4) da ase o e di e en con ex condi ions ................... 71
6.6
Con ex -awa e: Unique messages used o e e o a andomly picked speci ic concep in
he D(3,4) da ase o e di e en con ex condi ions ..................... 71
6.7
Con ex -unawa e: Unique messages used o e e o a andomly picked speci ic concep
in he D(5,4) da ase o e di e en con ex condi ions ................... 72
6.8
Con ex -awa e: Unique messages used o e e o a andomly picked speci ic concep in
he D(5,4) da ase o e di e en con ex condi ions ..................... 72
7.1 Summa y o p edic ions and measu es ........................... 88
7.2 Mean accu acies on he aining, alida ion and es da ase s ............... 89
7.3 Mean en opy-based sco es, i.e. NMI, e ec i eness and consis ency ........... 90
7.4 Accu acies on he es da ase ................................. 93
7.5 Lexicon sizes .......................................... 95
7.6 Quali a i e examples ...................................... 99
7.7 S a is ical models i ed wi h b m o Expe imen 1 ..................... 114
7.8 S a is ical models i ed wi h b m o Expe imen 2 ..................... 115
7.9
Pos e io summa ies o he Bayesian hie a chical models p edic ing he en opy-based
sco es, i.e. NMI, e ec i eness and consis ency ....................... 116
7.10 Mean numbe o aining epochs ............................... 117
8.1 Ze o-sho es accu acies o bo h condi ions. ........................ 131
8.2 NMI sco es o bo h condi ions. ................................ 131
8.3
The a io o messages and concep s o he es da a spli
𝑀𝑡𝑒𝑠𝑡/𝐶𝑡𝑒𝑠𝑡
. Pe cen age o
eused and no el messages om he o al se o unique messages 𝑀𝑡𝑒𝑠𝑡.......... 134
8.4 To speci ic: T aining and alida ion accu acies ....................... 139
8.5 To gene ic: T aining and alida ion accu acies ....................... 139
8.6 To speci ic: Numbe o concep s and messages o he es spli and hei a io . . . . . 140
8.7 To gene ic: Numbe o concep s and messages o he es spli and hei a io . . . . . 140
8.8
To speci ic: One andom example o a speci ic concep om he es da a pe da ase , he
con ex condi ion in which i was p esen ed (in numbe o sha ed a ibu es) and he
message ha was used o e e o he concep ........................ 141
8.9
To gene ic: One andom example o a gene ic concep om he es da a pe da ase , he
con ex condi ion in which i was p esen ed (in numbe o sha ed a ibu es) and he
message ha was used o e e o he concep ........................ 141
8.10 To speci ic: Numbe o unique concep s in each da ase o each da ase spli . . . . . . 142
8.11 To gene ic: Numbe o unique concep s in each da ase o each da ase spli . . . . . . 143
Lis o abb e ia ions and e ms
AI A i icial In elligence 23,25,28
ANN A i icial Neu al Ne wo k 22,23,28
Backp opaga ion 22–25
BOSS Bank o S anda dized S imuli 7,9,34
DL Deep Lea ning 22–25
EGG Eme gence o lanGuage in Games 27
EOS end-o -sequence 26
FF-NN Feed- o wa d Neu al Ne wo k 22,25,27
G adien 23
GRU Ga ed Recu en Uni 23,26,55,168
Gumbel-So max 23,24
IL I e a ed Lea ning 20,21
LLM La ge Language Model 23,27,28,152,159,160,168
LSTM Long Sho -Te m Memo y 23,55
MARL Mul i-agen Rein o cemen Lea ning 22–24
ML Machine Lea ning 22,23,32,152
RE e e ing exp ession xi,35,38–41,44–46
Rein o ce 23
RL Rein o cemen Lea ning 22–24,27
RNN Recu en Neu al Ne wo k 23,25
RSA Ra ional Speech Ac s 5,18–20,30–32,78,121,122,167
SL Supe ised Lea ning 23
T ans o me 23,55,168
1 Mo i a ion and ou line
One ascina ing opic linking Cogni i e Science and Linguis ics lies in he in e play be ween
language and cogni ion. The e a e wo esea ch di ec ions: one in es iga es how language in luences
cogni ion, and he o he in es iga es how cogni ion in luences language. On he one hand, we can
ask: “How does language in luence cogni i e p ocesses such as pe cep ion and ca ego iza ion, bu
also highe -le el cogni i e p ocesses such as easoning o p oblem sol ing?” This ques ion implies
ha language is no me ely a ool o communica ion bu ha i also has a ole in cogni ion. On he
o he hand, we can ask: “How is language shaped by he cogni i e sys em and he en i onmen ?”
This ques ion a ge s language e olu ion and he p essu es and cons ain s ha in luence an
eme ging language.
1.1 Language and cogni ion
How language in luences pe cep ion and ca ego iza ion is a key opic in he a ea o esea ch
ha in es iga es he in e play be ween language and cogni ion. This kind o esea ch is linked
o he Sapi -Who hypo hesis and s a ed ou as a iew o linguis ic de e minism (Sapi , 1912;
Who , 1956). Who (1956), acing he ichness o linguis ic di e si y, p oposed ha language
de e mined pe cep ion, ca ego iza ion, and, ul ima ely, ac ions. This s ong iew o he Who ian
hypo hesis was abandoned in he ield (Bo odi sky, 2001). Howe e , he weake iew ha he e is
some in luence o language on pe cep ion is, again, a i id opic o esea ch unde he b anding
‘linguis ic ela i i y.’ Recen s udies in es iga e how language impac s pe cep ion and ca ego iza ion,
o example, in he domains o colo (Regie & Kay, 2009; Winawe e al., 2007), space (Majid
e al., 2004), ime (Bo odi sky, 2001; Bo odi sky e al., 2011), o kinship ca ego ies (Kemp & Regie ,
2012). Fo ins ance, Winawe e al. (2007) es ed English and Russian speake s in a speeded colo
disc imina ion ask in which pa icipan s had o decide which o wo colo chips ma ched a a ge
colo . C ucially, English and Russian di e in e ms o how hese languages di ide he colo
spec um. While English speake s do no disc imina e be ween ligh e blues and da ke blues, he
Russian language makes his dis inc ion compulso y because he e exis s a colo e m o ligh e
blues (“goluboy”) and one o da ke blues (“siniy”). The expe imen al indings e ealed a ca ego y
ad an age o speake s o Russian bu no English. Russian speake s we e as e a ca ego izing
colo s when he wo blue colo chips ell in o di e en linguis ic ca ego ies han when hey we e
om he same linguis ic ca ego y. This was aken as e idence ha linguis ic ep esen a ions play
a ole in asks ha equi e p esumably objec i e pe cep ual decisions (Winawe e al., 2007). The
e ec was also eplica ed wi h G eek and Ge man speake s and shows obus elec ophysiological
signa u es in he b ain (Maie & Abdel Rahman, 2018). While i is a om being uncon o e sial,
he claim ha language in luences pe cep ion has gained empi ical and heo e ical suppo in ecen
yea s (Fo de & Lupyan, 2019; Lupyan, 2012; Lupyan e al., 2020).
2 1 Mo i a ion and ou line
Mo ing om low-le el o high-le el cogni ion, his disse a ion aims o p oceed om he ole
o language in pe cep ion and ca ego iza ion owa ds he ole o language in abs ac ion and
gene aliza ion. Abs ac ion is one o he p ime highe -le el cogni i e abili ies. I is in ol ed in
easoning, p oblem sol ing, complex hough s, and i has close ies o language. While abs ac ion
lies a he co e o Cogni i e Science (Tenenbaum e al., 2011), su p isingly li le empi ical esea ch
has been dedica ed o abs ac ion un il ecen ly. One eason o his migh be he lack o an
in eg a i e de ini ion ac oss ields (Bu goon e al., 2013). In he ollowing, I d aw on de ini ions and
cha ac e iza ions p o ided by Yee (2019) and Bu goon e al. (2013) o elucida e how abs ac ion can
be de ined, and wha makes abs ac ion special o cogni ion.
While he no ion o abs ac ion has been used in di e en manne s in he li e a u e, one aspec
appea s o be cen al o all iews (Bu goon e al., 2013). Abs ac ion is he p ocess by which
we agg ega e indi idual expe iences wi h objec s, ac ions, e en s, and ideas.
1
We abs ac o e
indi idual (senso y o mo o ) expe iences and highligh wha is common o all expe iences ela ed
o one objec . By way o abs ac ion, we can make ou wha di e en objec s ha e in common, and
we can ca ego ize hem oge he in o concep s. To build concep ual knowledge in his way is one o
he co e unc ions o abs ac ion (Bu goon e al., 2013; Yee, 2019). Concep s, in u n, a e bene icial o
many cogni i e p ocesses. Fi s , hey help o o ganize and s uc u e he pe cei ed en i onmen (see
also Ba salou e al., 2018). Second, hey can help o educe cogni i e e o . When so ing simila
hings in o he same ca ego y, he cogni i e sys em can deal wi h he abs ac ion o concep , a he
han wi h each indi idual ins ance. Thi d, concep s suppo gene aliza ion (Bu goon e al., 2013;
Yee, 2019). Fo example, hink o an agen who encoun e s a new objec . Once he agen ecognizes
ha his objec belongs o a known ca ego y, he agen can o m a hypo hesis on how o in e ac
wi h he new objec . This means ha abs ac ion is he p ocess by which we build concep s, which
in u n enable highe -le el abs ac ion and gene aliza ion.
The iew ha concep s a e essen ially abs ac ed knowledge also has implica ions o ca ego y
s uc u e. Fo example, we can hink o he concep u ni u e o be mo e abs ac ed han he concep
chai , which is in u n mo e abs ac ed han he concep o ice chai . The cogni i e implica ions
o hese hie a chical, axonomic ela ionships be ween concep s we e in es iga ed in ea ly wo k
in psychology by Rosch (1978) and Rosch e al. (1976). We can also hink o he (abs ac ) concep s
o eedom, iendship, o emo ion as being mo e abs ac ed han he p e ious (conc e e concep )
examples. This is because hey lack g ounding in senso y-mo o expe ience. These di e en kinds o
abs ac ed knowledge can be cha ac e ized by he dimensions o speci ici y
2
, ha is, he ela ionship
be ween u ni u e,chai , and o ice chai , and conc e eness
3
, ha is, he quali a i e di e ence
be ween conc e e concep s like chai and abs ac concep s like eedom (Bolognesi e al., 2020).
Recen ly, esea che s began o belie e ha abs ac ness e lec s a con inuous dimension a he
han a dicho omy be ween conc e e and abs ac concep s, a iew ha has domina ed esea ch on
1
In he ollowing, I will use he e m objec only, bu he claims also hold o o he ypes o concep s, such as ac ions,
e en s, and ideas.
2
The dimension o speci ici y ela es o how speci ic o gene ic a concep is. Subo dina e concep s a e mo e speci ic and
supe o dina e concep s a e mo e gene ic.
3
The dimension o conc e eness, o abs ac ness, ela es o how conc e e o abs ac , in he sense o no di ec ly
pe cei able, concep s a e.
1 Mo i a ion and ou line 3
concep s in he p e ious decades (see also Banks e al., 2023; Ba salou e al., 2018; Bu goon e al.,
2013; Yee, 2019). Th ough he lens o he concep ual hie a chy o mo e and less abs ac ed concep s,
we can in es iga e he abs ac ion p ocess in ol ed. This p ocess o concep ual abs ac ion is he key
opic o his disse a ion.
Abs ac ion and language ha e a p oduc i e ela ionship. On he one hand, language is inhe en ly
abs ac (Lupyan & Win e , 2018). In Linguis ics, language is s udied on di e en le els o abs ac ion,
anging om phonology, o e mo phology, up o syn ax, p agma ics, and discou se. On he o he
hand, language and abs ac ion, especially in he sense o concep ual knowledge, a e in e wined.
Language is hough o play a ole in o ming concep s in a leas h ee di e en ways. Fi s ,
he e is empi ical e idence ha language is use ul o ca ego izing known concep s (e.g., Lupyan,
2005; Rissman & Lupyan, 2024). One example o his is he Russian blues s udy, which se ed as
a mo i a ing example a he beginning o his chap e (Winawe e al., 2007). Second, e idence
sugges s ha language plays a ole in o ming new concep s. Speci ically, i was demons a ed
ha language helps in lea ning (and eaching) new concep s (e.g., A unachalam & Waxman, 2010;
Callanan, 1985; Nazzi & Gopnik, 2001; Slou sky & Deng, 2019; Sume s e al., 2023). In cogni i e
de elopmen , he acquisi ion o language and concep s goes hand in hand (A unachalam & Waxman,
2010). No only does he lexicaliza ion o concep s help o collec in o ma ion abou a concep unde
one label, bu language is also a sou ce o acqui ing in o ma ion abou concep s in he i s place,
o example, in eading o e bal eaching (Slou sky & Deng, 2019; Waxman & Ma kow, 1995). The
ole o language seems o be la ge he mo e abs ac ed ( ha is, gene ic and/o abs ac ) a concep
is (e.g., Doble e al., 2024; Gen ne & Bo odi sky, 2001; M. Lewis e al., 2021; Yee, 2019). Fo example,
child en lea n supe o dina e concep s la e han basic-le el concep s, and hey equi e language
mo e o lea n hem (M. S. Ho on & Ma kman, 1980; Me is & C isa i, 1982). This is ela ed o he
g owing consensus ha language is needed o explain he g ounding o abs ac concep s, which
is belie ed o happen mainly ia linguis ic and social expe ience in in e ac ion (e.g., Banks e al.,
2023; Bo ghi & Binko ski, 2014; Bo ghi e al., 2017). Thi d, language was ound o suppo success ul
gene aliza ion. Fo example, Su ill e al. (2022) ela ed child en’s knowledge o supe o dina e wo ds
o hei induc i e easoning abili ies. Taken oge he , hese indings poin o a ole o language, and
mo e speci ically, in e ac i e communica ion, in concep ual abs ac ion.
Gi en ha p agma ics plays a ole when conside ing communica i e in e ac ions, I addi ionally
conside he ole o p agma ics in concep ual abs ac ion. P agma ics is he s udy o meaning
in con ex ha goes beyond wha was ac ually said, ha is, wha is seman ically encoded in an
u e ance. In in e ac ion, speake s ypically do no use he mos speci ic desc ip ions possible, bu
a he ely on he con ex o he communica i e in e ac ion and on he lis ene ’s abili y o in e
he in ended meaning. P agma ic mechanisms play a ole in communica ing concep s on di e en
ime scales, anging om language use o language acquisi ion and language e olu ion. On a
sho ime scale, p agma ics plays a ole in o mula ing and in e p e ing u e ances in conc e e
communica i e si ua ions and con ex s du ing language use (e.g., G a e al., 2016). On a la ge
ime scale, p agma ics is hough o help child en acqui e wo ds and concep s in de elopmen (e.g.,
Bohn & F ank, 2019). And on an e en la ge ime scale, p agma ics shapes an eme ging language
10 2 Theo e ical and empi ical backg ound
simila i y) and hey sha e his p ope y wi h basic-le el ca ego ies (Mu phy & Wisniewski, 1989;
Rosch e al., 1976). Despi e he impo ance o supe o dina e e ms ha ing some imes been o e looked
in p io esea ch (G eene & Rohan, 2025), unde s anding he comp ehension and p oduc ion o
supe o dina e e ms is essen ial o unde s anding concep ual abs ac ion. Con a y o G eene
and Rohan (2025), I ake he ac ha supe o dina e e ms a e acqui ed las by child en (Me is &
C isa i, 1982) no o mean ha hey a e unimpo an , bu ins ead ha hey a e challenging o acqui e.
Mo eo e , empi ical e idence ques ions he p imacy o he basic-le el ad an age (Macé e al., 2009;
Mu phy & Wisniewski, 1989; Ponce & Fab e-Tho pe, 2014). Fo example, i was shown ha when
ca ego izing objec s in isual scenes, he basic-le el e ec dec eases (Mu phy & Wisniewski, 1989),
o is e en e e sed, leading o a supe o dina e ad an age (Macé e al., 2009; Ponce & Fab e-Tho pe,
2014). One explana ion o hese indings is ha he basic-le el ad an age is a lexical-seman ic e ec
and ha he pa hway o pu e isual ca ego iza ion is om coa se o ine, a o ing he p ocessing o
supe o dina es (Macé e al., 2009; Ponce & Fab e-Tho pe, 2014). Howe e , also in a ask close o
he o iginal objec ca ego iza ion ask p esen ed by Rosch e al. (1976) ha ea u es lexical-seman ic
in o ma ion in he o m o a label cue, he basic-le el ad an age dec eases i supe o dina es a e
ca ego ized in scenes (Mu phy & Wisniewski, 1989). This sugges s ha he ad an age o basic-le el
ca ego ies o e supe o dina e ca ego ies in speed and accu acy dec eases when supe o dina e
ca ego ies a e p esen ed in hei na u al habi a , such as when e e ing o single objec s in a scene
o collec ions o objec s (Wisniewski & Mu phy, 1989, see also). These esul s demons a e ha he
ad an age o he basic o e he supe o dina e le el migh no be as s aigh o wa d as sugges ed in
he li e a u e and ha i migh be wo h looking in o supe o dina e concep s mo e.
2.2 Labeling concep s in e e en ial communica ion
The s udies in es iga ing concep ual abs ac ion and he basic-le el e ec in oduced in he p e ious
sec ion ypically in es iga e he unde lying concep s ia e e ence. Fo example, in naming s udies,
pa icipan s a e asked o name objec s depic ed in images (Rosch e al., 1976). In ca ego iza ion asks,
pa icipan s hea o ead a label and hen judge whe he an objec is an ins ance o he ca ego y ha
he label e e s o. While ca ego iza ion has also been in es iga ed, o example, wi h a so ing ask,
whe e pa icipan s so objec s in o ca ego ies wi hou an explici label (Rosch e al., 1976), much o
he esea ch on concep ual abs ac ion and ca ego iza ion uses linguis ic labels o in es iga e he
men al and psychological cons uc s. Especially in on ogeny, i becomes clea ha language and
concep ual de elopmen a e in ica ely linked. Child en lea n wo ds and concep s simul aneously
and in a ecip ocal ela ionship whe e lea ning one acili a es lea ning he o he (A unachalam &
Waxman, 2010). Wi h he goal o his disse a ion in mind, o in es iga e concep ual abs ac ion,
I will examine he ex ensi e li e a u e on e e en ial communica ion, which essen ially in ol es
labeling concep s.
2 Theo e ical and empi ical backg ound 11
2.2.1 S udying e e en ial communica ion in in e ac ion
The s udy o e e en ial communica ion in Linguis ics and p agma ics is based on wo main
assump ions ha we e in oduced in he 1970s and 1980s. Fi s , G ice (1975) in oduced he idea
ha in e locu o s in a con e sa ion should be coope a i e. Speake s a e hypo hesized o adhe e
o se e al maxims o con e sa ion, which mus be sa is ied o communica ion o be success ul,
and o lis ene s o unde s and he implica ed meaning in he speake s’ u e ances. The quan i y
maxim s a es ha a speake ’s u e ance should be as in o ma i e as equi ed o he lis ene o
unde s and wha is mean , bu also no mo e in o ma i e. The maxim o quali y u ges speake s
o s ick o he u h and only u e wha hey belie e o be ue. The ela ion maxim is “[b]e
ele an ,” (G ice, 1975, p. 46), which means ha speake s should only say wha hey deem ele an
o he communica i e pu pose. Finally, he maxims o manne s a e ha speake s should be p ecise
and a oid leng hy o ambiguous exp essions i possible (G ice, 1975). Second, H. H. Cla k and
Wilkes-Gibbs (1986) in oduced he idea ha e e en ial communica ion is essen ially collabo a i e.
In e e en ial communica ion, speake s and lis ene s alk abou e e en s in he wo ld, o en using
de ini e e e ing exp essions such as he all dog. The wo pa ne s in a con e sa ion need o
nego ia e he e e en ial p ocess and es ablish a common pe spec i e and mu ual knowledge, which
is called he common g ound (H. H. Cla k & Wilkes-Gibbs, 1986). These p inciples o in e ac i e
communica ion guide he speake ’s p oduc ion and he lis ene ’s in e p e a ion o u e ances.
In mode n expe imen al p agma ics, he o mula ion o hese p inciples and obse a ions has
spa ked esea ch on e e ence esolu ion, ha is, he comp ehension o e e ing exp essions, among
o he opics (No eck & Reboul, 2008). Fo example, a ib an ield o esea ch e ol es a ound he
ques ion o how a sha ed pe spec i e o common g ound is es ablished be ween in e locu o s and
how he (absence o ) such mu ual knowledge in luences communica ion (e.g., B ennan & Cla k,
1996; B own-Schmid & Hanna, 2011; H. H. Cla k e al., 1983; G odne & Sedi y, 2011; Ka sos e al.,
2023; Keysa e al., 2000; Rich e e al., 2020). One cen al deba e in expe imen al p agma ics is
whe he de e mining he speake ’s in ended meaning equi es e o on he lis ene ’s side (No eck &
Reboul, 2008). The e o in a con e sa ion can also lie wi h he speake . Theo ies o audience design
posi ha speake s a e willing o pu in mo e e o in a con e sa ion and ca e hei u e ances o
he lis ene ’s needs (e.g., Fe ei a, 2019; S. O. Yoon & B own-Schmid , 2019). Resea ch in es iga ing
hese communica i e e ec s in in e ac ion can shed ligh on hese di e en accoun s and heo ies.
The mos commonly used pa adigm o s udying e e en ial communica ion in in e ac ion is he
e e ence game
1
. In a e e ence game, a speake and a lis ene in e ac in a isual con ex . The
speake ’s ask is o send a message ha helps he lis ene iden i y a pa icula a ge objec om he
isual con ex . The communica ion is success ul i he lis ene in e p e s he message co ec ly and
selec s he co ec a ge objec . Re e ence games can be used o s udy ei he one-sho e e ing,
whe e a speake has only one chance o o mula e a message, o epea ed in e ac ions, whe e
speake s and lis ene s es ablish a e e en o e mul iple u ns. While he o me is well-sui ed o
in es iga e he p agma ic conside a ions a play du ing in e p e ing and p oducing a speci ic kind
1Mo e in o ma ion on he e e ence game pa adigm will ollow in Sec ion 3.1.
12 2 Theo e ical and empi ical backg ound
o e e ing exp ession (e.g., Degen e al., 2020; Win e s e al., 2018), he la e is pa icula ly use ul
when in es iga ing he p ocess o con en ionaliza ion and en ainmen (e.g., Ba & K onmülle ,
2006; B ennan & Cla k, 1996; H. H. Cla k & Wilkes-Gibbs, 1986; Hawkins e al., 2020). Re e en ial
communica ion p o ides an ideal es bed o s udying he communica i e p inciples o mula ed by
G ice (1975) and H. H. Cla k and Wilkes-Gibbs (1986) in a sys ema ic and con olled way.
2.2.2 Concep ualiza ion: The p agma ics o lexical choice
Being in e es ed in concep ual abs ac ion, in his disse a ion I in es iga e a speci ic kind o
e e en ial communica ion whe e he e e en is a concep on a pa icula le el o abs ac ion.
Re e en ial communica ion, ha is, how speake s and lis ene s communica e abou objec s in he
wo ld, has been in es iga ed abundan ly. Howe e , e en hough he lexical choice o nouns is
c ucial o iden i ying an objec and languages such as English need a noun o o m a comple e
sen ence, esea ch on he choice o nominal e e ing exp essions is ela i ely sca ce (G a e al.,
2016). The li e a u e mos ly ocused on modi ied nominal e e ing exp essions, whe e nouns a e
modi ied wi h, o example, an adjec i e (e.g., Rubio-Fe nandez e al., 2019,2022; Sedi y, 2003,
2005; Tou ou i e al., 2017, bu see G a e al., 2016; Gualdoni e al., 2022 o ecen excep ions).
Following B ennan and Cla k (1996), he label ha a speake chooses o communica e an objec wi h
e lec s which concep he speake in ends o communica e. Fo example, when calling an objec a
loa e , speake s concep ualize he objec as a loa e and no as a shoe, o a piece o clo hing (B ennan
& Cla k, 1996). Rela ed o his, Ba and K onmülle (2006) ha e a gued ha con e sa ion, and
speci ically, e e en ial communica ion, is one o he p ime si es o ca ego y lea ning and ca ego y
use. The au ho s a gue ha he choice o a speci ic label o e ano he ca ies in o ma ion. They call
his he “’labeling as ca ego iza ion’ hypo hesis” (Ba & K onmülle , 2006, p. 5). Fo example, by
choosing he label mammal o e he label whale, a speake could highligh ha whales a e mammals
and sha e a ibu es wi h mammals, such as needing o come up o ai . In his disse a ion,
he ques ions o how speake s concep ualize a concep o communica e i and how lis ene s, in
u n, unde s and i , and which in o ma ion hey gain om he labeling in o ma ion a e o cen al
impo ance.
F om he li e a u e in cogni i e psychology and p agma ics, we know ha he e a e se e al ac o s
a play in he choice o a (communica i ely) app op ia e e e ing exp ession. In he classical
iew ep esen ed by Rosch e al. (1976), how an objec is ca ego ized and labeled depends on
ac o s such as in o ma i eness, lexical a ailabili y (i.e., he basic-le el ad an age), and pe cep ual
salience (B ennan & Cla k, 1996). These ac o s, pa icula ly in o ma i eness, a e well-sui ed o
a G icean p agma ics accoun . When conside ing which label (loa e ,shoe, o clo hing) o choose
o a pa icula objec , hese al e na i es compe e (B ennan & Cla k, 1996). The basic-le el e ec
li e a u e p edic s ha speake s choose a basic-le el e m (shoe) o e labels a he subo dina e (loa e )
o supe o dina e (clo hing) le el. Acco ding o p agma ic heo y, he mos app op ia e e e ing
exp ession depends on he con ex : I ano he ype o shoe is p esen , he objec should be e e ed
o wi h he subo dina e e m loa e . This was also ound in a ecen expe imen on nominal e e ing
2 Theo e ical and empi ical backg ound 13
exp essions conduc ed by G a e al. (2016). They ound ha speake s p e e ed basic-le el e ms
o e all, ha hey used subo dina e e ms p ima ily when ano he objec om he same basic-le el
ca ego y was p esen , and ha speake s la gely a oided p oducing supe o dina e e ms (G a
e al., 2016). Unde which ci cums ances a supe o dina e e m becomes (p agma ically) ele an
and app op ia e in a con e sa ion has long been o e looked. In his disse a ion, I ake inspi a ion
om cogni i e psychology, whe e i has been ound ha supe o dina e e ms a e used o e e o
collec ions o classes o objec s (Wisniewski & Mu phy, 1989), o hone in on his ques ion. This will
allow me o d aw a mo e comple e pic u e o he cogni i e and p agma ic ac o s a play when
choosing a subo dina e, basic-le el, o supe o dina e e e ing exp ession.
While he abo e-desc ibed ac o s do no depend on p e ious e e ence si ua ions and he cou se o
a con e sa ion, in na u al discou se, e e ing exp essions a e nego ia ed o e epea ed e e ences
in a collabo a i e se ing (Ba & K onmülle , 2006; B ennan & Cla k, 1996). Ba and K onmülle
(2006, p. 7) s a e ha “con e sa ional se ings shape speake s’ and lis ene s’ ca ego iza ions.” They
discuss he impo ance o he ac ha con e sa ions a e his o ical, mul imodal, and collabo a i e
o he p ocess o ca ego iza ion. In hei iew, ca ego iza ions a e nego ia ed in communica ion,
and he join goal o he in e locu o s is o es ablish a mu ual pe spec i e (Ba & K onmülle ,
2006). When b inging his iew oge he wi h he idea ha a e e ing exp ession e lec s he
concep ualiza ion o an objec (B ennan & Cla k, 1996; E. V. Cla k, 1997), he goal o a con e sa ion
can be o mula ed as aligning on a sha ed concep ualiza ion. When his goal is achie ed, speake s
and lis ene s achie e communica i e success and mu ual unde s anding. In his disse a ion, I
am in e es ed in how such sha ed concep ualiza ions a e o med du ing in e ac ion. In he case
s udies, we adop he s ance ha ca ego iza ions a e in luenced by con e sa ion and collabo a ion
and in es iga e e e en ial communica ion o concep s a di e en le els o abs ac ion in in e ac i e
expe imen s and mul i-agen models.
2.3 The ole o p agma ics in language and concep ual de elopmen
The iew ha con e sa ion and collabo a ion a e impo an o ca ego iza ion ex ends o language
acquisi ion. Child en acqui e new wo ds and hei meanings du ing con e sa ion and child-di ec ed
speech. In e disciplina y esea ch in cogni i e science p o ides a pa icula ly aluable amewo k o
in es iga e linguis ic and concep ual de elopmen join ly (A unachalam & Waxman, 2010). Resea ch
on ca ego y and wo d lea ning es ablished ha child en d aw on bo h linguis ic and obse a ional
in o ma ion o lea n he meaning o a no el wo d. This equi es child en o, on he one hand, iden i y
he concep being e e ed o, and on he o he hand, iden i y he wo d used o e e o he concep
wi hin a la ge con ex . Finally, child en need o es ablish a mapping be ween he wo d and he
concep (A unachalam & Waxman, 2010; Waxman, 1994). While concep ual de elopmen does no
depend on language de elopmen , he wo can in e ac and ein o ce each o he . In child en’s ypical
de elopmen , concep and language acquisi ion a e in ica ely linked, and wo ds and concep s
a e lea ned simul aneously. Fo example, language may help o di ec he lea ne ’s a en ion o
con as i e ea u es and hus in luence he ca ego ies being lea ned (A unachalam & Waxman,
14 2 Theo e ical and empi ical backg ound
2010; E. V. Cla k, 1997; Waxman & Ma kow, 1995; Welde & G aham, 2006). One p ominen ea u e
o he modeling case s udies p esen ed in his disse a ion is ha we in es iga e he eme gence o
concep s and language join ly in in e ac ion.
Cu en p oposals on language acquisi ion highligh he ole o p agma ics he ein (Bohn & F ank,
2019; E. V. Cla k, 1997; F ank e al., 2009; Papa agou, 2002; Tomasello & Akh a , 1995). Such social-
p agma ic heo ies s and in con as o associa ion-based and cons ain -based (Ma kman, 1994)
heo ies o language acquisi ion (see Tomasello, 2000, o an a gumen con as ing social-p agma ic
heo ies wi h associa ion- and cons ain -based heo ies). Fo example, Bohn and F ank (2019) a gue
ha communica ion and p agma ics a e cen al o language de elopmen in child en as young as
one yea old. A i s glance, his appea s o be a odds wi h empi ical e idence, which shows ha
child en s uggle wi h he p ime es case o expe imen al p agma ics, namely he de i a ion o
scala implica u es (Ho n, 2006; No eck, 2001; No eck & Reboul, 2008). The mos used example o
a scala implica u e is he implica u e on he some/all scale. Fo adul s, i has been eliably shown
ha hey ake a sen ence like in (2.2) o mean ha my niece a e some bu no all o he cookies.
My niece a e some o he cookies. (2.2)
This means ha adul s p agma ically en ich he meaning o some, which is seman ically compa ible
wi h all, o mean some bu no all (e.g., Bo & No eck, 2004; Ho n, 2006; No eck & Reboul, 2008;
Papa agou & Musolino, 2003). The easoning behind his meaning compu a ion is spelled ou as a
combina ion o he easoning abou al e na i es and he G icean quan i y maxim: I a speake could
ha e used he mo e in o ma i e e m all, hey would ha e done so. The e o e, he mo e in o ma i e
all mus be alse in his con ex . This easoning leads o he nega ion o he lexical al e na i e all in
he in e p e a ion o some. This mechanism can be gene alized om scala implica u es whe e lexical
al e na i es lie on he same scale, o example <some,all>, o con e sa ional implica u es whe e he
ele an al e na i es can also be al e na i e e e en s in a con ex (Go zne e al., 2020). Child en, as
opposed o adul s, s uggle o de i e scala implica u es (e.g., No eck, 2001; Papa agou & Musolino,
2003), which has been aken as e idence ha p agma ic de elopmen in child en is g adual and
comes much la e (a ound age 7) han seman ic de elopmen . How can hese esul s be econciled
wi h he iew ha child en a age one al eady use p agma ics o lea n new wo ds and concep s?
Bohn and F ank (2019) e iew e idence o he iew ha p agma ics in he o m o in e ence abou
he speake ’s goals al eady helps e y young child en o acqui e language. They p opose ha e en
in an s al eady unde s and he communica i e unc ion o language. Child en expec speake s o
use wo ds in en ionally, ha is, o achie e a pa icula goal. Fo example, Vouloumanos e al. (2012)
show ha 12-mon h-old in an s al eady a ibu e a communica i e unc ion o language. They le
he in an s wa ch a ideo whe e an ac o ailed o achie e hei goal. Subsequen ly, an obse e
used unknown wo ds s. non-linguis ic sounds (such as coughing). The in an s assumed ha he
unknown wo ds, bu no he non-linguis ic sounds, we e ins uc ions and p edic ed ha he ac o
would achie e hei goal in he subsequen ial (Vouloumanos e al., 2012).
In line wi h his, E. V. Cla k (1997) e iews e idence suppo ing he iew ha one-yea -old child en
use p agma ic cues o lea n new wo ds. They ocus on child en’s lea ning o mul iple wo ds o he
2 Theo e ical and empi ical backg ound 15
same objec , such as dalma ian,dog, and animal (E. V. Cla k, 1997). Lea ning mul iple wo ds ha
lie on he same axonomic hie a chy is undamen al o his disse a ion’s esea ch on concep ual
abs ac ion. E. V. Cla k (1997) p oposes ha child en can lea n mul iple pe spec i es, ha is, many
names o he same objec , om he beginning o language acquisi ion. This iew con adic s a
one-pe spec i e accoun ha assumes ha child en ha e a bias o lea ning only one wo d pe objec .
This ela es o he mu ual exclusi i y bias, a lea ning bias ha leads child en o assume ha a no el
wo d e e s o a no el objec and no o a known objec (Ma kman & Wach el, 1988). In con as o
his, he e idence e iewed by E. V. Cla k (1997) shows ha as long as he wo ds di e sligh ly in
meaning and speake s gi e su icien p agma ic cues, child en ha e no di icul y lea ning mul iple
names o he same objec . E. V. Cla k (1997) concep ualizes mul iple names o he same objec as
mul iple concep ual pe spec i es a speake can ake and communica e. One main conclusion is
ha p agma ic cues a e especially ele an o lea ning hese mul iple pe spec i es. The idea ha
child en need su icien cues o make use o p agma ic easoning is in line wi h ecen e idence o
he ea ly compu a ion o con e sa ional implica u es in young child en. Fo example, Sko dos and
Papa agou (2016) and Go zne e al. (2020) show ha i lexical and con ex ual al e na i es (such as
all o he some-scala -implica u e o al e na i e objec s in a con ex o exhaus i i y implica u es) a e
made accessible and ele an , 4- o 5-yea -old child en a e able o de i e implica u es. In summa y,
concep ual and wo d lea ning a e in ica ely linked, and p agma ics plays a c ucial ole in language
and concep ual de elopmen .
3 Expe imen al and modeling amewo ks
3.1 Communica ion and e e ence games
Re e en ial communica ion has been in es iga ed empi ically wi h a communica ion ask as ea ly as
he 1960s (K auss & Glucksbe g, 1969; K auss & Weinheime , 1964). Based on Wi gens ein (1959)’s
idea ha he meanings o wo ds ha e hei o igins in hei use in language and a e de e mined in
so-called language-games (o ig. Sp achspiele), esea che s ha e de ised an expe imen al pa adigm
o in es iga e e e ence in in e ac ion (K auss & Glucksbe g, 1969; K auss & Weinheime , 1964).
1
A he same ime, he so-called signaling game was in oduced as a game- heo e ic model o how
agen s es ablish communica i e con en ions (D. K. Lewis, 1969).
The Lewis signaling game can be o malized as ollows. The e a e wo playe s, a sende and a
ecei e . The e is a wo ld ha can be in one s a e om he se o possible s a es
𝑆={𝑠1, 𝑠2, ..., 𝑠𝑛}
.
The sende knows he cu en s a e o he wo ld, bu he ecei e does no . The speake ’s ask is
o choose a signal, o message, om a se o possible messages
𝑀={𝑚1, 𝑚2, ..., 𝑚𝑘}
and use i o
signal he wo ld’s s a e o he ecei e . The ecei e ’s ask is o ecei e he signal and choose an
ac ion om a se o possible ac ions
𝐴={𝑎1, 𝑎2, ..., 𝑎𝑖}
based on he signal. The sende hus maps
s a es o messages as a unc ion
𝑓:𝑆→𝑀
, and he ecei e maps messages o ac ions as a unc ion
𝑔:𝑀→𝐴
. The e is exac ly one co ec ac ion
𝑎𝑐
o each s a e
𝑠𝑛
. The join payo , o u ili y, o
sende and ecei e is
1
i he ecei e selec s he co ec ac ion
𝑎𝑐
and
0
i he ecei e does no . O e
epea ed in e ac ions, messages become con en ionalized ia coo dina ion (D. K. Lewis, 1969).
The expe imen pa adigm known as he communica ion game, he di ec o -ma che ask, o he
e e ence game, ollows he same gene al idea (see Figu e 3.1). The speake , o di ec o , knows he
ue s a e o he wo ld and has o communica e i o he lis ene , o ma che , who does no know he
ue s a e o he wo ld (H. H. Cla k & Wilkes-Gibbs, 1986). In H. H. Cla k and Wilkes-Gibbs (1986),
di ec o s and ma che s a e allowed o make as many u ns as hey like o es ablish mu ual knowledge.
Re e ence games also exis in e sions whe e he meaning o a signal is al eady es ablished, and he
use o his signal is in es iga ed in a one-sho e e ence. F anke and Degen (2016) ha e p oposed ha
e e ence games a e ins ances o signaling games whe e he meaning is al eady con en ionalized.
Bu , in cu en esea ch, he e m e e ence game is employed in bo h use cases, ha is, asks ha
in es iga e language use wi h es ablished con en ions (see e.g., Achimo a e al., 2022; G a e al.,
2016), and hose ha in es iga e he con e gence on meanings du ing epea ed in e ac ions (see e.g.,
Boyce & F ank, 2023; Hawkins e al., 2020; Laza idou e al., 2018; Ohme , Duda, & B uni, 2022). In
his disse a ion, e e ence games a e used o s udy bo h language use wi h es ablished meanings
and he con e gence on meanings du ing epea ed in e ac ions.
1
This communica ion game was la e called he di ec o -ma che ask (H. H. Cla k & Wilkes-Gibbs, 1986) and is
nowadays o en called he e e ence game in he con ex o e e en ial communica ion.
18 3 Expe imen al and modeling amewo ks
Speake
Lis ene
“blue”
Figu e 3.1: Visualiza ion o he e e ence game. The speake ’s ask is o communica e a a ge (he e p esen ed in a g een
ame) o a lis ene who has o selec he co ec objec om he isual display.
3.2 Modeling language use wi h he Ra ional Speech Ac s amewo k
In p agma ics, language use can be modeled wi h he Ra ional Speech Ac s (RSA) amewo k.
The RSA modeling amewo k in oduced in F ank and Goodman (2012) and Goodman and
S uhlmülle (2013) can be used o model language use as social cogni ion, ha is, as in e ence o e
he speake ’s in en ions and he lis ene ’s likely in e p e a ion o he speake ’s message. This and
ela ed app oaches can be subsumed unde he umb ella e m p obabilis ic p agma ics (F anke &
Jäge , 2016). They a e based on game heo y, such as he Lewis signaling game in oduced abo e,
and Bayesian cogni i e modeling (Degen, 2023; F ank & Goodman, 2012; F anke, 2013; F anke &
Jäge , 2016; Tenenbaum e al., 2011). Speci ically, he RSA amewo k models p agma ic language
use in in e ac ion, whe e speake s eason abou he lis ene ’s likely in e p e a ion o hei u e ance
in language p oduc ion, and lis ene s eason abou he speake ’s in ended meaning in language
unde s anding. Fo mo e de ails on he o maliza ion o his amewo k, gi en an example, see
Box 1.
Box 1. The Ra ional Speech Ac s modeling pa adigm
Conside he e e ence si ua ion in Figu e 3.2 and le us assume ha you only ha e a es ic ed
numbe o wo ds in you lexicon. You can choose one o he ollowing exp essions o e e o
he objec s in he display:
{“blue”, “g een”, “squa e”, “ci cle”}
. Wha is he bes exp ession o
desc ibing he a ge e e en ? Choosing ei he “blue” o “squa e” migh seem like a good op ion
o you because hei li e al seman ic meaning ma ches he a ge e e en . S ill, you migh no be
ully sa is ied wi h ei he o he wo desc ip ions, as some ambigui y abou he in e p e a ion o
he u e ance emains. Speci ically, “blue” could also e e o he ci cle, and “squa e” could also
e e o he g een squa e. As a p agma ic speake , you would like o maximize he p obabili y
3 Expe imen al and modeling amewo ks 19
ha you in e locu o selec s he co ec e e en . This means ha you need o eason abou a
lis ene ’s likely in e p e a ion o you possible u e ances in his con ex . Assuming ha you
in e locu o is coope a i e, hey will, in u n, eason abou you in en ions when p oducing an
u e ance, which means hey will y o in e he speake ’s in ended e e en in he con ex .
Figu e 3.2: An example o a e e ence game ial wi h one a ge displayed in he g een ame and wo dis ac o s.
This easoning p ocess can be modeled wi h Bayes’ ule (see Equa ion
(3.1)
). A p agma ic lis ene
ies o in e he s a e o he wo ld
𝑤
gi en a speake ’s u e ance
𝑢
by easoning abou he a
p io i likely wo ld s a es
𝑃(𝑤)
and he likelihood
𝑃𝑆(𝑢|𝑤)
ha a speake chooses a pa icula
u e ance 𝑢 o communica e a wo ld s a e 𝑤:
𝑃𝐿(𝑤|𝑢) ∝ 𝑃𝑆(𝑢|𝑤)𝑃(𝑤).(3.1)
The speake is assumed o be a ional (wi h an op ional a ionali y pa ame e
𝛼
), ha is, o
choose hei u e ance based on he u ili y 𝑈 hey expec o gain om a pa icula u e ance:
𝑃𝑆(𝑢|𝑤) ∝ exp(𝛼𝑈(𝑢;𝑤)).(3.2)
The expec ed u ili y
𝑈
is ypically based on he G icean assump ion ha speake s y o be
in o ma i e. This means ha hei u ili y is high i hei u e ance enables he lis ene o selec
he co ec s a e o he wo ld 𝑤:
𝑈(𝑢;𝑤)=log 𝑃𝐿0(𝑤|𝑢).(3.3)
𝐿0
is he base case o he ecu sion o lis ene s easoning abou speake s who a e in u n easoning
abou lis ene s. Usually, he base lis ene
𝐿0
is modeled as a li e al lis ene who in e p e s any
u e ance as hei li e al seman ic meaning (Goodman and F ank, 2016, see also Degen, 2023
and Scon as e al., 2021 o an in oduc ion o RSA models). Wha is he u e ance wi h he
highes u ili y in his con ex ? A a ional, p agma ic speake is expec ed o choose he u e ance
“blue”. This can be modeled ia he ecu si e in e ence as desc ibed abo e, bu he p oblem also
has an in ui i ely appealing solu ion: Two objec s in he isual display a e blue. Howe e , i he
speake had wan ed o e e o he blue ci cle, a mo e in o ma i e u e ance would ha e been
a ailable, namely, “ci cle”. This o m o p agma ic easoning is simila o he easoning abou
al e na i es in scala implica u e de i a ion discussed abo e.
The RSA amewo k is g ounded in assump ions on gene al cogni ion, as well as p agma ic
assump ions abou meaning in e p e a ion in con ex and in e ac ion. I has been shown o i he
26 3 Expe imen al and modeling amewo ks
3.3.2 The eme gen communica ion model used in his disse a ion
In his disse a ion, I use a compu a ional eme gen communica ion pa adigm o model he
eme gence o language and concep s as an in e linked and in e ac i e p ocess. I use he same basic
model h oughou case s udies 2-4 p esen ed in his disse a ion. The models p esen ed in he
di e en case s udies di e om he basic model only in ways ha help o answe he espec i e
esea ch ques ion o he case s udy. Using he same basic model also allows me o compa e he
models used in he case s udies wi h each o he and d aw conclusions o he gene al esea ch
ques ion o his disse a ion. The base model is p esen ed in Figu e 3.3.
Speake
neu al ne wo k
(GRU)
Lis ene
neu al ne wo k
(GRU)
message
[4,11,7,0]
p edic ion and aining
1
2
3
4
5
6
Figu e 3.3: Base model o language eme gence (case s udies 2-4).
The model consis s o wo agen s, a speake
𝑆
(1) and a lis ene
𝐿
(2). Bo h agen s a e implemen ed
as a GRU (Cho e al., 2014). We choose he e e ence game pa adigm because i is also used in he
ele an expe imen al li e a u e. As is common in he e e ence game pa adigm, he speake has o
send a message o he lis ene ha helps hem o iden i y he in ended e e en . The messages (3)
ha ou agen s can send consis o disc e e symbols om a ocabula y
𝑉
ha consis s o p imi i e
disc e e symbols, anging om
0
o
𝑉
: "0", "1", "2", e c. (Laza idou e al., 2018). Using a disc e e
communica ion channel has he ad an age (o e con inuous communica ion) ha agen s canno
obse e di ec ly he inne s a es o hei in e locu o s. This makes disc e e communica ion mo e
human-like han con inuous communica ion (whe e he message is ep esen ed by a con inuous
ec o ) as human communica ion is also disc e e (Hocke , 1960; Laza idou & Ba oni, 2020). One
addi ional ad an age is ha he combina ion o disc e e symbols can esul in composi ional
linguis ic s a egies (Laza idou & Ba oni, 2020). Speake agen s in ou simula ions can s ing up
o
𝑀
symbols oge he , ha is, hey can p oduce messages wi h a ying leng hs up o a ixed
maximum message leng h
𝑀
. The p oduc ion and p ocessing o hese messages is handled by
he GRUs. A each ime s ep, a p obabili y dis ibu ion o e he nex possible symbol om he
ocabula y is compu ed based on he sequence so a , and he symbol wi h he highes p obabili y
is selec ed o esume he sequence. The symbol
0
is ea ed as an end-o -sequence (EOS) symbol
and can end a message be o e
𝑀
is eached. In he e e ence game we employ, speake s and
3 Expe imen al and modeling amewo ks 27
lis ene s communica e no abou single a ge s, bu a he abou a ge concep s ha consis o
mul iple objec s. The speake inpu (4) and he lis ene inpu (5) consis o
𝑔
a ge objec s and
𝑔
dis ac o objec s. We call
𝑔
he game size and se
𝑔=10
in ou simula ions. Speake s ecei e
10
a ge objec s and
10
dis ac o objec s in an o de ed ashion, while lis ene s ecei e he same
numbe o a ge s and dis ac o s shu led. Bo h speake s and lis ene s mus lea n he concep s in
he wo ld, ha is, which a ge objec s belong oge he , alongside lea ning o communica e abou
hem. We use isualiza ions o objec s o di e en shapes, colo s, and sizes h oughou he case
s udies. Howe e , he agen s ecei e symbolic, no isual, inpu da a. We encode symbolic objec s
as uples o
𝑛
a ibu es ha can each ake
𝑘
alues. The speake ge s as inpu only he se o
objec s. The lis ene ecei es he se o objec s and a message gene a ed by he speake . We use
simple FF-NNs o embed he symbolic inpu da a. The main eason why we use symbolic inpu
da a is ha we wan o con ol and manipula e di e en aspec s o he inpu sys ema ically. Ou
goal is o sys ema ically cons uc inpu concep s a di e en le els o abs ac ion (dalma ian,dog,
animal) and p esen hem in sys ema ically manipula ed con ex s. Cons uc ing hese inpu s as
symbolic inpu s is a i s s ep, bu ou wo k can also be ex ended o isual inpu s. Ano he c i ical
ad an age o symbolic inpu s is, howe e , ha agen s ha e been shown o ocus on low-le el isual
ea u es in pas simula ions. In hese cases, agen s ha e no lea ned he ele an concep s, bu a he
ha e con e ged on a communica ion p o ocol ha , o example, communica es abou wo pixels
ins ead o he en i e image (e.g., Bouchacou & Ba oni, 2018; Laza idou & Ba oni, 2020). While
his s a egy is a guably e y e icien (Laza idou & Ba oni, 2020), his is some hing ha we do no
wan in ou case s udies because we a e in e es ed in s udying how concep s a e communica ed.
We can p e en he agen s om communica ing abou low-le el isual ea u es ins ead o concep s
by con olling all aspec s o he inpu . This would no be possible (o a leas much ha de ) wi h
isual image da a. We ain (6) he agen s in a RL pa adigm. The p ima y eaching signal in RL is a
ewa d o succeeding o ailing a gi en ask. In ou model, he ask is success ul communica ion,
and he ewa d is posi i e i lis ene s succeed in selec ing he co ec a ge objec s. We can call
his modeling amewo k ‘mul i-agen ’, as he agen s ea each o he as pa o he en i onmen
and a e no able o access he in e nal s a es o o he agen s (Laza idou & Ba oni, 2020). We use
he Eme gence o lanGuage in Games (EGG) oolki (Kha i ono e al., 2019) o implemen ing he
eme gen communica ion models. The abo e-p esen ed modeling pa adigm allows us o s udy
language and concep eme gence join ly as an in e ac i e p ocess. While he agen s do no know he
a ge concep s, o which objec s o m a a ge concep , no do hei messages ha e meaning, he
agen s lea n o communica e abou a ge concep s ia epea ed in e ac ions. A he beginning o
aining, he speake agen s send andom messages in esponse o he da a inpu , and he lis ene
agen s make andom p edic ions abou he objec s in hei inpu . Du ing hei join aining on
communica i e success, agen s lea n he concep s and de elop a lexicon wi h mappings be ween
concep s and messages ha a e commonly used o e e o hem.
Reade s migh ask why we use such small-scale models in imes when La ge Language Models
(LLMs) a e commonly a ailable. Ou choice is mo i a ed by h ee main conside a ions. Fi s , ou
main objec i e is o use compu a ional modeling o gain insigh s and gene a e hypo heses abou
communica ion sys ems unde con olled expe imen al condi ions. The languages eme ging in ou
28 3 Expe imen al and modeling amewo ks
case s udies a e no in ended o be ull- ledged languages. Ins ead, we examine speci ic p ope ies
o hese sys ems in simpli ied se ings unde ca e ully con olled condi ions. By con as , LLMs
ope a e in mo e complex na u al language, and a e he subjec o in es iga ions hemsel es (see e.g.,
Hol e man & an Deem e , 2023). While ou ul ima e aim is o de elop explana ions ha apply
b oadly o communica ion sys ems, s a ing wi h con olled, small-scale models allows us o isola e
ele an ac o s be o e add essing he ull complexi y o na u al language. Second, om a heo e ical
iewpoin , ou modeling amewo k closely pa allels expe imen al pa adigms used o s udy
language e olu ion, such as he a i icial language lea ning pa adigm. This alignmen acili a es
compa isons wi h exis ing indings. Thi d, om a p ac ical pe spec i e, he small ne wo ks we use
a e su icien o he inpu da a and esea ch ques ions we add ess. Using la ge models would no
yield addi ional insigh s bu would equi e as ly mo e compu a ional esou ces. LLMs equi e
subs an ial amoun s o compu e ime, no only du ing aining, bu also du ing in e ence (Bha dwaj
e al., 2025), wi h signi ican implica ions o sus ainabili y. Such esou ces should be employed
only when necessa y - a iew consis en wi h he classic Occam’s azo app oach o compu a ional
modeling, whe e he simples model is p e e ed o e he mo e complex model i o he ele an
pa ame e s a e equal. A good compu a ional model should balance complexi y and simplici y. This
means ha he model should be complex enough o cap u e he phenomenon o in e es (in ou case,
he eme gence o a communica ion sys em), and i should be simple enough o be in e p e able ( o
a ce ain deg ee). Impo an ly, ou analysis is conduc ed a Ma ’s compu a ional le el (1982). We
examine how he inpu s and condi ions we design change he obse able and analyzable p ope ies
o an eme gen communica ion sys em. We do no a emp o explain, a he algo i hmic le el,
how he ANNs compu e solu ions. Ou ocus is on using compu a ional modeling as a ool o
unde s anding linguis ic and cogni i e phenomena. As a gued by an Rooij e al. (2024), aluable
insigh s in Cogni i e Science can s ill be gained by employing mo e adi ional AI me hods o
compu a ional modeling in conjunc ion wi h c i ical heo izing.
4 Disse a ion o e iew
This disse a ion in es iga es he ole o language and p agma ics in concep ual abs ac ion - he
p ocess o o ming and communica ing concep s a di e en le els o a concep ual hie a chy. I
builds on p io esea ch in ca ego iza ion, lexical choice in e e en ial communica ion, ha is,
concep ualiza ion, and he ole o p agma ics and social in e ac ion in language de elopmen and
language e olu ion. In he p e ious chap e s, I e iewed he li e a u e on concep ual abs ac ion as
well as he expe imen al and compu a ional modeling amewo ks ha ha e been used o ackle
ques ions on concep ual abs ac ion o will be used in he case s udies p esen ed in his disse a ion.
Fi e main conclusions om his e iew in o m he design o he case s udies:
▶
Cogni i e economy and p agma ic in o ma i i y conside a ions play a ole in concep ual abs ac-
ion.
▶
Re e en ial communica ion in ol es concep ualizing and labeling concep s and he e o e p o ides
a con olled en i onmen o es ing ques ions abou concep ual abs ac ion.
▶
Collabo a i e and G icean p agma ic in e ac ion plays a ole in o ming sha ed concep ualiza ions
and con en ionalized labels.
▶
Concep ual abs ac ion is shaped no only by en i onmen al ac o s bu also by communica i e
needs a ising om in e ac ion.
▶
P agma ics plays an impo an ole in language use, in language and concep ual de elopmen ,
and in language e olu ion.
Acco dingly, we examine how cogni i e and p agma ic mechanisms in luence bo h he ac ual
use o language in conc e e communica i e si ua ions and he eme gence o s uc u ed linguis ic
and ca ego y sys ems. In he case s udies, we employ expe imen al and compu a ional modeling
me hods using he e e ence game pa adigm o s udy e e en ial communica i e in e ac ion. We
in es iga e collabo a i e in e ac ion and examine he ex en o which p agma ic mechanisms a e
needed o es ablish e ec i e and e icien sha ed concep ualiza ions and con en ions. We go beyond
in es iga ing he basic-le el e ec and also ocus on how in o ma i i y conside a ions make he
sub- and supe o dina e le els o abs ac ion ele an o communica ion. We in es iga e language
and concep ual e olu ion join ly o accoun o he assump ion ha concep ual abs ac ion is
shaped by communica i e needs and no only en i onmen al ac o s. Finally, o gain a mo e holis ic
unde s anding o he ole o language and p agma ics in concep ual abs ac ion, we in es iga e his
ole ac oss di e en imescales, om language use o language e olu ion.
The esea ch p esen ed in his disse a ion is o ganized in o ou case s udies, each aking a pa icula
angle on he b oade esea ch ques ion: “Wha is he ole o language and p agma ics in concep ual
abs ac ion?” Case s udy 1 in es iga es language use in in e ac ion and, speci ically, how people
communica e abou concep s a di e en le els o abs ac ion in a concep -le el e e ence game. In
case s udies 2 o 4, we use agen -based compu a ional modeling o in es iga e p agma ic ac o s
ha in luence he s uc u e o languages ha eme ged om in e ac i e communica ion abou
30 4 Disse a ion o e iew
concep s a di e en le els o abs ac ion. In case s udy 2, we a e in e es ed in he ole o con ex
in he eme ging language and whe he he a ailabili y o con ex leads o i s use in an eme ging
communica ion sys em. Case s udy 3 ex ends he s udy o p agma ic con ex by s udying ecu si e
p agma ic easoning abou he in en ions o he in e locu o modeled wi h he RSA pa adigm. I
b ings oge he he s udy o language use and language eme gence. Finally, in case s udy 4, we
s udy abs ac ion as gene aliza ion ia linguis ic s a egies. We ask how well agen s gene alize o
no el le els o abs ac ion and which linguis ic s a egies help hem achie e such gene aliza ion
and abs ac ion.
The ollowing p esen s an o e iew o he speci ic esea ch ques ion, he me hods used, and he
imescales in es iga ed in each o he ou case s udies. Box 3 below o e s sho de ini ions o he
e minology ha can also se e as a e e ence o la e chap e s.
▶
Case s udy 1: This case s udy in es iga es he cogni i e and p agma ic ac o s a play when
speake s and lis ene s communica e abou concep s a di e en le els o abs ac ion in an
in e ac i e se ing. We in oduce he concep -le el e e ence game, which is based on he classic
e e ence game pa adigm. This expe imen in es iga es language use in conc e e communica i e
si ua ions, manipula ing concep ual and con ex ual in o ma i i y.
▶
Case s udy 2: In his case s udy, we a e in e es ed in whe he and how con ex a ailabili y shapes
an eme ging language. We in oduce ou agen -based model ha is based on he eme gen
communica ion pa adigm and he concep -le el e e ence game. This amewo k ackles ques ions
o language e olu ion by s udying language eme gence be ween agen s.
▶
Case s udy 3: This is a comp ehensi e s udy di ided in o wo expe imen s ha examine he
in luence o di e en p agma ic mechanisms on e icien language use and language eme gence.
In Expe imen 1, we use ou amewo k o s udy he coe olu ion o ca ego y and linguis ic sys ems
and in es iga e he ole o con ex -based p agma ics he ein. In Expe imen 2, we in es iga e he
ole o con ex -based and u ili y-based p agma ics modeled wi h he RSA amewo k on e icien
language use. This s udy b ings oge he eme gen communica ion and RSA models in he wo
imescales o language e olu ion and language use.
▶
Case s udy 4: This case s udy explo es he linguis ic s a egies ha agen s use when hey need o
communica e abou concep s a no el le els o abs ac ion in a ze o-sho gene aliza ion ask. We
use ou agen -based eme gen communica ion model and manipula e he aining and es da a
spli s o manipula e whe he agen s need o communica e abou lowe o highe le els o he
concep ual hie a chy du ing language eme gence and language use.
Box 3. Ope a ionaliza ions and modeling ing edien s
▶
Communica i e need - A ises om he ele an dis inc ions ha need o be made in a language
du ing in e ac i e communica ion.
▶
Concep - Concep s can be ega ded as abs ac ed knowledge, which hey help o s uc u e
and o ganize. Concep s a e ope a ionalized as g oups o mul iple objec s in expe imen al and
modeling s udies.
4 Disse a ion o e iew 31
▶
Concep -le el e e ence game - A e e ence game whe e concep s need o be communica ed
ins ead o single objec s.
▶
Concep ual hie a chy: expe imen - De ined by inclusion ela ions, such as
dalma ian ⊂
dog ⊂animal
, whe e subo dina e concep s lie on a lowe le el and supe o dina e concep s lie
on a highe le el o he hie a chy (see Figu e 4.1).
▶
Concep ual hie a chy: modeling - Ope a ionalized pa allel o he expe imen by he numbe
o sha ed p ope ies, o a ibu es, among membe s o he same concep . Speci ic concep s,
whe e many a ibu es a e sha ed, lie on a lowe le el and gene ic concep s, whe e only one
a ibu e is sha ed, lie on a highe le el o he concep ual hie a chy (see Figu e 4.1).
▶
Con ex : expe imen - The dis ac o objec s in he e e ence game a e de ined by whe he
hey lie on he same axonomic hie a chy as he a ge s. Fo example, in a ine con ex , he
dis ac o s sha e he same basic-le el ca ego y wi h he a ge objec s (e.g., a dalma ian in
he con ex o ano he ype o dog).Inacoa se con ex , he dis ac o s do no lie on he same
axonomic hie a chy (e.g., an animal in a con ex o a ehicle).
▶
Con ex : modeling - Ope a ionalized pa allel o he expe imen by how many a ibu es he
dis ac o s sha e wi h he a ge concep . A con ex is ine i many a ibu es a e sha ed be ween
he dis ac o objec s and he a ge concep . Con e sely, a con ex is coa se i no a ibu es a e
sha ed be ween he dis ac o objec s and he a ge concep .
▶Con ex -based p agma ics - The speake agen ’s access o a (sha ed) con ex .
▶
E iciency - A concep om in o ma ion heo y (Shannon, 1948). E icien communica ion is
achie ed when he message con eys he in ended in o ma ion wi h he leas possible amoun
o esou ces, such as he leng h o he message o he size o he lexicon (e.g., Pian adosi e al.,
2012).
▶
Eme gen communica ion - A compu a ional amewo k o modeling language e olu ion,
see sec ions 3.3.1-3.3.2.
▶
In o ma i i y - How much in o ma ion can be gained om a label o message ha desc ibes a
concep a a pa icula le el o he concep ual hie a chy. Lowe hie a chy le els a e associa ed
wi h highe in o ma i i iy (in o he wo ds, we lea n mo e abou he concep ) and highe
hie a chy le els a e associa ed wi h lowe in o ma i i y (see Figu e 4.1).
▶
Leng h cos - A cos applied o longe messages. Fo malized as a loss p essu e ha penalizes
longe messages by mul iplying he symbol’s posi ion in he message by a cos ac o .
▶
Le els o abs ac ion - A e m coined by Rosch e al. (1976) ha is used o e e o di e en
le els o a concep ual hie a chy.
▶Lexicon - A mapping be ween concep s and messages.
▶
Polysemy - A one- o-many mapping be ween messages and concep s whe e one message can
e e o mo e han one concep .
▶
Re e ence game - An expe imen al pa adigm o s udying e e en ial communica ion, see
sec ion 3.1.
▶
Ra ional Speech Ac s (RSA) - A modeling pa adigm o modeling e e en ial communica ion
based on Bayesian cogni i e modeling, see sec ion 3.2.
▶
Sha ed con ex - A con ex is conside ed o be sha ed be ween speake s and lis ene s when
32 4 Disse a ion o e iew
he a ibu e(s) ha di e be ween he dis ac o objec s in he con ex and he a ge concep
a e he same in he speake ’s and he lis ene ’s inpu .
▶
Synonymy - A many- o-one mapping be ween messages and concep s whe e mo e han one
message can be used o e e o he same concep .
▶
U ili y - How use ul a message is o communica ion in a pa icula si ua ion o con ex . The
u ili y o an u e ance is in luenced by he lis ene ’s likely in e p e a ion o an u e ance (see
Sec ion 3.2). O he objec i es, such as a cos o longe u e ances, can be included in he u ili y
unc ion o an RSA model as well.
▶
U ili y-based p agma ics - Ope a ionalized by implemen ing RSA speake s, ha is, speake
agen s ha maximize hei u e ances’ u ili y ia an RSA-based p oduc ion s a egy conside ing
he lis ene ’s likely in e p e a ion o each po en ial u e ance.
▶
Ze o-sho gene aliza ion - A pa adigm o es ing ML models on unseen ca ego ies, asks, o
condi ions. In con as o s anda d es ing, he da a does no come om he same dis ibu ion
as he aining da a.
low
high
in o ma i i y
high
low
concep ual
hie a chy
Dalma ian
Animal
Dog
supe o dina e
basic
subo dina e
ew
many
gene ic
speci ic
a ibu es
sha ed
(wi hin concep )
Labels
Expe imen al s imuli
Modeling s imuli
Figu e 4.1: A compa ison o how he concep ual hie a chies a e ope a ionalized in he expe imen and modeling
s udies. Concep s a highe le els o he concep ual hie a chy a e supe o dina e o gene ic concep s, which a e
associa ed wi h low in o ma i i y. Concep s a lowe le els o he concep ual hie a chy a e subo dina e o speci ic
concep s, which a e associa ed wi h high in o ma i i y.
In he ollowing chap e s, he case s udies will be p esen ed. Each case s udy is p eceded by a
high-le el in oduc ion o he subjec . A e each case s udy, I sum up he key indings and desc ibe
he implica ions o he b oade esea ch ques ion ega ding he ole o language and p agma ics o
concep ual abs ac ion. The inal chap e o his disse a ion is a discussion o he key con ibu ions
and he ami ica ions o each case s udy o he ole o language and p agma ics in concep ual
abs ac ion. I will also delibe a e on he implica ions o he p esen ed case s udies o ongoing and
u u e esea ch in Cogni i e Science, (Compu a ional) Linguis ics, and p agma ics, and inally d aw
a gene al conclusion abou he he e p esen ed wo k.
5 Case s udy 1: The cogni i e and p agma ic
ac o s in e e ing o concep s a di e en le els
o abs ac ion
This chap e p esen s case s udy 1. I s a s wi h a high-le el in oduc ion ollowed by he con en
o he publica ion: Kob ock, K., Uhlemann, C., & Go zne , N. (2024). Supe o dina e e e ing
exp essions in abs ac ion: In oducing he concep -le el e e ence game. P oceedings o he Annual
Mee ing o he Cogni i e Science Socie y,46, 518–525. h ps://eschola ship.o g/uc/i em/31n5d3p6
The chap e ends wi h a b ie summa y o he main con ibu ions o he publica ion and a discussion
o i s implica ions o he b oade esea ch ques ion o his disse a ion.
5.1 High-le el in oduc ion
Du ing na u al con e sa ion, speake s ace choices abou which wo ds o use o communica e
wha hey in end. Acco ding o p agma ic heo y and he classical G icean maxims (G ice, 1975),
speake s a e coope a i e and wan o be unde s ood by hei in e locu o . This means ha hey
y o communica e as much in o ma ion as necessa y o he lis ene o unde s and wha he
speake means. Speake s also y o minimize hei e o and communica e no mo e han
necessa y (G ice, 1975). These p inciples guide he selec ion o wo ds du ing con e sa ion and
ensu e ha communica ion is success ul. In his wo k, we look a a speci ic kind o choice speake s
ace when hey sea ch o he igh wo d o con ey a pa icula concep . We ocus on nominal
e e ing exp essions ha a e used o communica e concep s a di e en le els o abs ac ion (Rosch
e al., 1976). Fo example, speake s can use he subo dina e e m dalma ian, he basic-le el e m
dog, o he supe o dina e e m animal o e e o he same dog in he eal wo ld (see Figu e 5.1).
In his publica ion, we expe imen ally in es iga e he speake ’s choice o e e ing exp ession
o concep s a di e en le els o abs ac ion. We pi agains each o he he speake ’s need o
being in o ma i e wi h hei need o low e o . Maximizing in o ma i eness, a speake would
always use he mos speci ic e m ha is possible o use in a gi en con ex . Minimizing e o ,
speake s would communica e e ms ha a e easies and as es o hem o p oduce. These a e
e ms a he basic le el, such as dog,ca , and chai (Rosch e al., 1976). Ou hypo hesis is ha he
speake ’s p e e ence o basic-le el e ms compe es wi h he speake ’s desi e o be in o ma i e. We
cons uc e e ence si ua ions in which speake s ha e o use a e m a a ce ain le el o abs ac ion
o success ully communica e an in ended e e en . These a e isualized in Figu e 5.1. Fo example,
i he speake wan s o e e o a dalma ian, bu ano he ype o dog is p esen , he speake has o
choose he subo dina e e e ing exp ession dalma ian because he basic-le el e m dog would no be
in o ma i e enough. To make also supe o dina e e e ing exp essions such as animal in o ma i e
in ce ain con ex s, we implemen si ua ions whe e wo kinds o animals need o be communica ed
34 5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion
oge he . Fo example, i speake s wan o communica e a pa o and a dog, hey would need o
use he supe o dina e e e ing exp ession animal in o de o be bo h in o ma i e, and economical
(exp essions using conjunc ions such as dog and pa o a e possible bu equi e mo e e o because
hey a e longe ). We look a bo h sides o a con e sa ion, speake s and lis ene s, because we wan o
unde s and how lis ene s comp ehend he e e ing exp essions chosen by he speake s. By looking
a he lis ene side o he con e sa ion, we can de e mine he communica i e success o a speake ’s
u e ance, and in es iga e whe he lis ene s also ha e a p e e ed le el o abs ac ion a which hey
can unde s and he in ended a ge concep s be e . I lis ene s p ocess e e ing exp essions a he
basic le el as e han o exp essions a he subo dina e o supe o dina e le els, his would indica e
ha lis ene s also ha e a p e e ence o basic-le el e ms and ha speake s migh choose basic-le el
e ms pa ly because hey make comp ehension easie and help hem achie e communica i e
success. Con e sely, i he basic-le el is no p ocessed as e han sub- o supe o dina e e ms,
hen his means ha lis ene s do no bene i om he speake s using basic-le el e ms. This could
mean ha he usage o basic-le el e ms is only ad an ageous o he speake s and ha hey use
i mainly o minimize hei own p oduc ion e o . Howe e , e e y ime a speake de ia es om
he basic le el o he sake o communica i e success, i would indica e ha speake s a e willing
o p oduce exp essions ha a e mo e e o ul o hem. One explana ion o his beha io could
be ha speake s a e ying o be coope a i e. They know ha a sub- o supe o dina e e e ing
exp ession can enhance communica i e success in ce ain e e ence si ua ions, allowing he lis ene s
o selec he in ended e e en s. Ou expe imen al design allows us o in es iga e hese ac o s in
communica i e in e ac ion.
basic-le el ad an age
concep ual in o ma i i y
con ex ual in o ma i i y
supe o dina e
Dalma ian
Animal
Dog basic
subo dina e
Figu e 5.1: Visualiza ion o he le els o abs ac ion and he cogni i e and communica i e p essu es o he concep ual
hie a chy
dalma ian ⊂dog ⊂animal
. The image o a dalma ian is aken om he BOSS image da abase licensed CC BY
4.0 (B odeu e al., 2014). The o he s imuli a e s imuli employed in he s udy and eused om G a e al. (2016).
5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion 35
5.2 Abs ac
We s udy e e en ial communica ion abou concep s a di e en le els o abs ac ion in an in e ac i e
concep -le el e e ence game. To be e unde s and p ocesses o abs ac ion, we in es iga e supe -
o dina e e e ing exp essions (animal). P e ious wo k iden i ied wo main ac o s ha in luence
speake s’ choice o e e ing exp essions o concep s: he immedia e con ex and he basic-le el
e ec , i.e. a p e e ence o basic-le el e ms such as dog. He e we in oduce a new concep -le el
e e ence game ha allows us o s udy di e ences in he basic-le el e ec be ween comp ehension
and p oduc ion and o elici supe o dina e e e ing exp essions expe imen ally. We ind ha
supe o dina e e e ing exp essions become ele an o g oups o objec s. Fu he , we ep oduce
he basic-le el e ec in p oduc ion bu no in comp ehension. In conclusion, e en hough basic-le el
e ms a e mos eadily accessible, speake s ailo hei exp essions o he con ex , allowing he
lis ene o iden i y he a ge concep .
Keywo ds: e e ence game, concep s, ca ego iza ion, supe o dina e le el, abs ac ion
5.3 In oduc ion
Concep s allow us o make sense o he wo ld. They help us o s uc u e and o ganize knowledge,
and o gene alize om one ins ance o a class o objec s ha sha e simila p ope ies h ough a
p ocess ha is commonly called “abs ac ion” (Rosch, 1978; Yee, 2019). We use e e ing exp essions
a di e en le els o abs ac ion, anging om subo dina e e ms like dalma ian o supe o dina e
ones like animal, o communica e abou concep s a di e en le els o abs ac ion.
P e ious wo k sugges s wo main ac o s ha in luence he choice o e e ing exp essions (REs)
people use o e e o concep s a di e en le els o abs ac ion. On he one hand, Rosch e al. (1976)
amously ound ha basic ca ego ies a e special because hese a e “ he mos inclusi e ca ego ies o
which a conc e e image o he ca ego y as a whole can be o med” (Rosch e al., 1976). I has been
shown ha child en acqui e basic-le el e ms like dog i s (J. M. Cla k & Johnson, 1994; Me is
& C isa i, 1982) and ha objec s can be ca ego ized as e a he basic le el han a he sub- o
supe o dina e le els (Mu phy & Smi h, 1982). On he o he hand, he G icean maxim o quan i y
p edic s ha speake s p o ide as much in o ma ion as equi ed o he lis ene o iden i y a a ge in
a gi en con ex and no mo e (G ice, 1975). This means ha speake s should ailo hei u e ances o
he communica i e si ua ion a hand, conside ing bo h he concep hey would like o communica e
and he con ex o hei u e ance. I has also been shown empi ically ha con ex plays a ole in he
selec ion o REs in e e en ial si ua ions (see o example Hawkins e al., 2018,2020; Konopka &
B own-Schmid , 2014; Sedi y, 2005). Ou goal is o pi he wo ac o s di ec ly agains each o he . We
use a e e ence game simila o G a e al. (2016), whe e a speake desc ibes an objec and a lis ene
needs o iden i y his objec om a se o dis ac o s. G a e al. (2016) showed ha while speake s
ailo hei u e ances o he con ex , hey also p e e basic-le el exp essions (e.g, dog) o e all.
42 5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion
98.93% and 90.29%, espec i ely. We ind subs an ial e idence o he di e ences be ween bo h he
basic and ine concep ual con ex (M=1.39, C I=[0.34, 2.34], pd=99.08%, ROPE=[-0.18, 0.18, 0% in
ROPE]) and be ween he coa se and ine concep ual con ex (M=2.3, C I=[0.33, 4.35], pd=99.11%,
ROPE=[-0.18, 0.18], 0% in ROPE).
6
The e ec o concep ual con ex on he e e ence le el is u he
suppo ed by a Bayes Fac o o 13.5 in a o o he model ha includes concep ual con ex as a
p edic o agains a null model, p o iding s ong e idence o he e ec o concep ual con ex on he
choice o he e e ence le el.7
5.5.4 Response imes
Fo he esponse ime analyses, we we e in e es ed in esponse imes o ials in which he lis ene
selec ed he co ec a ge objec s, i.e. communica ion was success ul, and in which he speake
chose an u e ance on he app op ia e le el o abs ac ion. The exclusion o da a on inapp op ia e
le els led o a da a loss o 11.12%.
Speake esponse imes
Ou second hypo hesis was ha speake s choose u e ances on he basic le el mo e quickly han
u e ances on sub- o supe o dina e le els. This hypo hesis was mo i a ed by he basic-le el e ec
ound in he li e a u e (Rosch e al., 1976; Tanaka & Taylo , 1991). We hus p edic ed ha speake
esponse imes would be sho e when he p oduced u e ance was on he basic le el compa ed o
he o he wo le els. As p e egis e ed, ials wi h esponse imes o e 2.5 s anda d de ia ions abo e
he mean (cu -o : 22,360 ms) we e excluded, leading o he exclusion o 2.61% o he emaining
da a. Figu e 5.5 shows he da a means and boo s apped con idence in e als o he cleaned da a.
Indeed, esponses on he basic le el a e sho e han esponse imes on he o he wo le els.
We an a Bayesian model wi h a logno mal link unc ion, p edic ing speake esponse imes by
concep ual con ex and including g oup-le el e ec s o he pa icipan pai and i em ca ego y.
We speci ied weakly in o ma i e p io s o enhance model i . We had o de ia e om one p io
speci ied in he p e egis a ion: P io p edic i e checks showed ha we o e es ima ed he e ec size,
and he p io we p e egis e ed o he popula ion-le el e ec was oo wide. We hus changed he
p io ’s s anda d de ia ion om 2 o 0.5 o enhance model con e gence and i .
8
The model esul s
6
A he sugges ion o ou e iewe s, we also an a model on he da a wi hou excluding unsuccess ul ials. This model
p o ides simila e idence o he o iginal model: basic- ine: M=1.63, C I=[0.35, 2.81], pd=98.65%, ROPE=[-0.18, 0.18],
0% in ROPE; coa se- ine: M=2.58, C I=[0.44, 4.79], pd=99.16%, ROPE=[-0.18, 0.18], 0% in ROPE.
7
The BF models an o 20,000 i e a ions wi h a wa m-up pe iod o 2,000 as ecommended o BF es ima ion (Nicenboim
e al., 2023).
8
The p io s we e speci ied as ollows: in e cep p io : no mal(8.65, 0.5), popula ion-le el e ec s slop p io : no mal(0,
0.5), g oup-le el e ec s s anda d de ia ion p io : no mal(0, 0.1), g oup-le el e ec s co ela ion p io : lkj(2), sigma
p io : no mal(0, 0.5).
5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion 43
Figu e 5.5: Speake esponse imes wi h boo s apped 95% CIs.
con i m ha speake s espond subs an ially as e on he basic le el han on he o he wo le els o
e e ence (M=-0.22, C I=[-0.36, -0.10], pd=99.86%, ROPE=[-0.01, 0.01], 0% in ROPE).910
Lis ene esponse imes
Ou hi d hypo hesis was ha we would also ind di e ences be ween le els in he lis ene esponse
imes, showing ha ei he a) lis ene s selec he co ec a ge s mo e quickly when a basic-le el
e m was p oduced o b) lis ene s selec he co ec a ge s mo e quickly when a supe o dina e
e m was p oduced. These hypo heses we e based on he basic-le el e ec s li e a u e (Rosch e al.,
1976) and mo e ecen s udies deba ing he basic-le el e ec s in ce ain asks (Macé e al., 2009). The
da a in Figu e 5.6 shows ha we do no ind ei he o he expec ed pa e ns.
Figu e 5.6: Lis ene esponse imes wi h boo s apped 95% CIs.
9This pos e io di e ence was con as -coded as p e egis e ed: basic s. ( ine + coa se)/2.
10
A he sugges ion o ou e iewe s, we also an a model including leng h as a p edic o . This model p o ides smalle
e idence o he di e ence o he basic le el compa ed o he o he wo: M=-0.10, C I=[-0.19, 0.00], pd=98.07%,
ROPE=[-0.01, 0.01], 0.62% in ROPE. This is mos ly d i en by he di e ence be ween basic and coa se being no as
p onounced as he di e ence be ween basic and ine.
44 5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion
We an a Bayesian model wi h he same model speci ica ions and p io s as he speake esponse
ime model.
11
We employed he same da a exclusion c i e ia as o he speake esponse ime model
wi h one excep ion: We decided o exclude he highes esponse ime da a poin and de ined he
cu -o o 2.5 s anda d de ia ions abo e he mean a e his exclusion o ge a mo e sensible cu -o
(17,076 ms). The model p edic ions do no o e enough eason o belie e ha ei he o he wo
p oposed pa e ns is a play: The di e ence be ween he basic le el and he o he wo le els was
es ima ed a M=-0.04 wi h a C I o [-0.12, 0.04] (pd=84.96%, ROPE=[-0.01, 0.01], 12.28% in ROPE).
And he di e ence be ween he coa se le el and he o he wo le els was es ima ed a M=0.04 wi h
a C I o [-0.07, 0.14] (pd=79.07%, ROPE=[-0.01, 0.01], 11.81% in ROPE).12
5.6 Discussion and conclusion
We s udied he exp essions speake s use o e e o concep s a di e en le els o abs ac ion in a
new in e ac i e concep -le el e e ence game. We ha e shown ha he le el o abs ac ion o he
u e ances speake s choose o communica e a ce ain concep mainly depends on he concep and
con ex in ques ion. Speake s ailo hei u e ances o he con ex , p oducing subo dina e e ms in
ine concep ual con ex s, basic-le el e ms in basic concep ual con ex s and supe o dina e e ms in
coa se concep ual con ex s. In ac , we ound ha i speake s p oduce a RE on he expec ed le el o
e e ence, his educes e o s in he a ge selec ion by he lis ene s, o , in o he wo ds, i inc eases
communica i e success by abou 14%. We also ep oduced he basic-le el e ec on he p oduc ion
side, i.e. speake s a e as e in p oducing a basic-le el e m han in p oducing a e m on he o he
wo le els. In e es ingly, e en hough speake s show a basic-le el ad an age, hey s ill ailo he
u e ances o i he con ex and make i easie o he lis ene o iden i y he co ec a ge s. We did
no ind e idence o an ad an age o basic-le el p ocessing on he comp ehension side, i.e. lis ene s
a e equally quick in selec ing he a ge s ega dless o he le el o abs ac ion o he u e ance hey
ecei e. A possible eason o his is ha he basic-le el ad an age is mos ly d i en by accessibili y
in p oduc ion. In comp ehension, on he o he hand, lis ene s migh be as quick in ca ego izing
objec s a o he le els han a he basic le el.
Despi e he basic-le el e ec , speake s use sub- and supe o dina e e ms equen ly in na u al
con e sa ion. In he case o subo dina e e ms, his is usually de e mined by he con ex . When, o
example, a dalma ian is he a ge and a g eyhound is he dis ac o , he basic-le el e m dog does
no su icien ly disc imina e he a ge om he dis ac o . I has been shown ha con ex wa an s
he use o a mo e cos ly, i.e. usually longe and less equen , subo dina e e m (G a e al., 2016). In
he case o supe o dina e e ms, howe e , he con ex does no su icien ly explain why hese e ms
migh be used because pa icipan s could always use he basic-le el e m o e e o single objec s
e en in a coa se concep ual con ex . Ou concep -le el e e ence game shows ha supe o dina e
11
The only excep ion was he p io on he in e cep ha was p e egis e ed o depend on he da a dis ibu ion: no mal(8.15,
0.38).
12
These pos e io di e ences we e con as -coded as p e egis e ed: basic s. ( ine + coa se)/2 and coa se s. (basic +
ine)/2.
5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion 45
e ms become ele an when dealing wi h mul iple a ge objec s, o when communica ing he
idea o a mo e gene ic class. The longe esponse imes we see when speake s use a supe o dina e
compa ed o a basic-le el e m migh be an indica o o a p ocess o abs ac ion ha speake s
unde go when ying o ind which supe o dina e class he wo a ge s ha e in common and
e ie ing he espec i e supe o dina e RE. This could be a good s a ing poin o u he esea ch
on abs ac ion.
One limi a ion o ou cu en s udy se up is ha so a , we only in es iga e REs in a con ex ha is
e y close o he a ge concep , i.e. ha includes a dis ac o om he same basic ca ego y in he
ine concep ual con ex o om he same supe o dina e ca ego y in he basic concep ual con ex .
This means ha we canno accoun o o e in o ma i e REs (see o example Degen e al., 2020)
because he con ex makes a ce ain le el o e e ence necessa y o disambigua ion. Fu u e s udies
can ex end ou se up and include con ex condi ions ha make a ce ain le el o e e ence only
su icien o disc imina ion, by using wide con ex s ha , o example, only include un ela ed
dis ac o s. Such a manipula ion would allow he in es iga ion o o e - and unde speci ica ion in
he concep -le el e e ence game. Howe e , e en in ou cu en se -up, we do see some u e ance
choices ha a e a odds wi h ou p edic ed le el o e e ence o each concep ual con ex . Fo
example, in he ine concep ual con ex , basic-le el exp essions a e p oduced almos 20% o he
ime. A close look a hese p oduc ions e eals ha speake s ei he unde speci y, i.e. p oduce dog
o dalma ian, o hey use a modi ied basic-le el exp ession, i.e. p oduce spo ed dog. The high
p opo ion o hese men ions p o ides u he e idence o a s ong basic-le el e ec on he speake
side.
The esponse ime esul s in which we ind a basic-le el e ec only o p oduc ion, bu no o
comp ehension, lead o an in e es ing obse a ion: Speake s ailo hei u e ances o he concep ual
con ex e en when i esul s in highe p ocessing cos s o hem. On he comp ehension side,
howe e , we do no ind highe p ocessing cos s o sub- o supe o dina e e ms. This could sugges
ha speake s a e willing o bea a highe cos because hey know ha i would make iden i ica ion
o he a ge objec s easie o he lis ene . This phenomenon has been discussed in he li e a u e as
audience design (see o example Gann & Ba , 2014; W. S. Ho on & Ge ig, 2002). We should no e,
howe e , ha he comp ehension esponse imes we e logged when lis ene s had clicked on bo h
a ge s. Thus, ou measu e is a he o line, and we canno comple ely ule ou ha he e a e mo e
immedia e di e ences ac oss le els in comp ehension ha migh be e ealed by mo e sensi i e
measu es. I howe e he di e ences we obse ed o p oduc ion and comp ehension a e no jus
due o such me hodological aspec s, his could indica e ha he basic-le el ad an age is ela ed o
lexical accessibili y and no ca ego iza ion i sel . On he lis ene ’s side, basic le el ca ego ies may
no ha e a p i ileged ep esen a ion.
In conclusion, he concep -le el e e ence game allows us o es hypo heses on he use o supe -
o dina e REs and abs ac ion. While we see di e ences in esponse imes be ween he basic and
supe o dina e le els on he p oduc ion side, we do no see he same di e ences on he comp e-
hension side. This opens up exci ing possibili ies o u u e esea ch on audience design and cos s
46 5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion
associa ed wi h abs ac ion in p oduc ion and comp ehension. He e, we showed ha supe o dina e
REs become ele an when a speake needs o desc ibe mo e han one a ge objec .
Acknowledgmen s
We would like o hank Il a Ho emann o help wi h p og amming he expe imen in Lab anced,
Be i Reise o help wi h p og amming he no ming s udy in pcIBEX, and Elli Tou ou i o help ul
discussions on he expe imen design. We also hank h ee anonymous e iewe s o hei help ul
commen s and eedback.
K is ina Kob ock is suppo ed by he DFG- unded Resea ch T aining G oup “Compu a ional
Cogni ion” (DFG-GRK 2340).
Au ho Con ibu ions:
K is ina Kob ock: Concep ualiza ion, Me hodology, So wa e, Valida ion, Fo mal analysis, In es iga-
ion, W i ing - O iginal D a , Visualiza ion. Cha lo e Uhlemann: Concep ualiza ion, Me hodology,
So wa e, W i ing - Re iew & Edi ing. Nicole Go zne : Concep ualiza ion, Me hodology, W i ing -
Re iew & Edi ing, Supe ision.
5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion 47
5.7 Summa y o he key con ibu ions
Wein oducedano elexpe imen alpa adigm oin es iga ehowspeake sandlis ene scommunica e
abou concep s a di e en le els o abs ac ion. We buil on he classic e e ence game pa adigm,
which is commonly used o s udy e e ing exp essions, and ex ended i o concep -le el nominal
e e ence. Ou indings con ibu e o he li e a u e on cogni i e mechanisms, such as he basic-le el
e ec , as well as p agma ic ac o s in p oducing and p ocessing e e ing exp essions, including
con ex and in o ma i i y conside a ions. Fi s , we eplica ed he well-known basic-le el e ec in
p oduc ion bu no in comp ehension. Second, we ound e idence ha supe o dina e e e ing
exp essions a e used when ele an , ha is, when hey a e concep ually in o ma i e. Thi d, when
conside ing he adeo be ween he basic-le el ad an age and in o ma i i y conside a ions,
we ound ha he p agma ic ac o s o concep ual and con ex ual in o ma i i y de e mine he
speake s’ choice o e e ing exp essions, e en i hey go agains he basic-le el e ec . The speake s’
p e alen s a egy o using e e ing exp essions a he app op ia e le el o e e ence, licensed by
he concep ual le el and con ex , leads o highe communica i e success han when speake s use
e e ing exp essions a an inapp op ia e le el o e e ence.
5.8 Implica ions o he b oade esea ch ques ion
We add essed open ques ions in he empi ical li e a u e on he ole o p agma ics in nominal
e e ence (e.g. Degen e al., 2020; G a e al., 2016). Ou s udy goes beyond p e ious wo k by
in es iga ing e e ence o a ge concep s ins ead o single a ge objec s. Mo eo e , as a as we
know, we a e he i s o s udy he basic-le el e ec in in e ac ion. This opens up he possibili y
o s udy how speake s communica e a ge concep s a he sub-, supe o dina e, and basic le els,
and how lis ene s comp ehend such a ge concep s. We ound ha speake s ade o p agma ic
ac o s, speci ically he desi e o be in o ma i e, wi h cogni i e ac o s, which ela e o he ease o
e ie ing basic-le el e e ing exp essions. In ou esul s, we see ha speake s use a basic-le el
e m whe e possible, bu hey also de ia e om he basic-le el p oduc ion s a egy and use e ms
a he sub- and supe o dina e le els o abs ac ion when hese a e mo e app op ia e. I he a ge
concep is a supe o dina e concep , hen he concep ual in o ma i i y makes he supe o dina e
e m mo e in o ma i e han he basic-le el e m. I he a ge concep is subo dina e and p esen ed
wi h a dis ac o objec om he same basic-le el ca ego y, hen he con ex ual in o ma i i y
makes he subo dina e e m mo e in o ma i e han he basic-le el e m. We ound ha speake s
ake hese concep ual and con ex ual in o ma i i y conside a ions in o accoun when e e ing o
concep s. Speake s addi ionally show a basic-le el ad an age in esponse imes which means ha
hey end o p oduce basic-le el e ms as e han e ms a he sub- o supe o dina e le els. This
co obo a es he well-known basic-le el e ec when naming concep s in an in e ac i e se ing in
p oduc ion (Jolicoeu e al., 1984; Rosch e al., 1976). In e es ingly, a po en ial s a egy o he speake s
in he supe o dina e condi ion would be o communica e bo h basic-le el objec s ha comp ise he
a ge concep in a conjunc ion. Howe e , he expe imen al da a sugges ha he supe o dina e e m
48 5 Case s udy 1: Re e ing o concep s a di e en le els o abs ac ion
is used mo e consis en ly and mo e success ully (men ions a he app op ia e le el o e e ence
inc ease communica i e success by 14%). This is in line wi h p edic ions o cogni i e economy: The
supe o dina e e m is usually sho e han men ioning wo basic-le el e ms. Thus, i he e is no
addi ional cos associa ed wi h e ie ing he supe o dina e e m, i should be cogni i ely mo e
economical o p oduce i . In comp ehension, we did no ind any di e ences in esponse imes
when lis ene s in e p e ed app op ia e e e ing exp essions. Lis ene s seem o be equally able o
iden i y he a ge concep s independen o hei le el o abs ac ion. This could mean ha lis ene s
do no ace an abs ac ion e o when iden i ying supe o dina e concep s o a disc imina ion e o
when disc imina ing subo dina e concep s in con ex s wi h close dis ac o s du ing in e ac ion.
Toge he , hese indings sugges ha speake s migh use e e ing exp essions a le els o he han
he basic le el p ima ily o achie e communica i e success in in e ac ion when he e is a need o be
mo e in o ma i e. In conclusion, we showed how cogni i e mechanisms and p agma ic ac o s a e
pi ed agains each o he in in e ac ion. Ou no el expe imen al pa adigm allowed us o ackle he
ques ion o how concep s a di e en le els o abs ac ion a e communica ed and unde s ood. In
he ollowing s udies, we ake his pa adigm o he nex s ep and use i o model how languages
e ol e du ing in e ac i e communica ion abou concep s.
6 Case s udy 2: The ole o con ex a ailabili y in
eme gen communica ion abou concep s
This chap e p esen s case s udy 2. The chap e s a s wi h a high-le el in oduc ion ollowed by
he con en o he publica ion: Kob ock, K., Ohme , X., B uni, E., & Go zne , N. (2024). Con ex
Shapes Eme gen Communica ion abou Concep s a Di e en Le els o Abs ac ion. In N. Calzola i,
M.
-
Y. Kan, V. Hos e, A. Lenci, S. Sak i, & N. Xue (Eds.), P oceedings o he 2024 Join In e na ional
Con e ence on Compu a ional Linguis ics, Language Resou ces and E alua ion (LREC-COLING 2024)
(pp. 3831–3848). ELRA and ICCL. h ps://aclan hology.o g/2024.l ec-main.339 The chap e
ends wi h a b ie summa y o he main con ibu ions o he publica ion and a discussion o i s
implica ions o he b oade esea ch ques ion o his disse a ion.
6.1 High-le el in oduc ion
In hep e iouscases udy,weexaminedwhich e e ingexp essionsspeake schoose ocommunica e
concep s in in e ac i e communica ion and how lis ene s comp ehend such e e ing exp essions.
The goal o he ollowing s udy is o in es iga e communica ion abou concep s a di e en le els
o abs ac ion in a compu a ional model o language e olu ion. This allows us o gene alize ou
esea ch indings om language use in speci ic communica i e si ua ions, as in es iga ed in case
s udy 1, o he ques ion o how a language e ol es o encode concep s a di e en le els o abs ac ion.
The ac ha we can e e o he same objec wi h di e en names ha lie on he same axonomic
hie a chy is puzzling o esea che s s udying language de elopmen and language e olu ion. F om
he pe spec i e o language de elopmen , child en ypically expec one- o-one mappings be ween
objec s and wo ds (Ma kman & Wach el, 1988). F om he pe spec i e o language e olu ion, i
emains an open esea ch ques ion which p essu es in luence he e olu ion o mul iple names o he
same objec . The goal o case s udy 2 is o add ess his ques ion. To achie e his aim, we implemen
a compu a ional model o language e olu ion ha enables us o es se e al hypo heses ega ding
how a language e ol es o acili a e e e ence o di e en le els o abs ac ion. The key componen s
o ou model a e: Fi s , we model communica ion in in e ac ion. We ain a i icial agen s, one wi h
he speake ole and one wi h he lis ene ole, on he same communica i e ask as was used in case
s udy 1, he concep -le el e e ence game. Second, ollowing ou indings om case s udy 1, we
model concep ual and con ex ual in o ma i i y. We use a da ase consis ing o hie a chical concep s
o he communica i e ask. We o malize concep s as g oups o objec s ha sha e a ce ain numbe
o a ibu es. We hypo hesize ha his concep ual hie a chy should allow a language o eme ge ha
encodes concep s a di e en le els o abs ac ion. We implemen con ex ual in o ma i i y by adding
dis ac o objec s. The con ex c ea ed by hese dis ac o objec s is cha ac e ized by he numbe
o a ibu es sha ed be ween a ge and dis ac o objec s. I many a ibu es a e sha ed be ween
a ge and dis ac o objec s, hen he con ex equi es ine dis inc ions and a speci ic desc ip ion o
50 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
he a ge concep . Con e sely, i ew a ibu es a e sha ed be ween a ge and dis ac o objec s,
coa se dis inc ions and mo e gene ic desc ip ions o he a ge concep a e su icien o e ec i e
communica ion. In case s udy 2, ou i s aim is o in oduce his compu a ional model and he
da ase , manipula ing concep ual and con ex ual in o ma i i y. Ou second aim is o in es iga e he
ole o he con ex and, speci ically, whe he he a ailabili y o con ex shapes an e ol ing language.
We achie e his aim by implemen ing wo expe imen al condi ions ha manipula e whe he he
speake agen has access o con ex in o ma ion o no .
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 51
6.2 Abs ac
We s udy he communica ion o concep s a di e en le els o abs ac ion and in di e en con ex s
in an agen -based, in e ac i e e e ence game. While playing a concep -le el e e ence game, he
neu al ne wo k agen s de elop a communica ion sys em om sc a ch. We use a no el symbolic
da ase ha disen angles concep ype ( anging om speci ic o gene ic) and con ex ( anging om
ine o coa se) o s udy he in luence o hese ac o s on he eme ging language. We compa e wo
game scena ios: one in which speake agen s ha e access o con ex in o ma ion (con ex -awa e)
and one in which he speake agen s do no ha e access o con ex in o ma ion (con ex -unawa e).
Fi s , we ind ha he agen s lea n highe -le el concep s om he objec inpu s alone. Second, an
analysis o he eme gen communica ion sys em shows ha only con ex -awa e agen s lea n o
communica e e icien ly by adap ing hei messages o he con ex condi ions and elying on con ex
o unambiguous e e ence. C ucially, his beha io is no explici ly incen i ized by he game, bu
e icien communica ion eme ges and is d i en by he a ailabili y o con ex alone. The eme ging
language we obse e is eminiscen o e olu iona y p essu es on human languages and highligh s
he pi o al ole o con ex in a communica ion sys em.
Keywo ds: eme gen communica ion, concep s, con ex
6.3 In oduc ion
Re e ing o hings in he wo ld is c ucial o e ec i e communica ion. When choosing a e e ing
exp ession, speake s ecu o wha hey know abou he e e en ’s unde lying concep and choose
o communica e he concep a a le el o abs ac ion ha i s well wi h hei communica i e
in en ions. Fo example, a e e ence o he objec in Figu e 6.1 can be made a a ious le els o
abs ac ion, anging om he mo e speci ic concep ‘wa e melon’ o he mo e gene ic concep ‘ ood’.
When communica ing a mo e gene ic concep , speake s and lis ene s need o abs ac away om
p ope ies o he indi idual objec s and ocus on wha all objec s belonging o a concep ha e in
common. By choosing he u e ance ‘ ui ’, o example, a speake abs ac s away om i ele an
p ope ies ( he size, colo e c. o he speci ic objec ) and s esses he p ope ies ha wa e melons
sha e wi h o he i ems belonging o he concep ‘ ui ’, o example ha hey a e edible.
“wa e melon”
“melon”
“ ui ”
“ ood”
gene ic
speci ic
Figu e 6.1: Example e e ing exp essions a di e en le els o abs ac ion.
58 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
when he lis ene co ec ly iden i ies mos objec s pe game. Fi s , we obse e e y high aining and
alida ion accu acies o all game se ings and da ase s (mean aining and alida ion accu acies
ac oss uns > 0.96 o all da ase s and bo h se ings).
4
Mean es accu acies ac oss uns on concep s
ha he agen s ha e ne e encoun e ed du ing aining a e 0.89 (SD=0.07) o con ex -unawa e
and 0.87 (SD=0.11) o con ex -awa e agen s. This sugges s ha he agen s lea n o success ully
communica e abou concep s on a ious le els o abs ac ion and in a ious con ex condi ions.
5
To ge a be e unde s anding o he agen s’ s a egies and whe e communica ion is especially
(un)success ul, we pe o med an addi ional analysis o he e o s, i.e. hose cases whe e he lis ene
agen s p edic some o he labels w ongly.
6
We ind ha mos e o s occu when a ge s and
dis ac o s sha e many a ibu es, making i mo e likely ha hey a e con used wi h each o he . In
o he wo ds, mos mis akes happen in he ine con ex condi ions.7
6.6.2 Quali a i e communica ion analysis
Second, we use a quali a i e analysis o he messages o see whe he we can obse e di e ences
be ween con ex -unawa e and con ex -awa e se ings. Tables 6.2 and 6.3 show he esul s o ou
quali a i e analysis on he D(4,4) da ase o con ex -unawa e and con ex -awa e, espec i ely. We
epo all unique messages o a andomly chosen speci ic concep (
[0,0,0,3]
, all a ibu es ixed)
and each con ex condi ion. Con ex -unawa e agen s end o use he same messages in all con ex
condi ions (in his case, “
[11,1,11,14,0]
” is used consis en ly ac oss con ex s). On he o he hand,
con ex -awa e agen s use a la ge se o messages ( ou unique messages o e all games), and hey
end o a y he messages mo e depending on con ex . In coa se con ex s, he se o messages
used o desc ibe he a ge concep is la ge han in he ines con ex , whe e he bes s a egy is o
communica e he mos speci ic concep . We obse e he same pa e n o o he andomly selec ed
concep s ac oss di e en da ase s.8
Table 6.2: Con ex -unawa e: Unique messages used o e e o a andomly picked speci ic concep in he D(4,4) da ase
o e di e en con ex condi ions.
Objec Con ex (# Sha edA ibu es) UniqueMessages
[0,0,0,3]
0 “[11,1,11,14,0]”
1 “[11,1,11,14,0]”
2 “[11,1,11,14,0]”
3 “[11,1,11,14,0]”
4
I is impo an o no e ha achie ing such high sco es is in en ional. Only wi h a high success sco e does he es o he
e alua ion become meaning ul. This ensu es ha he language we analyze can be assumed o e ec i ely communica e
wha is in ended in he e e en ial game.
5De ailed accu acy sco es can be inspec ed in Table 6.4 in Appendix A.
6Plo s o hese analyses can be inspec ed in Appendix B.
7Addi ional plo s o he dis ibu ion o alse posi i e and alse nega i e e o s can be ound in Appendix B.2 and B.3.
8Mo e examples a e gi en in Appendix C o show ha hese a e no che y-picked.
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 59
Table 6.3: Con ex -awa e: Unique messages used o e e o a andomly picked speci ic concep in he D(4,4) da ase o e
di e en con ex condi ions.
Objec Con ex (# Sha edA ibu es) UniqueMessages
[0,0,0,3]
0 “[6,2,10,14,0]”
“[6,2,14,10,0]”
1“[6,2,10,14,0]”
“[6,2,14,10,0]”
2
“[6,2,10,1,0]”
“[6,2,10,14,0]”
“[6,2,10,5,0]”
3 “[6,2,10,5,0]”
6.6.3 Quan i a i e communica ion analysis
Mappings be ween concep s and messages
Thi d, we use in o ma ion- heo e ic sco es and compa e he con ex -unawa e o he con ex -awa e
se ing o quan i y he esul s we ob ained om ou quali a i e analysis. In he con ex -unawa e
se ing, we obse e high o e all in o ma ion sco es (NMI sco es anging om 0.94 [0.9, 0.98]
9
o
D(5,4) o 0.97 [0.96, 0.98] o D(3,8)). This sugges s ha concep s and messages end o ha e one- o-
one mappings. While he mu ual in o ma ion be ween messages and concep s is also ela i ely high
o con ex -awa e agen s, i is sligh ly lowe han o con ex -unawa e agen s (NMI sco es anging
om 0.86 [0.78, 0.92] o D(3,16) o 0.9 [0.88, 0.93] o D(3,8)). This could mean ha con ex -awa e
ained agen s adap o he con ex , making s ic one- o-one mappings imp ac ical.
Figu e 6.4 shows o he con ex -unawa e se ing how he mu ual in o ma ion a ies when i is
calcula ed o all concep and con ex condi ions o da ase D(4,4).
10
He e, we obse e wo pa e ns:
On he one hand, he NMI inc eases wi h he numbe o ixed a ibu es. In o he wo ds, he mo e
speci ic he concep s a e, he mo e one- o-one mappings be ween concep s and messages eme ge.
On he o he hand, he NMI sco es s ay ela i ely cons an ac oss di e en numbe s o sha ed
a ibu es. This sugges s ha con ex -unawa e ained speake agen s adap hei choice o e e ence
o a concep ’s le els o abs ac ion, bu no o he con ex (o which hey a e no awa e).
When looking a he NMI o he con ex -awa e se ing in Figu e 6.5, we obse e he opposi e
pa e n: While changes in he concep le el (i.e., he numbe o ixed a ibu es) a e no e lec ed in
changing NMI sco es, we do obse e inc easing NMI sco es wi h an inc easing numbe o sha ed
a ibu es. In o he wo ds, he ine he con ex , he mo e one- o-one mappings be ween concep s
and messages can be ound in he agen s’ communica ion sys em.
9The in e als epo ed he e a e boo s apped 95% Con idence In e als.
10Plo s o all da ase s a e a ailable in Appendix D.
60 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
Figu e 6.4: Con ex -unawa e: Mean NMI sco es ac oss all da ase s o di e en concep (# ixed a ibu es) and con ex
condi ions (# sha ed a ibu es). F om op o bo om con ex becomes ine and om le o igh concep s become mo e
speci ic.
Figu e 6.5: Con ex -awa e: Mean NMI sco es ac oss all da ase s o di e en concep (# ixed a ibu es) and con ex
condi ions (# sha ed a ibu es). F om op o bo om con ex becomes ine and om le o igh concep s become mo e
speci ic.
E ec o he le el o abs ac ion
We will now look i s a he e ec o a concep ’s le el o abs ac ion and hen a he e ec o he
con ex on he eme ging language in mo e de ail. The e ec o a concep ’s le el o abs ac ion on
he eme ging language is isualized in Figu es 6.6 and 6.7 which plo he en opy-based sco es
o e di e en concep le els agg ega ed o e all da ase s and simula ion uns o con ex -unawa e
and con ex -awa e, espec i ely. In Figu e 6.6, we obse e ha he NMI is la gely cons an o
mo e speci ic concep s ( h ee ixed a ibu es and mo e) and sligh ly d ops owa d mo e gene ic
concep s wi h one o wo ixed a ibu es. This e ec is la gely d i en by a co esponding d op in
he consis ency sco e when i comes o mo e gene ic concep s, which sugges s ha mo e han one
unique message is used o e e o he same gene ic concep . We can hink o wo easons o his:
One eason migh be ha he agen s a e o e ly speci ic when e e ing o he gene ic a ge concep ,
o example, hey migh use “ ed ci cle” o “blue ci cle” o e e o “ci cle”. Ano he eason is ha
he eme ging language con ains mo e synonymous wo ds ha e e o mo e gene ic concep s, o
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 61
example he in en ed messages “1, 1, 2” and “2, 3, 4” bo h mean “ci cle”.
Figu e 6.6: Con ex -unawa e: Mean en opy sco es ac oss all da ase s o di e en concep le els indica ed by he numbe
o ixed a ibu es. F om le o igh concep s become mo e speci ic. E o ba s indica e boo s apped 95% con idence
in e als.
Figu e 6.7 shows ha we obse e a d op in he consis ency sco e when i comes o mo e gene ic
concep s also o languagesde eloped bycon ex -awa eagen s.Addi ionally,we ind ha consis ency
dec eases again o mo e speci ic concep s (i.e., when he numbe o ixed a ibu es is la ge han
h ee). This can be explained by he a ailabili y o con ex in he con ex -awa e se ing: Fo mo e
speci ic concep s wi h h ee o mo e a ibu es, he e a e mo e con ex condi ions possible, i.e.
𝑛−1
con ex condi ions. Thus, con ex -awa e ained speake s adap o use di e en messages o e e o
he same concep s when hey ake con ex in o accoun . The bu e ly shape we obse e in Figu e
6.7, whe e e ec i eness inc eases o speci ic and o gene ic concep s and consis ency, on he o he
hand, dec eases o speci ic and o gene ic concep s, can hus be explained by he wo ac o s ha
he agen s ake in o accoun when cons uc ing messages, bo h concep speci ici y and con ex .
Figu e 6.7: Con ex -awa e: Mean en opy sco es ac oss all da ase s o di e en concep le els indica ed by he numbe
o ixed a ibu es. F om le o igh concep s become mo e speci ic. E o ba s indica e boo s apped 95% con idence
in e als.
We used Bayesian es ima ion o s a is ically analyze hese obse ed di e ences be ween condi ions
ac oss all i e uns, ollowing K uschke (2013). We ind no subs an ial di e ence in NMI sco es
be ween he con ex -unawa e (M=0.92, C I=[0.9, 0.93]
11
) and he con ex -awa e (M=0.89, C I=[0.88,
0.91]) se ing o gene ic concep s wi h an es ima ed di e ence in means o M=0.023 (C I=[-0.003,
11C edible In e als (C Is) we e compu ed on he pos e io ia he Highes Densi y In e als.
62 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
0.048], pd=95.9%
12
,6%in ROPE
13
). The di e ence o speci ic concep s be ween he con ex -unawa e
(M=0.94, C I=[0.92, 0.97]) and he con ex -awa e (M=0.87, C I=[0.84, 0.91]) se ing on he o he hand
is subs an ial wi h an es ima ed di e ence in means o M=0.07 (C I=[0.026, 0.109], pd=99.4%, 0%
in ROPE). While hese e ec s a e a he small, we do ind eliable di e ences. These esul s a e in
line wi h ou obse a ions abo e, speci ically ha speci ic concep s can appea in a wide ange o
con ex s (coa se o ine). Thus, con ex -awa e agen s use a wide ange o messages o e e o he
same speci ic concep han con ex -unawa e agen s because hey can make use o he con ex .
E ec o he con ex
The e ec o he con ex on he eme ging language is especially e iden when we compa e Figu e
6.8 and Figu e 6.9 which plo he en opy-based sco es o e di e en con ex condi ions o con ex -
unawa e and con ex -awa e se ings. In he con ex -unawa e se ing, he NMI s ays a a cons an
le el ac oss di e en con ex condi ions. We obse e a small d op in consis ency and an inc ease
in e ec i eness o ine con ex s (i.e., o 3 o 4 sha ed a ibu es) in he da ase s wi h a leas 4
a ibu es. These esul s a e in line wi h he hypo hesis ha con ex -unawa e speake s communica e
concep s on he mos speci ic le el in all con ex s, including coa se con ex s. This beha io can
be e e ed o as o e in o ma i e om he lis ene ’s pe spec i e. Fo example, in a coa se con ex
whe e no o he ci cles a e p esen , communica ing a speci ic concep like “ ed ci cle” is conside ed
o e in o ma i e.
Figu e 6.8: Con ex -unawa e: Mean en opy sco es ac oss all da ase s o di e en con ex condi ions indica ed by he
numbe o sha ed a ibu es. F om le o igh con ex becomes ine . E o ba s indica e boo s apped 95% con idence
in e als.
When agen s a e ained con ex -awa e, on he o he hand, we obse e ha he in o ma ion- heo e ic
sco es di e mo e be ween con ex condi ions (see Figu e 6.9). Speci ically, we obse e a pa e n
whe e he coa se he con ex (i.e., he ewe sha ed a ibu es), he lowe he NMI and he ine he
con ex (i.e. he mo e sha ed a ibu es), he highe he NMI. When agen s de elop ewe one- o-one
mappings be ween messages and concep s in he coa se con ex condi ions, his migh indica e
12
The p obabili y o di ec ion (pd) can be in e p e ed as he p obabili y ha a pa ame e ’s pos e io dis ibu ion is s ic ly
posi i e o nega i e (Makowski, Ben-Shacha , Chen, & Lüdecke, 2019).
13
The Region O P ac ical Equi alence wi h ze o (ROPE) was calcula ed by using one- en h o he s anda d de ia ion o
he esponse a iable a ound he null ollowing ecommenda ions by K uschke, 2018: ROPE = [-0.004, 0.004].
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 63
ha hey adap mo e o he con ex which makes one- o-one mappings imp ac ical. The eason o
his migh be ha in coa se con ex s, bo h mo e and less speci ic messages can be success ul (e.g.,
“ci cle” can mean ‘ ed ci cle’, ‘blue ci cle’ e c.) because when less speci ic messages a e used, he
a ge concep can s ill be disambigua ed by he con ex . In ine con ex s, on he o he hand, he
messages need o con ain mo e in o ma ion on mo e speci ic le els o abs ac ion o be su icien ly
disc imina i e in he con ex , which in ui i ely esul s in mo e one- o-one mappings (e.g., a mo e
speci ic u e ance like “ ed ci cle” is only used o he mo e speci ic concep ‘ ed ci cle’).
Figu e 6.9: Con ex -awa e: Mean en opy sco es ac oss all da ase s o di e en con ex condi ions indica ed by he
numbe o sha ed a ibu es. F om le o igh con ex becomes ine . E o ba s indica e boo s apped 95% con idence
in e als.
In line wi h hese obse a ions, we ind a subs an ial di e ence in NMI sco es be ween he con ex -
unawa e (M=0.95, C I=[0.94, 0.96]) and con ex -awa e se ing (M=0.89, C I=[0.87, 0.9]) only o
coa se con ex s wi h a di e ence in means o M=0.064 (C I=[0.046, 0.811], pd=100%, 0% in ROPE). Fo
ine con ex s, he di e ence in NMI sco es be ween he con ex -unawa e (M=0.95, C I=[0.94, 0.97])
and he con ex -awa e se ing (M=0.95, C I=[0.92, 0.97]) is no signi ican (M=0.008, C I=[-0.026,
0.041], pd=70.1%, 20% in ROPE).
Looking a e ec i eness and consis ency sco es in he con ex -awa e se ing, we obse e highe
consis ency and lowe e ec i eness sco es o coa se con ex s and highe e ec i eness and lowe
consis ency sco es o ine con ex s. This means ha agen s end o consis en ly use he same
messages o e e o he same concep s (i.e. no synonyms) in coa se con ex s and ha agen s end o
e ec i ely use messages ha uniquely iden i y he a ge concep (i.e. non-polysemous exp essions)
in ine con ex s. This makes sense because he ine he con ex ge s, he mo e i is necessa y o
dis inguish he a ge concep s om he dis ac o s.
6.7 Discussion
Wi h ou in e ac i e agen -based model, we we e able o gene a e h ee main insigh s abou concep
communica ion in a ious con ex s and how his se up shapes an eme ging language.
64 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
Fi s , we show ha a i icial agen s can lea n o communica e success ully abou concep s a di e en
le els o abs ac ion and in di e en con ex s in a concep -le el e e ence game. P e ious wo k has
explici ly encoded concep in o ma ion in he o m o ele ance ec o s (Ohme , Duda, & B uni,
2022) o p o o ype embeddings (Mu & Goodman, 2021). Fo humans, howe e , abs ac ing he
ele an concep , o le el o e e ence, happens wi hou such explici in o ma ion. He e, we show
ha agen s can lea n highe -le el concep s om he objec inpu s alone, p o iding a mo e na u al
model o he eme gence o abs ac ion.
Second, we ind ha only con ex -awa e agen s lea n o communica e e icien ly by adap ing hei
messages o he con ex condi ions. While con ex -unawa e agen s use he same messages o e e
o concep s in all con ex condi ions, con ex -awa e agen s adap hei messages success ully o he
con ex . O e in o ma i e communica ion, in he sense ha speci ic concep s a e communica ed
also in coa se con ex s whe e hey con ain mo e in o ma ion han necessa y o disambigua ion,
is educed in he con ex -awa e game scena io. This migh indica e ha con ex -awa e agen s
communica e mo e e icien ly (Pian adosi e al., 2012). I should be no ed, hough, ha hese
agen s do no sha e he same biases as humans. Fu u e wo k should ocus on he biases and
p essu es ha shape he eme ging language be ween a i icial agen s owa ds he kind o e icien
o e in o ma i e communica ion we o en obse e in humans (e.g., Degen e al., 2020; K eiss e al.,
2017; Rubio-Fe nandez, 2021; Tou ou i e al., 2019).
Thi d, we conclude ha he a ailabili y o con ex alone shapes he eme ging language owa ds
being mo e e icien (i.e. less o e in o ma i e) wi hou addi ional p essu es. The agen s we e no
explici ly incen i ized o use he con ex bu hey sha e he same a chi ec u e and aining p ocedu e
wi h he con ex -unawa e agen s, he only di e ence being ha hey also ecei e dis ac o objec s
as inpu . Because we ha e no incen i ized he con ex -awa e agen s o use con ex , hey could
ollow he same s a egy as con ex -unawa e agen s and be maximally speci ic all he ime. Ins ead,
we ind ha he agen s de elop a s a egy ha makes use o he con ex in which hey communica e.
Al hough he di e ences we obse e be ween he con ex -awa e and con ex -unawa e se ings
a e a he small, hey a e eliable and hey do indica e ha he me e p esence o con ex al eady
d i es i s use in communica ion. Fu u e wo k can in es iga e whe he p essu es, such as inc easing
cogni i e load o longe messages, would e en in ensi y hese di e ences.
Ou esul s a e in line wi h p e ious wo k on how an eme ging ocabula y depends on he con ex s
in which he a ge s a e p esen ed. Hawkins e al. (2018) ound a simila pa e n in an a i icial
language lea ning pa adigm wi h human pa icipan s: The ine he con ex , he mo e one- o-one
mappings a e es ablished in an eme ging language, and he coa se he con ex , he mo e synonyms
can be ound. Fu he , hey also ound ha an eme ging language con ains mo e wo ds ha e e o
only one concep and ewe ha e e o mo e han one concep when pa icipan s only encoun e
ine con ex s.
Ou modeling esul s add o his e idence and highligh he ole o con ex om a di e en angle.
We ea neu al ne wo k models as es beds o hypo heses on human cogni ion. He e, we show
ha con ex in i sel is a p essu e ha d i es e iciency in an e ol ing language. E en hough
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 65
ou neu al ne wo k agen s lack human cogni i e biases, hey de elop mo e e icien p o ocols
when hey can (bu do no ha e o!) access in o ma ion abou he con ex compa ed o when
hey canno . This inding demons a es ha he p esence o con ex alone may d i e aspec s o
p agma ic communica ion wi hou any addi ional p essu es and cogni i e p e equisi es. We can
ake his as e idence o he ole o ex e nal ac o s such as con ex o he eme gence o an e icien
communica ion sys em. In line wi h ha , Pian adosi e al. (2012) a gue ha ambigui y, as we
see i in he eme ging communica ion sys em in he con ex -awa e se ing, makes a language
e icien because i can usually be esol ed by con ex . Ou simula ions p o ide e idence o his
hypo hesis.
In conclusion, he he e p esen ed models and analyses con ibu e o ou unde s anding o e e en ial
communica ion and he ole o p agma ics in communica ing concep s h ough a sys ema ic
manipula ion o communica i e needs. Ou esul s show ha he speake ’s access o he con ex
shapes he eme ging communica ion sys em, ep oducing a pa e n ha was obse ed in humans
(e.g., Hawkins e al., 2018; Win e s e al., 2015,2018). These indings ha e implica ions bo h o
linguis ics esea ch wi h he ques ions o how human language e ol ed and how we make use
o language e icien ly, as well as o eme gen communica ion esea ch wi h he ques ion o how
we can build a i icial models ha communica e in a human-like way. Mo e gene ally, ou wo k
illus a es how language eme gence simula ions wi h neu al ne wo k agen s can be used o explo e
ques ions abou human cogni ion.
Acknowledgemen s
We hank h ee anonymous e iewe s o hei help ul commen s and eedback.
The simula ions we e un on a high-pe o mance compu ing clus e unded by he Deu sche
Fo schungsgemeinscha (DFG, Ge man Resea ch Founda ion) - 456666331. K is ina Kob ock is
suppo ed by he DFG- unded Resea ch T aining G oup “Compu a ional Cogni ion” (DFG-GRK
2340).
Au ho Con ibu ions:
K is ina Kob ock: Concep ualiza ion, Me hodology, So wa e, Valida ion, Fo mal analysis, In es-
iga ion, W i ing - O iginal D a , Visualiza ion. Xenia Ohme : Concep ualiza ion, Me hodology,
So wa e, W i ing - Re iew & Edi ing, Visualiza ion. Elia B uni: Concep ualiza ion, Me hodology,
W i ing - Re iew & Edi ing, Supe ision. Nicole Go zne : Concep ualiza ion, Me hodology, W i ing
- Re iew & Edi ing, Supe ision, P ojec Adminis a ion.
66 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
Appendix
A Accu acy sco es ac oss all da ase s
Table 6.4: Accu acy means o agen s ained in he con ex -unawa e and con ex -awa e se ing a e aged o e i e uns
wi h s anda d de ia ions.
Da ase s Condi ion Accu acy
aining alida ion es
D(3,4) con ex -unawa e 0.995 (0.002) 0.99 (0.003) 0.84 (0.036)
con ex -awa e 0.993 (0.003) 0.983 (0.004) 0.784 (0.035)
D(3,8) con ex -unawa e 0.993 (0.003) 0.989 (0.003) 0.778 (0.068)
con ex -awa e 0.984 (0.006) 0.977 (0.006) 0.686 (0.061)
D(3,16) con ex -unawa e 0.981 (0.007) 0.979 (0.008) 0.896 (0.005)
con ex -awa e 0.969 (0.005) 0.968 (0.006) 0.874 (0.007)
D(4,4) con ex -unawa e 0.992 (0.002) 0.989 (0.002) 0.922 (0.028)
con ex -awa e 0.995 (0.003) 0.993 (0.005) 0.942 (0.048)
D(4,8) con ex -unawa e 0.961 (0.011) 0.961 (0.011) 0.943 (0.012)
con ex -awa e 0.984 (0.004) 0.982 (0.006) 0.976 (0.007)
D(5,4) con ex -unawa e 0.98 (0.011) 0.979 (0.012) 0.964 (0.014)
con ex -awa e 0.985 (0.007) 0.984 (0.008) 0.979 (0.01)
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 67
B E o s ac oss all da ase s
B.1 E o s pe game ound
These plo s show he e o s on he alida ion da ase ac oss all da ase s o di e en concep (# ixed
a ibu es) and con ex condi ions (# sha ed a ibu es). Game ounds in which a leas one objec
was inco ec ly classi ied coun as e o s and a e no malized wi h he numbe o occu ences o he
speci ic condi ion in he da ase . This means ha a alue o 1.0 indica es ha lis ene s inco ec ly
classi ied a leas one objec in each game ound in his condi ion.
Figu e 6.10: Con ex -unawa e: Mos e o s occu on he diagonal om op le o bo om igh , i.e. in he ines possible
con ex condi ions.
As can be seen in he igu es, e o s occu mos ly in ine con ex condi ions, i.e. whe e he maximally
possible numbe o a ibu es is sha ed be ween a ge s and dis ac o s. Some o hese e o s a e
alse posi i es, i.e. dis ac o s a e inco ec ly classi ied as a ge s, and some o hese e o s a e alse
nega i es, i.e. a ge s a e inco ec ly classi ied as dis ac o s. We ind ha alse nega i e e o s
occu mainly wi h mo e gene ic concep s and ine con ex s. This is p obably due o he a ge
concep s being e y he e ogenous and hus, ha de o disc imina e agains dis ac o s. False posi i e
e o s, on he o he hand, occu in he ines con ex s when he concep is e y speci ic. This can be
explained by he dis ac o s being e y simila o he a ge s in hese condi ions. In o he wo ds,
alse posi i e e o s migh indica e ha he lea ned a ge concep is a bi oo wide, and alse
nega i e e o s migh indica e ha he lea ned a ge concep is a bi oo na ow.
74 6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s
6.8 Summa y o he key con ibu ions
In his case s udy, we p o ided e idence o he ole o p agma ics in he e olu ion o a commu-
nica ion sys em. To s udy his e olu ion, we employed an agen -based eme gen communica ion
modeling pa adigm aimed a achie ing maximal con ol and sys ema ici y. We in oduced ca e ully
cons uc ed symbolic da ase s ha enabled us o manipula e bo h concep ual and con ex ual
in o ma i i y. Speci ically, we in es iga ed how agen s communica e a ge concep s wi h a ying
speci ici y anging om speci ic o gene ic, in con ex s wi h a ying g anula i y anging om ine
o coa se. Going beyond p e ious wo k (Mu & Goodman, 2021; Ohme , Duda, & B uni, 2022), we
demons a ed ha agen s can asce ain a ge concep s based solely on objec -based inpu s, wi hou
any addi ional explici in o ma ion. Ou key inding om compa ing speake s wi h access o he
con ex o speake s wi hou access o he con ex was ha he me e p esence o con ex d i es
i s use in communica ion. We u he ound ha communica ion s a egies di e depending on
whe he speake s had access o he con ex du ing aining o no . Con ex -awa e agen s end o a y
hei messages mo e in esponse o con ex . They use a b oade se o messages in coa se con ex s,
whe eas in ine con ex s, he mos success ul s a egy is o communica e he mos speci ic concep .
This sugges s ha in coa se con ex s, speake s a e mo e likely o come up wi h mul iple names o
he same concep . Con ex -unawa e speake s, in e u n, need o communica e all a ibu es ha a e
ele an o he a ge concep dis ega ding he con ex . While his can be conside ed an e ec i e
seman ic solu ion, i may also be o e in o ma i e om he lis ene ’s pe spec i e, who in e p e s he
messages wi hin con ex . Al hough hese con ex -unawa e languages a e sensi i e o he le els o
he concep ual hie a chy anging om speci ic o gene ic, hey lack con ex -dependen di e ences
in e e ence. In o he wo ds, he e, agen s do no come up wi h mul iple names o he same concep ,
bu a he use di e en names o each concep . We conclude ha he p agma ic ac o o con ex
plays a c ucial ole in shaping he eme gen communica ion sys em, con ingen on i s a ailabili y o
he speake .
6.9 Implica ions o he b oade esea ch ques ion
We in oduced a compu a ional model o in es iga e he eme gence o language du ing in e ac i e
communica ion abou concep s. This model will also be used in he ollowing wo case s udies.
One goal o his s udy was o in oduce he model and da ase , demons a ing how hey help us
in es iga e he ole o p agma ic ac o s on language e olu ion. Ou esul s indica e ha a i icial
agen s can lea n o communica e abou he hie a chically o ganized concep s in ou symbolic
da ase s. No ably, his is achie ed wi hou any explici encoding o in o ma ion abou he concep ual
hie a chy. We hus mo e beyond p e ious esea ch whe e such explici in o ma ion was p o ided
in he o m o ele ance o p o o ype ec o s (Mu & Goodman, 2021; Ohme , Duda, & B uni, 2022).
We can conclude om hese esul s ha he concep -le el e e ence game is an e ec i e ask o
highligh ing he concep ual s uc u e o he inpu da a. This is consis en wi h he indings om case
s udy 1, in which we ound ha he concep -le el e e ence game enables pa icipan s o a e se he
6 Case s udy 2: The ole o con ex a ailabili y in eme gen communica ion abou concep s 75
concep ual hie a chy and lexibly selec app op ia e e e ing exp essions om di e en le els o
his hie a chy. The second goal o his s udy was o in es iga e he ole o p agma ics in he e olu ion
o a communica ion sys em. We showed ha he a ailabili y o con ex in luences how concep s a
di e en le els o abs ac ion a e e e ed o in a language. P e ious expe imen al wo k using an
a i icial language lea ning pa adigm has also p o ided e idence o he ole o con ex in shaping
communica ion sys ems (Hawkins e al., 2018; Win e s e al., 2015,2018). Howe e , ou esea ch
highligh s ha e en wi hou explici incen i es o u ilize con ex , i s a ailabili y d i es i s use in
communica ion, he eby shaping he eme ging language. Fu he mo e, we p o ided ini ial e idence
ha languages which eme ged when con ex was a ailable a e mo e e icien han languages which
eme ged when con ex was no a ailable o he speake s. We will ollow up on hese indings in
case s udy 3. Addi ionally, we disco e ed ha he concep ual hie a chy and he con ex g anula i y
a e e lec ed di e en ly in p ope ies o he language, depending on he con ex a ailabili y. In he
con ex -unawa e condi ion, languages mo e closely e lec he concep ual s uc u e. We can ela e
his o he seman ics o a language whe e he lexicon consis s o mappings be ween concep s and
wo ds. Adding p agma ics in he o m o con ex in he con ex -awa e condi ion, leads o languages
e lec ing he con ex g anula i y. This co esponds o a language whe e e e ing exp essions can be
lexibly and p agma ically adap ed o di e en con ex s. Impo an ly, all obse a ions a e eme gen
ea u es o he model, a ising om he s uc u e o he inpu da a and he communica i e needs
du ing in e ac ions, wi hou any explici incen i es o he agen s o encode concep speci ici y
o con ex g anula i y. Thus, we conclude ha he dis inc i e linguis ic p ope ies we obse ed
eme ged om he en i onmen and communica i e needs.
7
Case s udy 3: The ole o p agma ic mechanisms
in language use and he eme gence o linguis ic
and concep ual s uc u e
This chap e p esen s case s udy 3. The chap e s a s wi h a high-le el in oduc ion ollowed by
he con en o he publica ion: Kob ock, K., Ohme , X., B uni, E., & Go zne , N. (2025b). The ole o
p agma ic mechanisms in e e en ial communica ion and ca ego iza ion: An eme gen communica ion model.
[unde e iew, PsyA Xi ]. h ps://doi.o g/10.31234/os .io/kb ua_ 1 The chap e ends wi h a b ie
summa y o he main con ibu ions o he publica ion and a discussion o i s implica ions o he
b oade esea ch ques ion o his disse a ion.
7.1 High-le el in oduc ion
Case s udy 2 p o ided ini ial e idence o he ole o language and p agma ics in concep ual
abs ac ion and, speci ically, in shaping an e ol ing language ha is used o communica ing
concep s a di e en le els o abs ac ion. The goal o case s udy 3 is o in es iga e his ole u he .
We explo e he possibili y o a join e olu ion o language and ca ego y sys ems h ough wo
expe imen s. In Expe imen 1, we aim o in es iga e he in luence o a sha ed con ex on language
eme gence. In Expe imen 2, we in es iga e he ole o di e en p agma ic mechanisms on language
use in conc e e e e ence si ua ions. Fo his pu pose, we change ou main aspec s wi h espec o
he model used in he p e ious case s udy: Fi s , we in oduce a sha ed con ex be ween speake s
and lis ene s ha is mo e in o ma i e han he con ex modeled in case s udy 2. In he p e ious
model, only he numbe o a ibu es sha ed be ween a ge and dis ac o objec s was conside ed
when cons uc ing he con ex . He e, speake s and lis ene s see he a ge concep s in he same
con ex , ensu ing ha he a ibu es ha a e sha ed wi h he a ge concep a e he same o bo h
in e locu o s. Fo ins ance, he speake needs o dis inguish small blue iangles om small ed
iangles. In his example, he a ge concep has h ee a ibu es ha need o be communica ed,
small, blue, and iangle. I is p esen ed in a ine con ex , whe e wo o he a ibu es, small and
iangle, a e sha ed be ween a ge s and dis ac o s. The disc imina i e a ibu e is hus he colo .
Likewise, he lis ene needs o disc imina e small blue iangles om small iangles wi h a
di e en colo , o example small g een iangles. This alignmen ensu es ha he same a ibu es
a e ele an o communica ion. We hypo hesize ha in oducing his sha ed con ex will make
he con ex mo e use ul o communica ion in compa ison o case s udy 2, and his will, in u n,
impac he eme ging language. Second, o enhance he eme gence o e icien ca ego y and linguis ic
sys ems, we add a leng h cos p essu e o he aining objec i e. This p essu e penalizes longe
messages by adding a cos o he aining loss. Consequen ly, speake -lis ene pai s ecei e a highe
ewa d o in e ac ions in which speake s use a sho e message han o hose in which speake s
78 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
use a longe message, p o ided he communica i e success, ha is, he lis ene ’s co ec selec ion
o he a ge objec s, emains he same. We hypo hesize ha he leng h cos p essu e will make
he eme ging language mo e e icien . Agen s migh be mo e inclined o communica e only he
necessa y a ibu es o a concep when communica ion is mo e cos ly. Thi d, in Expe imen 2,
we e alua e no only he mappings be ween messages and concep s ha eme ged as a unc ion
o aining, bu we e alua e language use a e aining. This is done by es ing he agen s on
a es da ase consis ing o held-ou concep s. In es iga ing how agen s use he languages hey
de eloped du ing epea ed in e ac ions wi h hei in e locu o s p o ides a new window o analysis
o ou esea ch ques ion. By in eg a ing language use and language eme gence, we can accoun o
p agma ic ac o s on bo h e olu iona y and si ua ion-based ime scales. Speci ically, in es iga ing
language use in conc e e e e ence si ua ions will be o cen al impo ance o d awing in e ences
om he linguis ic o he ca ego y sys em and will help us analyze how concep s a di e en le els o
abs ac ion a e communica ed in e e ence-based si ua ions. Finally, we add wo new expe imen al
condi ions o manipula e no only speake s’ access o con ex bu also hei abili y o conside he
expec ed u ili y o hei u e ance by easoning abou he lis ene ’s likely in e p e a ion o each
po en ial u e ance. This is modeled ia he RSA amewo k in oduced in Sec ion 3.2.RSA speake s
aim o maximize he u ili y o hei u e ances, which is in luenced by (1) he likelihood o he
lis ene selec ing he co ec a ge objec s, and (2) he cos associa ed wi h an u e ance in e ms o
i s leng h, whe e sho e messages yield highe u ili y. We expec agen s in Expe imen 1 o de elop
mo e e icien languages in he con ex -awa e compa ed o he con ex -unawa e condi ion consis en
wi h case s udy 2. Rega ding he compa ison o language use in Expe imen 2, we expec ha bo h
con ex a ailabili y and he a ailabili y o RSA-like easoning abou he u e ances expec ed u ili y,
will lead o speake s making mo e e ec i e and e icien use o hei language.
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 79
7.2 Abs ac
We model p agma ic mechanisms o e e en ial communica ion and ca ego iza ion in a mul i-agen
amewo k o eme gen communica ion. P agma ic heo ies and expe imen al wo k p edic ha
speake sconside he con ex in hei choiceo e e ingexp essions. Inaddi ion o his con ex -based
easoning, u ili y-based p agma ic easoning abou he lis ene ’s likely in e p e a ion o an u e ance
in luences he speake ’s p oduc ion choices. We aim o in es iga e hese wo ac o s and hei
ole in e e ing exp ession gene a ion and ca ego iza ion in a compu a ional model o language
eme gence and language use. We model communica ion in in e ac ion and conside an e iciency
adeo be ween speake and lis ene u ili ies. Ou esul s show ha an eme ging language becomes
mo e e ec i e, ambiguous, and e icien when speake s and lis ene s communica e in a sha ed
con ex . This is achie ed by an e icien adeo be ween p oduc ion and comp ehension whe e
languages can a o d o be simple in p oduc ion when hey a e su icien ly in o ma i e in con ex ,
placing mo e bu den on he lis ene ’s side. We u he demons a e ha inco po a ing u ili y-based
p agma ics, as modeled wi h he Ra ional Speech Ac s amewo k, imp o es he linguis ic e iciency
o language use only in languages ha eme ged wi h a sha ed con ex be ween in e locu o s, bu no
in languages ha eme ged wi hou such con ex ual in o ma ion du ing aining. We conclude ha
con ex -based p agma ics plays a ole in e e en ial communica ion and ca ego iza ion by shaping
an eme ging language. E icien e e ence in a communica i e si ua ion can bene i especially om
u ili y-based p agma ics i he language ha is being used has eme ged in con ex . This migh
sugges ha u ili y-based p agma ics hinges on mechanisms ha na u ally eme ge when con ex is
a ailable du ing he e olu ion o a language. In summa y, we show ha human-like language and
ca ego y sys ems eme ge as an op imized adeo be ween speake and lis ene needs in in e ac ion,
i.e. unde e iciency conside a ions o simplici y and in o ma i eness.
Au ho summa y
This s udy in es iga es p agma ic mechanisms o e e en ial communica ion and ca ego iza ion in a
mul i-agen amewo k. We s udy how speake s and lis ene s in e ac o c ea e an e icien language,
balancing he need o clea communica ion wi h he cos o p oducing messages. Ou esul s show
ha an eme ging language becomes mo e e ec i e, ambiguous, and e icien when speake s and
lis ene s communica e in a sha ed con ex . This mechanism o con ex -based p agma ics plays a
c ucial ole in shaping he language, allowing speake s o be simple in p oduc ion when lis ene s
can in e he co ec meaning in con ex . Addi ionally, we disco e ed ha u ili y-based p agma ics
modeled wi h he Ra ional Speech Ac s amewo k impac s language e iciency, pa icula ly
when he language has eme ged in con ex . Ou indings sugges ha u ili y-based p agma ics is
closely ied o he mechanisms ha na u ally eme ge when con ex is a ailable du ing language
e olu ion. The esul ing language and ca ego y sys ems esemble hose o human languages. This
esea ch con ibu es o ou unde s anding o how e icien communica ion sys ems eme ge in social
in e ac ions, and based on communica i e need.
80 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
7.3 In oduc ion
The languages we speak in luence which ca ego ies we know and alk abou . The idea ha language
shapes cogni ion in his way has become popula as he (weak) Sapi -Who hypo hesis, o linguis ic
ela i i y hypo hesis (Sapi , 1912; Who , 1956). One ins ance o he claim ha language shapes
cogni ion is ha con e sa ion shapes ca ego y s uc u e (Ba & K onmülle , 2006). A amous
example is he popula ized claim om Boas (1911) and Sapi (1912) ha Inui
1
ha e mo e wo ds o
subca ego ies o snow han languages spoken in wa me clima es. This claim has been ecen ly
e isi ed in a la ge-scale analysis o linguis ic and me eo ological da a (Regie e al., 2016). The
au ho s ound ha languages ha use he same linguis ic o m o bo h snow and ice end o be
spoken in wa me coun ies. They ela e his inding o he lowe communica i e need o alk abou
snow and ice in hese egions (Regie e al., 2016). This is in line wi h he e icien communica ion
hypo hesis which p edic s ha human languages a e op imized o e icien communica ion, a claim
ha has gained ac ion in many sub ields o linguis ics such as language e olu ion (Kanwal e al.,
2017; Win e s e al., 2018), p agma ics (K eiss e al., 2017; Peloquin e al., 2020; Rubio-Fe nandez,
2021; Rubio-Fe nandez e al., 2021) and mo e gene al p oposals (Gibson e al., 2019; Gualdoni &
Boleda, 2024; Le shina, 2018; Pian adosi e al., 2012; Regie e al., 2016; Zip , 1949).
Following he e icien communica ion hypo hesis, he main idea o how language shapes ca ego y
s uc u e is ha languages ha e wo ds o hings ha a e ele an , i.e. whe e he e is a commu-
nica i e need (Kemp & Regie , 2012; Regie e al., 2016). This can be exempli ied by he ollowing
ela ionship (Regie e al., 2016):
En i onmen →Communica i e Need →Ca ego y Sys em (7.1)
I he e is a communica i e need o alk abou subca ego ies in an en i onmen , hen he ca ego y
sys em will be mo e ine-g ained han i he e is no such need (Regie e al., 2016). Mo e ine-g ained
ca ego ies a e mo e in o ma i e han b oade ca ego ies, bu hey a e also mo e complex and
equi e mo e s o age space in memo y. I has been p oposed ha , due o cogni i e cons ain s,
bo h languages and ca ego y sys ems end o ind an op imal adeo be ween simplici y and
in o ma i eness (Kemp & Regie , 2012; Pian adosi e al., 2012; Regie e al., 2016; Rosch, 1978; Rosch
e al., 1976; Win e s e al., 2018; Zip , 1949).
How his adeo is op imized can be be e unde s ood when conside ing in e ac i e commu-
nica ion. Le ’s assume ha a lexical sys em e ol es in a simple con e sa ional se up in which a
speake desc ibes an objec o a lis ene . I his sys em was op imized o he speake , hen in an
ex eme case, he sys em would only include one wo d ha can be used o desc ibe all objec s. This
op imizes simplici y and keeps memo y s o age o he speake o a minimum. A sys em ha is
op imized o he lis ene , on he o he hand, includes one label o each objec . This op imizes
in o ma i i y o he lexicon and i makes i easy o he lis ene o iden i y he co ec a ge when
hea ing a wo d. This p inciple was p oposed as he P inciple o Leas E o by Zip (1949) and he
1
The ea ly wo ks and deba e he ea e ha e equen ly used he e m ‘Eskimo’ which is la gely disp e e ed by he
na i e popula ions o Alaska due i s colonial he i age (Kaplan, 2025).
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 81
idea has been also aken up la e (Le inson, 2000; Pian adosi e al., 2012). Rela edly, Rosch e al.
(1976) and Rosch (1978) ha e p oposed a simila p inciple o ca ego y s uc u e: The P inciple o
Cogni i e Economy s a es ha a ca ego y sys em should p o ide maximum in o ma ion wi h he
leas cogni i e e o . An op imal ca ego y sys em would hus op imize a adeo be ween ha ing
many ca ego ies wi h e y ine-g ained dis inc ions be ween hem and ha ing ewe ca ego ies
wi h coa se dis inc ions be ween hem. A ca ego y sys em wi h many ca ego ies is op imized o
in o ma i i y: Many p ope ies o one ca ego y can be p edic ed by knowing only one p ope y,
e.g. he label. A ca ego y sys em wi h ew ca ego ies is op imized o cogni i e economy because
ewe ca ego ies need o be lea ned and s o ed in memo y (Rosch, 1978; Rosch e al., 1976). This
adeo can also be hough o as a adeo be ween speake and lis ene needs in a communica i e
si ua ion, speci ically in a e e en ial communica i e si ua ion. Ba and K onmülle ha e a gued
ha e e en ial communica ion is especially c i ical o ca ego iza ion, speci ically ca ego y lea ning
and ca ego y use (Ba & K onmülle , 2006). Following his, we use e e en ial communica ion o
s udy ca ego iza ion in a communica i e se ing whe e speake s and lis ene s in e ac . This allows
us o s udy communica i e and ca ego iza ion e iciency as a adeo be ween speake and lis ene
needs.
The ques ion ha we ackle in his pape is he ollowing: Can he e icien communica ion hypo hesis
p edic ca ego y s uc u e? E idence o his has been ound in he con ex o kinship (Kemp &
Regie , 2012) and colo (Fide & Koma o a, 2019; Ohme , Ma ino, e al., 2022; Zasla sky e al., 2018,
2022) sys ems. Howe e , his hypo hesis has no ye been es ed o mo e gene al ca ego ies o e en
o he mos ubiqui ous ca ego y sys em ha we use: How we ca ego ize he hings a ound us in
hie a chical s uc u es ha a e also called le els o abs ac ion (Rosch e al., 1976), e.g. dalma ians,
dogs and animals. To s udy ca ego iza ion, we look a he e e en ial communica ion o ca ego ies
a di e en le els o abs ac ion. We use a modeling app oach wi h con olled symbolic da ase s
ha include hie a chical ca ego ies.
One cen al puzzle o heo ies o e icien communica ion is how o accoun o ambigui y and
edundancy obse ed in human language (Degen e al., 2020; Pian adosi e al., 2012; Wasow,
2015). Ambigui y in human languages has been ela ed o p agma ic conside a ions (Achimo a
e al., 2022; Fe ei a, 2008; Pian adosi e al., 2012). P agma ic heo ies assume ha language is
in e p e ed in con ex and in a e e en ial si ua ion whe e speake s use an unde in o ma i e o
ambiguous exp ession, con ex usually helps o disambigua e he in ended meaning (Pian adosi
e al., 2012; Wasow, 2015). Con ex -based p agma ics can hus econcile ambigui y wi h he e icien
communica ion hypo hesis. Bu , humans ha e also been shown o equen ly o e speci y e e ing
exp essions wi h a ypical o salien p ope ies such as colo e en in si ua ions whe e hese p ope ies
a e no needed o unambiguously iden i y a a ge e e en in a gi en con ex . On i s sigh , his is a
odds wi h he e icien communica ion hypo hesis and G icean maxims o con e sa ion, especially
he quan i y maxim which s a es ha speake s should p o ide as much in o ma ion as needed
and no mo e (G ice, 1975). Bu ecen p oposals ha e shown ha he edundan exp essions
p o ide in o ma ion ha helps lis ene s o iden i y he a ge e e en , e.g. by men ioning a ypical
ea u es o p o iding salien in o ma ion in complex scenes ha acili a e isual sea ch (Degen
82 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
e al., 2020; K eiss e al., 2017; Rubio-Fe nandez, 2021; Rubio-Fe nández, 2016; Tou ou i e al., 2019).
This can be explained by he idea ha speake s ake in o accoun he lis ene ’s needs and wan
o guide he lis ene owa d iden i ying he co ec e e en as quickly as possible (Degen e al.,
2020; Tou ou i e al., 2019). Reasoning abou he lis ene ’s likely in e p e a ion is a classic case o
u ili y-based p agma ics and i can be s aigh o wa dly modeled wi h he Ra ional-Speech-Ac
amewo k (Degen e al., 2020; F ank & Goodman, 2012; F anke & Degen, 2016). U ili y-based
p agma ics can hus explain how edundancy can be conside ed e icien and econcile edundancy
wi h he e icien communica ion hypo hesis.
The main objec i e o his wo k is o ind ou how ca ego iza ion in luences an eme ging language
and wha ole language and p agma ics play in ca ego iza ion. The goals o ou s udy a e hus
wo old: Fi s , we in es iga e he in luence o ou manipula ions on he (use o he) eme ging
language. We ask whe he e icien communica ion, o malized as a adeo be ween speake
and lis ene needs, eme ges om in e ac ion and p agma ic conside a ions, when communica ing
concep s a di e en le els o abs ac ion. We use an eme gen communica ion pa adigm o model
he eme gence o ca ego ies and ca ego y labels in an in e ac i e communica i e amewo k ha
akes in o accoun speake and lis ene needs. Second, we d aw in e ences om he eme ged
linguis ic sys em o he ca ego y sys em and analyze how concep s a di e en le els o abs ac ion
a e communica ed in he language. We use he ained eme gen communica ion models and
es hem in speci ic communica i e si ua ions and unde conside a ion o di e en p agma ic
mechanisms, such as con ex -based and u ili y-based p agma ics.
7.3.1 Modeling app oach
The o e a ching goal o his esea ch is o model con ex -based and u ili y-based p agma ic
mechanisms in a communica i e in e ac ion ha in ol es ca ego iza ion, and o ind ou unde
which ci cums ances hey become especially use ul and make he (use o he) eme ging language
pa icula ly e icien . Ou esea ch makes use o an eme gen communica ion model (Kob ock,
Ohme , e al., 2024; Laza idou e al., 2017,2018; Ohme , Duda, & B uni, 2022). The a chi ec u e o
he model is isualized in Fig7.1 and implemen s ou communica i e p inciples ha a e ou lined
below.
Fi s , he model is in e ac i e, i.e. i akes in o accoun and models he ac ha communica ion
happens in in e ac ion be ween a speake and a lis ene agen . This means ha ou model should
ha e a leas a basic one-way in e ac ion be ween a speake who sends a message and a lis ene who
ecei es a message. The model ha we use u ilizes such a basic in e ac ion which is also u ilized in
s anda d linguis ic heo e ical and expe imen al amewo ks (F ank & Goodman, 2012; Hawkins
e al., 2018).
Second, he model should make as li le p io assump ions abou communica ion as possible.
The e o e we do no use a model o language use on an exis ing language, o example English, bu
a he le a p o o-language eme ge in in e ac ion be ween a i icial agen s. The goal is o obse e
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 83
Speake
neu al ne wo k
(GRU)
Lis ene
neu al ne wo k
(GRU)
message
[4,11,7,0]
p edic ion and aining
1
2
3
4
Figu e 7.1: Model a chi ec u e.
wha a language migh look like gi en he cons ain s we implemen . I in ou agen -based models
we obse e ha a language eme ges ha sha es ce ain cha ac e is ics wi h human language, we
can conclude ha hese cha ac e is ics can eme ge e en in simple neu al-ne wo k-based models ha
implemen some cen al cha ac e is ics o communica ion, such as a e e ence-based in e ac ion, bu
do no implemen o he uniquely human cons ain s, such as language uni e sals o a pa icula
g amma (Chaabouni e al., 2021).
Thi d, he model akes in o accoun ha communica ion should be g ounded. Fo g ounded
communica ion, i is c ucial ha communica ion is si ua ed in a wo ld - his can be a oy wo ld o
keep he complexi y o he model small. This is done by simpli ying communica ion o e e ing.
The agen s need o sol e a e e ence game whe e he speake has o desc ibe a a ge and he lis ene
has o selec he a ge om a se o dis ac o s. The eme ging language is a meaning ul mapping
be ween messages and objec s o he wo ld ha a e being e e ed o. This simple ask and se ing
ha e an addi ional ad an age: We ha e ull con ol o e he en i onmen , i.e. we can cons uc he
p ope ies o he objec s and he eby con ol o e e y pa ame e o he model: The objec s ha
should be e e ed o and he con ex in which he e e ence akes place. In o he wo ds, when
con olling o he a ge s and he con ex , we can con ol o how much in o ma ion needs o be
communica ed o achie e success ul communica ion. The amoun o in o ma ion ha needs o be
communica ed ela es closely o e iciency.
Fou h, he model should allow us o in es iga e ca ego ies a di e en le els o abs ac ion. To
achie e hisgoal,weimplemen aconcep -le el e e encegame ha he agen s need osol e(Kob ock,
Ohme , e al., 2024; Kob ock, Uhlemann, & Go zne , 2024; Mu & Goodman, 2021). The concep -le el
e e ence game is simila o a simple e e ence game. The only di e ence is ha ins ead o a single
a ge objec , a a ge concep needs o be communica ed. This a ge concep is cons uc ed by
combining se e al a ge objec s. Fo example, i has been shown in expe imen al wo k ha when
speake s obse e wo a ge s, a pa o and a dog, hey use he supe o dina e e e ing exp ession
“animal” o e e o bo h (Kob ock, Uhlemann, & Go zne , 2024). The concep -le el e e ence game
ensu es ha he agen s in ou model communica e abou concep s a di e en le els o abs ac ion
90 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
(a) Con ex -unawa e (b) Con ex -awa e wi h sha ed con ex
Figu e 7.4: Message leng hs pe concep hie a chy le el. Concep speci ici y, o he amoun o in o ma ion ha needs o
be communica ed, inc eases wi h he numbe o ixed a ibu es ha is sha ed among a ge objec s om mo e gene ic
concep s (wi h ewe ixed a ibu es) on he le , o mo e speci ic concep s (wi h mo e ixed a ibu es) on he igh .
Bayesian hie a chical models p edic ing he espec i e sco e (NMI, e ec i eness, consis ency)
by he condi ion. We ound a main e ec o condi ion in all h ee models, i.e. p edic ing NMI
(M=0.14, C I=[0.13, 0.16], pd=100%, ROPE=[-0.02, 0.02], 0% in ROPE), p edic ing e ec i eness
(M=0.22, C I=[0.20, 0.25], pd=100%, ROPE=[-0.02, 0.02], 0% in ROPE), and p edic ing consis ency
(M=0.04, C I=[0.03, 0.05], pd=100%, ROPE=[-0.01, 0.01], 0% in ROPE). This means ha he amoun
o one- o-one mappings be ween messages and concep s was lowe , and ha he eme ging language
con ained mo e synonyms and polysemous messages, in he con ex -awa e condi ion compa ed o
he con ex -unawa e baseline.
Table 7.3: Mean en opy-based sco es, i.e. NMI, e ec i eness and consis ency.
Con ex -unawa e Con ex -awa e sha ed con ex
Da ase s NMI e ec i eness consis ency NMI e ec i eness consis ency
(3,4) 0.81 ±0.01 0.90 ±0.05 0.74 ±0.01 0.59 ±0.03 0.52 ±0.04 0.68 ±0.01
(3,8) 0.87 ±0.00 0.99 ±0.00 0.77 ±0.00 0.58 ±0.01 0.47 ±0.02 0.75 ±0.00
(3,16) 0.86 ±0.04 0.83 ±0.10 0.90 ±0.07 0.59 ±0.09 0.48 ±0.11 0.77 ±0.01
(4,4) 0.85 ±0.00 0.97 ±0.02 0.77 ±0.01 0.54 ±0.03 0.44 ±0.04 0.69 ±0.01
(4,8) 0.86 ±0.02 0.89 ±0.05 0.84 ±0.05 0.55 ±0.06 0.45 ±0.07 0.73 ±0.01
(5,4) 0.83 ±0.08 0.88 ±0.16 0.79 ±0.01 0.51 ±0.02 0.41 ±0.02 0.69 ±0.01
Mean en opy sco es and s anda d de ia ions o e i e uns a e calcula ed o each da ase and
bo h condi ions.
How is i ha he con ex -awa e agen s had such an ambiguous language bu s ill achie ed e y
high pe o mance? We looked a en opy-based sco es calcula ed o each le el o he concep ual
hie a chy o add ess his ques ion. Fig 7.5A-B shows he in o ma ion- heo e ic sco es compa ing
he con ex -unawa e o he con ex -awa e condi ion. The e ec i eness sco e is closely ela ed
o communica i e success, i.e. how likely he lis ene is o selec he co ec a ge objec s. I
e ec i eness is high, his means ha meaning unce ain y is low. E ec i eness is gene ally highe in
he con ex -unawa e han in he con ex -awa e condi ion as shown in Table 7.3 and in Fig 7.5A-B.
In e es ingly, in bo h con ex -awa e and con ex -unawa e condi ions, gene ic concep s wi h only
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 91
one ixed a ibu e we e e e ed o almos wi h maximal e ec i eness. Thus, he e was no meaning
unce ain y o messages ha we e used o e e o gene ic concep s, i.e. hey e ec i ely singled ou
he a ge s. The e ec i eness sco e had he endency o d op he mo e speci ic a concep was which
means ha he e we e mo e polysemous messages and meaning unce ain y he mo e speci ic a
concep . While in he con ex -unawa e condi ion he e ec i eness sco e d opped g adually and
had a linea end, in he con ex -awa e condi ion, he e ec i eness sco e d opped exponen ially
wi h each le el o he concep ual hie a chy. The consis ency sco e can be iewed as a measu e o o
he speake ’s endency o selec he same message o a gi en a ge concep in epea ed i e a ions.
I he consis ency sco e is high, his means ha he e is low signal unce ain y. In o he wo ds,
he speake ’s choice o message is s aigh o wa d and easy. In he con ex -unawa e condi ion,
we obse ed highe consis ency, i.e. lowe signal unce ain y, he mo e speci ic he concep s. In
he con ex -awa e condi ion, we obse ed lowe consis ency, i.e. highe signal unce ain y, he
mo e speci ic he concep s. The NMI cu es o bo h condi ions e lec ed he desc ibed ends:
In he con ex -unawa e condi ion, he NMI sco e was mo e closely o ien ed o he consis ency
sco e, i.e. he e we e mo e one- o-one mappings he mo e speci ic a concep . In he con ex -awa e
condi ion, NMI ollowed he e ec i eness cu e and dec eased wi h inc easing numbe o ixed
a ibu es, sugges ing ha he e we e ewe one- o-one mappings be ween messages and concep s
wi h speci ic han wi h gene ic concep s. The abo e desc ibed ends we e suppo ed by h ee
hie a chical Bayesian models p edic ing he espec i e sco es by he in e ac ion o condi ion and
concep hie a chy le el, i.e. ixed a ibu es. We ound a main e ec o condi ion and concep
hie a chy le el, as well as an in e ac ion e ec o hese wo p edic o s in models p edic ing he
NMI, e ec i eness and consis ency (see Table 7.9 in S2 Appendix: S a is ical models).
(a) Con ex -unawa e (b) Con ex -awa e wi h sha ed con ex
Figu e 7.5: NMI, consis ency and e ec i eness sco es o each le el o he concep ual hie a chy.
Discussion
In summa y, we obse ed ha a mo e ambiguous and e icien language eme ged in he con ex -
awa e condi ion han in in he con ex -unawa e condi ion which is in line wi h ou p edic ions. This
inding was suppo ed by measu es o communica i e success (accu acy), o e iciency (message
leng h), and o ambigui y (en opy-based sco es). We ound ha communica i e success on he
alida ion da ase was subs an ially highe in he con ex -awa e han in he con ex -unawa e
92 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
condi ion. Messages we e subs an ially sho e in he con ex -awa e han in he con ex -unawa e
condi ion. Toge he wi h he high pe o mance, his means ha he languages ha eme ged in he
con ex -awa e condi ion a e likely simple which makes hem well in e p e able and well sui ed
o gene aliza ion pu poses. When in e p e ing he en opy-based sco es we ob ained o each
language, i was e iden ha he languages ha eme ged in he con ex -awa e condi ion we e mo e
ambiguous. We ound ewe one- o-one mappings be ween messages and concep s shown by he
subs an ially lowe NMI sco e in he con ex -awa e compa ed o he con ex -unawa e condi ion.
The subs an ially lowe e ec i eness sco es in he con ex -awa e condi ion mean ha he e is high
meaning unce ain y and ha lis ene s ha e o choose he mos likely in e p e a ion om mo e
han one candida e in e p e a ion. Toge he wi h he high accu acies, his sugges s ha lis ene s
success ully make sense o he polysemous messages hey ecei e. In he con ex -unawa e condi ion,
e ec i eness sco es we e qui e high, sugges ing ha he messages ha he lis ene s ecei e we e no
ambiguous and i is hus s aigh o wa d o he lis ene s o in e p e he messages and iden i y he
co ec a ge s. The consis ency sco es sugges ed ha he e is a ce ain amoun o signal unce ain y,
o synonymous messages, in he eme ging languages o bo h condi ions wi h li le, bu subs an ially,
mo e synonymous messages in he con ex -awa e han in he con ex -unawa e condi ion. The
e ec i eness and consis ency sco es can also p o ide insigh in o how he languages eme ged as
a adeo be ween speake and lis ene needs. The e icien communica ion hypo hesis p edic s
ha an e icien language o he speake will ha e as ew dis inc messages as possible ha can be
used in mul iple con ex s. This ansla es in o a high consis ency sco e, i.e. low meaning unce ain y.
An e icien language o he lis ene , on he o he hand, will ha e many dis inc messages because
ha makes in e p e a ion easie . This ansla es in o a high e ec i eness sco e, i.e. low meaning
unce ain y. In he con ex -unawa e condi ion, we obse ed highe e ec i eness han consis ency.
This means ha he eme ging languages in his condi ion we e mo e op imal o lis ene s han o
speake s. The languages ha e messages ha can be in e p e ed unambiguously which makes he
lis ene ’s ask especially easy. In o he wo ds, he adeo be ween p oduc ion and comp ehension
is balanced by pu ing bu den on he speake who has o selec a message om a se o mul iple
possible u e ances and send qui e long messages ha make a ge iden i ica ion easy o he
lis ene s. In he con ex -awa e condi ion, we obse ed he opposi e pa e n, namely ha consis ency
is highe han e ec i eness. This pu s mo e bu den on he lis ene ’s side who has o in e p e
he meaning o polysemous messages in con ex . The speake , on he o he hand, has a lowe
bu den han he lis ene . This is due o he lexicon ha ing e ol ed owa ds a simple lexicon wi h
polysemous signals. This ambigui y makes he esul ing language mo e e icien . This is only
possible in he con ex -awa e condi ion because he con ex is in o ma i e and can help he lis ene
o esol e he ambigui y o he polysemous messages.
In he e e en ial communica ion ask we employ, agen s ha e o communica e abou concep s
anging om speci ic o gene ic p esen ed in a con ex o a ying g anula i y. Looking a how
concep s o a ying speci ici y a e encoded in he eme ging languages p o ides insigh s in o
he ca ego y s uc u e lea ned by he agen s. We ha e conduc ed a concep -le el analysis o he
en opy-based sco es o assess his. In bo h condi ions, we obse ed ha gene ic concep s wi h
only one ixed a ibu es a e e e ed o wi h almos maximal e ec i eness, i.e. e y low meaning
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 93
unce ain y. Mo ing owa ds mo e speci ic concep s, he e ec i eness d ops g adually in he
con ex -unawa e condi ion and exponen ially in he con ex -awa e condi ion. The exponen ial d op
in he con ex -awa e condi ion e lec s he ac ha he numbe o possible con ex s, and hus
he possibili y o exploi hese con ex s o e icien communica ion, inc eases exponen ially, he
mo e a ibu es a e ixed in a a ge concep . This end can be also obse ed in na u al languages
whe e he same speci ic concep can be e e ed o wi h di e en labels, e.g. “animal”, “dog”, o
“dalma ian”, depending on he con ex . The mos gene ic concep s, on he o he hand, e.g. animal
o hing, end o ha e only one label. We conclude ha he a ailabili y o a sha ed con ex no only
makes he eme ging language mo e e icien , bu also shapes he ca ego y sys em and how concep s
a di e en le els o abs ac ion a e e e ed o.
7.4.2 Expe imen 2: U ili y-based p agma ic easoning
The goal o he second expe imen was o in es iga e he in luence o u ili y-based easoning on he
si ua ional e e ence o concep s a di e en le els o abs ac ion in di e en con ex s. We looked
a si ua ional e e ence a e a language has eme ged in he con ex -unawa e and con ex -awa e
condi ions om he i s expe imen . This means ha we in es iga ed he in e ac ions on a es
da ase wi h concep s ha ha e no been included in he ain and alida ion da ase s.
Fi s , we compa ed accu acies collec ed on he es da ase o con ex -unawa e agen s, con ex -
unawa e ained agen s wi h RSA, con ex -awa e agen s, and con ex -awa e ained agen s wi h
RSA. Table 7.4 shows ha es accu acies we e highe in he con ex -awa e han in he con ex -
unawa e condi ion. This means ha agen s gene alized be e in he con ex -awa e condi ion.
Rega ding u ili y-based p agma ic easoning, accu acies in he con ex -unawa e condi ion did no
imp o e wi h RSA bu RSA seemed o wo sen pe o mance o con ex -unawa e agen s. In he
con ex -awa e condi ion, howe e , u ili y-based p agma ics lead o simila pe o mance. These
e ec s we e suppo ed by a Bayesian hie a chical model p edic ing es accu acies by he in e ac ion
o con ex -based (con ex -unawa e s. con ex -awa e) and u ili y-based (wi hou RSA s. wi h
RSA) p agma ics. We ound a main e ec o con ex -based p agma ics (M=-0.08, C I=[-0.09, -0.07],
pd=100%, ROPE=[-0.01, 0.01], 0% in ROPE), a main e ec o u ili y-based p agma ics (M=0.03,
C I=[0.02, 0.04], pd=100%, ROPE=[-0.01, 0.01], 0% in ROPE) and an in e ac ion e ec (M=0.03,
C I=[0.02, 0.04], pd=100%, ROPE=[-0.01, 0.01], 0% in ROPE).
Table 7.4: Accu acies on he es da ase .
Da ase Con ex -unawa e + RSA Con ex -awa e + RSA
(3,4) 0.79 ±0.02 0.87 ±0.04 0.93 ±0.02 0.96 ±0.02
(3,8) 0.87 ±0.03 0.82 ±0.03 0.97 ±0.01 0.99 ±0.01
(3,16) 0.95 ±0.04 0.72 ±0.03 0.96 ±0.07 0.95 ±0.09
(4,4) 0.88 ±0.02 0.75 ±0.04 0.98 ±0.01 0.98 ±0.01
(4,8) 0.89 ±0.07 0.66 ±0.02 0.97 ±0.02 0.96 ±0.04
(5,4) 0.88 ±0.09 0.67 ±0.07 0.99 ±0.01 0.98 ±0.03
Mean accu acies and s anda d de ia ions o e i e uns a e calcula ed o each da ase .
94 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
Nex , we looked a he e iciency o he messages in ela ion o he amoun o in o ma ion ha
needed o be communica ed, i.e. he numbe o ixed a ibu es in a concep (Fig 7.6A-D). In he
con ex -unawa e condi ion (Fig 7.6A), he mean leng h o messages used o e e o gene ic concep s
was 10.5 and he mean leng h o messages used o e e o speci ic concep s was 17.2. When adding
RSA (Fig 7.6B), messages in he con ex -unawa e condi ion became much sho e , i.e. hey used
messages o 5.5 symbols on a e age when e e ing o gene ic concep s and messages o 10 symbols
on a e age when e e ing o speci ic concep s. In he con ex -awa e condi ion (Fig 7.6C), message
leng hs we e e y sho in gene al. In e e ence o gene ic concep s, con ex -awa e agen s used
messages wi h 3.6 symbols on a e age and messages wi h 3.9 symbols on a e age in e e ence
o speci ic concep s. Adding RSA educed he mean leng hs o messages used by con ex -awa e
ained agen s by abou 1 symbol in e e ence o bo h speci ic and gene ic concep s (Fig 7.6D).
O e all, RSA imp o ed e iciency and lead o sho e messages being chosen by he speake agen s.
These obse a ions we e in line wi h he p edic ions o a Bayesian hie a chical model p edic ing
message leng h by he ac o s con ex -based p agma ics (con ex -unawa e s. con ex -awa e),
u ili y-based p agma ics (wi hou RSA, s. wi h RSA) and concep ual hie a chy le el (numbe o
ixed a ibu es). We ound a main e ec o con ex -based p agma ics (M=1.88, C I=[1.28, 2.46],
pd=100%, ROPE=[-0.52, 0.52], 0% in ROPE), a likely exis ing main e ec wi h undecided signi icance
o u ili y-based p agma ics (M=0.70, C I=[0.09, 1.30], pd=98.78%, ROPE=[-0.52, 0.52], 27.82% in
ROPE), an in e ac ion e ec be ween con ex -based p agma ics and concep ual hie a chy (M=0.87,
C I=[0.58, 1.18], pd=100%, ROPE=[-0.52, 0.52], 0% in ROPE) and a possibly exis ing h ee-way
in e ac ion be ween con ex -based p agma ics, u ili y-based p agma ics and concep ual hie a chy
wi h undecided signi icance (M=0.28, C I=[-0.02, 0.59], pd=97.02%, ROPE=[-0.52, 0.52], 97.05% in
ROPE).
Nex , we looked a he lexical e iciency o he concep -message mappings. We calcula ed lexicon
sizes and in o ma i eness on he es in e ac ions, i.e. he messages ha ha e been p oduced when
agen s ha e been p esen ed wi h concep s in he es da ase . Table 7.5 shows he numbe o concep s
in he es da ase s and he numbe o unique messages p oduced in he ou di e en condi ions. In
a hie a chical Bayesian model p edic ing lexicon size by con ex -based and u ili y-based p agma ics
and hei in e ac ion, we ound no main e ec o con ex -based p agma ics (M=-9.18, C I=[-26.92,
8.37], pd=84.67%, ROPE=[-26.42, 26.42], 99.68% in ROPE), i.e. no di e ence be ween con ex -
unawa e and con ex -awa e aining wi h ega ds o he lexicon sizes. Howe e , we obse ed a main
e ec o u ili y-based p agma ics (M=89.59, C I=[71.74, 107.88], pd=100%, ROPE=[-26.42, 26.42],
0% in ROPE) sugges ing ha RSA leads o smalle lexicon sizes. An in e ac ion e ec be ween
con ex -based and u ili y-based p agma ics (M=-41.42, C I=[-58.62, -24.36], pd=100%, ROPE=[-26.42,
26.42], 2.00% in ROPE) ha was p obably signi ican , sugges ed ha adding RSA educed lexicon
sizes mainly o con ex -awa e ained agen s bu no o con ex -unawa e ained agen s.
How did agen s in he di e en condi ions op imize he adeo be ween lexicon in o ma i eness
and size? In Fig 7.7A-D, we plo he adeo be ween lexicon in o ma i eness and size o each
condi ion. We calcula ed size-concep a ios by di iding he numbe o unique messages sen by he
numbe o unique concep s in he da ase . In he con ex -unawa e condi ion (Fig 7.7A), he messages
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 95
(a) Con ex -unawa e (b) Con ex -unawa e + RSA
(c) Con ex -awa e (d) Con ex -awa e + RSA
Figu e 7.6: Dis ibu ion o message leng hs o di e en le els o he concep ual hie a chy.
Table 7.5: Lexicon sizes.
Da ase # concep s Con ex -unawa e + RSA Con ex -awa e + RSA
(3,4) 250 99.6 ±4.51 51.6 ±4.34 153.8 ±32.34 42.4 ±5.55
(3,8) 1460 517.2 ±10.13 379.8 ±28.55 902.4 ±116.91 155.4 ±14.86
(3,16) 100 24.6 ±9.4 56.4 ±9.99 40.0 ±17.39 39.0 ±9.77
(4,4) 1250 736.6 ±38.81 265.0 ±63.53 821.6 ±117.53 155.6 ±25.21
(4,8) 100 42.0 ±17.25 79.4 ±3.05 66.4 ±24.5 54.4 ±6.02
(5,4) 100 43.8 ±13.59 52.6 ±21.7 82.6 ±12.3 51.4 ±4.1
Mean numbe o unique messages and s anda d de ia ions o e i e uns a e calcula ed o each
da ase .
ha he agen s sen we e e y in o ma i e, i.e. hey singled ou speci ic concep s. No mo e han six
messages we e needed o e e o en concep s on a e age. The lexicon o con ex -awa e agen s was
less in o ma i e han ha (Fig 7.7C): Lexicon in o ma i eness was much lowe and anged om 3.21
o 4.11 o he di e en da ase s. In a Bayesian hie a chical model p edic ing lexicon in o ma i eness
by con ex -based p agma ics (con ex -unawa e s. con ex -awa e), u ili y-based p agma ics (wi hou
RSA s. wi h RSA) and hei in e ac ion, we ound a main e ec o con ex -based p agma ics
(M=0.50, C I=[0.39, 0.61], pd=100%, ROPE=[-0.11, 0.11], 0% in ROPE), a main e ec o u ili y-based
p agma ics (M=0.43, C I=[0.32, 0.54], pd=100%, ROPE=[-0.11, 0.11], 0% in ROPE) and an in e ac ion
96 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
e ec (M=0.44, C I=[0.33, 0.55], pd=100%, ROPE=[-0.11, 0.11], 0% in ROPE). These e ec s suppo
he abo e obse a ions. The mean lexicon size-concep a ios o each da ase we e smalle in he
con ex -unawa e han in he con ex -awa e condi ion. Howe e , when adding RSA, his pa e n
was e e sed (Fig 7.7B,D). In a Bayesian hie a chical model p edic ing lexicon size-concep a io by
con ex -based p agma ics (con ex -unawa e s. con ex -awa e), u ili y-based p agma ics (wi hou
RSA s. wi h RSA) and hei in e ac ion, we ound a main e ec o con ex -based p agma ics
(M=-0.03, C I=[-0.04, -0.01], pd=99.90%, ROPE=[-0.02, 0.02], 33.63% in ROPE) wi h undecided
signi icance, a main e ec o u ili y-based p agma ics (M=0.08, C I=[0.06, 0.09], pd=100%, ROPE=[-
0.02, 0.02], 0% in ROPE) and an in e ac ion e ec (M=-0.09, C I=[-0.10, -0.07], pd=100%, ROPE=[-0.02,
0.02], 0% in ROPE). These e ec s sugges ed ha adding RSA imp o ed he lexicon size-concep
a io o con ex -awa e bu no o con ex -unawa e agen s.
(a) Con ex -unawa e (b) Con ex -unawa e + RSA
(c) Con ex -awa e (d) Con ex -awa e + RSA
Figu e 7.7: T adeo be ween lexicon size and in o ma i eness. The lexicon size is no malized by he numbe o concep s
in a da ase .
Lexicon-le el e iciency o a language can also be quan i ied by how well he language ollows a
Zip ’s law like dis ibu ion o messages. The i s pa o Zip ’s law s a es ha an e icien language
con ains ew exp essions which a e used e y equen ly in a language (Zip , 1949). In Fig 7.8A-B,
we p esen he equency dis ibu ion o he messages ha ha e been sen in he ou di e en game
scena ios, con ex -awa e and con ex -unawa e wi h and wi hou RSA.
5
Messages we e o de ed by
hei equency ank on he x-axis and hei ela i e equency in he p o ocol is p esen ed on he
y-axis. We obse ed ha in he con ex -unawa e baseline, almos all messages we e used wi h simila
5F equency dis ibu ions o all da ase s can be ound in S5 Appendix: F equency ank dis ibu ions o all da ase s.
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 97
equency. In he con ex -awa e condi ion, some messages we e used wi h e y high equency
compa ed o he es o he messages leading o he cha ac e is ic loga i hmic ela ionship be ween
ela i e equency o messages in he p o ocol and hei espec i e equency ank. Adding RSA
mo ed he espec i e dis ibu ions close o a dis ibu ion esembling Zip ’s p oposed ela ionship
and na u al languages, speci ically he equency dis ibu ions o English and A abic co po a. The
dis ibu ion o con ex -awa e + RSA is closes o na u al language. We conclude ha he mo e
we allowed o p agma ic mechanisms o be exploi ed in he simula ions, he mo e he message
dis ibu ion esembled a Zip ’s law like dis ibu ion o messages whe e he e we e ew messages
which we e used e y equen ly and many messages which we e used in equen ly.
(a) Con ex -unawa e and con ex -awa e
(b) Con ex -unawa e + RSA and con ex -awa e + RSA
Figu e 7.8: Zip ’s law like dis ibu ion o message equency plo ed o da ase D(4,4). D(4,4) is a medium-sized
da ase whe e objec s consis o ou a ibu es which can each ake ou di e en alues.
The second pa o Zip ’s law s a es ha languages maximize hei e iciency by using sho e
exp essions mo e equen ly han longe exp essions (Zip , 1935). In Fig 7.9A-B, we p esen he
dis ibu ion o messages as a ela ion be ween equency ank on he x-axis and message leng h on
he y-axis. We obse ed again ha he mo e we allowed o p agma ic mechanisms o be exploi ed
in he simula ions, he mo e he message dis ibu ions esembled Zip ’s law and na u al languages
wi h sho e messages being used mo e equen ly han longe messages.
(a) Con ex -unawa e and con ex -awa e
(b) Con ex -unawa e + RSA and con ex -awa e + RSA
Figu e 7.9: Zip ’s law like dis ibu ion o message leng h plo ed o da ase D(4,4).
Las ly, we p esen a quali a i e analysis o messages ha we e used in he ou di e en es
condi ions, con ex -unawa e, con ex -unawa e + RSA, con ex -awa e and con ex -awa e + RSA.
Table 7.6 shows examples o he andomly picked speci ic es concep (2,0,3) in da ase D(3,4),
which is he smalles da ase whe e objec s ha e h ee a ibu es ha can each ake ou di e en
alues. We use a uple no a ion o concep s, e.g. (2,0,3), whe e only ixed a ibu es a e speci ied
and a ibu es which a e no ixed by a concep bu can ake any alue a e ep esen ed as “_”,
e.g. (2,_,0). Fo a mo e in ui i e in e p e a ion o he esul s, we gi e na u al language examples.
98 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
Fo example, (2,0,3) could mean small blue ci cle i he a ibu es we e size, colo and shape.
These na u al language examples a e only possible in e p e a ions o he concep s, he agen s we e
ained on symbolic objec ec o s speci ying only nume ical a ibu es. We show examples o
a coa se con ex , whe e he concep has been p esen ed in a con ex whe e he dis ac o objec s
di e ed om he a ge concep in each a ibu e and a ine con ex , whe e he dis ac o objec s
in he con ex sha ed all bu one a ibu e wi h he a ge concep . We lis he messages ha we e
sen and analyze hei o me use du ing aining as well as which a ibu es o he concep he
message p obably communica ed. In he con ex -unawa e condi ion, he speake s a e no awa e
o he con ex . This means ha o be communica i ely success ul, hey should communica e all
a ibu es ha a e ele an o he a ge concep . Indeed, he message ha was used o e e o (2,0,3)
(small blue ci cle) du ing es ing (“1,1,1,1”) was used o e e o simila speci ic concep s du ing
aining, such as (2,1,3) (small ed ci cle) and (2,3,3) (small g een ci cle), o o e e o ela ed
mo e gene ic concep s, i.e. (2,_,3) (small ci cle) and (2,0,_) (small blue). We obse ed no di e ence
be ween he coa se and he ine condi ion in he con ex -unawa e baseline. When adding RSA o
he con ex -unawa e baseline, he messages ended o be sho e and a mo e di e se se o messages
was used. This was likely due o di e en aspec s o he a ge concep being communcica ed. We
deduced his om he usage o he messages du ing aining. By looking a he se o concep s
ha a message was used o e e o and hei simila i y, we easoned abou which a ibu es o he
concep we e likely communica ed wi h his message. In he con ex -awa e condi ion, we obse ed
ha he agen s used di e en messages depending on he con ex condi ion ( ine s. coa se) and he
speci ic dis ac o objec s in he con ex . A likely s a egy seemed o be ha he agen s compa ed
he a ge s and dis ac o s and communica ed he one a ibu e ha di e ed in he ine con ex
condi ion o jus one o he a ibu es ha di e ed be ween a ge s and dis ac o s in he coa se
con ex condi ion. In he con ex -awa e + RSA condi ion, agen s used e y simila messages o he
messages ha we e used in he con ex -awa e condi ion. The only di e ence was ha hey ended
o use sho e messages whe e possible and, o example, educed he messages “1,1,1,1” and “1,1,1”
o “1,1”6.
Discussion
O e all, he esul s o Expe imen 2 sugges ha in no el communica i e si ua ions, he agen s ha
had been ained in he di e en con ex condi ions in Expe imen 1, used di e en app oaches
o communica e he no el concep s and did so wi h a ying deg ees o success. The s a egies
o con ex -awa e agen s and con ex -awa e + RSA agen s lead o almos op imal communica i e
success, whe eas con ex -unawa e and con ex -unawa e + RSA agen s exhibi ed lowe accu acies
on he es da ase . These di e ences in pe o mance we e likely due o he di e en s a egies
ha we e used: Con ex -unawa e speake s which had been ained wi hou access o a sha ed
con ex , came up wi h a s a egy whe e hey ended o communica e all ixed a ibu es o a a ge
concep as p edic ed. This was shown by he quali a i e analysis o messages and has implica ions
6
Sending he message “1” would in his case be e en mo e e icien , bu no possible in his scena io because he RSA
agen s a e only able o use messages ha ha e been p oduced du ing aining a leas once.
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 99
Table 7.6: Quali a i e examples.
Game
scena io
ixed
indices
ixed
alues
con ex
condi ion
dis ac o s
communi-
ca ed
a ibu es
messages
use
du ing
aining
Con ex -
unawa e
(1,1,1) (2,0,3) 0 (coa se
con ex )
(1,3,0) (2,0,3) [1,1,1,1] (2,1,3),(2,3,3)
(0,1,0) (2,_,3), (2,0,_)
(0,1,1)
2 ( ine
con ex )
(0,0,3) (2,0,3) [1,1,1,1] see abo e
(2,3,3)
(2,1,3)
Con ex -
unawa e
+ RSA
(1,1,1) (2,0,3) 0 (1,3,0) (2,0,_) [1,2,1] (2,0,_)
(0,1,0) (_,_,3) [1,1] (0,2,3),(0,_,3)
(0,1,1) (_,_,3) [1,1] (2,3,3),(2,_,3)
2 (0,0,3) (2,_,_) [1,4,1,3] (2,1,_)
(2,3,3) (_,0,_) [2] (1,0,3),(1,0,_)
(2,1,3) (_,0,_) [2] (_,0,_)
Con ex -
awa e
(1,1,1) (2,0,3) 0 (1,3,0) (2,_,_) [4] (2,_,_)
(0,1,0) (2,_,_) [4]
(0,1,1) (_,0,_) [1] None
2 (0,0,3) (2,_,_) [4]
(2,3,3) (_,0,_) [1,1,1,1] (_,0,_)
(2,1,3) (_,0,_) [1,1,1] (_,0,_)
Con ex -
awa e +
RSA
(1,1,1) (2,0,3) 0 (1,3,0) (2,_,_) [4] (2,_,_)
(0,1,0) (2,_,_) [4]
(0,1,1) (2,_,_) [4]
2 (0,0,3) (2,_,_) [4]
(2,3,3) (_,0,_) [1,1] (_,0,_)
(2,1,3) (_,0,_) [1,1]
Quali a i e examples o all ou condi ions and he da ase D(3,4) in esponse o he speci ic
concep (2,0,3) (all a ibu es ixed).
o he e iciency o he language used in his condi ion. Con ex -unawa e agen s used a smalle
bu e y in o ma i e lexicon compa ed wi h he con ex -awa e agen s because hey did no adap
hei messages o he con ex condi ions. Thei messages ended o be longe , he mo e a ibu es
we e ixed in a a ge concep , sugges ing ha hey communica ed mo e a ibu es when he a ge
concep was mo e speci ic. The messages used in he con ex -unawa e condi ion did no ollow Zip ’s
law, i.e. equen ly used messages we e no sho e no used in mo e communica i e si ua ions.
When adding u ili y-based p agma ic abili ies in he o m o RSA, con ex -unawa e speake s sen
messages ha we e mo e e icien and sho e . This e ec was mo e p onounced when agen s had
been ained con ex -unawa e han when hey had been ained con ex -awa e. This was likely
due o he con ex -unawa e condi ion being he one in which messages we e e y long in he i s
place. E en hough he messages p oduced by con ex -awa e speake s we e al eady almos op imal
wi h ega ds o hei message leng h, adding RSA s ill imp o ed he e iciency o he messages
chosen by con ex -awa e speake s ha easoned abou hei in e locu o ’s likely in e p e a ion wi h
RSA. Tha RSA imp o ed he e iciency o messages is unsu p ising gi en ha he u ili y unc ion
106 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
side. This means ha he communica ion sys em can become simple by being less in o ma i e
because lis ene s can eason abou he con ex and esol e he meaning o polysemous messages
in con ex . In u n, his does no happen when speake s a e ained wi hou access o a (sha ed)
con ex . Hence, ou esul s p o ide e idence o he ole o p agma ics in shaping ca ego y s uc u e
and communica ion abou ca ego ies. We can conside he ela ionship be ween en i onmen ,
communica i e need and he ca ego y sys em o be modi ied by he a ailabili y o con ex . As
p edic ed by Rosch e al. (1976) and Rosch (1978), he ca ego y sys em o a language can a o d o be
simple when he meanings o ca ego ies a e in e p e ed in con ex . In addi ion o he in luence o
he a ailabili y o con ex -based p agma ics on he eme gence o a language, we in es iga ed he ole
o u ili y-based p agma ics in no el communica i e si ua ions. We ound ha speake s choose mo e
e icien messages when hey eason abou he u ili y o he message in e ms o he lis ene ’s likely
in e p e a ion ia RSA-based easoning. In summa y, we ha e shown how con ex -based p agma ics
shapes a ca ego y sys em when concep s a di e en le els o abs ac ion need o be communica ed
in in e ac ion. U ili y-based p agma ics has been shown o be especially e ec i e when used o an
eme ging language ha includes ambigui y in he con ex -awa e condi ion, and i s main unc ion
in ou model is o make he communica ion abou concep s a di e en le els o abs ac ion mo e
e icien om he speake pe spec i e. Ou indings sugges ha e icien ca ego y sys ems ha a e
e lec ed in an e icien linguis ic sys em can eme ge in agen -based models o communica ion in
in e ac ion. This has impo an implica ions o he ques ion how human ca ego y and language
sys ems migh ha e e ol ed and a e con inuously shaped by he en i onmen and a communica i e
need, namely he idea ha p agma ics, in he o m o con ex and u ili y-based easoning, plays a
c ucial ole in his p ocess.
7.6 Me hods
7.6.1 Fo maliza ion
We use a mul i-agen model o eme gen communica ion. In eme gen communica ion models, a
speake and a lis ene agen need o communica e abou a a ge in a gi en con ex . In one ound
o he game, speake and lis ene obse e an inpu each, and hen he speake sends a message
o he lis ene ha desc ibes he a ge . The lis ene in e p e s he message and selec s he objec
hey belie e is he a ge , i.e. he one hey assign he highes p obabili y. Du ing se e al i e a ions
o aining, he agen s con e ge on a language-like sys em, i.e. a consis en mapping be ween
a ge s and messages (Laza idou e al., 2017,2018). To p e en agen s om communica ing abou
lowe -le el ea u es ha a e no ele an om a human pe spec i e, we use a concep -le el e e ence
game o aining (Kob ock, Ohme , e al., 2024; Mu & Goodman, 2021). This means ha he agen s
do no communica e abou a single objec , bu a he abou a concep . We achie e his by c ea ing
concep s ha consis o mul iple a ge objec s (Kob ock, Ohme , e al., 2024; Mu & Goodman, 2021),
see Fig 7.1.
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 107
The model is o malized as a communica ion game
𝐺=(𝑇𝑆, 𝐷𝑆, 𝑇𝐿, 𝐷𝐿)
be ween a speake agen
𝑆
and a lis ene agen
𝐿
. Bo h agen s ecei e hei own se o inpu objec s, comp ised o game size
𝑔
a ge objec s
𝑇={𝑡1, ..., 𝑡𝑔}
and
𝑔
dis ac o objec s
𝐷={𝑑1, ..., 𝑑𝑔}
. The speake agen ecei es
hei own se s o a ge s
𝑇𝑆
and dis ac o s
𝐷𝑆
o de ed such ha a ge s come i s .
7
The ecei e
agen ecei es hei own se s o a ge s
𝑇𝐿
and dis ac o s
𝐷𝐿
, bu hey a e shu led and he lis ene
does no know which objec s a e a ge s and which a e dis ac o s. This means hey ecei e an inpu
𝑋𝐿={𝑥𝐿
1, ..., 𝑥𝐿
𝑖}
, whe e
𝑖=2·𝑔
. The indi idual objec s ha comp ise he a ge concep and he
con ex a e no necessa ily sha ed be ween lis ene s and speake s, bu hey belong o he same
a ge concep and con ex condi ion. Fo example, i he a ge concep is blue, hen wha
𝑇𝑆
and
𝑇𝐿
ha e in common is ha all objec s in hese se s a e blue. Howe e , i does no ma e , whe he
he a ge objec s a e blue ci cles, blue squa es, blue iangles e c. These a e andomly sampled
om he se o blue objec s. Simila ly, in a con ex condi ion whe e one concep -de ining a ibu e
is allowed o be sha ed be ween
𝑇𝑆
and
𝐷𝑆
and be ween
𝑇𝐿
and
𝐷𝐿
, hen he speci ic objec s in
he se s
𝐷𝑆
and
𝐷𝐿
, may di e . Howe e , bo h dis ac o s p esen ed o he speake and dis ac o s
p esen ed o he lis ene di e in he same concep -de ining a ibu es om he a ge concep . Fo
example, blue ci cles a e p esen ed in a con ex comp ised o ci cles o di e en colo s (see Fig 7.2B).
The lis ene ’s ask is o p edic a label
𝑦𝐿
𝑖∈ {0,1}
(0: dis ac o , 1: a ge ) o each objec
𝑥𝐿
𝑖
based on
a message
𝑚
ha i ecei es om he speake . The speake gene a es a message
𝑚
by choosing up
o
𝑀
symbols om a ocabula y
𝑉
. The ocabula y is comp ised o p imi i e disc e e symbols,
anging om
0
o
𝑉
, i.e. "0", "1", "2", e c. (Laza idou e al., 2018). The symbol
0
is de ined as he
end-o -sequence (EOS) symbol ha can be used o e mina e a message be o e
𝑀
, he maximum
message leng h, is eached (Kob ock, Ohme , e al., 2024; Mu & Goodman, 2021; Ohme , Duda, &
B uni, 2022). All o he symbols can be sen wi hou such implica ions. They do no ha e a meaning
in he beginning o aining, bu hei meaning is nego ia ed and con e ged upon h ough aining,
i.e. h ough se e al i e a ions o playing he concep -le el e e ence game. We ain he speake and
lis ene agen s wi h a join bina y c oss en opy (BCE) loss maximizing he p obabili y ha he
lis ene agen co ec ly iden i ies he a ge objec s in hei inpu :
L
𝐵𝐶𝐸(𝑆, 𝐿, 𝐺)=−X
𝑖
log 𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖,ˆ
𝑚),(7.2)
whe e ˆ
𝑚∼𝑝𝑆(𝑚|𝑇𝑆, 𝐷𝑆)and 𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖,ˆ
𝑚)=ReLu(GRU𝐿(ˆ
𝑚) · embed(𝑥𝐿
𝑖)).
We implemen he game using he EGG amewo k (Kha i ono e al., 2019) de eloped o imple-
men ing eme gen communica ion games. Bo h ou speake and lis ene agen s a e implemen ed
as neu al ne wo ks wi h eed- o wa d (dense) laye s o embedding he inpu objec s and Ga ed-
Recu en -Uni (GRU, Cho e al., 2014) cells o encode (speake ) o decode (lis ene ) he messages.
GRU cells a e a speci ic ype o Recu en Neu al Ne wo ks (RNNs) which ha e been p o en use ul
o dealing wi h sequence da a such as language. The GRUs consis o a single hidden laye and a
ga ing mechanism ha weighs how much p e ious in o ma ion is conside ed when p ocessing he
cu en inpu . They a e hus well-sui ed o de ec ing dependencies and pa e ns in language-like
7
I agen s a e ained con ex -unawa e, he speake agen does no p ocess he dis ac o objec s, bu only he a ge
objec s.
108 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
inpu .
Modeling a p agma ic speake wi h he Ra ional Speech Ac s (RSA) amewo k
We model an RSA speake ollowing he Ra ional Speech Ac s (RSA) amewo k (F ank & Goodman,
2012; F anke & Degen, 2016; Goodman & F ank, 2016). The RSA model p o ides a amewo k o
model di e en lis ene and speake ypes ha a y by he in o ma ion hey ake in o accoun
o choose o in e p e a message. The base le el speake
𝑆0
chooses a message ha desc ibes he
a ge s. The base le el lis ene
𝐿0
in e p e s he message li e ally, i.e. i p edic s which a e he
a ge s based on he lea ned meaning o a gi en message. The RSA speake
𝑆RSA
we use in his
model is a le el-1 speake
𝑆1
, also called G icean speake (F anke & Degen, 2016), which maximizes
ele an in o ma ion as o allow he hea e o choose he co ec a ge . This means ha based on
he assump ion ha a li e al lis ene
𝐿0
chooses all a ge s o which he message is ue wi h equal
p obabili y, he RSA speake chooses a message ha maximizes he p obabili y o he lis ene o
selec he a ge s. Fo example, i he speake needs o communica e a blue ci cle in a ine con ex
consis ing o ci cles o di e en colo s han blue, hen
𝑆0
would choose he messages “blue” o
“ci cle” wi h equal p obabili y.
𝐿0
would hen choose a ge s o which he message is ue, leading
o communica ion success i he message was “blue” and o communica ion ailu e i he message
was “ci cle”. An
𝑆1
speake ha we call
𝑆RSA
in ou model, would choose he message “blue” wi h
highe p obabili y han “ci cle” because hey eason abou he lis ene ’s likely in e p e a ion o he
message. We implemen
𝑆RSA
as a speake ha maximizes a u ili y unc ion ollowing s anda d
RSA models (F ank & Goodman, 2012; F anke & Degen, 2016; Goodman & F ank, 2016):
𝑆RSA(𝑇𝑆, 𝐷𝑆)=a gmax
𝑚∈M𝑈(𝑚|𝑇𝑆, 𝐷𝑆),(7.3)
whe e
𝑈
is he u ili y unc ion, ypically de ined as
𝑈(𝑚|𝑤)=log 𝑝𝐿0(𝑤|𝑚)
(Goodman & F ank,
2016). This means ha speake s a e ewa ded i hey choose a message ha maximizes he log-
likelihood o lis ene s in e p e ing he wo ld
𝑤
co ec ly gi en message
𝑚
. O en his basic u ili y
unc ion is e ined, o example by adding a cos e m ha penalizes long messagesScon as e al.,
2021. In ou implemen a ion, we ollow p e ious wo k (Fang e al., 2022) and de ine he u ili y
unc ion as:
𝑈(𝑚|𝑇𝑆, 𝐷𝑆)=log 𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖, 𝑚) − 𝐶(𝑚),(7.4)
whe e
𝐶(𝑚)
is a cos applied o he message leng h and
𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖, 𝑚)
is he log-likelihood o he
lis ene o selec he co ec a ge s, i.e. o p edic he co ec labels
𝑦𝑖
based on a message and
inpu objec s
𝑥𝑖=(𝑇𝑆, 𝐷𝑆)
. The speake e alua es he u ili y based on an in e nal lis ene model
which is he ained agen om he i s expe imen
𝐿0
. We de ine
log 𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖, 𝑚)
as he logi s
ou pu om he in e nal lis ene model a e aged o a ge s and dis ac o s. To accoun o he ac
ha in ypical RSA models, he messages ha maximize he p obabili y o choosing he co ec
a ge s a he same ime minimize he p obabili y o selec ing he dis ac o s, we calcula e he o e all
u ili y as
P𝑦𝑖=1
|𝑦𝑖=1|−P𝑦𝑖=0
|𝑦𝑖=0|
. Impo an ly, he RSA speake
𝑆RSA
only has access o hei a ge s
𝑇𝑆
and
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 109
dis ac o s
𝐷𝑆
and hus is only able o in e how he lis ene would in e p e a message e e ing o
he a ge concep he speake obse es in a gi en con ex . This p ese es some na u alness o he
communica i e scena io by p e en ing speake s om being omniscious. Finally, he accu acy is
calcula ed on how he lis ene ac ually in e p e s a message in hei own con ex
𝑥𝑖
consis ing o
𝑇𝐿
and
𝐷𝐿
. The cos unc ion
𝐶(𝑚)=𝜆|𝑚|
penalizes messages based on hei leng h, i.e. he numbe
o symbols in a message. We use a cos ac o o 1 o ou simula ions.
7.6.2 Da ase
We cons uc six symbolic da ase s ha con ain concep s wi h a ying le els o speci ici y ( anging
om speci ic o gene ic) depending on how many a ibu es a e sha ed wi hin a concep , i.e. be ween
he di e en a ge objec s. Objec s a e symbolic ec o s comp ised o
𝑛
a ibu es whe e each
a ibu e can ake
𝑘
alues. Fo example, (1,2,1) is an objec om he da ase D(3,4). We use he
no a ion D(
𝑛
,
𝑘
) o deno e da ase s o di e en sizes, anging om h ee o i e a ibu es which
can each ake be ween ou and 16 alues. Ta ge objec s ha belong o a speci ic concep sha e all
a ibu es. Ta ge objec s ha belong o a gene ic concep sha e only one a ibu e. This means ha
he mo e speci ic a concep is, he mo e in o ma ion needs o be communica ed when desc ibing
he concep . Each concep is p esen ed in a con ex . Depending on how many a ibu es a e sha ed
be ween he a ge objec s and he dis ac o objec s, we compa e con ex s anging om ine con ex s,
in which all bu one a ibu es a e sha ed, o coa se con ex s, in which one a ibu e is sha ed
be ween a ge s and dis ac o s. The con ex g anula i y also ela es o how much in o ma ion
needs o be communica ed o desc ibe a gi en concep . As a gene al ule o humb, he ine he
con ex , he mo e in o ma ion needs o be communica ed. In Fig 7.10A-C, we p esen examples o
concep and con ex combina ions in he da ase s. We use shapes wi h di e en colo s and sizes
only o isualiza ion pu poses. The agen s a e ained on symbolic ec o s as desc ibed abo e. 60%
o he concep s a e assigned o he ain spli and 20% o he concep s a e assigned o he alida ion
spli o he da ase . Each concep is p esen ed in exac ly one con ex , whe e he numbe o sha ed
a ibu es is andomly sampled and sha ed be ween speake and lis ene inpu s. The da a spli used
o es ing (20% o he da a) con ains no el concep s in andomly sampled con ex s.
Figu e 7.10: Da ase examples. The h ee objec s in he op ow a e he a ge objec s ha o m a a ge concep oge he .
The h ee objec s in he bo om ow a e he dis ac o objec s ha o m he con ex . A: An example o a speci ic concep
la ge blue ci cle in a ine con ex ( wo a ibu es sha ed). B: An example o a speci ic concep la ge blue ci cle in a
coa se con ex (no a ibu e sha ed). C: An example o a gene ic concep ci cle in a coa se con ex .
110 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
7.6.3 Hype pa ame e s and aining
We use he ollowing hype pa ame e s: The agen s ha e in e nal GRUs wi h a hidden size o 128
o encode and decode messages. We ained wi h a ba ch size o 16 and a lea ning a e o 0.001.
The da ase s we e c ea ed wi h game size 10, i.e. 10 a ge s and 10 dis ac o s pe game ound, and
scaling ac o 10, i.e. each concep is ep esen ed in he da ase 10 imes. Fo aining, we make use
o he Gumbel-So max unc ion ha makes i possible o use backp opaga ion (Jang e al., 2017), a
loss p essu e ha penalizes he leng h o messages and an ea ly s opping c i e ion o make aining
mo e e icien . De ails can be ound in S1 Appendix: T aining speci ics and hype pa ame e s.
7.6.4 E alua ion
We will e alua e he models and eme ging languages acco ding o he ollowing c i e ia: Fi s ,
we will epo accu acies on he ain, alida ion and es da ase s. Accu acies on he ain and
alida ion da ase s gi e us a measu e o communica i e success, i.e. how well do he lis ene agen s
pe o m, o , in o he wo ds, how o en do hey selec he co ec a ge objec s. Accu acies on he
es da ase s addi ionally can be used as a measu e o gene aliza ion abili ies. He e, we can see
whe he he eme ging language gene alizes well o unseen concep s.
Second, o in es iga e whe he ce ain game scena ios lead o he eme gence o mo e e icien
languages, we need a measu e ha cap u es he e iciency o an eme ging language. In he
li e a u e, e iciency has been ela ed o small ocabula y sizes and o sho messages leng hs while
communica ing he same amoun o in o ma ion (Pian adosi e al., 2012). While he ocabula y
size is in p inciple ixed in ou simula ions, we can measu e he message leng hs o an eme gen
language di ec ly a e aining. We will also calcula e he numbe o unique messages used o e e
o concep s as an app oxima ion o he agen s’ lexicon size (Laza idou e al., 2018). We will ela e he
size o he lexicon o he lexicon’s in o ma i eness. The in o ma i eness
𝐼
o a lexicon
𝐿
is de ined as
he a e age o e he message in o ma i eness
𝐼𝑚
o e
𝑁
in e ac ions ollowing (Gualdoni & Boleda,
2024):
𝐼𝐿=1
𝑁
𝑁
X
𝑖=1
𝐼𝑖
𝑚
wi h 𝐼𝑚=1
𝑆𝑚, whe e
𝑆𝑚=1
𝑁X
𝑖
X
𝑗≠𝑖
𝑑(𝐶𝑖, 𝐶𝑗)
is he sp ead o ea u es ha is calcula ed based on he a e age dis ance be ween concep s
𝐶
ha
ha e been e e ed o by
𝑚
. This means ha messages ecei e a lowe in o ma i eness sco e i hey
a e used o e e o concep s wi h highe dis ance and ha messages ecei e a highe in o ma i eness
sco e i hey a e used o e e o concep s wi h lowe dis ance, i.e. highe simila i y. This cap u es he
in ui i e idea ha wo ds in a language which e e o a highly speci ic concep such as dalma ian
a e mo e in o ma i e han wo ds which e e o a highly gene ic concep such as animal. A na u al
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 111
language lexical sys em con ains bo h highly speci ic and highly gene ic e e ences and op imizes
he adeo be ween size and speci ici y.
Thi d, ela ing back o he p oposed e iciency adeo be ween p oduc ion and comp ehension,
we expec ha an e icien language should con ain ambigui y, i.e. synonymous and polysemous
exp essions. To measu e whe he an eme gen communica ion p o ocol con ains synonymous
and polysemous messages, we calcula e in o ma ion- heo e ic sco es on he se o
𝐾
messages
𝑀={𝑚1, ..., 𝑚𝐾}
u e ed and he se o
𝐿
a ge concep s
𝐶={𝑐1, ..., 𝑐𝐿}
he agen s communica ed
abou du ing he simula ion. We calcula e hese sco es on he inal in e ac ions be ween speake s
and lis ene s in he las aining epoch.
Fig 7.11 shows he se s o concep s and messages (displayed as ci cles) wi h ins ances o he concep s
and messages displayed as do s. The ele an in o ma ion- heo e ic sco es a e calcula ed on hese
wo se s.
A language ha is maximally e icien only o he lis ene bu no o he speake should con ain
only one- o-one mappings be ween concep s and messages. The amoun o one- o-one mappings in
a language can be calcula ed wi h he no malized mu ual in o ma ion sco e be ween concep s and
messages:
𝑁𝑀𝐼(𝐶, 𝑀)=𝐻(𝑀) − 𝐻(𝑀|𝐶)
0.5· (𝐻(𝐶) + 𝐻(𝑀)).(7.5)
I his sco e is 1, hen he e iciency adeo is maximal o he lis ene and minimal o he speake .
I his sco e is 0, on he o he hand, we canno conclude ha e iciency o he speake is high; as i
migh be ha a comple ely andom mapping be ween concep s and messages ha e been lea ned
ha would also be highly ine icien o he speake . Thus, o add ess e iciency o he speake , we
conside how much ambigui y, i.e. how many synonymous and polysemous messages, an eme gen
p o ocol includes. To do his, we calcula e sco es based on condi ional en opies. The condi ional
en opy o messages gi en concep s,
𝐻(𝑀|𝐶)=−X
𝑐∈𝐶,𝑚∈𝑀
𝑝(𝑐, 𝑚)log 𝑝(𝑐, 𝑚)
𝑝(𝑐),(7.6)
measu es how much unce ain y emains abou he messages a e knowing he concep s, i.e.
he signal unce ain y. High signal unce ain y means ha he e is a one- o-many mapping o
meanings- o-signals (Win e s e al., 2015). In o he wo ds, when knowing he concep , i is highly
unce ain which message has been used o e e o i . This measu e cap u es he synonymy o
messages (Ohme , Duda, & B uni, 2022; Win e s e al., 2015). The consis ency sco e uses he
condi ional en opy
𝐻(𝑀|𝐶)
o measu e how much unce ain y emains abou he messages a e
knowing he concep s. I is calcula ed as ollows:
consis ency(𝐶, 𝑀)=1−𝐻(𝑀|𝐶)
𝐻(𝑀).(7.7)
112 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
Con e sely, he condi ional en opy o concep s gi en messages,
𝐻(𝐶|𝑀)=−X
𝑐∈𝐶,𝑚∈𝑀
𝑝(𝑚, 𝑐)log 𝑝(𝑚, 𝑐)
𝑝(𝑚),(7.8)
measu es how much unce ain y emains abou he concep s a e knowing he messages, i.e. he
meaning unce ain y. High meaning unce ain y means ha he e is a one- o-many mapping o
signals- o-meanings (Win e s e al., 2015). In o he wo ds, when knowing he message, i is highly
unce ain which concep i e e s o. This measu e cap u es he polysemy o messages (Ohme ,
Duda, & B uni, 2022; Win e s e al., 2015). The e ec i eness sco e uses he condi ional en opy o
concep s gi en messages
𝐻(𝐶|𝑀)
o measu e how much unce ain y emains abou he concep s
a e knowing he messages. I is calcula ed as ollows:
e ec i eness(𝐶, 𝑀)=1−𝐻(𝐶|𝑀)
𝐻(𝐶).(7.9)
I he e ec i eness sco e is maximal, i.e. 1.0, his means ha he language is highly e ec i e, o , in
o he wo ds, ha he e a e no polysemous exp essions in he lexicon.
Figu e 7.11: In o ma ion- heo e ic sco es and ambigui y in language. The se o concep s
𝐶
is he ed ci cle on he le
( ed and pu ple), he se o messages
𝑀
is he ull blue ci cle on he igh (blue and pu ple). The mu ual in o ma ion
𝐼(𝐶, 𝑀)
(only pu ple) cap u es one- o-one mappings be ween messages and concep s. The condi ional en opy
𝐻(𝑀|𝐶)
(only ed) is ela ed o synonymous mappings be ween messages and concep s. The condi ional en opy
𝐻(𝐶|𝑀)
(only
blue) is ela ed o polysemous mappings be ween messages and concep s.
Fou h, as ano he measu e o e iciency and o assess whe he he eme ging languages sha e
e iciency ela ed p ope ies wi h human languages, we analyze he equency dis ibu ions o
messages. We in es iga e he ollowing wo ela ionships p oposed by Zip ’s law (Zip , 1935; Zip ,
1949) which seem o be a common p inciple o na u al languages on he lexicon le el (Cancho &
Solé, 2003; Chaabouni e al., 2019; Pian adosi e al., 2011). Fi s , when o de ing messages acco ding
o hei equency, i.e. hei ank in a lexicon, i has been shown ha he ela i e equency o a
wo d in a co pus decays exponen ially wi h each ank. This means ha ew messages a e used
e y equen ly (Cancho & Solé, 2003). Second, wo d leng hs a e de e mined by hei equency,
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 113
whe e he mos equen wo ds end o be he sho es in a language (Pian adosi e al., 2011). We will
es hese wo p ope ies on he agen ’s p o ocols o messages used du ing in e ence on no el es
concep s and compa e he esul ing dis ibu ions in he di e en condi ions and o na u al language
da a om English and A abic. We use he Leipzig Co po a Collec ion (Goldhahn e al., 2012) and
speci ically he News copo a wi h 10K wo ds o English (2024) and A abic (2022) including wo ds
and hei equencies in he co pus. F equency anks a e de e mined by so ing he wo ds/messages
by hei equency o occu ence
𝐹={𝑓1, ... 𝑓𝑚}
. The ela i e equency
𝐹𝑟𝑒𝑙
o a wo d is calcula ed
as
𝐹𝑟𝑒𝑙 =𝑓𝑖
P𝑚
𝑗𝑓𝑗
,
whe e
𝑚
is he numbe o wo ds/messages and
𝑖
is he
𝑖
h wo d. We choose
𝑚=30
o bo h
a i icial simula ions and na u al languages, i.e. we calcula e he equency dis ibu ions o he
30 mos equen wo ds/messages. We calcula e message leng hs wi h a Py hon sc ip and he
PyA abic lib a y (Ze ouki, 2023).
Acknowledgmen s
The simula ions we e un on a high-pe o mance compu ing clus e unded by he Deu sche
Fo schungsgemeinscha (DFG, Ge man Resea ch Founda ion) - 456666331. K is ina Kob ock was
suppo ed by he DFG- unded Resea ch T aining G oup “Compu a ional Cogni ion” (DFG-GRK
2340).
We hank Muhip Tezcan, Eosand a G und, and Isabella del Pozo o hei assis ance in implemen ing
pa s o he so wa e used in his p ojec .
Au ho Con ibu ions:
K is ina Kob ock: Concep ualiza ion, Me hodology, So wa e, Valida ion, Fo mal analysis, In-
es iga ion, W i ing - O iginal D a , W i ing - Re iew & Edi ing, Visualiza ion. Xenia Ohme :
Concep ualiza ion, Me hodology, So wa e, W i ing - Re iew & Edi ing. Elia B uni: Concep ualiza-
ion, Me hodology, W i ing - Re iew & Edi ing, Supe ision. Nicole Go zne : Concep ualiza ion,
Me hodology, W i ing - Re iew & Edi ing, Supe ision.
Appendix
A S1 Appendix: T aining speci ics and hype pa ame e s.
Gumbel-So max elaxa ion We ain wi h he Gumbel-So max elaxa ion ha makes i possible
o use backp opaga ion (Jang e al., 2017). We use a empe a u e o 2.0 and a empe a u e upda e o
0.99.
114 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
Leng h cos p essu e We ain he agen s wi h a loss p essu e ha penalizes he leng h o messages
by mul iplying he symbol’s posi ion in a message wi h a cos ac o
𝑐𝑓
. We choose he cos ac o
𝑐𝑓=0.001
. We se he maximum message leng h
𝑀
o 20 symbols and he ocabula y size
𝑉
o
he numbe o alues in a da ase plus one o he end-o -sequence (EOS) symbol. The e a e wo
main easons o applying such a leng h p essu e du ing aining. The i s eason is ha he leng h
cos p essu e wo ks like an e iciency p essu e on he eme ging language. I we expec di e ences
be ween he con ex -unawa e and he con ex -awa e scena io, hen hese should be due o e iciency
easons. While human communica ion is ypically cos ly, wi h neu al ne wo k agen s sending bi s
o in o ma ion is cheap and we hus need o include a p essu e ha incen i izes he agen s o use
e icien language. The second eason is ha o he RSA simula ions o wo k, we wan he aining
o gene a e an eme ging language ha con ains messages o di e en leng hs. These can be be e
exploi ed by he RSA sende s. I messages gene a ed du ing aining we e all o he same leng h,
hen he cos e m in he RSA u ili y unc ion would ha e no e ec .
Ea ly s opping We use ea ly s opping o aining. We s op aining when a alida ion accu acy o
0.90 has been eached a leas once du ing aining and he alida ion loss has no changed mo e
han 0.001 du ing 10 epochs. In case hese c i e ia a e no me du ing 300 epochs o aining, we
s op ne e heless.
B S2 Appendix: S a is ical models
In Tables 7.7 and 7.8, we speci y he model syn ax and p io s used o he s a is ical models i ed
o Expe imen 1 and 2, espec i ely. Table 7.9 p esen s he pos e io summa ies o he Bayesian
hie a chical models p edic ing NMI, e ec i eness and consis ency o Expe imen 1.
Table 7.7: S a is ical models i ed wi h b m o Expe imen 1.
Model Syn ax P io s
1) Accu acy alida ion accu acy ∼
con ex + (1|da ase )
In e cep
:
uni o m(0,1)
,
sigma: uni o m(0,0.1)
2) Message leng h message leng h ∼con ex ∗
ixed a ibu es + (1|da ase )
In e cep
:
uni o m(0,20)
,
sigma: uni o m(0,5)
3)
En opy
NMI ∼con ex + (1|da ase )In e cep
:
uni o m(0,1)
,
sigma:
uni o m(0,0.1)
e ec i eness ∼con ex +
(1|da ase )
consis ency ∼con ex +
(1|da ase )
4)
En opy
* concep
hie a chy
NMI ∼con ex ∗
ixed a ibu es + (1|da ase )
In e cep
:
uni o m(0,1)
,
sigma:
uni o m(0,0.1)e ec i eness ∼con ex ∗
ixed a ibu es + (1|da ase )
consis ency ∼con ex ∗
ixed a ibu es + (1|da ase )
7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence 115
Table 7.8: S a is ical models i ed wi h b m o Expe imen 2.
Model Syn ax P io s
1) Accu acy es accu acy ∼con ex ∗
RSA + (1|da ase )
In e cep
:
uni o m(0,1)
,
sigma: uni o m(0,0.1)
2) Message leng h message leng h ∼con ex ∗
RSA ∗ ixed a ibu es +
(1|da ase )
In e cep
:
uni o m(0,20)
,
sigma: uni o m(0,5)
3) Lexicon size lexicon size ∼con ex ∗RSA +
(1|da ase )
In e cep
:
uni o m(0,1460)
,
sigma: uni o m(0,100)
4) Lexicon in o ma i eness lexicon in o ∼con ex ∗RSA+
(1|da ase )
In e cep
:
uni o m(0,9)
,
sigma: uni o m(0,2)
5) Lexicon size-concep a io lexicon size-concep a io ∼
con ex ∗RSA + (1|da ase )
In e cep
:
uni o m(0,1)
,
sigma: uni o m(0,0.1)
122 7 Case s udy 3: The ole o p agma ic mechanisms in language use and eme gence
Rega ding he ole o p agma ics in his join e olu ion and in he use o a language o conc e e
e e ences o concep s, we compa ed he a ailabili y o con ex -based and u ili y-based p agma ics
o a baseline whe e no p agma ic mechanisms we e a ailable. Conside ing he ole o con ex -
based p agma ics, we showed ha languages end o be mo e ambiguous and e icien when hey
eme ged du ing in e ac ions in a sha ed con ex . The speake agen s came up wi h a e y e icien
s a egy whe e hey made use o he con ex o communica e only he a ibu es ha we e needed o
disc imina e he a ge concep om he con ex . This led o a polysemous mapping be ween concep s
and messages, bu he ambigui y could be esol ed in con ex (as was also p oposed in Pian adosi
e al., 2012). In language use, we again ound inc eased e iciency a he message and lexicon
le els when a sha ed con ex was a ailable du ing he eme gence o he language. Conside ing
u ili y-based p agma ics, when speake s ook he expec ed u ili y o an u e ance in o accoun ,
his imp o ed he gene aliza ion accu acy and e iciency o he communica ion sys em. While
he message-le el e iciency can be a ibu ed o he cos e m in he RSA unc ion, he imp o ed
lexicon-le el e iciency is an eme gen ea u e. This likely ela es o he ac ha RSA speake s
can condense longe messages wi h he same meaning in o a single message. When compa ing
agen s equipped wi h di e en combina ions o con ex - and u ili y-based p agma ic abili ies, we
ound ha con ex - and u ili y-based p agma ic mechanisms yield e y simila esul s. Speci ically,
we obse ed ha u ili y-based p agma ics wo ks be e when combined wi h a language ha
eme ged when a sha ed con ex was a ailable. This migh mean ha u ili y-based p agma ics and
con ex -based p agma ics ely on simila mechanisms. One hypo hesis migh be ha p agma ically
e icien languages eme ge in in e ac ions whe e a sha ed con ex is a ailable, leading o he
con en ionaliza ion o e ms whe e much o he con ex in o ma ion is sha ed by bo h in e locu o s
and he e is li le unce ain y in ol ed. We can u he hypo hesize ha u ili y-based p agma ics
is needed only when he con en ionalized e ms all sho o he communica i e in en ions. Fo
example, i only he e ms g een and ci cle a e con en ionalized, bu he concep o be communica ed
is a blue ci cle ha should be disc imina ed om a g een ci cle, hen RSA-s yle ecu si e p agma ic
easoning can help o choose he u e ance wi h highes u ili y, ci cle, in his con ex .
In conclusion, we modeled he p essu es o simplici y and in o ma i eness as a adeo be ween
he speake ’s and he lis ene ’s needs in in e ac ion. The communica i e and cogni i e need o
e iciency likely plays a key ole in op imizing his adeo and, as a esul , in shaping linguis ic
and ca ego y sys ems. Ou modeling esul s p o ide e idence o his hypo hesis and show ha
he eme gence o e icien ca ego y and linguis ic sys ems can be a esul o e icien in e ac ion.
Mo eo e , we p o ided e idence o he key ole o p agma ics in shaping linguis ic and ca ego y
sys ems. I is widely accep ed ha p agma ics plays a key ole in conc e e e e ence si ua ions, and
also in shaping communica i e sys ems. He e, we a gue ha p agma ics also plays a key ole in
shaping ca ego y sys ems due o he same mechanisms and he p oposed join e olu ion o ca ego y
and linguis ic sys ems. This is a key inding o he iew ha language and cogni ion in e ac .
8 Case s udy 4: Linguis ic s a egies o
gene aliza ion and abs ac ion
This chap e p esen s case s udy 4. The chap e s a s wi h a high-le el in oduc ion ollowed by he
con en o he publica ion: Kob ock, K., Ohme , X., B uni, E., & Go zne , N. (2025a). Agen s gene alize
o no el le els o abs ac ion by using adap i e linguis ic s a egies. In W. Che, J. Nabende, E. Shu o a,
& M. T. Pileh a (Eds.), Findings o he Associa ion o Compu a ional Linguis ics: ACL 2025 (pp. 8685–
8699). Associa ion o Compu a ional Linguis ics. h ps://doi.o g/10.18653/ 1/2025. indings-
acl.455 The chap e ends wi h a b ie summa y o he main con ibu ions o he publica ion and a
discussion o i s implica ions o he b oade esea ch ques ion o his disse a ion.
8.1 High-le el in oduc ion
The goal o his case s udy is o in es iga e gene aliza ion and abs ac ion and how language can help
achie e hese abili ies. Fo his pu pose, we again use an eme gen communica ion model simila
o he one used in he p e ious wo case s udies. He e, he ocus is on a ze o-sho gene aliza ion
ask whe e agen s need o gene alize hei language and e e o unseen concep s a no el le els
o abs ac ion. We ope a ionalize concep s again as a combina ion o objec s ha sha e a ce ain
numbe o a ibu es. The mo e a ibu es sha ed wi hin membe s o he same a ge concep , he
mo e speci ic he concep is. The ewe a ibu es sha ed wi hin membe s o a a ge concep , he
mo e gene ic he concep is. We s udy he ze o-sho gene aliza ion o no el le els o abs ac ion
by manipula ing he aining and es da a spli s. In he “ o gene ic” condi ion, we ain agen
pai s o communica e abou mo e speci ic concep s, and es hem on he mos gene ic concep s.
In he “ o speci ic” condi ion, we ain speake and lis ene agen s o communica e abou mo e
gene ic concep s, and es hem on he mos speci ic concep s. Fi s , we compa e he gene aliza ion
pe o mance o he agen s in bo h condi ions. Nex , we in es iga e in de ail which messages he
agen s use o gene alize o concep s a no el le els o abs ac ion. We gain insigh s in o he agen s’
linguis ic s a egies by compa ing he se o messages ha agen s use o gene alize o he unseen
concep s wi h he se o messages ha we e es ablished du ing aining. Speci ically, we ask wha
he a io o no el messages is and how many messages agen s euse om aining. This helps
us o gain insigh in o whe he he agen s spon aneously use no el labels o e e o he no el
concep s. I agen s indeed use no el labels, hen he nex ques ion is, how hese no el messages
a e composed. We hypo hesize ha in he “ o speci ic” gene aliza ion condi ion, agen s migh
p oduce a la ge amoun o no el messages by composi ionally combining symbols om p e iously
es ablished messages. In he “ o gene ic” condi ion, howe e , we hypo hesize ha a composi ional
s a egy will be less use ul. We use quali a i e analyses and quan i a i e me ics o gain a de ailed
unde s anding o he ypes o messages ha he agen s send in he di e en condi ions and how
hese messages di e om hose sen du ing aining. This enables us o iden i y linguis ic s a egies
124 8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion
o gene aliza ion o mo e speci ic concep s in he “ o speci ic” condi ion, as well as o abs ac ion
o mo e gene ic concep s in he “ o gene ic” condi ion.
8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion 125
8.2 Abs ac
We s udy abs ac ion in an eme gen communica ion pa adigm. In eme gen communica ion, wo
a i icial neu al ne wo k agen s de elop a language while sol ing a communica i e ask. In his
s udy, he agen s play a concep -le el e e ence game. This means ha he speake agen has o
desc ibe a concep o a lis ene agen , who has o pick he co ec a ge objec s ha sa is y he
concep . Concep s consis o mul iple objec s and can be ei he mo e speci ic, i.e. he a ge objec s
sha e many a ibu es, o mo e gene ic, i.e. he a ge objec s sha e ewe a ibu es. We es ed wo
di ec ions o ze o-sho gene aliza ion o no el le els o abs ac ion: When gene alizing om mo e
gene ic o e y speci ic concep s, agen s u ilized a composi ional s a egy. When gene alizing om
mo e speci ic o e y gene ic concep s, agen s u ilized a mo e lexible linguis ic s a egy ha in ol es
eusing many messages om aining. Ou esul s p o ide e idence ha neu al ne wo k agen s can
lea n obus concep s based on which hey can gene alize using adap i e linguis ic s a egies. We
discuss how his esea ch p o ides new hypo heses on abs ac ion and in o ms linguis ic heo ies
on e icien communica ion.
8.3 In oduc ion
One o he mos undamen al goals o A i icial In elligence (AI) and Na u al Language P ocessing
(NLP) esea ch is o build models which can gene alize well o unseen da a. This is, a e all,
one o he c ucial abili ies obse ed in human in elligence. Abs ac ion has been a gued o be a
necessa y i s s ep owa ds achie ing gene aliza ion (Yee, 2019). Bu he e a e also al e na i e
iews such as he exempla -based model o ca ego ies whe e gene aliza ion is achie ed wi hou
abs ac ion (Amb idge, 2020; Daelemans, 2008). We belie e ha unde s anding abs ac ion and
how i in e ac s wi h gene aliza ion, is undamen al o building well-gene alizing models.
Humans na u ally use abs ac ion o sol e complex asks and o communica e abou s a egies and
solu ions. Well-designed AI and NLP sys ems canno only bene i om good abs ac ion abili ies in,
o example, easoning and sol ing complex asks (Ho e al., 2019; Zheng e al., 2024), bu in e ac i e
sys ems should also be able o deal wi h human language inpu s which in ol e abs ac ions (e.g.,
Lachmy e al., 2022). Many esea che s s udying human abs ac ion a gue o a ole o language
he ein (see e.g., Gen ne & Asmu h, 2019; Lupyan & Lewis, 2019; Slou sky & Deng, 2019; Yee, 2019).
The main idea o hese accoun s is ha he lexicaliza ion o concep s, i.e. ha ing a label o a concep ,
helps o acqui e and s uc u e in o ma ion we ob ain abou an en i y and o obse e commonali ies
wi hin membe s o a concep in he i s place. The ole o language in abs ac ion can also be es ed
in compu a ional sys ems. The goals o he cu en esea ch a e o unde s and how abs ac ion and
gene aliza ion in e ac and ul ima ely o in o m he imp o emen o AI and NLP sys ems owa ds
achie ing human-like abs ac ion and gene aliza ion abili ies.
S a ing om he assump ion ha language is use ul o abs ac ion, we s udy abs ac ion in a
communica i e se ing and in es iga e how abs ac ion is achie ed wi h he help o linguis ic
126 8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion
s a egies such as composi ionali y and he euse o p e iously es ablished messages. To gain
insigh s in o he p incipled mechanisms o abs ac ion and he ole o language o abs ac ion,
we use a language eme gence scena io. In language eme gence esea ch, he idea is o de ine a
se o assump ions and hen obse e how hese assump ions change a language-like sys em ha
eme ges du ing in e ac ion (see e.g., Chaabouni e al., 2020; Galke e al., 2022; Laza idou e al.,
2017; Rod íguez Luna e al., 2020). This modeling amewo k is ideal o in es iga ing whe he
human-like beha io and communica ion s a egies can eme ge e en in a compa a i ely simple
communica i e se up be ween wo a i icial neu al ne wo k agen s.
In ou model, wo a i icial neu al ne wo k agen s sol e a e e ence game, whe e a speake agen has
o communica e a a ge concep o a lis ene agen who needs o selec he co ec a ge concep in
a con ex . We ope a ionalize a concep as a se o a ge objec s ollowing p e ious wo k (Kob ock,
Ohme , e al., 2024; Kob ock, Uhlemann, & Go zne , 2024; Mu & Goodman, 2021). Ou a ge
concep s a e designed in a hie a chical ashion, anging om e y speci ic concep s consis ing
o objec s whe e all a ibu es (e.g., size, colo and shape) a e ixed o a ce ain alue, e.g., ‘small
blue ci cle’, o e y gene ic concep s consis ing o objec s whe e only one a ibu e is ixed, e.g.,
‘ci cle’. We can s udy abs ac ion by making use o his abs ac ion hie a chy. He e, we a e in e es ed
in a speci ic kind o abs ac ion, namely he ze o-sho gene aliza ion o concep s a no el le els
o he concep hie a chy, o , o concep s a no el le els o abs ac ion ( ollowing he e minology o
seminal esea ch om Cogni i e Psychology by Rosch e al., 1976). We will no only look a he
gene aliza ion pe o mance o he ained models, bu also a he linguis ic s a egies he agen s
employ. Speci ically, we in es iga e he p ope ies o he eme gen p o ocol and he use o no el s.
es ablished messages du ing abs ac ion.
While p e ious wo k in eme gen communica ion has highligh ed he ole o composi ionali y
o gene aliza ion (see e.g., Haz a e al., 2021; Ko u e al., 2017; Laza idou e al., 2018), in ou
expe imen s we disen angle wo di ec ions o gene aliza ion and p opose ha hey equi e di e en
linguis ic s a egies. We ind ha agen s use a composi ional s a egy only when gene alizing
o speci ic concep s, bu no when gene alizing o gene ic concep s. These esul s highligh ha
composi ionali y is no he only way o achie e gene aliza ion, which is in line wi h ecen indings
om Chaabouni e al. (2020) and Kha i ono and Ba oni (2020).
8.4 Me hod
8.4.1 Gene al se up
We use an eme gen communica ion pa adigm (e.g., Chaabouni e al., 2019; Laza idou e al., 2018)
and build on he concep -le el e e ence game de eloped in p e ious wo k (Kob ock, Ohme , e al.,
2024; Mu & Goodman, 2021). We ain wo a i icial neu al ne wo k agen s, one speake and one
lis ene agen . O e se e al i e a ions, hese agen s de elop a communica ion sys em by sol ing
he ollowing ask: The speake agen
𝑆
has o communica e a concep , i.e. a se o a ge objec s
8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion 127
𝑇={𝑡1, ..., 𝑡𝑔}
, o he lis ene agen
𝐿
whose ask is o iden i y he co ec a ge s among a se o
dis ac o s
𝐷={𝑑1, ..., 𝑑𝑔}
. We call he se o a ge objec s he concep and he se o dis ac o
objec s he con ex . The lis ene ’s ask is o iden i y he a ge concep in a ce ain con ex gi en a
message gene a ed by he speake . The message is a ec o o symbols gene a ed by he speake
neu al ne wo k which does no ha e a p e-speci ied meaning. Ra he , he meaning o a message
eme ges o e se e al in e ac ions be ween he agen s and is de ined by i s usage (see e.g., Laza idou
e al., 2017). Concep s a y in speci ici y, anging om speci ic, whe e all a ibu es a e sha ed
among he a ge objec s, o gene ic, whe e only one a ibu e is sha ed among he a ge s. Con ex s
can ange om being ine, whe e all bu one a ibu es a e sha ed be ween a ge s and dis ac o s,
o being coa se, whe e no a ibu e is sha ed be ween a ge s and dis ac o s. Bo h agen ne wo ks
a e ained in a Rein o cemen Lea ning pa adigm wi h he Gumbel-So max elaxa ion (Jang
e al., 2017) on a join loss ha depends on whe he he lis ene co ec ly iden i ies he a ge s and
dis ac o s gi en he speake -gene a ed message.
8.4.2 Ze o-sho condi ions and hypo heses
We es he ze o-sho gene aliza ion abili ies o he ained ne wo ks in wo condi ions (see Figu e
8.1): The i s condi ion, “ o speci ic”, es s whe he agen s a e able o gene alize o he mos speci ic
concep s when ha ing seen mo e gene ic concep s du ing aining. In his condi ion, we expec he
eme ging communica ion sys em o encode mo e gene ic concep s (such as “blue” o “ci cle”). Fo
a success ul ze o-sho gene aliza ion, hese mo e gene ic concep s would need o be combined o
desc ibe a speci ic concep (such as “blue ci cle”). He e, agen s will need o combine p e iously
lea ned a ibu es composi ionally o desc ibe a mo e speci ic concep . The second condi ion, “ o
gene ic”, es s whe he agen s a e able o gene alize o he mos gene ic concep s when ha ing seen
mo e speci ic concep s du ing aining. In his condi ion, we expec he eme ging communica ion
sys em o encode mo e speci ic concep s (such as “blue ci cle” o “o ange ci cle”). Fo a success ul
ze o-sho gene aliza ion, agen s will need o abs ac away om con ex ually i ele an ea u es
and ind he common a ibu e ha all a ge s sha e (e.g., “ci cle”).
8.4.3 Da ase
The agen s a e ained on six symbolic da ase s de eloped in p e ious wo k (Kob ock, Ohme , e al.,
2024). These da ase s con ain all possible concep s, anging om speci ic o gene ic, and con ex s,
anging om ine o coa se, o a gi en numbe o a ibu es and alues. Fo example, da ase D(3,4)
con ains all possible concep s and con ex s gi en ha objec s in his da ase ha e h ee a ibu es
and each a ibu e can ake ou di e en alues. I we hink o he h ee a ibu es as shape, colo
and size, an example o a speci ic concep would be “small blue ci cle” and an example o a gene ic
concep would be “squa e”. In a ine con ex , objec s belonging o he concep “small blue ci cle”
would need o be disc imina ed agains objec s ha a e also small and blue. In a coa se con ex ,
dis ac o objec s do no sha e any a ibu es wi h he a ge concep . This also means ha he e a e
128 8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion
…
aining
es ing
…
aining
es ing
1) o speci ic:
2) o gene ic:
Figu e 8.1: Examples o speake inpu s o aining and es ing in he wo ze o-sho es condi ions “ o speci ic” and “ o
gene ic”. Each inpu consis s o a ge s (i.e., concep s) in he g een bounding box and dis ac o s (i.e., con ex ).
mo e possible con ex s o speci ic concep s han o gene ic concep s and he da ase s e lec his
ela ionship. We use a scaling ac o o 10 o cons uc he da ase s, i.e. each concep is included in a
da ase 10 imes.1This ensu es ha he da ase s con ain enough aining da a.
Fo he ze o-sho da ase gene a ion, we manipula e he aining, alida ion and es spli s o he
da a. In he “ o speci ic” condi ion, he es spli con ains all mos speci ic concep s a ailable, i.e.
hose whe e all a ibu es a e sha ed among he a ge s. The aining and alida ion spli s a e
composed o he emaining concep s which a e mo e gene ic wi h 75% o he da a used o aining
and 25% o he da a used o alida ion. In he “ o gene ic” condi ion, he es spli con ains all
mos gene ic concep s a ailable, i.e. hose whe e only one a ibu e is sha ed among he a ge s. The
aining and alida ion se s con ain he emaining mo e speci ic concep s wi h 75% o he da a used
o aining and 25% o he da a used o alida ion. Da ase sizes can be inspec ed in Tables 8.10
and 8.11 in Appendix Dand a e compa able be ween ze o-sho condi ions.
8.4.4 A chi ec u e and aining
A communica ion game be ween a speake
𝑆
and a lis ene
𝐿
is de ined as
𝐺=(𝑇𝑆, 𝐷𝑆, 𝑇𝐿, 𝐷𝐿)
,
whe e
𝑇𝑆={𝑡𝑆
1, ..., 𝑡𝑆
𝑔}
and
𝐷𝑆={𝑑𝑆
1, ..., 𝑑𝑆
𝑔}
a e he inpu s o he speake , i.e. se s o game
size
𝑔
a ge s and dis ac o s, and
𝑇𝐿
and
𝐷𝐿
a e he analogously de ined inpu s o he lis ene .
Fo hese inpu s,
𝑇𝑆≠𝑇𝐿
and
𝐷𝑆≠𝐷𝐿
hold, i.e. he a ge s and dis ac o s p esen ed o he
speake di e om he a ge s and dis ac o s p esen ed o he lis ene o ensu e communica ion
o highe -le el concep s (Kob ock, Ohme , e al., 2024; Mu & Goodman, 2021). In each ound o
1
We use his scaling ac o only o cons uc he ain and alida ion da ase spli s. The ze o-sho es is pe o med on a
es spli ha con ains he no el concep s only once.
8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion 129
he game,
𝑆
gene a es a message
𝑚=(𝑠𝑗)𝑗≤𝑀
, whe e
𝑠𝑗
is a symbol om ocabula y
𝑉
and
𝑀
is
he maximal message leng h
2
, based on he inpu s
𝑇𝑆
and
𝐷𝑆
.
𝐿
in u n, ecei es
𝑚
and an inpu
𝑋𝐿={𝑥𝐿
1, ..., 𝑥𝐿
𝑖}
, whe e
𝑖=2·𝑔
which con ains he a ge s
𝑇𝐿
and dis ac o s
𝐷𝐿
shu led.
𝐿
hen
p edic s a label
𝑦𝐿
𝑖∈ {0,1}
(0: dis ac o , 1: a ge ) o each objec
𝑥𝐿
𝑖
in i s inpu (see e.g., Kob ock,
Ohme , e al., 2024; Mu & Goodman, 2021; Ohme , Duda, & B uni, 2022). We isualize he se up in
Figu e 8.2.
Speake
neu al ne wo k
(GRU)
Lis ene
neu al ne wo k
(GRU)
message
[4,11,7,0]
p edic ion and aining
Figu e 8.2: A chi ec u e: Speake and lis ene neu al ne wo ks ecei e sepa a e inpu s whe e a ge objec s sa is y he
same a ge concep (he e “blue”) and dis ac o objec s (i.e., he con ex ) sha e he same numbe o a ibu es wi h he
a ge concep (he e 0). They a e ained on success ul communica ion, i.e. when he lis ene iden i ies he co ec a ge
objec s.
Fo heimplemen a ion
3
,weuse heEGG amewo k o eme gen communica iongames(Kha i ono
e al., 2019, MIT license). Bo h agen s a e implemen ed in a simila ashion: Feed- o wa d laye s
wi h 64 uni s se e as embedding laye s o he inpu objec s. The speake a ge s and dis ac o s
a e embedded sepa a ely and hen conca ena ed in o a join embedding. The lis ene inpu objec s
a e p ocessed by jus one embedding laye . Fo message encoding and decoding, bo h speake and
lis ene ne wo ks use single-laye Ga ed Recu en Uni s (GRU, Cho e al., 2014) wi h a hidden laye
size o 128 ha can deal wi h sequen ial inpu s o a ying leng hs. A speake -lis ene pai is ained
wi h bina y c oss en opy loss
L
𝐵𝐶𝐸(𝑆, 𝐿, 𝐺)=−X
𝑖
log 𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖,ˆ
𝑚),(8.1)
whe e
ˆ
𝑚∼𝑝𝑆(𝑚|𝑇𝑆, 𝐷𝑆)
and
𝑝𝐿(𝑦𝐿
𝑖|𝑥𝐿
𝑖,ˆ
𝑚)=ReLU(GRU𝐿(ˆ
𝑚) · embed(𝑥𝐿
𝑖))
maximizing he p ob-
abili y ha he lis ene co ec ly iden i ies a ge s and dis ac o s wi h a label
𝑦𝑖∈ {0,1}
(0:
dis ac o , 1: a ge ) o each objec
𝑥𝑖
. To ensu e di e en iabili y o backp opaga ion, we use he
s aigh - h ough Gumbel-So max ick (Jang e al., 2017) wi h empe a u e
𝜏=2
and a decay a e
o 0.99. These and o he hype pa ame e s we e de e mined in a g id sea ch ha we conduc ed o
all pa ame e s o e he di e en da ase sizes aiming o maximal alida ion accu acy. We ain
wi h ba ch size 32 and lea ning a e 0.001. Fo ou simula ions, we use game size 10, i.e. 10 a ge
2The end-o -sequence symbol 0can be used o e mina e a message be o e 𝑀is eached.
3All code and analysis sc ip s a e a ailable a h ps://gi hub.com/k is inakob ock/ze o-sho -abs ac ion.
130 8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion
objec s o m a concep and 10 dis ac o objec s o m he con ex . The maximum message leng h
𝑀
is de ined as he o al numbe o a ibu es in a da ase plus he End o Sequence (EOS) symbol
0
.
The ocabula y size o each da ase co esponds o he o al numbe o a ibu e alues p esen .
We es ablish a minimal ocabula y size o each da ase as he sum o he numbe o a ibu e
alues plus one addi ional symbol. This minimal ocabula y size is hen scaled by a ac o o
𝑓=3
,
as sugges ed by Ohme , Duda, and B uni (2022) o ensu e a su icien ly la ge communica ion
channel (Chaabouni e al., 2020).
8.5 Resul s
We ained he models on six symbolic da ase s wi h a ying numbe s o a ibu es and alues. In a
da ase
𝐷(𝑛, 𝑘)
, objec s ha e
𝑛
a ibu es which each can ake
𝑘
di e en alues. Fo all me ics, we
epo means and s anda d de ia ions o e i e indi idual uns pe da ase .
8.5.1 Gene aliza ion pe o mance
We e alua e he agen s’ pe o mance on he es da ase s o assess hei ze o-sho gene aliza ion
abili ies.
4
Accu acies a e calcula ed as a pe cen age o e he objec s ha he lis ene classi ies as
a ge s o dis ac o s. An accu acy o 0.9 means ha 90% o he objec s, i.e. 18 objec s wi h a game
size o 10, ha e been classi ied co ec ly as a ge s o dis ac o s. O , in o he wo ds, wo objec s
ha e been misclassi ied.
Table 8.1 summa izes he mean es accu acies o e he i e uns conduc ed on each da ase o bo h
condi ions. All ze o-sho es accu acies a e >=0.63 indica ing ha he lis ene s co ec ly iden i y
mo e han 60% o he 20 objec s as a ge s o dis ac o s. This co esponds o a numbe o 12 co ec ly
iden i ied objec s. Agen s achie e highe pe o mance in he “ o speci ic” condi ion compa ed o he
“ o gene ic” condi ion in all da ase s. Compa ing es accu acies be ween da ase s, gene aliza ion
pe o mance is be e o da ase s wi h mo e a ibu es. Speci ically, on da ase s wi h a leas ou
a ibu es, agen s achie e gene aliza ion accu acies o 0.82 o highe in bo h condi ions. This means
ha speake s choose exp essions o desc ibe he held-ou concep s a no el le els o abs ac ion ha
enable lis ene s o classi y a leas 16 o he 20 objec s co ec ly.
4
T aining and alida ion accu acies o bo h condi ions a e >=0.97 indica ing ha he agen s ha e lea ned he ask and
achie ed high pe o mance on bo h he aining and he alida ion da a spli s - a necessa y p e equisi e o a alid
in e p e a ion o he ze o-sho es accu acies (see Tables 8.4 and 8.5 in Appendix A).
8 Case s udy 4: Linguis ic s a egies o gene aliza ion and abs ac ion 131
Table 8.1: Ze o-sho es accu acies o bo h condi ions.
o speci ic o gene ic
D(3,4) 0.92 ±0.02 0.71 ±0.04
D(3,8) 0.85 ±0.01 0.68 ±0.07
D(3,16) 0.82 ±0.03 0.63 ±0.03
D(4,4) 0.95 ±0.00 0.82 ±0.02
D(4,8) 0.95 ±0.01 0.82 ±0.07
D(5,4) 0.96 ±0.01 0.84 ±0.06
8.5.2 Concep e e ence
We in es iga e he eme gen mappings be ween concep s and messages du ing aining wi h he
No malized Mu ual In o ma ion (NMI) sco e calcula ed o e messages 𝑀and concep s 𝐶:
NMI(𝐶, 𝑀)=𝐻(𝑀) − 𝐻(𝑀|𝐶)
0.5· (𝐻(𝐶) + 𝐻(𝑀)) ,(8.2)
The NMI sco e is maximal (i.e., 1.0) i o all messages and concep s seen du ing aining, e e y
message maps o exac ly one concep and ice e sa. In o he wo ds, a maximal sco e indica es
ha he agen s de eloped a p o ocol ha includes only one- o-one mappings be ween messages
and concep s, i.e. no ambigui y. We expec high bu no maximal NMI sco es which would indica e
ha he agen s ha e lea ned a s uc u ed bu no unambiguous mapping be ween concep s and
messages. The mean NMI sco es calcula ed o messages and concep s du ing aining in i e
uns ange be ween 0.84 and 0.95 in he “ o speci ic” condi ion, i.e. when ained on mo e gene ic
concep s, and be ween 0.77 and 0.87 in he “ o gene ic” condi ion, i.e. when ained on mo e speci ic
concep s (see Table 8.2). This indica es ha a s uc u ed communica ion p o ocol has eme ged in
bo h condi ions, while mo e ambigui y a ises when aining he agen s on mo e speci ic concep s in
he “ o gene ic” condi ion.
Table 8.2: NMI sco es o bo h condi ions.
o speci ic o gene ic
D(3,4) 0.93 ±0.03 0.87 ±0.04
D(3,8) 0.95 ±0.01 0.82 ±0.02
D(3,16) 0.87 ±0.01 0.77 ±0.02
D(4,4) 0.94 ±0.01 0.87 ±0.05
D(4,8) 0.84 ±0.03 0.83 ±0.03
D(5,4) 0.87 ±0.02 0.83 ±0.04
8.5.3 Gene aliza ion s a egies
When agen s gene alize o no el concep s in he ze o-sho es , he e a e wo concei able s a egies.
Fi s ly, agen s migh euse messages ha ha e been success ully used du ing aining also on he
es da ase . Secondly, agen s migh in en no el messages o desc ibe he no el concep s in he