Towards the computational analysis of a peripheral literary tradition: The case of Alsatian theater

Towards the computational analysis of a peripheral literary tradition: The case of Alsatian theater Pablo Ruiz Fabo Université de Strasbourg · UR 1339 LiLPa Institute of Contemporary History, Universidade NOVA, Lisbon November 2023 REWIND project workshop Outline ●Scientific context ○Large-scale analysis of cultural products ○Literary polysystems ●Digitally underrepresented traditions: Alsatian ○Automatic corpus encoding ○Knowledge engineering for metadata representation ○Analysis of non-standardized text: Handling variation ●Resource sharing Culturomics (2010) The “million books” trope (2013) 5 (2013) 6 (2018) (2019) 8 Large scale analysis of cultural products ●Relies on large (textual) corpora ●Subcorpus characterization ○Period, author group ●Based on concrete linguistic or structural traits ○Automatic annotation ○Quantitative analysis ●Complements close reading 9 Emotion analysis 16 Dennerlein et al (2022) Schmidt et al. (2019) 17 Macroanalysis of theater texts (French) ●French classical theater (Schöch, 2017) ●Subgenre characterization based on LDE topic models 18 Macroanalysis of theater texts (French) 19 (2017) ●Special issue of Revue d’Historiographie du Théâtre DraCor(Drama Corpora) : multilingual platform Fischer et al. (2019) 20 Situation for the most commonly studied traditions 21 Outline ●Scientific context ○Large-scale analysis of cultural products ○Literary polysystems ●Digitally underrepresented traditions: Alsatian ○Automatic corpus encoding ○Knowledge engineering for metadata representation ○Analysis of non-standardized text: Handling variation ●Resource sharing ●Proposed by Even-Zohar (1990) ●Method pays attention to ○Interaction between a core literary tradition and peripheral or emerging ones ○Canon formation practices 23 Literary polysystems Example: Galician literary system (in 1977) 24 ●Samartim (2010) ●Configuration of the Galician literary system 1974-1978 Source: Samartim (2010) Italian literary polysystem (including translated literature) ●Rybycki (2018) ●Position of translated authors in the polysystem (Italian) ●Clustering based on lexical cooccurrence 25 Image source Several challenges ●No previous electronic text collections for Alsatian theater ○MeThAL project first effort in this direction, covering 1870-1940 period ●No standard orthography for Alsatian varieties ○Statistical analyses requiring a common vocabulary across corpus documents hindered ●NLP resources limited ○But under construction (cf. DIVITAL and RESTAURE projects) 32 Benefits? ●Computational Literary Studies (CLS) ○New material, possibility to compare with results for “larger” traditions ●NLP ○Stimulate methods for small unstandardized, “code-switched” corpora ●More widely ○Promote vitality of language varieties involved 33 34 Ongoing project Goals ●Large electronic corpus creation and quantitative analysis to complement existing knowledge of the Alsatian theater tradition ●Linguistic resources for the analysis of the corpus Corpus creation: Play selection ●Target: 50 plays (>500,000 tokens), 1870-1940 ●Prefer digitized sources (Ca. 250 plays at National Library Strasbourg) ●Prefer plays unanimously considered important in secondary literature ●Variety: ○Genres: comedy must predominate, but also include serious plays and tales ○Author origins and publishers ○Decades covered ○Author status (established or lesser know) ○Male/female authors 35 Corpus creation: Encoding ●TEI format (Text Encoding Initiative) ●Structural divisions ●Detailed bibliographic metadata ●Social variables to describe characters (TEI prosopography) 36 TEI Encoding Recommandations de la Text Encoding Initiative (TEI) 37 38 Source for our TEI versions: Image-mode digitizations by Bnu (National library in Strasbourg) TEI : Bibliographic metadata 39 TEI : Bibliographic metadata 40 41 TEI encoding workflow 48 TEI encoding workflow 49 TEI encoding workflow 50 Corpus creation: TEI encoding PLAYS’ BODY ●Initial 7 plays with rule-based methods ●Based on them (and their OCR), trained Conditional Random Fields to predict: ○act/scene ○stage direction ○speaker name ○speaker’s text ○verse line ●Performance ≥ 0.89 F1 for all classes ●Small training set (largest class 100K) ●Manual correction follows prediction 51 FRONTMATTER ●Too much variability for automation ●Manual transcription of ○Bibliographic metadata ○Dramatis personæ First, we OCR plays. Based on OCR output (after its manual correction) TEI 52 Crédits: Andrew Briand OCR to TEI Océrisation 53 Credits: Andrew Briand OCR to TEI Océrisation 54 Credits: Andrew Briand OCR to TEI OCR to TEI: Rule-based methods ●Rules defined on the basis of ○Lexical triggers ○Typographical and layout information (bounding boxes in OCR output) 55 OCR to TEI: Machine learning ●Supervised learning based on OCR output and its corresponding TEI encoding ●Main developer Andrew Briand ●Sequence labeling task ●Typical method: CRF (Conditional Random Fields) ●The GROBID Dictionaries tool often used (Khemakhem et al., 2018a) ○CRF cascade ○Each model in the cascade is used for a single TEI element ○Used for complex documents like dictionaries or encyclopedias (Khemakhem et al., 2018b) et catalogues (Gabay et al., 2021) 56 FETE: Fast Encoding of Theater in TEI FETE: Simple CRF ●The structure of theater plays is simpler than GROBID’s main use cases (dictionaries, encyclopedias) ●We implemented a single CRF (no cascade) with scikit-learn to predict all relevant TEI elements ●Predicts: ○Act / scene divisions ○Speakers for each speech turn ○Speech turn text ○Stage directions ○Verse lines (inside a speech turn) 57 TEI encoding workflow 64 Character lists provide a sociobiographic and dramatic overview (Wiedmer et al., 2020 i.a.) 65 Ruiz & Werner (2021) Character social variable annotation ●Links with earlier studies on the tradition, which have focused on the plays’ “social picture” ●Annotating dramatis personæ is a feasible effort while text is still under encoding ●It can give an overview of social groups in the plays and their evolution ○Allows us to find plays where these groups interact ○Can help predict conflicts depicted 66 Character social variable annotation ●Annotated profession, professional group, class, gender, age ○Missing: origin and languages ●Characteriseme taxonomy (cf. Galleron, 2017) with mimetic features (Phelan, 1989) ●TEI feature structures 67 Transcription et annotation des personnages 68 Socioprofessional groups ●12 groups ○professionals, scientific, technical ○intermediate professions ○service and sales ○crafts ○industry and transportation ○agriculture 69 ○elementary professions ○rentiers ○clergy ○military ○government officials ○associative world ●Inspired by historical profession taxonomies like HISCO (van Leeuwen et al., 2014) Ruiz & Werner (2021) 70 f general fs mimetic fsdDecl character_specification fs synthetic f socio_economic_status f relation_position f language fs professional _activities fs socio_economic _other fs relation_type f professional _category f occupation f social_class f family f status_cerf f personal f professional fs general_features 71 f general fs mimetic fsdDecl character_specification fs synthetic f socio_economic_status f relation_position f language fs professional _activities fs socio_economic _other fs relation_type f professional _category f occupation f social_class f family f status_cerf f personal f professional fs general_features 72 f general fs mimetic fsdDecl character_specification fs synthetic f socio_economic_status f relation_position f language fs professional _activities fs socio_economic _other fs relation_type f professional _category f occupation f social_class f family f status_cerf f personal f professional fs general_features 73 f general fs mimetic fsdDecl character_specification fs synthetic f socio_economic_status f relation_position f language fs professional _activities fs socio_economic _other fs relation_type f professional _category f occupation f social_class f family f status_cerf f personal f professional fs general_features <person xml:id="mtl-per-0890"> <bibl corresp="#mtl-090"/> <persName>Alice Sandel</persName> <note type="roleDesc">Dactylo</note> <occupation>Dactylo</occupation> <fs type="character_specification"> <f name="specification_type"> <fs type="mimetic_features"> <f name="general"> <fs type="general_features"> <f name="sex"><symbol value="F"/></f> </fs> </f> <f name="socio_economic_status"> <vColl> <fs type="professional_activities"> <f name="occupation"> <symbol value="typist"/> </f> <f name="professional_category"> <symbol value="intermediate_professionals"/> </f> </fs> <fs type="socio_economic_other"> <f name="social_class"> <symbol value="lower_class"/> </f> </fs> </vColl> </f> <f name="language"> <default/> </f> </fs> </f> </fs> </person> TEI personography: Data volume ●Plays: 231 ●Characters: 2,386 ●Professions: Vocabulary with 350 unique professions (divided into 12 groups) 81 82 Group evolution: High-change groups 83 Group evolution: Stable groups 84 Group evolution: Stable groups 85 Group evolution: Discontinuous groups 86 Group evolution: Female characters Character gender distribution 87 ●Crossing bibliographic metadata with character annotations, characterization differences emerge between female/male authors 88 Female authors Male authors Characters with a profession in the dramatis personæ 35.71% 48% - Female characters among these 45.71% 11.3% - Male characters among these 54.29% 85.43% Characterization differences 89 Social groups and subgenre: Gender Outline ●Scientific context ○Large-scale analysis of cultural products ○Core and periphery ●Digitally underrepresented traditions: Alsatian ○Automatic corpus encoding ○Knowledge engineering for metadata representation ○Analysis of non-standardized text: Handling variation ●Resource sharing What about text-based analyses? ELAL: An Emotion Lexicon for the Analysis of Alsatian Theatre Plays ●Text-based analyses require handling scriptolinguistic variation ●ELAL (work by Delphine Bernhard, 2022) ○Creating an emotion lexicon that handles variation ○Based on French-Alsatian bilingual lexica ○Based on existing emotion lexica for French and German ■FEEL (Abdaoui et al, 2017) ■NRC lexica (Mohammad, 2020) ●All source lexica are aggregated into a large network ●Variants are identified with string similarity + the double metaphone algorithm 97 What about text-based analyses? ELAL: An Emotion Lexicon for the Analysis of Alsatian Theatre Plays ●Manual correction to detect incorrect variants ●Lexicon: 3,273 entries ▫ 11,920 Alsatian forms ▫ 3.64 variants/entry Delphine Bernhard (2022) 98 What about text-based analyses? ELAL: An Emotion Lexicon for the Analysis of Alsatian Theatre Plays Bernhard (2022) 99 doi.org/10.34847/nkl.40cex998 Results ●Best model overall: Ridge Classifier, using all the features. ●DeezyMatch (deep learning, Coll et al., 2020) obtains slightly lower results (due to the amount of training data?) (experiments by Bernhard, 2022) 100 101 Potential use: Variant-based search EDYTHA: Emotion DYnamics in THeater in Alsatian ●Applies ELAL lexicon to determine emotion evolution patterns in ●Based on Tweet Emotion Dynamics tool (Vishnubhotla & Mohammad 2022) ●Main developer: Qinyue Liu ●Configurable term-weighting schemes to give more importance to emotion terms that are discriminative in a given text-unit ○tf-idf weights, with idf per play ○tf-idf weights, with idf per speech turn 102 EDYTHA: Emotion dynamics in theater in Alsatian ●Sanity check: positive correlation between anger and fear negative correlation between anger and joy 103 EDYTHA: Emotion dynamics in theater in Alsatian ●Relative position of dramas vs. comedies according to sadness vs. joy vocabulary 104 105 git.unistra.fr/methal Liu et al (2023) 112 113 114 115 Conclusion 116 Conclusion 117 Conclusion ●The rich Computational Literary Studies results available for major literary traditions will be possible for lesser studied ones ●A lot of effort is needed to create the resources that enable such results ●As more NLP becomes available for Alsatian varieties—cf. corpora developed in DIVITAL (Bernhard et al., 2022), the range of CLS questions we can address will increase ●Metadata-based studies a possibility when other resources are lacking 118 New project ●Comparison is based on character lists, play settings, song names + bibliographic metadata ●Assess influence of German and French popular genres on Alsatian theater ●We will attempt to represent plays as feature vectors and apply similarity metrics 119 Corpus Target data volume 120 plays authors characters Alsatian 211 103 2,099 German 109 44 1,826 French 142 69 1,477 Already annotated(≈40%) ●For each tradition ○300 plays ○25 plays per decade (1800 - 1929) First trends 121