Arbitraliųjų lietuvių kalbos kolokacijų nustatymas

JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ Vytautas Magnus University IDENTIFICATION OF ARBITRARY LITHUANIAN COLLOCATIONS KEYWORDS: arbitrary collocations, Lithuanian, restricted lexical combinability, vector method, automatic identification, manual linguistic identification. ANNOTATION The object of the study presented in this article is arbitrary collocations in Lithuanian. Their combinability is governed not by semantic but by lexical restrictions; such collocations do not directly reflect extralinguistic reality. The relationship between their words is arbitrary: although, given a similar meaning, more than one collocate could theoretically be chosen, speakers tend to use one particular option. The article describes methods for identifying arbitrary collocations, tested as part of the project “A Manual Method of Linguistic Analysis and an Automated Method of Computational Linguistics (ARKA)”. The data for the study were drawn from the Database of Lithuanian Fixed Word Combinations. ABSTRACT The article aims to describe the methodological approaches of arbitrary collocation (AC) identification developed within the framework of the project “Arbitrary Collocations of Lithuanian: Identification, Description and Usage (ARKA)”. The source of the research was the “Database of Lithuanian Multiword Expressions”. Arbitrary collocations were determined by combining manual and semi-automatic methods of computational linguistics. INTRODUCTION As is well known, collocations make up a large proportion of fixed word combinations. However, when these are automatically identified in corpora using statistical or hybrid methods (for a more detailed discussion, see Bielinskienė et al. 2019: 15–22), many combinations are also identified as collocations BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 2 doi.org/10.35321/bkalba.2021.94.06 because the frequency with which their constituent words occur together is automatically assessed as statistically significant, e.g.: six years, a sunny day, a beautiful woman, and so on. Thus, false collocations arise: these are formed like ordinary free word combinations on the basis of semantic rather than lexical constraints (Marcinkevičienė 2010: 90–91). Some researchers consider only genuine collocations (also called unmotivated or arbitrary collocations) to be collocations: conventional word combinations that do not directly reflect extralinguistic reality, with an arbitrary relationship between their words (for a detailed discussion of arbitrariness, see Kovalevskaitė et al. 2021). The subject of this study is precisely arbitrary collocations (hereafter ACs), that is, collocations characterized by limited lexical combinability (lexical restrictedness): although, in theory, more than one collocate could be chosen on the basis of a similar meaning, usage tends to favor one particular collocate, e.g., understand a hint rather than *perceive a hint, within arm’s reach rather than *graspable by hand. Rūta Petrauskaitė writes: “The absence of a direct relationship with the reality being named means that not only does the combinability of the same word differ greatly across languages, but even within the same language one member of a collocation cannot be replaced with another synonymous one without risking an unusual-sounding phrase, e.g., time goes, runs, passes, but not *walks, scurries, crawls” (Marcinkevičienė 2010: 91). According to R. Petrauskaitė, the collocates of the node in genuine (arbitrary) collocations are unpredictable; the basis of their combinability is lexical constraints (Marcinkevičienė 2010: 90). ACs have so far received little attention from Lithuanian language researchers, although identifying them would help distinguish these collocations as the core of Lithuanian collocations, particularly for purposes of translation, teaching Lithuanian as a foreign language, and lexicography. ACs are being analyzed for these purposes as part of the project “Arbitrary Lithuanian Collocations: Identification, Description, Usage (ARKA)”1, which aims to describe Lithuanian ACs and develop lexicographic and teaching resources for them. The aim of the article is to describe the methods used in the ARKA project to identify Lithuanian ACs and to compare the results obtained using different methods. ACs were analyzed using manual linguistic analysis and automated computational linguistics methods. 1 The project is funded by the Research Council of Lithuania (No. S-LIP-20-18); for further details, see https://arka.pastovu.vdu.lt. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 3 doi.org/10.35321/bkalba.2021.94.06 Sources: 1) The Database of Lithuanian Fixed Word Combinations (hereafter PŽJDB)2, which contains approximately 12,000 collocations; 2) the Delfi.lt corpus3, consisting of 70 million words and covering texts published in 2014–2016 in various sections of the Delfi.lt news portal (for more on the corpus, see Bielinskienė et al. 2019: 14–15). The research material consists of collocations formed with 97 nouns (place, environment, way, man, woman, thing, etc.4). Their structures are adjective / participle used adjectivally + noun, verb + noun, and noun + noun. The study identified approximately 2,400 collocations as AK. Most are two-word collocations, but some are three-word collocations; the latter are usually formed from verbs, adjectives, and nouns (the order of the constituents may vary). The first part of the research material, namely collocations with the structure adjective + noun, was discussed in an earlier article (Kovalevskaitė et al. 2021). It presented an experiment in identifying AK using one of the methods of computational linguistics—the automated (vector) method—and described this method in detail. This article summarizes the entire study on AK identification. The first section presents both methods used: the non-automated method (see Section 1.1), including a description of the criteria and tests applied, and the automated method, whose application is illustrated using a group of collocations with the structure verb + noun (see Section 1.2). The second section draws on data from all stages of the study to discuss and compare the results of applying the methods, their potential and shortcomings, which are important in evaluating the list of identified AK and planning further research on AK. 1. METHODS FOR IDENTIFYING ARBITRARY COLLOCATIONS In writing about AK, it is noted that the boundary between arbitrary and so-called trivial collocations could be drawn at the boundary of abstract nouns: „<...> collocations based on abstract nouns can be considered genuine, unmotivated, 2 http://resursai.pastovu.vdu.lt/paieska/paprastoji 3 https://klc.vdu.lt/pastovuSearch.html 4 Thus, nouns—both concrete and abstract—that form fixed word combinations, identified from the Delfi.lt corpus, are discussed in greater detail by Bielinskienė et al. 2019. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 4 doi.org/10.35321/bkalba.2021.94.06 “while fixed expressions with concrete nouns are generally based on extralinguistic reality and are trivial” (Marcinkevičienė 2010: 91). Thus, one of the criteria identified in studies of collocations is noun semantics. This study found that most ACs are indeed formed with names of abstract concepts; however, arbitrariness was also observed in collocations with concrete nouns, cf.: limit speed, do exercises, take a bribe, leave the square, take out a loan, talk on the phone, stage a play, etc. The challenge of this study was not only to apply the criteria for identifying ACs, but also to choose a methodology suitable for working with large amounts of data, so that it could be used to compile lists of Lithuanian ACs after examining a large number of collocations (for this purpose, material from the PŽJDB was used). In the course of the project, two methods for identifying ACs were tested and applied: a manual and an automated method. First, the manual approach was applied, using the criteria for identifying ACs (see Section 1.1). Since the PŽJDB contains a large amount of data, one of the project’s objectives was also to test, alongside the manual method, an approach that would allow the process of identifying ACs to be at least partly automated (see Section 1.2). 1.1. Manual identification of ACs Two main criteria were applied to identify ACs manually: 1) the criterion of restricted lexical combinability, assessed using the synonym substitutability method and/or the method of lexical semantic groups of nouns; 2) the criterion of figurative meaning. If a collocation met at least one criterion, it was considered arbitrary. The following discusses how each criterion was applied to assess whether a collocation should be considered arbitrary. 1.1.1. Identifying ACs on the basis of restricted lexical combinability As mentioned in the introduction, collocations with restricted lexical combinability are considered arbitrary, that is, an adjective or verb used with a particular meaning in a phrase with a noun is characterized by a certain, more or less arbitrary, restriction on substitutability (Nesselhauf 2003, 2005). Thus, based on the concept of restricted lexical combinability, any component of a collocation cannot, in a particular meaning, be BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 5 doi.org/10.35321/bkalba.2021.94.06 used with all syntactically and semantically possible collocates. Although it is theoretically possible to find substitutes for the collocates, they would either be entirely unsuitable or unusual, e.g.: favourable vs. *useful environment; multilayered vs. *layered novel. Linguists tested the criterion of restricted lexical collocability using two tests: 1) using verb or adjective synonyms (i.e. whether they can be replaced by other words of similar meaning that can be regarded as synonyms); 2) using the lexical-semantic groups of nouns. Since ACs are considered to be collocations in which one of the components generally cannot be replaced by another word—a synonym of similar meaning—the synonym-substitution test was used to identify ACs (adapted from Nesselhauf 2003): if one component of a collocation cannot be replaced by another synonym of similar meaning, that collocation was considered an AC. For example, in assessing the arbitrariness of the collocations disobey the law, implement a system, work put in, possible verb synonyms were tested *violate the law, *introduce a system, *inserted / loaded work, and it was determined whether substitution was possible; if not, the collocations were classified as ACs. The test was conducted by three linguists using the introspection method, but it was based on collocations identified in the Delfi.lt corpus. Nevertheless, it must be acknowledged that finding a synonym was often quite difficult, so the second test—that of the lexical-semantic groups of nouns—was also frequently applied. Restricted lexical collocability was also assessed by considering the diversity of lexical-semantic groups: on the basis of the collocations listed in the PŽJDB, it was determined how many lexical-semantic groups of nouns a collocation’s adjective or verb combines with, and the extent of lexical collocability was assessed accordingly. To determine as accurately as possible which words a noun can be used with, its entire paradigm was analysed, i.e. all collocations formed with that noun. However, although the research material comprised approximately 12,000 collocations available in the PŽJDB, it was sometimes apparent that the information was insufficient. If the research material contained only one collocation with a particular adjective/verb, the linguists filled out the paradigm on the basis of their linguistic intuition as native speakers; if the database contained several collocations, the tests for restricted lexical collocability were applied to the available paradigm. The extent of collocability was assessed on the basis of the lexical-semantic groups of nouns, taking the following conditions into account: BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 6 doi.org/10.35321/bkalba.2021.94.06 a) Collocations were not considered ACs if an adjective/verb combines with many nouns belonging to a wide range of lexical-semantic groups and those groups are large. Such adjective/verb collocability was considered broad, without any clear lexical restrictions; cf. the abundance of collocations with good: good result, work, world, person, time, life, way, thing, country, team, day, decision, year, weather, place, side, season, opportunity, price, match, car, food, sign, health, road, man, child, product, question, beginning, means, etc.; cf. also buy: buy a car, food, a product, property, land; b) If an adjective/verb combines with many nouns belonging to a wide range of lexical-semantic groups, but one group stands out or only a few nouns from a particular semantic group are used, the collocation could be considered an AC. For example, although to park has broad collocability, this verb cannot be used with the names of all means of transport; cf. park a car (as well as park a bicycle, a motorcycle), but moor a ship, land a plane. Thus, due to its restricted collocability, the collocation park a car is classed as an AC. The verb to have is considered to have broad lexical collocability, since it can form many and varied collocations; however, one collocation stands out across its entire paradigm: have a matter—it is considered arbitrary. Words from two lexical-semantic groups—natural phenomena and feelings—are used with to turn foul, but collocations are formed only with the nouns the weather turned foul and the mood turned sour, so both cases are ACs; c) Collocations belonging to one large semantic group are not considered ACs. For example, only one collocation, deduct taxes, was found in the PŽJDB, but linguists can fill out the paradigm with other words on the basis of their native-speaker intuition (one can deduct money, a loan, and other financial items), and therefore do not recognize deduct taxes as an AC. This is a large lexical-semantic group that is represented by only one collocation in the PŽJDB. Restricted lexical collocability was tested by assessing both words in a collocation: the adjective/verb cannot be replaced by other synonyms (test 1), and/or the noun paradigm is restricted (test 2). For example, the collocation food served would suggest that there may be many nouns in the food group (one can serve food, drinks, dinner, lunch...). Thus, it does not meet the condition of test 2, since there is only one semantic group; however, it meets condition b, because serve cannot be replaced by hand over. Thus, according to test 2, the collocation is classified as arbitrary. Another example: BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas 7 doi.org/10.35321/bkalba.2021.94.06 The verb surengti ‘to hold’ is used with nouns denoting events (surengti rinkimus, koncertą... ‘to hold an election, a concert...’); therefore, such collocations should not be regarded as arbitrary according to the criterion of the size of the semantic group. However, they can be classified as ACs under Test 1, since in this case surengti cannot be replaced by the synonym suruošti. The synonymy test cannot always be applied. For example, it cannot be applied when there are several AC variants, cf.: kainos kyla ir kainos auga, didėja; klausimas kamuoja ir klausimas kankina; duoti galimybę ir suteikti galimybę; įgyti teisę ir įgauti teisę. Nor can this test be applied in cases such as sumokėti mokesčius, apgyvendintas miestas—it is difficult to suggest even a hypothetical synonym for the words sumokėti, apgyvendinti. If no synonym can be found, substitutability cannot be tested, since there must be at least a theoretical possibility of choosing between two (or more) synonyms. When checking for restricted lexical collocability using both the synonymy test (Test 1) and the lexical-semantic group test (Test 2), the linguists relied on their competence as native speakers. However, the results of such work are inevitably affected by subjectivity (on this problem of identifying synonyms, see Divjak, Gries 2006: 24–26). Therefore, an automated method was also tested for assessing the substitutability of synonyms (see Section 1.2). 1.1.2. Identifying ACs on the basis of figurative meaning Kai kuriuose ankstesnio laikotarpio tyrimuose (plg. Howarth 1998) tokios kolokacijos, kurių vienas iš dėmenų – perkeltinės reikšmės, vadinamos ribotosiomis kolokacijomis (angl. restricted collocations). Pasak Annos Reder (2006: 82), kolokacijos su perkeltinės reikšmės žodžiu yra riboto junglumo junginiai, kuriuose nenuspėjama ryši tarp kolokacijos narių lemia konvencija, o metaforiškumas yra konvencijos dalis. Todėl šios kolokacijos svarbios kalbos mokymui (Lewis 2000: 29). Tarkim, su anglų kalbos būdvardžiu heavy vartojami įvairūs daiktavardžiai, kurie reiškia dalykus su požymiu „sunkus“, bet žodžių junginyje heavy smoker (liet. užietėjęs rūkalius) minimas ne objektas, o asmuo, todėl mokant kalbos šią kolokaciją verta išskirti kaip įsidėmėtiną žodžių junginį. Adjectives used figuratively generally have a broader range of collocability, and collocations with adjectives that have broader collocability are more difficult for learners (Nesselhauf 2005: 238). Given their relevance to language teaching and translation BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 8 doi.org/10.35321/bkalba.2021.94.06 (Volungevičienė 2010; Stulpinaitė et al. 2016), the criterion of figurative meaning was used to identify Lithuanian ACs. In the non-automated analysis of collocations, the use of one of the components of a collocation (an adjective, verb or noun) in a non-literal sense was taken as an indicator of an AC. This was established by analysing the entire paradigm of a given word, generated from the Delfi.lt corpus data (see column 1 of Table 1). Such a paradigm shows which other adjectives or verbs a particular noun forms collocations with, and what meanings both the noun and its collocates—adjectives and verbs—have in those collocations. Sometimes only a broader context makes it possible to determine whether a word is used literally or figuratively; in such cases, the PŽJDB and the Delfi.lt corpus were consulted, cf. žalias + automobilis: ‘green + car’: The representative of „Unicredit Leasing“ in Lithuania notes a couple more trends in the green-car market. The green car driven by Marius Slaboševičius finished half a second <...> behind. The „Tesla Model S“ has won several best-car awards, one of which is the 2013 World Green Car. Interesting concept models of green cars. The examples show that the adjective žalias ‘green’ describes a car as environmentally friendly, but does not specify its colour5. This collocation was considered arbitrary. Of course, the Delfi.lt corpus, from whose data the PŽJDB was compiled, contains many instances in which žalias describes the colour of a car. Noun phrases with adjectives and verbs were considered ACs if at least one component was used figuratively. In such cases, the tests described above (see Section 1.1.1) were not applied. Figurative use of adjectives and verbs is particularly characteristic of collocations involving abstract nouns, as well as a few nouns denoting concrete things, for example, collocations formed with the nouns amžius, aplinka, apsauga, būdas, dalykas, dalis, dėmesys, duomenys, galimybė, gyvenimas, informacija, įstatymas, įvykis, kalba, karas, kartas, klausimas, laikas, metas, pabaiga, pagalba, pradžia, reikalas, rezultatas, saugumas, sprendimas, sveikata, teisė, turtas. 5 We note that attributive forms also signal distinctive usage. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 9 doi.org/10.35321/bkalba.2021.94.06 Collocations with abstract nouns, as established by R. Petrauskaitė (see Marcinkevičienė 2010), can form conceptual metaphors (e.g.: gentle measure, sharp criticism, dry result, leaked information, fuel rumors, data leak, loophole in the law6). As R. Petrauskaitė states, “the most important criterion of metaphoricality is the transfer of meaning, which emerges from the opposition between a concrete and an abstract noun. When the properties of a concrete thing are used as a basis for bringing an abstract thing into a context and forming a word combination with it, a metaphor is born <...> one thing, usually abstract, is perceived as another, concrete, by means of the latter’s terminology <...>” (Marcinkevičienė 2010: 178). Adjectives (and, less often, verbs) used figuratively in combinations with concrete nouns give them semantically atypical properties (e.g.: fast food, prominent woman, visible person). As with combinations involving abstract nouns, the meaning of the word combination is often understood not from the individual words but from the combination as a whole (e.g.: iron man, golden age, wild woman; see also the examples given earlier with the noun automobile). In studies of collocations in other languages (e.g., Strakatova, Hinrichs 2019), lexicalized collocations include those with a figurative-meaning collocate in which the figurative meaning emerges only with a limited number of words from specific lexical-semantic groups. For example, the meaning of the adjective heavy in combinations with the nouns smoker, rain, traffic (Lith. heavy smoker, heavy rain, heavy traffic) is non-prototypical, since it denotes not weight but intensity. It has been established that words with medium lexical combinability7 have high metaphorization potential, for example the verbs paralyze (paralyze traffic), run out (time is running out), destroy (destroy a life). Among adjectives, there are many that, when used in literal meanings 6 For spatial, ontological, and structural metaphors in collocations, see also: Drūlienė 2008a, b; Šilinytė 2008. The aim of this article and of the project as a whole is not to classify the identified metaphors; therefore, this is not explored further, although it is understood that AC may be an interesting subject of study for specialists in cognitive linguistics. 7 Based on the formula presented by Susane Handl in 2008, collocativity scores for the components of the collocations under analysis were calculated using the Delfi.lt corpus. These studies will continue. For now, it can be said that AC should be combinations of medium strength, i.e., the collocativity score of neither the first nor the second (or third) component should be very low or very high. If, say, the score of the first component is high and that of the second is low, the first component may be used with few words, while the second may be used with very many words, e.g.: psychoactive substance, autistic child, fresh water, shady place, trilateral agreement, minor child, carbonated / distilled water, passing car. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 10 doi.org/10.35321/bkalba.2021.94.06 meanings, denote physical properties (e.g.: slippery matter, clean work, dirty work), but adjectives may also have other meanings (precious time, troubling question, major event). Application of the figurative-meaning criterion is illustrated by an analysis of the paradigms of the adjective (not)clean, the verb leak, and the noun euro (see Table 1): TABLE 1. Application of the figurative-meaning criterion Paradigm AC identification 1. clean sport, work; dirty work, matter, business 2. clean (-ly) water, environment, air, city, place, car; dirty water 1. Considered AC because of the figurative meaning of the adjective in combinations with nouns denoting abstract concepts. 2. Not considered AC, because in these collocations the adjective is used literally. 1. leak information 2. leaked (-ly) information, material, data 1–2 considered AC, because a figurative meaning of the verb is realized in combinations with abstract nouns. 1. euro weakening, strengthening 2. euro exchange rate, coin, cent 3. euro change 1. Considered AC because the euro is attributed with properties characteristic of a living object. 2. These are terms; not considered AC. 3. Not considered AC because there is no figurative meaning. As the examples in Table 1 show, in a collocational paradigm an adjective or verb used literally—or sometimes even the entire collocation with a particular adjective/verb—does not always have a figurative meaning. However, figurative meaning is clear in cases such as: painful decision / matter / event; closed-off person / disposition / life; a road / market / view opens up; win the right / a place, and so on. Such collocations are considered arbitrary. In summary, it can be argued that the figurative-meaning criterion is particularly relevant when analyzing adjective–noun collocations—it has often been relied on as BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 11 doi.org/10.35321/bkalba.2021.94.06 the main and clearest indicator of arbitrariness. In analyzing verb–noun collocations, however, the restricted lexical combinability criterion (and the semantic-group test) was applied more often. Since very few ACs consisting of two nouns were found, it is difficult to generalize about the characteristics of collocations of this structure (see Section 1.1.1). 1.2. Automated identification of ACs One of the criteria for ACs—restricted lexical combinability—was also tested automatically: that is, synonyms required for the substitutability test were identified using the vector method. The substitutability test and the application of the vector method are discussed in greater detail in an earlier article (Kovalevskaitė et al. 2021). That article describes the application of the vector method to evaluate the arbitrariness of adjective + noun collocations; therefore, this article provides a detailed account only of the evaluation of arbitrariness in verb + noun collocations. First, we will briefly recall the main stages in applying the vector method (Section 1.2.1); then we will discuss the selection of verb synonyms using the vector method (see Section 1.2.2) and summarize the results of classifying verb–noun collocations as arbitrary (see Section 1.2.3). 1.2.1. Identifying ACs on the basis of restricted lexical combinability using the vector method The vector method was used to automate AC identification and reduce subjectivity in data evaluation. The vector method was applied to identify restricted lexical combinability in the following sequence. First, related words were automatically extracted from the entire Delfi.lt corpus8. For this purpose, GloVe, a model of word vectors (embeddings), was used (Pennington et al. 2014). The vector method (Word Embedding Approach) belongs to the group of methods for automatically analyzing words according to similarities in their distributions, or contexts (distributional similarity measures). These methods are used to automatically identify various relations (hyperonymy, synonymy, antonymy, and meronymy) and are based on 8 In the previous project (“Automatic Recognition of Fixed Word Combinations in Lithuanian (PASTOVU),” see https://pastovu.vdu.lt), the vectors identified are presented here: https://clarin.vdu.lt/xmlui/handle/20.500.11821/26. They are used in this study to select the synonyms needed to identify ACs. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 12 doi.org/10.35321/bkalba.2021.94.06 the hypothesis that words with similar meanings and/or semantically related words occur in similar contexts (see Grigonytė 2010). First, vector sets (hereafter VS) of the adjectives and verbs used in the automatically compiled collocations in the PŽJDB were created9. When automatically selecting related words, adjectives and verbs with broad lexical combinability or those used non-autonomously were excluded, e.g.: geras, naujas, puikus; būti, reikėti, pradėti. Adjectives or verbs grouped into VSs were selected on the basis of contextual similarity. About 800 VSs containing adjectives and more than 3,800 containing verbs were compiled. Although the VSs included related words, the data also contained words related through other semantic relations (e.g., antonymy), or even unrelated words. Linguistic editing was therefore required to establish which related words could be considered synonyms and used to assess how restricted the combinability of one or another synonym was with a particular noun. Synonyms were defined as words belonging to the same part of speech and having very similar denotative (referential) meanings, but partly or completely different lexical combinability10. The edited verb VSs were used to assess the arbitrariness of verb + noun collocations. The aim was to determine which collocations do not allow the verb component to be replaced by a verb with a similar meaning. Collocations with a particular noun were listed for each pair or group of verb synonyms. If a collocation with that noun was listed with only one synonym from a pair or group, it was concluded that the noun under analysis combined with only one verb in that synonym pair or group, and the collocation was considered arbitrary. If the noun under analysis formed collocations with all the verbs in a synonym pair or group, the collocations were not classified as arbitrary (see Table 2). 9The related words identified by the vector method would most accurately be called sets of possible synonyms identified by the vector method. For brevity, hereafter we use vector sets, or, more precisely, the abbreviation VS. 10The concept of synonym used in this article may be associated with the more common term partial synonyms in Lithuanian lexicography (see Jakaitienė 2010). BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 13 doi.org/10.35321/bkalba.2021.94.06 TABLE 2. Stages in identifying ACs using VSs Recognized as an AC Not recognized as an AC Stage 1: a VS containing two synonym pairs, gauti–įgyti and duoti–teikti, is selected; Stage 2: this VS is used to check how many of these synonyms collocate with the noun apdovanojimas; two collocations are identified: gauti apdovanojimą (one synonym in the pair gauti–įgyti) and teikti apdovanojimą (one synonym in the pair duoti–teikti); Stage 3: it is decided that gauti apdovanojimą and teikti apdovanojimą should be considered ACs. Stage 1: a VS containing the two synonyms įkurti and įsteigti is selected; Stage 2: this VS is used to check how many of these synonyms collocate with the noun asociacija; two collocations are identified: įkurti asociaciją and įsteigti asociaciją; Stage 3: it is decided that įkurti asociaciją and įsteigti asociaciją are not ACs. For example, the noun apdovanojimas was listed with the collocations gauti apdovanojimą, teikti apdovanojimą, but not with *įgyti apdovanojimą, *duoti apdovanojimą. Thus, with the synonym pairs gauti–įgyti and duoti–teikti, the noun apdovanojimas has only one option in each case; therefore, the collocations gauti apdovanojimą and teikti apdovanojimą were classified as arbitrary. Among verb collocations, those involving semantically bleached (delexicalized) verbs were often considered arbitrary (for further discussion, see Kovalevskaitė et al. 2020), e.g.: atlikti tyrimą, duoti rezultatą, imti mokestį. The arbitrariness of the collocations with the noun asociacija was assessed using a VS containing two synonyms, įkurti and įsteigti; since this noun formed collocations with both synonyms, they were rejected as non-arbitrary. As with the assessment of collocations with adjectives, verb synonyms were tested manually in the experimental study, since the meanings of both the synonym and the noun being tested must be taken into account. If the meanings of the synonyms in the VS do not match the meaning of the verb in the collocation being assessed, the verb substitutability test BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 14 doi.org/10.35321/bkalba.2021.94.06 cannot be applied. For example, for assessing the arbitrariness of the collocation build a car, the VEs with build were unsuitable because of the meaning ‘build a building’ (166 VEs11). Thus, the stages of applying the automated method are as follows: a) automatic (VEs are generated using the vector method); b) non-automatic (a linguist reviews and, if necessary, revises them); c) partially automatic (collocations are automatically linked to specific VEs and, after review, are either confirmed or not confirmed as AKs). 1.2.2. Selection of verb synonyms in vector sets Because the generated VEs contained much redundant, irrelevant data, or so-called noise (noise), a linguist reviewed each VE and deleted unnecessary words that were not related by synonymy, and removed some VEs altogether. Different cases of linguistic editing of VEs are discussed below. The number of verb VEs was much greater than that of adjective VEs (3,800 and 800 VEs, respectively), because the list of verbs itself was also larger. A VE is generated automatically for each verb, so some VEs are similar; however, the number of synonyms in a similar VE may vary depending on which verb is the anchor. The anchor verb is the word to which the other words listed are similar; in other words, it is the word for which possible synonyms were sought (similarity may be established through a hypernymic, synonymic, antonymic, or partonymic relation; this word is underlined in synonym sets). For example, in VE 151, the anchor verb is organize, whereas VE 40 is much longer, and its anchor word is carry out: VE 151: organize, prepare, carry out VE 40: carry out, implement, prepare, organize, perform, complete, terminate As already mentioned, the automatically generated VEs still had to be reviewed by a linguist. After this work was completed, it was observed that, as with adjectives, the automatic process had collected many long VEs for verbs, containing several pairs of verb synonyms, and few containing only one synonym pair or three synonyms. Sometimes, after review, it became clear that the anchor verb on which a VE was based had no synonyms, although that VE might contain another 11 Here and below, the numbers of the VEs of potential synonyms identified using the vector method and consulted for this article are given. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 15 doi.org/10.35321/bkalba.2021.94.06 a suitable synonym pair; for example, the base verb kisti in the VE has the synonym pair paaštrėti, pablogėti. As with adjectives, antonyms were responsible for the “noise” in the verb VEs; cf. VE pasivyti–aplenkti, sutrumpinti–pailginti, griežtinti–švelninti were made up of members of a single lexical-semantic group that are not synonyms, e.g.: pasnigti: palynoti, nugriaudėti; nepasiekti: pargrūti, iššausėti, patinti, pakrypti, sušlubuoti, sudužti; vakarieniauti: pietauti; pavirti: troškinti. When refining the VEs, it often became clear that one or more verbs were unsuitable because they denoted an overly specific aspect of a similar action. For example, the VE for the word remtis included the following words: vadovautis, pasikliauti, apsiriboti, kliautis. apsiriboti was deleted, and only vadovautis, pasikliauti, kliautis. were considered synonyms. Cases were also common in which one VE word had a broader meaning, for example, paveikti, while other words assigned to the VE named aspects of negative impact more specifically, for example, pakenkti, sutrikdyti. Therefore, only the latter were considered synonymous. Of the VE pagrįsti, įvardinti, paaiškinti, pateisinti, pagrįsti and paaiškinti were recognized as synonyms, whereas pateisinti and įvardyti were not considered synonyms, since additional semantic features make their meanings too dissimilar from those of the first two words. In some such cases, it was necessary to discard the entire VE, for example, atmesti, ignoruoti, uždrausti. Sometimes even VEs for base verbs were entirely unsuitable, although one might think that synonyms should certainly be identifiable automatically, e.g.: suprastėti (a possible synonym is pablogėti, but it was not in its VE); sąlygoti (a possible synonym is lemti, but it was not listed; pakentėti was included instead). VEs for verbs containing negation were inaccurate and therefore rejected; for example, the following base verbs were excluded: negydyti, nesuskaičiuoti, nepridėti. It should be noted that VEs often had to be refined because of the derivational characteristics of the verbs as well. For example, imperfective and perfective verbs were not considered synonyms (tirti ir ištirti), nor were verbs that had acquired a different meaning through prefixation, such as palyti ir palynoti, naršyti ir panaršyti, or reflexive and non-reflexive verbs (klausyti ir klausytis). For instance, if the base verb in a VE was reflexive, the non-reflexive form was excluded. For example, the VE išsirinkti, atsirinkti, pasirinkti, nuspręsti, rinktis, the verbs išrinkti, parinkti were excluded. Derivational variants (e.g., vardinti ir vardyti, lukterėti ir luktelėti) were considered synonyms. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 16 doi.org/10.35321/bkalba.2021.94.06 Refining the VEs also revealed another feature of automatic selection: because the vector method collects words according to the similarity of their contexts, a VE may consist of synonyms that seem unsuitable only at first glance. In other words, without knowing the specific nouns that led two or more verbs to be grouped into a single VE, the quality of some VEs may be judged incorrectly. For example, VE 60 included the following words: rodyti, demonstruoti, jausti. Deleting the word jausti immediately would have been a mistake, since it appeared here because of the context shared by these verbs, which concerns feelings; cf. demonstruoti, rodyti jausmus and, for example, jausti pyktį, jausti nusivylimą, and so on. The verbs sumažėti, pigti or padidėti, ūgtelėti may be synonymous when referring to prices. All three verbs in the VE apsispręsti, priimti, pasirinkti are synonyms, while the verb imti, which is usually deleted, appeared here because it denotes priimti sprendimą; in other words, this VE should be read as follows: apsispręsti, priimti sprendimą, pasirinkti12. These examples show that more and/or different synonyms can be identified automatically, because the vector method evaluates similarity of context rather than similarity of individual word meanings. It is unlikely that, considering the individual words, we would link the verbs atsikratyti, numesti, sulieknėti as synonyms, yet they occur in the same VE because they were used in a shared context that also included the noun svoris. More examples can be given; cf.: trukti, užtrukti, gyvuoti, egzistuoti, užsitęsti; domėtis and užsiimti; gerti and vartoti; nusiteikti and pasiruošti; skambinti, kreiptis, pranešti; atkurti, atstatyti, atgaivinti; peržiūrėti, koreguoti, atnaujinti, keisti; išprovokuoti, lemti, peraugti; šokiruoti, sukresti, priblokšti; atlaikyti, ištverti, įveikti. In a sports context, all four verbs—užsitikrinti, iškovoti, pasiekti, išsaugoti—are synonymous, since winning something is sometimes expressed as follows: užsitikrino sau vietą finale or išsaugojo titulą. These findings show that assessing synonymy requires looking beyond individual words, since members of a synonym group may be either single words or fixed word 12 VEs may contain words that are syntagmatically related, e.g., įžeisti and nenorėti. Since nenorėti often occurs in a context shared with įžeisti (when it appears in the phrase nenorėjo įžeisti, bet...), the two were grouped into one VE. Another case is: liautis: nustoti, verkti. The verb verkti appeared here because it was selected from its frequent context liovėsi verkęs / verkusi. In other words, very long (and unsuitable) VEs are generated for those base verbs that are used as auxiliaries and, owing to their grammatical properties, have broad collocational range, e.g.: daryti, imti, pradėti, norėti, reikėti, privalėti, reikėti, stengtis, nepajėgti. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 17 doi.org/10.35321/bkalba.2021.94.06 combinations. The vector method, which compares contextual similarity, is precisely what brings out the associations between individual words and their contexts. It would be difficult for a linguist to cover so much data (for more on the use of a broad spectrum of distributional measures for automatic synonym analysis, see Divjak, Gries 2006). It was observed that the vector set (VS) of a verb with a more specific meaning and narrower combinatorial range is often shorter and cleaner, presumably because of its narrower combinatorial range. A comparison of the VSs of the synonyms nerimauti, jaudintis, būgštauti and nuogąstauti shows that the VS of nerimauti is longer and less clean, containing words with broader meanings, for example, bijoti (nerimauti: jaudintis, baimintis, bijoti), whereas the VS of būgštauti contains only those four words: būgštauti: nerimauti, jaudintis, nuogąstauti. It was also observed that VSs based on verbs with more concrete meanings (duoti, bėgti, išimti, atsistoti, medžioti, kvėpuoti, sugesti, kalėti) may contain few or no synonyms, because the words collected on the basis of a verb with a specific meaning are thematically rather than synonymically related; cf.: išimti: išmesti, išbraukti, išnešti; medžioti and žvejoti; kvėpuoti and vaikščioti; virti and kepti; kalėti and nuteisti; sugesti: skilti, perkaitti, sprogti, užsikimšti, užsidegti. Such VSs are also composed of two words, but may be longer. They can sometimes be very long; for example, the VS with the base verb sumenkti contained as many as 26 verbs. In summary, it can be noted that after refinement, 2,634 of the 3,800 VSs (70 percent) were obtained. Thus, the automatically generated VSs are not sufficiently accurate to be used without editing in the subsequent stage of AC identification. Such errors may be due to characteristics of the Delfi.lt corpus—although it is large (about 70 million words), lexical diversity may be lower in texts of a single style. A corpus with a greater variety of functional styles would be needed to generate more accurate VSs. 1.2.3. Results: VS quality and recognition of arbitrariness A comparison of adjective and verb VSs before and after review showed that in both cases a fairly large number of VSs were discarded as uninformative: 1) of the 1,927 adjective VSs, 800 remained after review (41 percent); 2) of the 3,800 verb VSs, 1,166 remained after review (30 percent). The remaining VSs also needed editing. It should be noted that no assessment was made of how many words are usually deleted from a single VS. Often, auxiliary or not fully independent words had to be deleted BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 18 doi.org/10.35321/bkalba.2021.94.06 verbs: daryti, imti, gyventi, vertėti, reikėti, privalėti, pradėti, gauti, norėti, stengtis, pavykti. Many of the words to be deleted were at the end of a VE, where words with the lowest applied score threshold (0.5; the highest score is 1) occurred, but there were also quite a few cases in which a word with a high score, listed in positions 2–5 in a given VE (a VE can contain a maximum of 20 words), was deleted. It should be emphasised that the evaluation was based on the VE and was limited to the words in that VE. If there may be other synonyms, but they were not in the VE used for evaluation (they were not selected by the vector method, were not used in the Delfi.lt corpus, etc.), some collocations may have been classified as non-arbitrary (or vice versa). For example, in evaluating the collocations krintanti žvaigždė and kylanti žvaigždė, the VE used contained the verbs augti, kristi, kilti, mažėti, vystytis, didėti. This VE does not include griūti, which could be synonymous with kristi; consequently, krintanti žvaigždė was not recognised as an AC. The study also revealed advantages of automatic data selection: 1) as mentioned in Section 1.2.2, the vector method can identify more synonyms, and different ones, than the introspection method alone; 2) the vector method is suitable for processing large quantities of data, whereas it would hardly be possible to assess the contextual similarity of such a large number of words manually. After applying the substitutability test to the VEs of verb synonyms, 291 collocations were recognised as arbitrary. Clearly, the synonyms in the VE determine which collocations will be recognised as ACs. On the one hand, this is a limitation of the method; on the other, it makes it possible to automate the process of identifying arbitrariness. When the recognition of arbitrariness depends solely on a particular VE, it can often happen that some collocations are not selected based on one VE but are selected based on another. For example, when assessing the arbitrariness of the collocations kurti įvaizdį and formuoti įvaizdį using VE 368, kurti, formuoti, neither of the two collocations would be classified as an AC. However, the verbs kurti, formuoti also occur in other VEs to which the same noun įvaizdis is assigned; therefore, the collocation kurti įvaizdį is recognised as an AC based on another VE (45): kurti, sukurti, vystyti, plėtoti, įgyvendinti, rengti, įkurti. Another case is also possible: the same collocations may be recognised as arbitrary according to more than one VE. For example, atlikti mankštą, daryti mankštą can be evaluated as ACs both according to VE 11 (atlikti, vykdyti, daryti), and VE 15 (padaryti, atlikti, sukurti; BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 19 doi.org/10.35321/bkalba.2021.94.06 nuveikti). The problem arises when an arbitrary collocation is not regarded as arbitrary if this cannot be established based on any VE. This happened quite often when evaluating collocation variants. For example, both kelti and didinti are used with the noun mokesčiai, and both verbs occur in VE 52, which is used to assess arbitrariness; therefore, these collocations will not be classified as ACs. However, they are variants of ACs (cf. auga, kyla, didėja kainos). The same applies to atlikti / vykdyti / daryti tyrimą (according to VE 11, these combinations were not recognised as ACs). The same VE also contained gauti and įgyti, so gauti diplomą and įgyti diplomą were not recognised as ACs. These cases show that, when evaluating the arbitrariness of collocations according to the criterion of restricted lexical combinability, both the testing tools used and the fact that a non-automatic method is more suitable for identifying AC variants are important. Because testing the substitutability of synonyms is so closely tied to the words in a particular VE, it can also happen that collocations with similar but opposite meanings are not selected. For example, the collocation kyla nepasitenkinimas was selected, whereas mažėja nepasitenkinimas was not. Manually, both would be evaluated as ACs. Thus, when data assessed through linguistic introspection are used to evaluate arbitrariness, some of the ACs lost by Method 2 are preserved. However, the linguistic non-automatic method involves subjectivity. The automated method is more objective, although it must be acknowledged that the data are more fragmentary (testing uses only as many VEs, and of such kinds, as have been collected). In addition, not all collocations that are ACs are recognised as such. 2. RESULTS OF IDENTIFYING AK Looking at the numbers of identified AKs, it is clear that very few AKs identified by the manual and automated methods overlap (for details, see Sections 2.1 and 2.2); consequently, each method separately yields important data. The data collected using the manual analysis method are much more numerous than those identified automatically, but manual analysis applied not one but two selection criteria (restricted lexical combinability and figurative meaning), whereas the vector method applied only restricted BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 20 doi.org/10.35321/bkalba.2021.94.06 lexical combinability. Thus, applying more criteria yields more collocations with a wider variety of adjectives and verbs. Drawing on examples of specific AKs, this section will explain why these particular results were obtained, how they were determined by the criteria applied and the data used. First, we will discuss what the data show. All AKs formed with a particular adjective or verb can be identified in different ways: a) by only one of the two methods; b) some collocations by one method and some by the other; c) some collocations by one method, some by the other, and some by both methods. In the following subsections, we will discuss AKs with different structures by commenting on the results obtained using both methods. 2.1. Arbitrary collocations formed by adjectives and nouns A total of 1,048 AKs with the structure adjective + noun were identified. These collocations include 320 different adjectives. It is worth noting that several or a dozen AKs are formed with some adjectives (didelis, mažas, atviras, aukštas, modernus, išskirtinis, pagrindinis, and so on), while with others (e.g., netrumpas, nuodugnus, permainingas, pikantiškas) only isolated AKs are formed (on average, 3.3 collocations are formed with the same adjective; for details, see Table 3). TABLE 3. AKs identified using different methods (adjective + noun) AKs identified by the manual methodAKs identified by the automated methodAKs identified by both methodsNumber (percent)Number of different adjectives (average number of AKs formed with the same adjective)Number (percent)Number of different adjectives (average number of AKs formed with the same adjective)Number (percent)Number of different adjectives (average number of AKs formed with the same adjective)577 (55.06)230 (2.5)306 (29.2)86 (3.56)165 (15.74)42 (3.93) BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 21 doi.org/10.35321/bkalba.2021.94.06 As can be seen from Table 3, most ACs were identified using the non-automated method—more than half of all ACs (e.g., a ghastly sight, slave labor, a blunt question, a painful thing); about a third of ACs were identified automatically (e.g., valuable property, a unique project, a difficult day, an optimal method), while the smallest share (15.74 percent) consists of ACs identified by both methods (e.g., fratricidal war, intensive trade, a huge project, a low price). Similar proportions can also be seen among ACs with verbs (see Section 2.2). An analysis of the adjectives in ACs shows that some occur only in ACs identified by one of the methods (see the examples given above), whereas others occur in ACs identified by both methods (e.g., big, small, strong). At the beginning of the study, the hypothesis was proposed that the diversity of adjectives would be greater in ACs identified using the non-automated method and smaller in ACs identified using the automated vector method. This hypothesis was confirmed: 235 different adjectives occur in ACs identified by the first method, meaning that the same adjective forms, on average, 2.5 ACs. There are 86 different adjectives in ACs identified by the vector method, so an average of 3.56 ACs are formed with the same adjective. The smallest diversity of adjectives (42) is found among ACs identified by both methods, although this group also contains the fewest such collocations. In this group of ACs, an average of 3.93 collocations are formed with the same adjective. It is likely that this limited diversity of adjectives is due to the relatively small amount of data. There is a high probability that a collocation will be arbitrary if it consists of adjectives such as great, big, huge, enormous, small, tense, dangerous, serious, strong, difficult combined with abstract nouns, since these adjectives will be used figuratively. To discuss how the methods used led to these particular results, we would like to consider as an example the ACs with the adjective intelligent, since they illustrate the full range: some collocations were identified by one method, some by the other, and some by both. Of the eight collocations with the adjective intelligent (an intelligent decision, an intelligent city, an intelligent child, an intelligent man, an intelligent time, an intelligent way), six were classified as ACs: the adjective’s collocations with the nouns decision and city were selected non-automatically; its collocations with child and man were selected automatically; and its collocations with time and way were selected by both methods. Linguists found that the semantic group “persons” is typical for this adjective. Words in this group can optionally combine with adjectives of similar meaning (it can be BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 22 doi.org/10.35321/bkalba.2021.94.06 not only intelligent, but also a wise person), so intelligent child and intelligent man entered the AC set only as a result of automatic selection (the collocation a wise person was not found in the Delfi.lt corpus); they were later rejected during data review. The linguists observed that the other nouns (decision, time, way) mostly belong to the lexical-semantic group “abstracts”; nevertheless, collocations are formed only with certain abstract nouns, meaning that the options within this group are limited, which indicates arbitrariness. Therefore, all these collocations were classified as arbitrary. Only two of them (an intelligent time and an intelligent way) were identified as ACs automatically, because the vector method does not apply the lexical-semantic group test. ACs identified using the non-automated method. For example, nine collocations with the adjective high were classified as arbitrary: a high result, a senior official, a high price, etc. Here the adjective is used figuratively (and almost exclusively with abstract nouns), so the figurative-meaning criterion led the linguists to classify these collocations as ACs. These collocations were very often not detected by the automated method, because no VEs containing the corresponding adjective (high, closed, painful, weak, broad, comfortable, easy, free, long or fast) were available for the synonym-substitution test, making it impossible to assess substitutability. As mentioned in the previous section, the results of the automated method depend greatly on whether synonyms are available for testing substitutability and on the quality of the synonym list. This case illustrates that, when only the automated method is used, some ACs may remain unidentified, while linguistic competence can compensate for a lack of corpus data. Another reason why some ACs may not be identified automatically is that not all nouns may be associated with a synonym set of a particular meaning, since the contexts of some nouns may differ. For example, among the collocations with the adjective prominent, many were identified non-automatically (e.g., a prominent woman, a prominent person), and only one AC was identified by both methods—a prominent event. This means that a VE containing this adjective was available, but not all three nouns were automatically associated with it; only event was. ACs identified using the automated method. ACs with the adjective effective were identified only using the vector method, e.g., an effective measure, an effective result, an effective system, effective work. The VEs contained more synonyms (e.g., efficacious), but the collocations an efficacious measure, an efficacious result, etc., were not found in the corpus; therefore, after applying the impossibility BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 23 doi.org/10.35321/bkalba.2021.94.06 the substitutability test, the collocations with effective were classified as ACs. It is important to add here that, drawing on native speakers’ competence, linguists assessed this differently in this case: substitution is possible (it is possible to use both an effective means and an efficient means), and so it was decided not to consider these collocations ACs. The automated method classified many collocations as ACs whose adjectives are evaluative, cf.: exceptional (because it is synonymously associated with special), strange (associated with unusual), modern (associated with contemporary), main (associated with essential), complex (associated with difficult). Linguists were more inclined not to classify such cases as ACs. Collocations with stable, for example, stable life, stable price, stable week, stable state, were recognized as arbitrary only by the vector method, because substitutability was tested using the synonym constant, which was not used as a collocate with the corresponding nouns. Nevertheless, linguists did not recognize these collocations as arbitrary, because examining the collocational range of stable shows that it is indeed broad and no restrictions have been established, whereas constant was not considered a synonym of stable, cf.: stable work, stable state, stable price, stable result, stable life, stable week. ACs identified by both methods. Among the available data, there is not a single case in which all collocations identified as ACs with the same adjective were identified by both methods. In every case, only some of them were identified by both methods, and the adjectives in these cases most often include those that can be used figuratively, e.g.: famous, fast, sad, small, not easy, attractive, dangerous, serious, strong, difficult, firm. As already mentioned, the ACs involving most such adjectives were identified only by the non-automated method, for example, collocations with the adjective weak (see also the examples mentioned earlier). This is because, when ACs were identified manually, such cases were selected according to the figurative-meaning criterion, whereas under Method 2 they qualified only if they were members of the synonym sets used for the substitutability test. Examining how many collocations each method identifies for one of the most productive nouns, decision, reveals trends similar to those shown in Table 3: the non-automated method yielded the largest number of ACs, but a considerable proportion of the ACs were identified automatically. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 24 doi.org/10.35321/bkalba.2021.94.06 TABLE 4. ACs identified with the noun decision and adjectives ACs identified by non-automated methods ACs identified by automated methods ACs identified by both methods 1) responsible 2) unprecedented 3) drastic 4) bold 5) formal 6) strict 7) smart 8) wise 9) constructive 10) creative 11) stupid 12) temporary 13) decisive 14) fateful 15) easy 16) illogical 17) irrational 18) justified 19) convenient 20) advanced 21) practical 22) opposite 23) intelligent 24) rational 25) conscious 26) safe 27) independent 28) scandalous 29) painful 30) hasty 31) urgent 32) clever 33) peaceful 34) fair 35) sustainable 36) united 1) clear 2) categorical 3) strange 4) concrete 5) creative 6) useful 7) unexpected 8) optimal 9) simple 10) poor 11) real 12) complex 13) unique 1) fast 2) logical 3) difficult 4) serious 5) risky 6) difficult 7) harmful 8) solid BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 25 doi.org/10.35321/bkalba.2021.94.06 Using the example of the noun sprendimas ‘decision’, it is important to mention one more aspect related to the automated method: some collocations classified as arbitrary may be debatable, that is, they may seem to lie on the borderline, closer to the non-arbitrary ones. For example, one might wonder whether collocations such as aiškus sprendimas ‘clear decision’ or keistas sprendimas ‘strange decision’ should really be considered arbitrary. It might seem that the adjectives aiškus, keistas, kūrybiškas, naudingas, netikėtas (‘clear, strange, creative, useful, unexpected’) and others can combine not only with sprendimas ‘decision’, but also with many other nouns with which they can combine on semantic grounds; cf. all the collocations of these adjectives identified in the Delfi.lt corpus: aiškus ženklas, aiškus vaizdas, aiški informacija, aiškus rezultatas, aiški sistema, aiškus dalykas; keistas dalykas, keistas būdas, keistas klausimas, keista vieta, keistas sprendimas, keistas įvykis, keistas vaizdas. Linguists tested restricted lexical combinability not only using the synonymy test, but also by means of lexical-semantic groups; therefore, they did not classify the collocations aiškus sprendimas and keistas sprendimas as arbitrary. When assessing lexical restriction automatically, these collocations were classified as AK because testing synonym substitution showed that the noun sprendimas formed collocations with only one member of the synonym pair aiškus, suprantamas (‘clear, understandable’) and one member of the synonym pair keistas, neįprastas (‘strange, unusual’). These examples show that more accurate AK results can be expected when testing restricted lexical combinability by means of not only the synonymy test but also the lexical-semantic group test. 2.2. Arbitrary collocations consisting of verbs and nouns A total of 1171 AK consisting of a verb and a noun were identified. This is quite similar to the number of AK formed with adjectives, which is 1048. The AK identified using different methods are also distributed quite similarly to the collocations formed with adjectives (see Table 2) (see Table 3): approximately half of the AK were identified manually (e.g., apniukusi diena, aptarti galimybes, iškvies­ti apsaugą), about a third automatically (e.g., suskubti į pagalbą, žvelgti į gyvenimą, sugadintas turtas, kaupti turtą), and nearly a fifth were identified by both methods (e.g., atlikti tarnybą, duoti komandą, matyti reikalą). The lexical diversity of AK of this structure was also analyzed, that is, how many different verbs form AK identified using a given method. On average, 2.05 of the AK discussed in this subsection were formed with the same verb (see Table 5). BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 26 doi.org/10.35321/bkalba.2021.94.06 TABLE 5. AK identified using different methods (verb + noun) AK identified using manual methods AK identified using automated methods AK identified using both methods No. (percent) Number of different verbs (average number of AK formed with the same verb) No. (percent) Number of different verbs (average number of AK formed with the same verb) No. (percent) Number of different verbs (average number of AK formed with the same verb) 561 (47,91) 376 (1,5) 382 (32,62) 189 (2,02) 228 (19,47) 102 (2,24) The syntactic functions of verbs in AK are more varied than those of adjectives: finite forms can be used, taking nouns that function as objects, subjects, or adverbials (e.g., apipilti klausimais, kelti kainą, parodyti laikas13, pastatyti į vietą), or non-finite forms, most often participles, functioning as attributes (e.g., išpūsta kaina, einamieji metai, neapgalvotas sprendimas, klestinti šalis, negirdėtas dalykas, vykdomas darbas). In three-word AK, word order may vary, e.g., atskubėti į pagalbą / į pagalbą atskubėti; eiti į rinkimus / į rinkimus eiti. Compared with collocations formed by adjectives, there is greater diversity of verbs: in the AK identified using the non-automatic method, the same verb forms an average of 1.5 collocations; in the AK identified using the automated vector method, an average of 2.02 collocations; while in the AK identified by both methods, verb diversity is lowest—the same verb forms an average of 2.24 AK. This result was likely due to the fact that the fewest collocations were identified by both methods. On the other hand, the trends are similar: the greatest lexical diversity is found in the AK identified using the non-automatic method. 13 Sometimes the lemmas of subject collocations seem unusual, but the infinitive is given because a wide variety of inflected forms may be used, e.g.: time will tell, time told, time would tell. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 27doi.org/10.35321/bkalba.2021.94.06 The data show a great diversity of verbs (words with and without a prefix, and with different prefixes (e.g., saugoti, apsaugoti, išsaugoti) were treated as different verbs; reflexive and non-reflexive forms were treated as the same verb if they did not change the meaning), but some verbs are highly productive, forming several or even a dozen AK, e.g.: suteikti: provide protection, provide information, provide an opportunity, provide an opportunity to choose, provide information about; skirti: devote little attention, devote time, pay attention, pay a lot of attention, house arrest imposed, pay great attention, devote a lot of time, pay a lot of attention, attention needs to be paid; ateiti / ateinantis: come to the election, the euro is coming, the time is coming, the moment is coming, coming years, upcoming elections, the budget for the coming year, the beginning of the coming year, the coming month, the beginning of the coming week, the upcoming season, coming years, the coming day, the coming week. It can be seen that a considerable proportion of the AK formed with a single verb are variants of one collocation; compare the previously presented variants of AK skirti dėmesį. Nevertheless, any lexical or grammatical variations of collocations (word order changes, additional words are added) are counted as separate collocations. They are treated as separate collocations when the same verb is used in a finite form (in which case the lemma is usually given as the infinitive) or a non-finite form (e.g., fight for a place, fight to the end, a fighting group). Thus, the greater lexical diversity of verb collocations (compared with adjective collocations) is due to a characteristic of verbs as a part of speech: frequent use of verbs with the same root and different prefixes. On the one hand, such verbs can be considered different words; on the other hand, these data cannot be fully compared with data for adjectives, which in every case are different words rather than derivational variants of the same word. AK identified using the non-automatic method. Unlike adjective + noun collocations, collocations with verbs were often classified as arbitrary not only by applying the criterion of figurative meaning (e.g., ruin a life, open up an opportunity), but also by applying the criterion of restricted lexical combinability and the lexical-semantic group test. For example, collocations with vaizdas, būdas, informacija were evaluated with the following paradigms: BENDRINĖ KALBA 94 (2021)http://journals.lki.lt/bendrinekalbaISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 28 doi.org/10.35321/bkalba.2021.94.06 distort information distort the market distorted information distorted manner distorted image The collocability of the verbs iškraipyti, iškreipti (also used attributively) is not extensive: to the paradigm of collocations with iškraipyti one could also add iškraipyti žodžius, but the noun žodžiai would fall into the same group as informacija, so the decision was made to classify both collocations as AK. Three collocations are formed with the participle iškreiptas used attributively, and all were considered AK because, although the paradigm could be expanded (e.g., by adding iškreiptas mąstymas), it would still be limited. The nouns informacija, vaizdas, būdas, even though assigned to one abstract lexical-semantic group, number only three, and a small number of words is one of the conditions for applying the lexical-semantic group test. Applying the synonym substitutability test reveals that iškraipyti, iškreiptas cannot be replaced, for example, with išdarkyti / išdarkytas. AK identified using the automated method. It is important to emphasize one limitation of the automated method: some AK may go unrecognized because such collocations were not identified in the Delfi.lt corpus (although the word combination itself may have been used). For example, the three-word collocation aiškintis įvykio aplinkybes was recognized as AK on the basis of synonyms included in VE: analizuoti, aiškintis, tirti. Only a collocation with the verb aiškintis was found in the corpus, and it was therefore recognized as AK. On the other hand, another corpus might also contain the collocation tirti įvykio aplinkybes; thus, testing the same synonym group would lead to two collocations being recognized as arbitrary rather than one. Among the collocations identified using the automated method are cases that linguists did not recognize as AK precisely because, based on their intuition about the language, they did not think there was a tendency to select one synonym from among several possible ones. Examples include the collocations suprasti pasaulį, pažinti pasaulį, žiūrėti į pasaulį. Linguists did not recognize any of these collocations as arbitrary. The automated method identified all three as AK, because among the collocations in the Delfi.lt corpus there were no collocations with other members of the respective synonym groups: there was suprasti pasaulį, but not suvokti pasaulį; there was pažinti pasaulį, but not išmokti pasaulį; there was žiūrėti į pasaulį, but not žvelgti į pasaulį. Linguistic assessment established that the collocability of these three verbs does not exhibit the characteristics of limited collocability. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 29doi.org/10.35321/bkalba.2021.94.06 Arbitrary collocations were identified using both methods. It should be noted that the arbitrary collocations in this group were not necessarily identified according to the same criterion. When assessing arbitrariness using the non-automated method, an arbitrary collocation could be identified according to any one test, not necessarily the synonym-substitutability test. As noted concerning adjective–noun collocations, the same arbitrary collocations were selected manually according to the figurative-meaning criterion, and by the automated method according to the restricted lexical combinability criterion (the synonym-substitutability test); cf. traffic was disrupted and health deteriorated were selected by linguists because sutr būti combines with a restricted group of nouns, whereas the two methods identified these collocations as arbitrary because, among the synonyms sugedo and sutriko in the evaluative collocations, only one synonym was attested. Apga l v otas sprendimas was identified as an arbitrary collocation because no collocation with another synonym—apmąstytas sprendimas—was found. Thus, it is important to bear in mind that the collocations in this group were not identified according to the same criterion. TABLE 6. Arbitrary collocations identified with the noun sprendimas and verbs Arbitrary collocations identified using the non-automated method Arbitrary collocations identified using the automated method Arbitrary collocations identified using both methods 1) appeal a (court) decision 2) postpone a decision 3) make a decision 4) base a decision 5) seek a solution 6) final decision 7) ill-considered decision 8) not make a decision 9) fail to enforce a court decision 10) the decision proved effective 11) make an appropriate decision make the necessary decision 12) keep making a decision 13) agree with a decision 14) announce a decision 15) appeal a decision (to a court) 16) enforce a court decision 1) make a decision 2) demand a decision 3) proposed solution 4) await a decision 5) sign a decision 1) enforce a court decision 2) a decision made 3) a proposed decision 4) announce a decision 5) a considered decision BENDRINĖ KALBA 94 (2021)http://journals.lki.lt/bendrinekalbaISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 30 doi.org/10.35321/bkalba.2021.94.06 Having assessed how many collocations of the verb + noun type are identified by different methods for one of the most productive nouns, sprendimas ‘decision’ (see Table 6), it can be seen that the largest number were identified by the manual method, while only five ACs were identified by both methods and by the automated method. CONCLUDING REMARKS Having analysed two-word and three-word ACs comprising adjectives and nouns, as well as verbs and nouns, it can be stated that about half were identified by the manual method, about a third by the automated method, and approximately a fifth by both methods. Only a small proportion of the ACs identified by these methods overlap, which suggests that more ACs can be identified by using both methods. Thus, to identify ACs, it is worth making use of the capabilities offered by both manual and automated methods. It should be emphasised that the manual analysis applied not one but two selection criteria, which is why it yielded far more data than the automated method. The manual analysis included not only the figurative-meaning criterion but also the restricted lexical combinability criterion, tested using two tests (lexical-semantic groups and synonym substitutability), whereas the automated method applied only the restricted lexical combinability criterion, according to the synonym substitutability test. Because the number of criteria differs, the results obtained by the different methods cannot be compared: the first method identified more ACs in any case, while the number of overlapping collocations is small. It would be interesting to compare ACs identified automatically and manually using only the shared restricted combinability criterion (the synonym substitutability test). Such a comparison could show that some of the identified ACs are the same, while also making it possible to assess which method selected more ACs and how the non-overlapping ACs differ. The manual method yielded not only more data but also greater lexical variety. The ACs identified using this method comprise a wider variety of adjectives and verbs, whereas the ACs produced by the automated method have less lexical variety. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 31 doi.org/10.35321/bkalba.2021.94.06 When testing the automated method, the main aim was to reduce the subjectivity involved in assessing arbitrariness and to rely exclusively on data attested in the corpus. Given synonym pairs or groups identified using the vector method from a particular corpus, the evaluation procedure is simplified, because it is easier to decide whether a collocation component can be substituted: if the material under study contains collocations with all the words in a synonym pair or group, substitution is possible. Of course, the automated method also revealed some shortcomings. Since the arbitrariness of the collocations being evaluated is determined on the basis of synonyms in the lexical database, some ACs may go unrecognised, because the determination of arbitrariness depends greatly on the lexical database available. Testing uses only as many synonyms, and only those synonyms, as have been collected in the lexical database; consequently, not all collocations are recognised as ACs. Although the AC analysis process can be automated, linguists’ input is needed at its initial stage: in the lexical databases compiled using the vector method, they must distinguish synonyms from words related by other semantic relations, such as antonyms. An examination of how much of the automatically obtained data remains after linguistic review shows that many adjectives and verbs have been removed from the lexical databases, and that the remaining 41 per cent and 30 per cent of the data, respectively, also require editing and cleaning. Since the vector method identifies not only words related by synonymy, preparing synonym data for the substitutability test takes longer. Other methods for identifying synonym relations should therefore be tested. In assessing the suitability of the vector method for selecting synonyms, the following arguments demonstrate the value of the automatically collected data. First, the vector method can identify more synonyms and different synonyms than introspection alone. Second, the synonyms identified using this method (after non-synonyms have, of course, been filtered out) reveal a phenomenon important from the perspective of synonymy: not only individual words but also multiword expressions are synonymous. Third, the vector method is suitable for processing large volumes of data, that is, for identifying synonyms in a large corpus; by hand, it would be difficult to assess the contextual similarity of such a large number of words. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 32 doi.org/10.35321/bkalba.2021.94.06 LITERATŪRA Bielinskienė Agnė, Boizou Loic, Bumbulienė Ieva, Kovalevskaitė Jolanta, Krilavičius Tomas, Mandravickaitė Justina, Rimkutė Erika, Vilkaitė-Lozdienė Laura (sud.) 2019: Lietuvių kalbos kolokacijų žodynas, Kaunas: Vytauto Didžiojo universitetas. Prieiga internete: https://doi.org/10.7220/9786094673733. Divjak Dagmar, Gries Stefan T. 2006: Ways of Trying in Russian. Clustering and Comparing Behavioral Profiles. – Corpus Linguistics and Linguistic Theory 2(1), 23–60. Drūlienė Viltė 2008a: Erduvinės metaforos Daiktavardinių frazių žodyne: magistro darbas, Kaunas: Vytauto Didžiojo universitetas. Drūlienė Viltė 2008b: Lietuviškosios metaforikos vertikalumo matas. – Darbai ir dienos 50, 73–90. Grigonytė Gintarė 2010: Building and Evaluating Domain Ontologies: NLP Contributions: daktaro disertacija, Berlin: Logos Verlag Berlin. Handl Susanne 2008: Essential Collocations for Learners of English: The Role of Collocational Direction and Weight. – Phraseology in Foreign Language Learning and Teaching, eds. F. Meunier, S. Granger, Amsterdam, Philadelphia: John Benjamins Publishing Company, 43–66. Howarth Peter 1998: Phraseology and Second Language Proficiency. – Applied Linguistics 19 (1), 24–44. Jakaitienė Evalda 2010: Lietuvių kalbos leksikologija, Vilnius: Vilniaus universiteto leidykla. Kovalevskaitė Jolanta, Rimkutė Erika, Vilkaitė Laura 2020: Light Verb Constructions in Lithuanian: Identification and Classification. – Kalbų studijos 36, 5–16. Kovalevskaitė Jolanta, Rimkutė Erika, Vaičenonienė Jurgita 2021: Automatizuotas arbitraliųjų kolokacijų atpažinimas: būdvardžių ir daiktavardžių kolokacijos. – Kalbų studijos 39, 71–84. Lewis Michael 2000: Teaching Collocation: Further Developments in the Lexical Approach, London: Language Teaching Publications. Marcinkevičienė Rūta 2010: Lietuvių kalbos kolokacijos, Kaunas: Vytauto Didžiojo universitetas. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 33 doi.org/10.35321/bkalba.2021.94.06 Nesselhauf Nadja 2003: The Use of Collocations by Advanced Learners of English and Some Implications for Teaching. – Applied Linguistics 24 (2), 223–242. Nesselhauf Nadja 2005: Collocations in a Learner Corpus, Amsterdam: John Benjamins. Pennington Jeffry, Socher Richard, Manning Christopher D. 2014: GloVe: Global Vectors for Word Representation. – Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNL). Association for Computational Linguistic. Doha, Qatar, October 25–29, 2014, 1532–1543. Reder Anna 2006: Kollokationen in der Wortschatzarbeit. Praesens Verlag. Strakatova Yana, Hinrichs Erhard 2019: Semantic Modelling of Adjective-Noun Collocations Using FrameNet. – Proceedings of the Joint Workshop on Multiword Expressions and WordNet, 104–113. Stulpinaitė Miglė, Horbačauskienė Jolita, Kasperavičienė Ramunė 2016: Issues in Translation of Linguistic Collocations. – Kalbų studijos 29, 31–41. Šilinytė Justina 2008: Ontologinės ir struktūrinės metaforos Daiktavardinių frazių žodyne: magistro darbas, Kaunas: Vytauto Didžiojo universitetas. Volungevičienė Skaistė 2010: Metaforinių kolokacijų vertimo problemos. – Kalbų studijos 16, 23–27. Gauta 2021 11 14 Priimta 2021 12 15 BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 34 doi.org/10.35321/bkalba.2021.94.06 ARBITRALIŲJŲ LIETUVIŲ KALBOS KOLOKACIJŲ NUSTATYMAS Santrauka Šiame straipsnyje sprendžiama arbitraliųjų lietuvių kalbos kolokacijų atpažinimo problema. Tam skirtas projektas „Arbitrariosios lietuvių kalbos kolokacijos: atpažinimas, aprašymas, vartojimas (ARKA)“. Arbitraliosiomis kolokacijomis laikomos kolokacijos, kurių žodžių junglumo pagrindas yra ne semantiniai, o leksiniai ribojimai. Tokios kolokacijos tiesiogiai neatspindi nekalbinės tikrovės, o ryšys tarp jų žodžių – arbitralius: nors pagal panašią reikšmę teoriškai galima rinktis daugiau negu vieną junglumo partnerį, bet linkstama vartoti kurį nors vieną, pvz., suprasti užuominą, bet ne *subokti užuominą. Tyrimo šaltinis yra Lietuvių kalbos pastoviųjų žodžių junginių duomenų bazė, kurioje pateikiama apie 12 tūkst. kolokacijų; ištyrus kolokacijas su 97 daiktavardžiais, arbitraliosiomis priskirta apie 2400 kolokacijų, kurių struktūra yra būdvardis / būdvardiškai vartojamas dalyvis + daiktavardis, veiksmažodis + daiktavardis, daiktavardis + daiktavardis. Straipsnyje aprašyti du pagal minėtą projektą išbandyti arbitraliųjų kolokacijų nustatymo metodai: neautomatinis lingvistinės analizės ir automatizuotas kompiuterinės lingvistikos metodas. Nustatant AK neautomatiškai, taikyti du pagrindiniai kriterijai: perkeltinės reikšmės kriterijus ir riboto leksinio junglumo kriterijus, kuriam įvertinti taikyti veiksmažodžio ar būdvardžio sinonimų pakeičiamumo ir (arba) daiktavardžių leksinių semantinių grupių testai. Vienas iš AK nustatymo kriterijų – riboto leksinio junglumo – buvo tikrinamas ir automatizuotai, vektorių metodų nustatčius pakeičiamumo testui reikalingus sinonimus. Automatizuoto metodo taikymą sudarė šie etapai: 1) automatinis, kai vektorių metodu buvo surenkamos vektorių eilės su panašų kontekstą turinčiais žodžiais; 2) neautomatinis – vektorių eilių tvarkymas paliekant tik sinoniminiais ryšiais susijusius žodžius; 3) iš dalies automatinis etapas, kai kolokacijos buvo automatiškai susiejamos su konkrečiomis sinonimų eilėmis ir po peržiūros lingvistų patvirtinamos arba nepatvirtinamos kaip AK. Neautomatiniu metodu nustatyta apie pusė AK, automatizuotu metodu – apie trečdalis. Neautomatinei analizei taikytas ne vienas, o du atrankos kriterijai, todėl neautomatiškai nustatytų duomenų daug daugiau negu nustatytų automatizuotai. Nors BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitralijų lietuvių kalbos kolokacijų nustatymas | 35 doi.org/10.35321/bkalba.2021.94.06 automatizuotu metodu galima sumažinti subjektyvumą vertinant duomenis ir supaprastinti arbitralumo vertinimo procedūrą, tačiau pastebėta ir tai, kad arbitralumo pripažinimą lemia vektorių eilėje esantys sinonimai, taigi dalis AK gali likti neatpažintos. Vektorių metodu surinktus sinonimus reikia atskirti nuo kitais semantiniais ryšiais susijusių žodžių, tam reikalinga lingvisto patikra. Tačiau, nepaisant šio ribotumo, vektorių metodas yra tinkamas dideliam duomenų kiekiui apdoroti, o rankiniu įvertinti gausybės žodžių konteksto panašumą būtų sunku. Abiem metodais atpažintos kolokacijos sudaro tik penktadalį visų nustatytų AK, todėl yra pagrindo manyti, kad siekiant atpažinti AK svarbu taikyti tiek automatizuoto, tiek ir neautomatinio metodo teikiamas galimybes. BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 36 doi.org/10.35321/bkalba.2021.94.06 IDENTIFICATION OF LITHUANIAN ARBITRARY COLLOCATIONS Summary In this article, we describe the methodological approaches of arbitrary collocation (AC) identification developed within the framework of an ongoing project “Arbitrary Collocations of Lithuanian: Identification, Description and Usage (ARKA)”. The object of this research is arbitrary collocations of the Lithuanian language the collocability of which is determined by lexical rather than semantic constraints. Such collocations do not directly reflect the non-linguistic reality and the relation between the collocates is arbitrary: although there may be several close synonyms, a particular one is preferred in a certain word combination (e.g., suprasti užuominą vs. *suokti užuominą). The source of the research was the Database of Lithuanian Multiword Expressions encompassing over 12.000 collocations. The structural composition of the identified 2400 arbitrary collocations was as follows: adjective (participle)+noun, verb+noun, and noun+noun. Arbitrary collocations were determined by combining the manual linguistic analysis and semi-automatic methods of computational linguistics. The manual method included two major AC identification criteria: lexical restrictedness and (or) meaning transfer. Lexical restrictedness was measured using two tests: the synonym substitution of pre-modifier and (or) semantic field comparison of the head noun. Moreover, lexical restrictedness was also assessed using a semi-automatic approach by analyzing synonym pairs in the automatically generated vector strings. The semi-automatic approach consisted of three stages: (1) the automatic generation of vector strings with potential synonyms; (2) the manual vector string editing to reduce the noise, and (3) the semi-automatic process during which the collocations were compared to particular synonym pairs in vector strings and approved or not approved by linguists as arbitrary. Approximately half of ACs were detected by using the manual method based on two AC identification criteria, whereas about one-third of ACs were identified using the semi-automated method based on one criterion. On the one hand, the semi-automatic method can reduce the subjectivity in data evaluation and simplify the AC identification procedure. On BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351-7204 JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ. Arbitraliųjų lietuvių kalbos kolokacijų nustatymas | 37 doi.org/10.35321/bkalba.2021.94.06 the other hand, as arbitrariness is determined on the basis of the available synonyms in the vector string, some ACs may not be identified because of the insufficient scope of the generated synonyms. Also, the automatically generated synonyms need to be manually analyzed to eliminate semantically unrelated words. Despite these limitations, the method is suitable for processing large amounts of the data. It would be difficult to manually assess the contextual similarity of such a large number of words. About one fifth of ACs were identified using a combination of both methods. These results suggest that different methodological approaches allowed the researchers to detect more arbitrary collocations. It is maintained that taking into consideration the mentioned advantages and disadvantages, the best results in the Lithuanian arbitrary collocation identification are acquired when combining both manual and semi-automatic methods. KEYWORDS: arbitrary collocations, Lithuanian, lexical restrictedness, vector space model, semi-automatic approach, linguistic manual approach. JOLANTA KOVALEVSKAITĖ, ERIKA RIMKUTĖ, JURGITA VAIČENONIENĖ Vytauto Didžiojo universitetas Kristijono Donelaičio g. 58, 44248 Kaunas [email protected] [email protected] [email protected] BENDRINĖ KALBA 94 (2021) http://journals.lki.lt/bendrinekalba ISSN 2351–7204