Syn2Vec: Synset Colexification Graphs for Lexical Semantic Similarity

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5259 - 5270 July 10-15, 2022 ©2022 Association for Computational Linguistics Syn2Vec: Synset Colexification Graphs for Lexical Semantic Similarity John Harvill1, Roxana Girju2, Mark Hasegawa-Johnson1 1Department of Electrical and Computer Engineering, 2Departments of Linguistics and Computer Science, University of Illinois at Urbana-Champaign {harvill2, girju, jhasegaw}@illinois.edu Abstract In this paper we focus on patterns of colexifi- cation (co-expressions of form–meaning map- ping in the lexicon) as an aspect of lexical- semantic organization, and use them to build large scale synset graphs across BabelNet’s typologically diverse set of 499 world lan- guages. We introduce and compare several approaches: monolingual and cross-lingual colexification graphs, popular distributional models, and fusion approaches. The models are evaluated against human judgments on a se- mantic similarity task for nine languages. Our strong empirical findings also point to the im- portance of universality of our graph synset embedding representations with no need for any language-specific adaptation when evalu- ated on the lexical similarity task. The in- sights of our exploratory investigation of large- scale colexification graphs could inspire signif- icant advances in NLP across languages, es- pecially for tasks involving languages which lack dedicated lexical resources, and can ben- efit from language transfer from large shared cross-lingual semantic spaces. 1 Introduction Distributional models like word embeddings have been widely used in Natural Language Processing (NLP) (Iacobacci et al., 2016; Devlin et al., 2014; Hewlett et al., 2016). They operate under the as- sumption that words appearing in similar contexts have similar meanings, and thus close representa- tions. However, as do other unsupervised learning models, they suffer from classic limitations – i.e., there is no guarantee that all context words con- tribute to the meaning of the target word, while, in fact, it is possible that some low frequency words, with poorly-trained embeddings, are highly seman- tically connected. Also, they don’t distinguish be- tween topically related words and near synonyms. Dictionaries and thesauruses, on the other hand, have traditionally offered an alternative approach, through their discrete lists of fine-grained senses, textual definitions, and relationships with other senses. Given a sufficiently large dictionary of many fine-grained sense representations in many of the world’s languages, one could perform so- phisticated semantic tasks on word senses (Conia and Navigli, 2020). In fact, investigating univer- sal and areal cross-linguistic variations in the lex- icon has been the focus of lexical typology. One increasingly popular empirical method of investi- gating senses based on cross-linguistic compari- son in typological studies has been that of colex- ification patterns. “A given language is said to colexify two functionally distinct senses if, and only if, it can associate them with the same lexical form” (François, 2008), reflecting natural seman- tic connections (Haspelmath, 2003). For example, ’fire’ and ’firewood’ are colexified in Kamoro (New Guinea language) as ’uta’ and in Wayuu (Arawakan language) as ’siki’, but receive distinct lexemes in English and Romanian. Each polysemous lexeme as a whole …

Content truncated to 500 words