Assessing Polyseme Sense Similarity through Co-predication Acceptability and Contextualised Embedding Distance

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 114–124 Barcelona, Spain (Online), December 12–13, 2020 114 Assessing Polyseme Sense Similarity through Co-predication Acceptability and Contextualised Embedding Distance Janosch Haber and Massimo Poesio Queen Mary University of London {j.haber|m.poesio}@qmul.ac.uk Abstract Co-predication is one of the most frequently used linguistic tests to tell apart shifts in po- lysemic sense from changes in homonymic meaning. It is increasingly coming under crit- icism as evidence is accumulating that it tends to mis-classify specific cases of polysemic sense alteration as homonymy. In this paper, we collect empirical data to investigate these accusations. We asses how co-predication ac- ceptability relates to explicit ratings of po- lyseme word sense similarity, and how well either measure can be predicted through the distance between target words’ contextualised word embeddings. We find that sense similar- ity appears to be a major contributor in deter- mining co-predication acceptability, but that co-predication judgements tend to rate less similar sense interpretations as being as un- acceptable as homonym pairs, effectively mis- classifying these instances. The tested contex- tualised word embeddings fail to predict word sense similarity consistently, but the similari- ties between BERT embeddings show a sig- nificant correlation with co-predication ratings. We take this finding as evidence that BERT embeddings might be better representations of context than encodings of word meaning. 1 Introduction Polysemy is a form of lexical ambiguity which occupies a unique middle ground between mono- semy –word forms with exactly one interpretation– and homonymy –word forms associated with two or more completely unrelated interpretations. Un- like monosemes, polysemes can evoke different interpretations, but unlike homonyms, polysemic sense interpretations are thought to be closely re- lated to each other (Lyons, 1977). It is commonly This work is licensed under a Creative Commons Attri- bution 4.0 International Licence. Licence details: http: //creativecommons.org/licenses/by/4.0/. assumed that most words in natural language are in fact polysemous to some degree (Falkum and Vicente, 2015), and the question whether there in fact are any proper monosemes has been the source of ongoing debate (see for example Jackendoff, 1989; Fodor, 1998). Homonyms have been a driv- ing factor in developing contextualised language models (e.g. Peters et al., 2018; Devlin et al., 2018; Radford et al., 2019) in order to account for the dif- ferent unrelated meanings some words can evoke in different contexts: a. The match burned my fingers. b. The match ended without a winner. Comparing these uses of the word match to the various closely related interpretations of canonical polyseme school illuminates the conceptual differ- ence between the two phenomena of lexical ambi- guity:1 a. The school [building] is on fire. b. The school [rules] has prohibited wearing hats in the classroom. c. I have talked to the school [director, staff] about it already. d. The school [participants] went for a visit to the cathedral. Although the distinction is clear in theory, distin- guishing monosemy, polysemy and homonymy in practice proves exceedingly difficult: At what point are interpretation nuances pronounced …

Content truncated to 500 words