Team dina at SemEval-2022 Task 8: Pre-trained Language Models as Baselines for Semantic Similarity

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 145 - 150 July 14-15, 2022 ©2022 Association for Computational Linguistics UAlberta at SemEval 2022 Task 2: Leveraging Glosses and Translations for Multilingual Idiomaticity Detection Bradley Hauer, Seeratpal Jaura, Talgat Omarov, Grzegorz Kondrak Alberta Machine Intelligence Institute, Department of Computing Science University of Alberta, Edmonton, Canada {bmhauer,seeratpa,omarov,gkondrak}@ualberta.ca Abstract We describe the University of Alberta systems for the SemEval-2022 Task 2 on multilingual idiomaticity detection. Working under the as- sumption that idiomatic expressions are non- compositional, our first method integrates in- formation on the meanings of the individual words of an expression into a binary classi- fier. Further hypothesizing that literal and id- iomatic expressions translate differently, our second method translates an expression in con- text, and uses a lexical knowledge base to de- termine if the translation is literal. Our ap- proaches are grounded in linguistic phenom- ena, and leverage existing sources of lexical knowledge. Our results offer support for both approaches, particularly the former. 1 Introduction In this paper, we describe the University of Alberta systems for the task of classifying multi-word ex- pressions (MWEs) in context as either idiomatic or literal (Tayyar Madabushi et al., 2022). Each instance in the data includes a MWE (e.g., closed book), its language, and its context, composed of the three surrounding sentences. We participate in both the zero-shot and one-shot settings. While the exact definitions of the two key terms are not stated explicitly in the task description1, it is suggested that idiomatic is synonymous with non-compositional. The Pocket Oxford Dictionary defines idiomatic as “not immediately comprehen- sible from the words used,” and literal as “taking words in their basic sense.” Therefore, we adopt the following MWE compositionality criterion literal ≡ compositional ≡ ¬ idiomatic where the three terms are considered to be Boolean variables. In addition, the shared task considers all proper noun MWEs (e.g., Eager Beaver) as literal. 1https://sites.google.com/view/ semeval2022task2-idiomaticity Figure 1: An example of defBERT input. Our goal is to explore the idea that glosses and translations of word senses can help decide whether the meaning of a given MWE occurrence is compo- sitional. Based on the above-stated compositional- ity criterion, this in turn could facilitate idiomatic- ity detection. In particular, we hypothesize that at least one of the words in any idiomatic expression is used in a non-standard sense. Following the intu- ition that a traditional word sense disambiguation (WSD) system can only identify senses that are included in a given sense inventory, we propose two methods that indirectly detect non-standard senses by leveraging either glosses or translations of senses from such an inventory. Our gloss-based method follows from the intu- ition that the meaning of a given MWE occurrence is related to any of the existing sense glosses of its component words only if the expression is composi- tional. Therefore, the addition of the glosses to the context of the expression should help the classifier in deciding whether the MWE is used in a literal or …

Content truncated to 500 words