WiC = TSV = WSD: On the Equivalence of Three Semantic Tasks

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2478 - 2486 July 10-15, 2022 ©2022 Association for Computational Linguistics WiC = TSV = WSD: On the Equivalence of Three Semantic Tasks Bradley Hauer, Grzegorz Kondrak Alberta Machine Intelligence Institute, Department of Computing Science University of Alberta, Edmonton, Canada {bmhauer,gkondrak}@ualberta.ca Abstract The Word-in-Context (WiC) task has attracted considerable attention in the NLP community, as demonstrated by the popularity of the re- cent MCL-WiC SemEval shared task. Systems and lexical resources from word sense disam- biguation (WSD) are often used for the WiC task and WiC dataset construction. In this paper, we establish the exact relationship be- tween WiC and WSD, as well as the related task of target sense verification (TSV). Build- ing upon a novel hypothesis on the equivalence of sense and meaning distinctions, we demon- strate through the application of tools from theoretical computer science that these three semantic classification problems can be pair- wise reduced to each other, and therefore are equivalent. The results of experiments that in- volve systems and datasets for both WiC and WSD provide strong empirical evidence that our problem reductions work in practice. 1 Introduction This paper answers an open question about the the relation between two important tasks in lexical se- mantics. Word sense disambiguation (WSD) is the task of tagging a word in context with its sense (Navigli, 2009). The word-in-context (WiC) prob- lem is the task of deciding whether a word has the same meaning in two different contexts (Pilehvar and Camacho-Collados, 2019). A crucial differ- ence between the two tasks is that WSD depends on a pre-defined sense inventory1 while WiC does not involve any identification or description of word meanings. Despite ongoing interest in both tasks, there is substantial disagreement in the literature as to whether WiC is a re-formulation of WSD (e.g. Levine et al. (2020)) or an entirely distinct task (e.g. Martelli et al. (2021)). 1For the purposes of this paper, we assume that the WSD sense inventory, the discrete enumeration of the senses of each content word, is the WordNet sense inventory (Fellbaum, 1998), which is a standard practice in WSD (Raganato et al., 2017). By establishing that WSD and WiC are equiv- alent, we construct a theoretical foundation for the transfer of resources and methods between the two tasks. WSD has been intensively studied for decades, while WiC has recently attracted con- siderable attention from the research community. For example, the MCL-WiC SemEval shared task (Martelli et al., 2021) attracted 48 teams, and WiC instances have been integrated into the SuperGLUE benchmark (Wang et al., 2019). Understanding how the two tasks relate to each other allows us to correctly interpret and confidently build upon those results, including prior work on using WSD systems for WiC (e.g. Loureiro and Jorge (2019)). We establish the theoretical equivalence of WiC and WSD by specifying reduction algorithms which produce a solution for one problem by ap- …

Content truncated to 500 words