Multilingual Word Sense Disambiguation with Unified Sense Representation

Proceedings of the 29th International Conference on Computational Linguistics, pages 4193–4202 October 12–17, 2022. 4193 Multilingual Word Sense Disambiguation with Unified Sense Representation Ying Su1, Hongming Zhang1,2, Yangqiu Song1, Tong Zhang1 1HKUST, 2Tencent AI lab, Seattle ysuay@connect.ust.hk, {hzhangal,yqsong}@cse.ust.hk, tongzhang@ust.hk Abstract As a key natural language processing (NLP) task, word sense disambiguation (WSD) eval- uates how well NLP models can understand the lexical semantics of words under specific contexts. Benefited from the large-scale an- notation, current WSD systems have achieved impressive performances in English by com- bining supervised learning with lexical knowl- edge. However, such success is hard to be replicated in other languages, where we only have limited annotations. In this pa- per, based on the multilingual lexicon Ba- belNet describing the same set of concepts across languages, we propose building knowl- edge and supervised-based Multilingual Word Sense Disambiguation (MWSD) systems. We build unified sense representations for multiple languages and address the annotation scarcity problem for MWSD by transferring annota- tions from rich-sourced languages to poorer ones. With the unified sense representations, annotations from multiple languages can be jointly trained to benefit the MWSD tasks. Evaluations of SemEval-13 and SemEval-15 datasets demonstrate the effectiveness of our methodology. 1 Introduction As a critical natural language understanding task, word sense disambiguation (WSD) aims at classi- fying words into pre-defined senses. With such a disambiguation process, machines can understand the precise meanings of words. Previous researches have demonstrated that a sound WSD system could benefit many downstream NLP tasks, such as ma- chine translation (Pu et al., 2018; Liu et al., 2018) and information extraction (Bovi et al., 2015). Existing researches on word sense disambigua- tion mostly focus on English only. By leverag- ing lexical knowledge such as gloss (Iacobacci et al., 2016; Luo et al., 2018; Huang et al., 2019; Blevins and Zettlemoyer, 2020) or graph structure WSD Example. Language: English Context: Detailed studies of the plan were well underway. plan#NOUN: plan%1:09:00:: plan%1:09:01:: plan%1:06:00:: MWSD Example. Language: French Context: Le groupe des Nations_Unies a des projets de plans pour la réduction des émissions. plan#NOUN: bn:00062759n bn:00062766n bn:00005439n Figure 1: Examples of WSD and Multilingual WSD (MWSD) task. The target words are indicated with the bold font in contexts. Candidate sense keys are listed below each context, and the one in blue is the correct sense. (Banerjee et al., 2003; Kumar et al., 2019; Bevilac- qua and Navigli, 2020) and supervised training over large-scale annotations, these models have achieved impressive performance on the standard English WSD task. However, though the English WSD task (Raganato et al., 2017) and multilingual WSD (MWSD) task (Navigli et al., 2013; Moro and Navigli, 2015) are of the same form as shown in Figure 1, this progress can not be easily applied across languages as the paucity of annotated train- ing data and immense labor in handling diverse lex- ical knowledge of multiple languages separately. BabelNet (Navigli and Ponzetto, 2012) is a mul- tilingual semantic lexicon and contains a set of multilingually lexicalized concepts. Similar to WordNet (Miller, …

Content truncated to 500 words