MCL@IITK at SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation using Augmented Data, Signals, and Transformers

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 511–520 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 511 MCL@IITK at SemEval-2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation using Augmented Data, Signals, and Transformers Rohan Gupta Jay Mundra∗ Deepak Mahajan∗ Indian Institute of Technology Kanpur (IIT Kanpur)Ashutosh Modi {rohangpt, jaym, dipakam}@iitk.ac.in ashutoshm@cse.iitk.ac.in Abstract In this work, we present our approach for solving the SemEval 2021 Task 2: Multilin- gual and Cross-lingual Word-in-Context Dis- ambiguation (MCL-WiC). The task is a sen- tence pair classification problem where the goal is to detect whether a given word common to both the sentences evokes the same mean- ing. We submit systems for both the settings - Multilingual (the pair’s sentences belong to the same language) and Cross-Lingual (the pair’s sentences belong to different languages). The training data is provided only in English. Consequently, we employ cross-lingual trans- fer techniques. Our approach employs fine- tuning pre-trained transformer-based language models, like ELECTRA and ALBERT, for the English task and XLM-R for all other tasks. To improve these systems’ performance, we propose adding a signal to the word to be dis- ambiguated and augmenting our data by sen- tence pair reversal. We further augment the dataset provided to us with WiC, XL-WiC and SemCor 3.0. Using ensembles, we achieve strong performance in the Multilingual task, placing first in the EN-EN and FR-FR sub- tasks. For the Cross-Lingual setting, we em- ployed translate-test methods and a zero-shot method, using our multilingual models, with the latter performing slightly better. 1 Introduction A key challenge in lexical semantics is to identify or to encode the different senses of an ambigu- ous word. The Word Sense Disambiguation task (WSD) (Navigli, 2009) is a framework used to eval- uate systems in their ability to identify different senses of the word. The task involves selecting the correct sense (meaning) of a target word from a list of senses listed in a sense inventory like WordNet (Fellbaum, 2012). Pilehvar and Camacho-Collados ∗ Authors equally contributed to this work. (2018) proposed a novel benchmark (WiC - Word in Context Disambiguation) for the task casting the problem as a binary classification task, wherein it has to be identified whether a word common to a sentence pair is used in the same sense or not. The WiC task frees up the word sense disambiguation task from being tied to any sense inventory. The SemEval 2021 Task 2: Multilingual and Cross-lingual Word-in-Context Disambiguation (Martelli et al., 2021) extends the WiC frame- work proposed by Pilehvar and Camacho-Collados (2018) to more languages. The task is divided into two subtasks - the Multilingual task and the Cross- Lingual task. The sentence pair, with a word in common, which is to be disambiguated, is drawn from the same language in the MultiLingual task, whereas the pair is drawn from two different lan- guages in the Cross-Lingual Task. The task is posed as binary classification task over a pair of sentence wide contexts sent1 and sent2, …

Content truncated to 500 words