LexSubDis:Knowledge-Integrated Lexical Substitution via Discriminative Ranking

ACL ARR 2025 May Submission410 Authors

Abstract: Lexical substitution, a fundamental task in natural language processing, aims to replace target words with semantically equivalent or synonymous substitutes while preserving original sentence meaning. Although extensively explored, existing methods exhibit two major limitations: 1) inadequate investigation of embedding representations when target word contain subwords, and 2) excessive hyperparameters and computational complexity from multi-metric evaluation in candidate ranking. To address these issues, we propose LexSubDis, which constructs more MLM-compatible substitution mechanisms by averaging subword embeddings of target words and combining them with synonym embeddings. Moreover, we pioneer the introduction of discriminator models to assess semantic impacts of substitutions. Experimental results demonstrate that LexSubDis significantly reduces hyperparameters while achieving state-of-the-art performance under unsupervised learning on CoInCo dataset's $ootm$ metric, offering novel insights and solutions for lexical substitution research.

Paper Type: Long

Research Area: Semantics: Lexical and Sentence-Level

Research Area Keywords: lexical relationships, lexical semantic change, semantic textual similarity

Languages Studied: English

Submission Number: 410