Lexical Substitution: Enhancing Language Generation with Word Replacement

Abstract

Lexical substitution (LS) is a technique used in natural language processing (NLP) to replace words or phrases in a sentence while preserving the original meaning. Recent advancements in LS based on pretrained language models have shown promising results in suggesting suitable replacements for target words by considering their context. This article explores LS approaches using neural language models (LMs) and masked language models (MLMs) such as context2vec, ELMo, BERT, and XLNet. The study demonstrates the effectiveness of injecting target word information into these models and analyzes the semantic links between targets and substitutes. Lexical substitution plays a vital role in enhancing language generation models and is widely used in NLP tasks such as data augmentation, paraphrase generation, word sense induction, and text simplification. While earlier methods relied on manual lexical resources, the emergence of contextual language models like BERT, ELMo, and XLNet has revolutionized LS by incorporating contextual information and achieving state-of-the-art results. Ongoing research aims to develop more context-aware approaches to address the challenges of lexical substitution, ultimately advancing the capabilities of NLP systems.