Exploring Synonym Generation for Lexical Simplification: A Comparative Analysis of Static and Contextualized Word Embeddings

In this age of the Internet, users are inundated with information ubiquitously. So, the ability to understand acquire knowledge is essential. Aiming to enhance the text’s accessibility to diverse users, we focused on the simplification of the text at the lexical level. According to earlier research findings, the emphasis is on synonyms or substitute replacements for and a complex word, which is a sub-task of lexical simplification. The generation of relevant substitutes pivotally improves the overall coherence while preserving the original meaning of the text. Currently, research has centered on extracting semantic relations from word embeddings since they capture relatedness and similarity between words using neural language models. We conduct an assessment by comparing static and contextualized word embeddings to investigate how these models can generate relevant substitutes for a given word based on context. Our experiments reveal that contrast senses are also generated for a target word due to shared context in mostly static word embedding models, a phenomenon not observed in ChatGPT’s zero-shot prompting. Experimental analysis shows that the micro-precision has increased by 15.52 and 16.72% on BenchLS and NNSevel datasets by GPT embeddings.