Leveraging large language models and embedding representations for enhanced word similarity computation.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Current word similarity methods struggle with fine-grained semantics and context.
• Generative semantic representations often exhibit biases (e.g., part-of-speech) and redundancy, impacting accuracy.

Purpose of the Study:

• To propose WSLE, a framework that integrates large language models (LLMs) with vector embeddings for improved word similarity computation.
• To address challenges in LLM-based semantic representation generation, including bias and redundancy.

Main Methods:

• WSLE applies constraints to lexical items, grammatical categories, semantic descriptions, and prompt length to refine LLM semantic generation.
• Generated semantic representations are converted into high-dimensional vector embeddings using a deep semantic embedding module.
• Evaluation employs Pearson's (r) and Spearman's (ρ) correlation coefficients on benchmark datasets.

Main Results:

• WSLE effectively mitigates part-of-speech bias, semantic ambiguity, and informational redundancy in LLM-generated representations.
• The framework demonstrates superior performance compared to existing methods on RG65, MC30, YP130, and MED38 datasets.
• WSLE achieves notable advantages in accuracy and robustness for word similarity measurement.

Conclusions:

• WSLE offers a robust and accurate approach to word similarity computation by enhancing LLM semantic representations.
• The proposed method overcomes limitations of traditional techniques, paving the way for more precise semantic analysis.