Reproduce “Evolution of Semantic Similarity -- A Survey” on OpenResearch

About this paper

Dhivya Chandrasekaran, Vijay Mago

A survey of 118 studies traces semantic similarity from lexical databases and distributional statistics to neural, transformer, and hybrid representations. It organizes methods by their semantic assumptions, reviews major benchmarks and evaluation metrics, and finds that contextual and ensemble models generally achieve stronger sentence-level correlations while retaining challenges in interpretability, resource demands, multilingual coverage, and word-sense disambiguation.

Key insights

• Knowledge-based systems provide interpretable synonymy and taxonomy relations but depend on complete lexical resources and transfer poorly across languages and specialized domains.
• Distributional and neural representations adapt more readily to data and context, yet can conflate word senses, require substantial corpora or computation, and offer limited interpretability.
• Hybrid and ensemble systems combine explicit semantic knowledge with learned representations; their strong SemEval performance supports integrating complementary signals rather than relying on one representation type.

Summary by alphaXiv.