Abstract:

Sentence embeddings are fundamental to natural

language processing, as they enable models to capture semantic

meaning beyond surface-level word similarity. The ability to

represent sentences and paragraphs in dense vector spaces

facilitates tasks such as semantic search, paraphrase detection,

and textual inference. In this work, we present a comparative

analysis of eight representative models—paraphrasedistilroberta,

msmarco-roberta, paraphrase-mpnet, paraphrasexlm-

r, LaBSE, e5-base, gte-base, and bge-base—evaluated across

four benchmark datasets (MRPC, QQP, PAWS, and VISLA). Our

experiments highlight strengths and limitations of each model,

providing insights into their effectiveness across diverse semantic

similarity tasks.