Performance Improvement of Semantic Search Using Sentence Embeddings by Dimensionality Reduction

Abstract

Semantic search, which searches for sentences with a high similarity in meaning to that of queries, allows a user to search for the desired sentences even when they cannot think of the appropriate keywords for a lexical search. Moreover, the search function can appropriately handle synonyms and spelling variations. We previously reported a semantic search method for Japanese sentences using sentence embeddings that appropriately processed queries in which sentences were combined using the logical operators AND, OR, and NOT. Reducing the dimensionality of sentence embeddings is expected to make semantic search more robust to noise in the embeddings, resulting in improved search accuracy and faster semantic search computation. In this study, we experimentally verified the improvement in semantic search performance by reducing the dimensionality of sentence embeddings generated by Japanese SimCSE. We also evaluated the runtimes for generating sentence embeddings and reducing dimensionality with PCA.