Semantic Search for Undergraduate Theses Using Sentence Embeddings

Abstract

Students searching a university thesis repository often fail to find relevant work because keyword search does not match synonyms or paraphrases. We implemented a semantic search service that encodes thesis titles and abstracts with a multilingual sentence embedding model and retrieves results by cosine similarity, combined with the existing keyword index. On a benchmark of 300 queries written by students, the hybrid approach improved normalised discounted cumulative gain at rank ten from 0.52 to 0.71. Gains were largest for conceptual queries written in everyday language. Indexing 21,000 theses took under an hour on a single server, and queries return in under 200 milliseconds. The service is now offered alongside traditional search in the library catalogue.