QUERY-TYPE SENSITIVITY AND SELECTIVE BENEFITS OF HYBRID LEXICAL- SEMANTIC RETRIEVAL: AN EXPERIMENTAL ANALYSIS WITH BM25 AND MULTILINGUAL SBERT

Abstract

Hybrid lexical-semantic retrieval combines signals derived from explicit term matching and dense semantic representations. However, an aggregate retrieval score can conceal heterogeneous behavior across information needs. This study examines that heterogeneity at query and predefined query-type levels using a controlled university information-retrieval benchmark of 50 queries and 19 documents, corresponding to 950 query-document pairs. Four retrieval conditions are evaluated descriptively: BM25, multilingual SBERT, Reciprocal Rank Fusion (RRF), and normalized lexical-semantic score fusion with a fixed weight a=0.5.