SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval

Multilingual information retrieval models often favor documents in the same language as the query, leading to language bias. A new method, SHIFT, aims to mitigate this issue by correcting language-specific offsets during indexing.

Multilingual Information Retrieval (MLIR) enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query^[1]. However, recent multilingual dense retrieval models often exhibit a strong preference for documents in the same language as the query. SHIFT, a training-free method, utilizes parallel translation pairs to estimate a relative language vector for each target language with respect to a source language and corrects the language-specific offset by subtracting this relative language vector from document embeddings during indexing^[1]. A new benchmark, MMed-Bench-IR, has been designed to evaluate the capabilities of MLIR models, consisting of three structurally heterogeneous tasks: cross-lingual medical QA retrieval, concept discrimination, and multilingual evidence retrieval^[2]. Evaluation of ten systems across six paradigm families using MMed-Bench-IR reveals severe cross-lingual failure, with a significant drop in performance from English to Japanese^[2].