Distillation for Multilingual Information Retrieval
Eugene Yang, Dawn Lawrie, James Mayfield
TL;DR
Multilingual information retrieval (MLIR) requires ranking documents across languages, which is challenging for models trained on monolingual data. The paper introduces Multilingual Translate-Distill (MTD), a distillation-based training framework that translates training passages into all document languages and uses a teacher-student setup to learn a multilingual ColBERT-X. Across four MLIR benchmarks, MTD consistently surpasses the previous state-of-the-art (MTT), with substantial gains in nDCG@20 and MAP, and shows robustness to how languages are mixed in training batches. The work provides practical guidance on language-mixing strategies and demonstrates that training in native languages yields strong benefits, especially in morphologically diverse collections; code and models are publicly available.
Abstract
Recent work in cross-language information retrieval (CLIR), where queries and documents are in different languages, has shown the benefit of the Translate-Distill framework that trains a cross-language neural dual-encoder model using translation and distillation. However, Translate-Distill only supports a single document language. Multilingual information retrieval (MLIR), which ranks a multilingual document collection, is harder to train than CLIR because the model must assign comparable relevance scores to documents in different languages. This work extends Translate-Distill and propose Multilingual Translate-Distill (MTD) for MLIR. We show that ColBERT-X models trained with MTD outperform their counterparts trained ith Multilingual Translate-Train, which is the previous state-of-the-art training approach, by 5% to 25% in nDCG@20 and 15% to 45% in MAP. We also show that the model is robust to the way languages are mixed in training batches. Our implementation is available on GitHub.