Neural Approaches to Multilingual Information Retrieval
This work addresses the problem of efficiently retrieving documents across multiple languages for users needing cross-lingual access, presenting an incremental improvement over earlier MLIR techniques.
This paper tackled the challenge of multilingual information retrieval (MLIR) by investigating neural document translation and pretrained multilingual language models, finding that using XLM-R with native language indexing achieved 98% of the best Mean Average Precision while reducing indexing time by 84%, with no statistically significant difference in effectiveness.
Providing access to information across languages has been a goal of Information Retrieval (IR) for decades. While progress has been made on Cross Language IR (CLIR) where queries are expressed in one language and documents in another, the multilingual (MLIR) task to create a single ranked list of documents across many languages is considerably more challenging. This paper investigates whether advances in neural document translation and pretrained multilingual neural language models enable improvements in the state of the art over earlier MLIR techniques. The results show that although combining neural document translation with neural ranking yields the best Mean Average Precision (MAP), 98% of that MAP score can be achieved with an 84% reduction in indexing time by using a pretrained XLM-R multilingual language model to index documents in their native language, and that 2% difference in effectiveness is not statistically significant. Key to achieving these results for MLIR is to fine-tune XLM-R using mixed-language batches from neural translations of MS MARCO passages.