Multilingual Generative Retrieval via Cross-lingual Semantic Compression

2025-10-09

AI Summary

Addressing two key challenges in multilingual generative retrieval—cross-lingual token misalignment and token inflation—this paper proposes a cross-lingual semantic compression framework. It achieves semantic alignment of multilingual keywords via shared atomic representation learning and introduces a dynamic multi-step constrained decoding strategy to jointly optimize semantic consistency and decoding efficiency within a unified, low-dimensional token space. This work is the first to deeply integrate cross-lingual semantic alignment with token-space compression, thereby enhancing retrieval robustness and generation controllability. On the mMARCO-100k and mNQ-320k benchmarks, the method improves retrieval accuracy by 6.83% and 4.77%, respectively, while reducing average token length by 74.51% and 78.2%. These results demonstrate its comprehensive advantages in accuracy, efficiency, and compactness.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Data Mining & Knowledge Management: Data Compression

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Generative Information Retrieval is an emerging retrieval paradigm that exhibits remarkable performance in monolingual scenarios.However, applying these methods to multilingual retrieval still encounters two primary challenges, cross-lingual identifier misalignment and identifier inflation. To address these limitations, we propose Multilingual Generative Retrieval via Cross-lingual Semantic Compression (MGR-CSC), a novel framework that unifies semantically equivalent multilingual keywords into shared atoms to align semantics and compresses the identifier space, and we propose a dynamic multi-step constrained decoding strategy during retrieval. MGR-CSC improves cross-lingual alignment by assigning consistent identifiers and enhances decoding efficiency by reducing redundancy. Experiments demonstrate that MGR-CSC achieves outstanding retrieval accuracy, improving by 6.83% on mMarco100k and 4.77% on mNQ320k, while reducing document identifiers length by 74.51% and 78.2%, respectively.

Problem

• Addressing cross-lingual identifier misalignment in generative retrieval
• Compressing identifier space to reduce redundancy and inflation
• Improving multilingual retrieval accuracy and decoding efficiency

Innovation

• Unifies multilingual keywords into shared semantic atoms
• Compresses identifier space to reduce redundancy
• Uses dynamic multi-step constrained decoding strategy