Multilingual Generative Retrieval via Cross-lingual Semantic Compression

TL;DR

This paper tackles cross-lingual identifier misalignment and DocID inflation in multilingual Generative Information Retrieval by introducing MGR-CSC, which unifies semantically equivalent multilingual keywords into shared semantic atoms and uses dynamic constrained multi-step decoding. The method operates in three stages: multilingual keyword extraction, semantic-atom based DocID construction, and atom-by-atom constrained decoding guided by the query, reducing decoding space from $O(N^m)$ to $O(C^m)$ and compressing the identifier space. Empirical results on mMarco100K and mNQ320K show consistent improvements in Recall@10 (e.g., $6.83\%$ and $4.77\%$), along with substantial DocID length reductions of $74.51\%$ and $78.2\%$, respectively. The work demonstrates strong cross-lingual retrieval performance and decoding efficiency, with ablations confirming the necessity of both semantic compression and the constrained decoding strategy; limitations include dependency on multilingual PLMs and the current lack of multimodal integration. Overall, MGR-CSC offers a principled route to scalable, cross-lingual generative retrieval with practical implications for multilingual search and retrieval systems.

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.

Figures (6)

• Figure 1: (a) illustrates the language-independent keyword approach, where the model biases toward query-language DocIDs during decoding. In contrast, (b) demonstrates MGR-CSC successfully retrieves target documents via semantic clustering, enabling more reliable identification.
• Figure 2: An overview of MGR-CSC's DocID allocation and reasoning. (a) shows how to extract keywords of documents; (b) shows clustering is performed based on cross-lingual semantic similarity of keywords, with each cluster represented by an atom and each document assigned a unique DocID; (c) shows MGR-CSC's reasoning strategy, which returns the identifier corresponding to the query, and narrows the decoding range at each step.
• Figure 3: Distribution of the number of DocIDs across different languages for various methods on the mMarco100k and mNQ320k datasets. For SE-DSI, DSI-QG, and DSI, the stacked bar segments represent the distribution of retrieved DocIDs across languages. The overall DocID count for each method is indicated to the right of the corresponding bar.
• Figure 4: Recall@10 performance of target-language document retrieval with varying source query languages in the mNQ320k dataset
• Figure 5: Performance in Recall@10 and decoding time under varying keyword quantities $m$.