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

Abstract: With the rapid expansion of massive multilingual corpora, Multilingual Information Retrieval (MLIR) has emerged as a critical technology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query. However, recent multilingual dense retrieval models often exhibit a strong preference for documents in the same language as the query. This leads to severe language bias, where top-ranked results are dominated by documents of specific languages, even when documents in other languages contain more semantically relevant information. To address this issue, we propose SHIFT, a training-free method applicable in the indexing stage. Specifically, SHIFT utilizes parallel translation pairs to estimate a relative language vector for each target language with respect to a source language. Subsequently, SHIFT corrects the language-specific offset by subtracting this relative language vector from document embeddings during indexing. Our comprehensive evaluation across four MLIR benchmarks and diverse dense retrieval models confirms that SHIFT can effectively mitigate language bias and enhance MLIR performance.

• The paper introduces a training-free index-side feature transformation that realigns document embeddings, mitigating language bias in multilingual retrieval.
• The paper quantifies improvements using Target-Languages Recall@k, demonstrating up to 16.4% gain and enhanced semantic access across various language pairs.
• The paper shows that SHIFT's design preserves embedding geometry and reduces inference overhead, enabling fair and efficient multilingual information systems.

SHIFT: Mitigating Language Bias in Multilingual Information Retrieval

Background and Motivation

The widespread adoption of multilingual corpora has made Multilingual Information Retrieval (MLIR) a foundational component in global information access and retrieval-augmented applications. Despite advancements in multilingual dense retrieval architectures, current models demonstrate pronounced language bias: retrieved documents are overwhelmingly in the query's language, often prioritizing linguistic identity over semantic relevance. This discrepancy impacts not only user experience and fairness but also downstream tasks such as Retrieval-Augmented Generation (RAG), where models may be deprived of diverse and semantically relevant multilingual contexts (2606.18801).

Prior Approaches and Limitations

Existing solutions for addressing language bias fall into two main categories: (1) enhanced linguistic alignment via additional training (e.g., contrastive learning) and (2) pipeline-level debiasing such as query translation and post-hoc normalization. The former incurs significant computational costs and lacks flexibility toward different language combinations. The latter, including language-wise centering [Libovický et al., 2020] and projection-based methods [Yang et al., 2021], typically increases system complexity and inference-time latency, sometimes at the expense of semantic signal preservation.

SHIFT: Index-side Feature Transformation

SHIFT (Semantic Harmonization via Index-side Feature Transformation) is proposed as a training-free, indexing-stage method to realign document embeddings within the multilingual retrieval embedding space. It estimates a relative language vector for each target language based on parallel translation pairs from corpora (mMARCO), quantifying the representational offset between languages. During indexing, SHIFT subtracts the appropriate language vector from document embeddings—pulling multilingual documents toward the source-language semantic subspace.

The process preserves the underlying high-dimensional geometry of the embeddings, ensures zero additional overhead at retrieval time, and provides robustness across a broad range of dense retriever models. Notably, a global scale factor $\alpha$ allows for adjustable debiasing intensity, offering either uniform correction or explicit language-wise control (2606.18801).

Quantifying Language Bias: Target-Languages Recall@k

Traditional evaluation metrics (Recall@20, nDCG@20) may mask the extent of bias since they reward source-language retrieval. The paper introduces Target-Languages Recall@k (TLR@k), quantifying retrieval effectiveness exclusively on non-query language documents. TLR@k enables deeper inspection of whether debiasing methods genuinely enhance cross-lingual semantic access or merely shuffle rankings within the query language.

Empirical Evaluation

Extensive experiments across four MLIR benchmarks (Belebele, XQuAD, MLQA, MultiEuP-v2) and diverse encoder/decoder-based dense retrievers demonstrate the effectiveness of SHIFT. Results show consistent improvements in nDCG@20, Recall@20, and especially TLR@20. For example, SHIFT yields up to a 16.4% relative gain in TLR@20, with pronounced improvements in models exhibiting severe baseline bias (multilingual-e5-large). The method generalizes robustly to non-English source queries (Chinese, Vietnamese, Hindi), correcting alignment even in models strongly optimized for English-centric retrieval.

Language Distribution and Qualitative Alignment

Heatmap analyses reveal that SHIFT disperses high-scoring document density across languages, mitigating query-language dominance in top retrievals and producing a more balanced and semantically grounded ranking. The method outperforms post-hoc debiasing baselines not only numerically but also in practical system deployment, as it can be applied exclusively at indexing time—reducing inference latency and operational complexity.

Robustness and Practical Considerations

SHIFT is resilient to imperfect language identification, maintaining debiasing efficacy even when relying on predicted document language labels. Performance degradation is minimal in realistic scenarios where ground-truth language is not available, and language vectors can be cached across indexing runs as long as the same retriever and anchor language are retained.

Sensitivity analyses on the scale factor $\alpha$ demonstrate its role as a controllable parameter: increasing $\alpha$ boosts target-language document exposure but can lead to trade-offs in overall retrieval balance—empowering adaptive system design for different deployment contexts.

Further experiments in Multilingual-to-Multilingual IR (M2MIR) settings showcase that SHIFT's alignment, when applied to both queries and documents, maximizes geometric congruence in the embedding space—yielding the best retrieval metrics in practical multilingual search scenarios.

Implications and Future Directions

From a theoretical perspective, SHIFT provides evidence that representational misalignment between languages in dense retrieval embedding spaces is a major source of language bias, and that linear correction via index-side transformations is both effective and computationally efficient. Its deployment can improve fairness, completeness, and semantic fidelity in international information systems.

Future research directions include relaxing dependence on explicit language labels, exploring unsupervised estimation of relative language vectors, and integrating SHIFT with further advances in retrieval architectures or embedding models. The controllability offered by SHIFT's scale factor hints at possibilities for user-adaptive retrieval, regional customization, and more granular semantic harmonization in the IR pipeline.

Conclusion

SHIFT offers a robust, training-free mechanism for harmonizing semantic representations across languages in MLIR. By correcting language-induced embedding offsets, it mitigates language bias and enhances semantic retrieval, validated across benchmarks, architectures, and language settings. The approach facilitates fair, effective multilingual information access and opens avenues for adaptive and scalable retrieval system design (2606.18801).