Paper deep dive
SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval
Youngjoon Jang, Seongtae Hong, Hyeonseok Moon, Heuiseok Lim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 6/21/2026, 3:04:44 AM
Summary
The paper introduces SHIFT (Semantic Harmonization via Index-side Feature Transformation), a training-free method designed to mitigate language bias in Multilingual Information Retrieval (MLIR). MLIR models often exhibit a preference for documents in the same language as the query, even when semantically relevant documents exist in other languages. SHIFT addresses this by estimating a 'relative language vector' using parallel translation pairs (from the mMARCO dataset) and subtracting this vector from document embeddings during the indexing stage. This process aligns target-language embeddings closer to the source-language embedding space. The authors also introduce a new metric, Target-Languages Recall@k (TLR@k), to specifically quantify the retrieval effectiveness of non-query language documents. Experimental results across various dense retrieval models (encoder-based like multilingual-e5-large and decoder-based like Qwen3-Embedding) and benchmarks (Belebele, XQuAD, MLQA, MultiEuP-v2) demonstrate that SHIFT significantly improves retrieval performance and enhances language diversity in top-ranked results without increasing query-time latency.
Entities (10)
Relation Signals (4)
SHIFT â addresses â language bias
confidence 100% · To address this issue, we propose SHIFT, a training-free method applicable in the indexing stage.
Target-Languages Recall@k â quantifies â language bias
confidence 100% · we introduce Target-Languages Recall@k (TLR@k), a novel metric that explicitly quantifies language bias.
SHIFT â uses â mMARCO
confidence 100% · To estimate these vectors, we utilize parallel documents consisting of sourceâ target translation pairs from mMARCO.
multilingual-e5-large â exhibits â language bias
confidence 90% · An illustration of language bias in MLIR using multilingual-e5-large on the Belebele dataset.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2606.18801v1
- Canonical: https://arxiv.org/abs/2606.18801v1
Trouble viewing inline? Open PDF directly â
Full Text
110,341 characters extracted from source content.
Expand or collapse full text
SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval Youngjoon Jang, Seongtae Hong, Hyeonseok Moon * , Heuiseok Lim * Department of Computer Science and Engineering, Korea University dew1701, ghdchlwls123, glee889, limhseok@korea.ac.kr Abstract With the rapid expansion of massive multi- lingual corpora, Multilingual Information Re- trieval (MLIR) has emerged as a critical tech- nology for global information access. MLIR enables users to retrieve semantically relevant documents from multilingual text collections using a single-language query. However, re- cent 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 re- sults are dominated by documents of specific languages, even when documents in other lan- guages contain more semantically relevant in- formation. 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. 1 Introduction Information Retrieval (IR) constitutes the backbone of modern information systems, from traditional search engines to Retrieval-Augmented Generation (RAG) (Singhal et al., 2001; Kobayashi and Takeda, 2000; Lewis et al., 2020). With the rapid expansion of massive multilingual corpora, Multilingual In- formation Retrieval (MLIR) has become essential for equitable information access. MLIR enables users to retrieve relevant information from mul- tilingual text collections using a single-language query (Peters and Braschler, 2002a). In real-world applications, such as global enterprise search and * Corresponding authors [1 st ] ...Germany code-named âOperation Sealionâ... [2 nd ] ...the invasion was called âOperation Dragoonâ... [3 rd ] ...se referiuaoataquecomoâOperaçãoSealionâ... [4 th ] ...aircraft headed to Irkutsk and was being... [14 th ] ... « OpĂ©rationLion de Mer » Ă cetteattaque. ... Q: âOperation Sealionâ was an invasion of which country? Top-14 Retrieved Documents Language Query Relevance ... :10/14 Language Composition :1/14:1/14:1/14:1/14 Figure 1: An illustration of language bias in MLIR using multilingual-e5-large on the Belebele dataset. The symbolsâandĂdenote semantic relevance. Despite the existence of semantically equivalent ground-truth documents across 14 languages, the top retrieved results are dominated by English documents (10/14). multilingual RAG, robust MLIR is essential to sup- ply LLMs with diverse multilingual knowledge, preventing them from generating answers based solely on particular languages. Unlike Cross-Lingual Information Retrieval (CLIR), which assumes a one-to-one language map- ping (Goswami et al., 2021), MLIR presents a more complex one-to-many challenge: simultaneously retrieving relevant documents across multiple tar- get languages. Consequently, naive CLIR adapta- tions such as query translation, fail to perform effec- tively across massive multilingual corpora (Granell, 2014; Peters et al., 2012). Ideally, an MLIR system should rank documents solely based on semantic similarity, invariant to linguistic surface forms. However, state-of-the-art multilingual dense retrievers suffer from severe lan- guage bias (Yang et al., 2024c; Hong et al., 2026). Rather than prioritizing semantic relevance, these models exhibit a disproportionate preference for documents written in the query language (Park and Lee, 2025; Sharma et al., 2025). As illustrated in Figure 1, retrieving with multilingual-e5-large on the Belebele dataset reveals a pronounced language 1 arXiv:2606.18801v1 [cs.IR] 17 Jun 2026 skew. Despite the existence of semantically equiva- lent ground-truth documents across 14 languages, the top ranks for an English query are saturated with semantically irrelevant English documents (10 out of 14). This indicates that language iden- tity dominates semantic signals in the embedding space, distorting MLIR ranking behavior. Conventional metrics often mask this bias by pre- dominantly rewarding source-language retrievals. To explicitly quantify this, we introduce Target- Languages Recall@k (TLR@k), a novel metric measuring the retrieval effectiveness of non-query language documents. Furthermore, to mitigate this bias, we propose SHIFT, a training-free method applied during the indexing stage. By estimating a relative language vector from parallel transla- tion pairs and subtracting it from document embed- dings, SHIFT corrects the language-specific offset and forces the model to rely purely on underlying semantic signals. In our experiments, SHIFT yields consistent improvements, achieving up to a 16.4% relative gain in TLR@20. The main contributions of this work are as follows: âąWe demonstrate that SHIFT is a universally ef- fective solution, consistently improving MLIR performance and mitigating language bias across diverse dense retrieval models. âąWe introduce Target-Languages Recall@k (TLR@k), a novel metric that explicitly quan- tifies language bias by measuring the retrieval of non-query language documents. âąThrough detailed analysis and experiments, we provide empirical evidence that SHIFT successfully realigns multilingual documents based on semantics, correcting the skewed ge- ometry of the embedding space. 2 Preliminaries 2.1 Related Works The concept of "multilingual" in Information Re- trieval (IR) can be interpreted in various contexts. For instance, multilingual IR is often referred to as monolingual retrieval across multiple lan- guages (Hull and Grefenstette, 1996; Blloshmi et al., 2021) or as CLIR tasks involving multiple languages (Braschler, 2002a,b; Lawrie et al., 2022; Mitamura et al., 2008). In this work, however, we adopt the definition of MLIR established by the Cross-Language Evaluation Forum (CLEF): a task where a user retrieves information from a mixed- language document collection using a query in a single language (Peters and Braschler, 2002b). Traditional MLIR approaches relied on Machine Translation (MT) to translate either the entire doc- ument collection or the query (Darwish and Oard, 2003; Kraaij et al., 2003; McNamee and Mayfield, 2002; Granell, 2014; Magdy and Jones, 2011; Pe- ters et al., 2012; Si et al., 2008; Tsai et al., 2008). However, these methods incur significant compu- tational costs or introduce systematic ranking bias due to disparate score distributions (Huang et al., 2023; Lawrie et al., 2023). Recently, dense retrieval models based on mul- tilingual Pre-trained Language Models (mPLMs) have become a standard, demonstrating robust per- formance in MLIR tasks (Conneau et al., 2020; Chen et al., 2024b; Yu et al., 2024). In MLIR scenarios, dense retrievers encode both queries and documents into a shared embedding space, en- abling efficient retrieval via Approximate Nearest Neighbor (ANN) search using scoring functions such as dot product or cosine similarity (Arya et al., 1998; Liu et al., 2004; Dehak et al., 2010). 2.2 Language Bias in MLIR However, the embedding spaces of dense retrievers fail to achieve perfect linguistic alignment. Con- sequently, documents representing identical mean- ings but written in different languages are often mapped to distant locations, leading to language bias that distorts the ranking (Yang et al., 2024c,a). An ideal MLIR system should rank documents solely based on semantic similarity, independent of the documentâs language. Yet, in practice, a recur- ring phenomenon is observed where documents in the same language as the query disproportionately occupy the top ranks (Park and Lee, 2025; Sharma et al., 2025). This bias increases the exposure of specific languages within multilingual collections, simultaneously undermining the fairness of infor- mation access and user experience. To mitigate this issue, prior research has largely adopted two approaches.The first involves strengthening the modelâs linguistic alignment through additional training (Hu et al., 2023; Yang et al., 2024c; Huang et al., 2023; Yang et al., 2024b). While effective, this approach requires high com- putational costs due to large-scale training and ne- cessitates retraining whenever the language combi- nation of the model or data changes. The second approach addresses language bias at the pipeline 2 level without modifying the model. A representa- tive line of work translates each query into multi- ple languages and fuses language-specific ranked lists (Lawrie et al., 2025a; Yang et al., 2025). Al- though effective, such pipelines typically increase system complexity and inference-time latency. An- other line mitigates bias by post-hoc normalization of multilingual embeddings (e.g., language-wise centering (LibovickĂœ et al., 2020) or projection- based methods (Yang et al., 2021)) without any additional training. While these strategies can improve retrieval performance, they may inadver- tently suppress semantic signal when enforcing language-agnostic representations, and often re- quire additional query-side processing, increasing inference-time latency. 3SHIFT In this work, we propose SHIFT, a training-free method designed to align multilingual represen- tations at the indexing stage. Unlike approaches that necessitate computationally expensive train- ing or impose additional inference latency, SHIFT employs a linear displacement strategy to correct spatial misalignment. This method preserves the high-dimensional structure of the original embed- dings to the greatest extent possible, thereby im- proving MLIR performance with zero additional query-time overhead. 3.1 Estimating Relative Language Vectors LetL tgt denote the set of target languages, and let â tgt be an element ofL tgt . The core mechanism of SHIFT is to estimate a relative language vector V â tgt for each target language, with respect to a source languageâ src . To estimate these vectors, we utilize parallel documents consisting of sourceâ target translation pairs from mMARCO (Bonifacio et al., 2022). For each target languageâ tgt , we collect a set of translation pairs: P â tgt = (D â src i , D â tgt i ) N i=1 , â tgt âL tgt Each pair consists of a source-language document D â src i and its translationD â tgt i , sharing identical semantic meaning. We first encode each document into the embedding space using the dense retrieval model f (·): z â src i = f (D â src i ), z â tgt i = f (D â tgt i ) The difference between the two embeddings, â â tgt i = z â tgt i â z â src i , approximates the representa- tional offset arising solely from the language shift Algorithm 1 SHIFT Indexing Require:Labeled DocumentsD =(d j ,â j ) M j=1 , Modelf, Source Languageâ src , Target Lan- guage SetL tgt , Relative Language Vectors V â ââL tgt , Scale factor αâ [0, 1] 1: for each (d,â)âD do 2: z â f (d) 3:if â = â src then 4: Ìz â z 5:else 6: Ìz â zâ αV â 7:end if 8:Add Ìz to the index 9: end for fromâ src toâ tgt within two semantically equiva- lent documents. We define the relative language vector of each target languageâ tgt with respect to the source language by averaging this offset over all N translation pairs: V â tgt = 1 N N X i=1 z â tgt i âz â src i ), ââ tgt âL tgt (1) 3.2 Index-side Language Shift After estimating the relative language vectors, we calibrate document embeddings during indexing via linear subtraction. Specifically, for each docu- ment pre-labeled with its languageâ, we subtract the corresponding vector ifâ Ìž= â src . This adjust- ment is controlled by a hyperparameterα, which scales the transformation magnitude: largerαsub- tracts more ofV â tgt , pulling target-language docu- ment embeddings closer to the source-language em- bedding space. We use a single globalαshared by all documents and sweepα â 0.1, 0.2,..., 1.0 to report sensitivity toα. We do not tuneαper dataset, language, or query set. The overall index- ing procedure is summarized in Algorithm 1. 3.3 Retrieval The retrieval process on the constructed index Ì Z = Ìz 1 ,..., Ìz M is performed identically to standard dense retrieval procedures. Given a queryqin the source language, we generate the query embedding z q = f (q)using the same embedding modelfused for indexing. The final retrieval score is calculated as the similarity between the query embedding and the calibrated document embedding Ìzin the index: score(q,d) = sim(z q , Ìz) 3 All models used in our experiments utilize cosine similarity forsim(·,·). Since the documents pro- cessed with SHIFT are already aligned to the source language space, no additional transformation is re- quired for the query. Consequently, our method introduces no additional latency at query-time com- pared to baseline models. 4 Experimental Setup 4.1 Models & Resources Relative Language Vector Estimation To es- timate relative language vectors, we use the mMARCO dataset (Bonifacio et al., 2022). mMARCO provides machine-translated versions of the original English content in 14 languages (via Google Translate), yielding aligned text pairs across languages. We leverage all 533k aligned passage pairs for language vector estimation. Multilingual Retrievers To verify the gener- alizability of SHIFT, we employ a diverse set of multilingual dense retrievers featuring differ- ent architectures.We include encoder-based models such as embeddinggemma-300m (Vera et al., 2025), multilingual-e5-large (Wang et al., 2024), and bge-m3 (Chen et al., 2024a), as well as decoder-based models including Qwen3- Embedding-0.6B (Zhang et al., 2025), llama- nemotron-embed-1b-v2 (Babakhin et al., 2025), and gte-Qwen2-1.5B-instruct (Li et al., 2023). Im- plementation details are provided in Appendix B. 4.2 Evaluation Datasets To rigorously evaluate the robustness of SHIFT, we employ four benchmarks adapted to the MLIR setting. In our main experiments (Section 5.1), we use English queries as the source language and retrieve from a multilingual corpus. We further evaluate Chinese, Vietnamese, and Hindi queries as source languages in Section 5.3. Below, we briefly describe each dataset and our MLIR construction procedure. Belebele & XQuAD Belebele (Bandarkar et al., 2024) is a fully parallel 122-language MRC bench- mark built on the FLORES-200 translation bench- mark (Team et al., 2022), and XQuAD (Artetxe et al., 2020) is a fully parallel 10-language transla- tion of a subset of SQuAD v1.1 (Rajpurkar et al., 2016). To adapt these datasets for MLIR, we con- struct a multilingual retrieval including all language versions of the context passages, inspired by the multilingual pool construction in LAReQA (Roy et al., 2020). Accordingly, for each source query, we treat the source-language gold context and all target-language versions of that context as posi- tive documents. In our experiments, we utilize the subset of languages that overlap with mMARCO, resulting in 14 languages for Belebele and 8 lan- guages for XQuAD. MLQAMLQA (Lewis et al., 2019) is an extrac- tive QA benchmark covering 7 languages. Unlike fully parallel datasets where every item exists in all languages, MLQA is constructed such that each instance is aligned across a subset of languages (typically four). For the MLIR setup, we aggre- gate contexts from all 7 languages in the document collection. For a given source query, all available contexts across different languages are treated as positive documents. This implies that the number of positive documents and their language composi- tion may vary per query, but all semantically linked multilingual contexts are considered as positives. MultiEuP-v2 MultiEuP-v2 (Yang et al., 2024c) is a dataset derived from European Parliament pro- ceedings, spanning 24 languages, using debate ti- tles as queries and individual speeches as docu- ments. We configure the MLIR task such that for a single source query (debate title), all speeches belonging to that debate topic are considered pos- itive documents. Thus, the positive document in- cludes both the speech in the original language and speeches in other languages within the same topic. Consistent with our training setup, we evaluate on the 6 languages that overlap with mMARCO. We summarize the dataset statistics in Table 1, with additional details in Appendix A. Data StatisticsBelebeleMLQAXQuADMultiEup-v2 # of Languages 14786 Fully ParallelOXOX # of Queries90011,5821,1901,781 # of Documents6,83236,7991,92035,444 # of d + /q143.65819.9 # of d + en /q1117.91 Table 1: Statistics of the MLIR benchmark datasets. # ofd + /qdenotes the number of positive documents per query, and # ofd + en /qindicates the number of English positive documents per query. 4.3 Evaluation Metrics We employ standard metrics, nDCG@20 and Re- call@20, to measure retrieval accuracy. However, 4 ArchiModel BelebeleMLQAXQuADMultiEup-v2AVG nDCGRecallnDCGRecallnDCGRecallnDCGRecallnDCGRecall encoder embeddinggemma-300m0.9240.9260.6640.7400.9470.9700.4950.4330.7580.767 + SHIFT0.9260.9280.6710.7480.9510.9720.5110.4430.7650.773 multilingual-e5-large0.8160.8120.4940.5230.8550.9000.3670.3020.6330.634 + SHIFT0.9100.9070.6490.7210.9440.9630.4420.3750.7370.742 bge-m30.8740.8710.6070.6790.9230.9470.4550.3870.7150.721 + SHIFT0.8760.8740.6090.6820.9250.9500.4600.3910.7170.724 decoder Qwen3-Embedding-0.6B0.8900.8870.6110.6880.9140.9400.4680.4030.7210.729 + SHIFT0.8950.8920.6110.6880.9160.9420.4760.4110.7250.733 llama-nemotron-embed-1b-v20.6880.6810.5230.6020.9270.9570.2920.2510.6080.623 + SHIFT0.7000.6970.5270.6090.9310.9600.3120.2730.6180.635 gte-Qwen2-1.5B-instruct0.9040.9010.6210.6910.9260.9460.4960.4300.7370.742 + SHIFT0.9140.9120.6390.7160.9360.9550.5190.4510.7520.758 Table 2: Top-20 retrieval performance (nDCG@20, Recall@20) of baseline models and SHIFT across four MLIR benchmarks, evaluated with English as the source query language. We conduct a global grid search overα â 0.1,..., 1.0 and report the best α for each model in subsequent experiments, as described in Section 5.4. these conventional metrics do not reveal whether relevant documents beyond the query language are retrieved. In MLIR benchmarks such as MultiEuP- v2, for instance, a model can obtain high nDCG and Recall scores by disproportionately retrieving source-language documents, despite the presence of relevant documents in other languages. To better capture this aspect, we introduce Target-Languages Recall@k (TLR@k). Essen- tially, this metric computes the recall score exclu- sively on the subset of relevant documents writ- ten in languages other than the query language. For a queryq, letD + q denote the set of rele- vant documents to the query. We define the sub- set of relevant documents in target languages as R tgt q =dâD + q | L(d)Ìž= L(q) . TLR@k is then calculated as the average recall of these documents: TLR@k = 1 |Q| X qâQ top k (q)â© R tgt q R tgt q A higher TLR@k indicates that the model success- fully retrieves target-language documents, comple- menting conventional accuracy metrics that may be dominated by source-language documents. 5 Results & Analysis 5.1 Main Results Retrieval Performance Table 2 presents the retrieval performance before and after applying SHIFT across four MLIR benchmarks. The re- sults demonstrate that SHIFT consistently achieves significant performance gains across all settings. For the encoder-based model, nDCG@20 for the embeddinggemma-300m model increased from 0.758 to 0.765 on average, while it significantly rose from 0.633 to 0.737 for multilingual-e5-large after applying SHIFT. This improvement extends to decoder-based models, suggesting that SHIFT effectively mitigates the representational discrep- ancies between source queries and multilingual documents across architectures and model scales. ModelBelebeleMLQAXQuADMultiEup-v2AVG embeddinggemma0.9220.7060.9660.3180.728 + SHIFT 0.9260.7360.9710.3710.751 mE5-large0.7980.3800.8860.0950.540 + SHIFT0.8980.6540.9560.2680.694 bge-m30.8640.6370.9420.2930.684 + SHIFT0.8680.6430.9450.3000.689 Qwen30.8800.6500.9330.3020.691 + SHIFT0.8850.6510.9340.3220.698 llama-nemotron0.6670.5540.9510.1230.574 + SHIFT0.6830.5530.9530.1560.586 gte-Qwen20.8930.6290.9390.3030.691 + SHIFT0.9080.6830.9510.3780.730 Table 3: Comparison of TLR@20 across different mod- els and datasets. Mitigation of Language Bias To verify that the performance improvement stems from increased exposure of relevant target-language documents rather than merely higher rankings of source- language documents, we analyze the TLR@20 met- ric in Table 3. Notably, standard metrics such as Recall can mask underlying language bias, as a model may achieve high scores solely by retrieving documents in the source language. For instance, on MultiEup-v2, multilingual-e5-large attains a 5 reasonable Recall@20 of 0.302 in Table 2, yet its TLR@20 is only 0.095 in Table 3. This gap suggests that the model predominantly retrieves relevant documents in the source language, under- exposing relevant target-language documents. Ap- plying SHIFT alleviates this issue, yielding a sub- stantial increase in TLR@20 across all models. This confirms that SHIFT bridges the representation gap between the query and document languages, effectively promoting relevant target-language doc- uments to the top ranks. We provide additional comparisons to representative debiasing baselines in Appendix C. 5.2 Language Distribution Analysis To complement the TLR analysis, Figure 2 visual- izes the language distribution of the top-20 docu- ments retrieved by multilingual-e5-large and gte- Qwen2-1.5B-instruct. The heatmap compares the density of retrieved documents by language before and after applying SHIFT, and thereby provides a qualitative view of the language composition and diversity of the retrieved results. Our experimental results demonstrate that exist- ing baseline models have a certain language pref- erence. Given an English source query, the mod- els assign disproportionately high similarity scores to English documents, filling the top ranks with English content. Conversely, documents in target languages are ranked lower, appearing as inactive or very faint regions on the heatmap. This indicates that the modelâs similarity space favors certain lan- guage regions even when relevant documents exist more broadly in target languages. After applying SHIFT, this concentration be- comes noticeably less pronounced. The density previously concentrated in English becomes more evenly dispersed across languages, and previously inactive regions become clearly visible. This sug- gests that SHIFT encourages stronger cross-lingual alignment between the source-language query and target-language documents, enabling the retriever to return a markedly more balanced language com- position among the top-ranked results. Impor- tantly, this qualitative trend is consistent with the TLR@20 improvements in Table 3, suggesting that higher TLR@20 coincides with reduced query- language dominance and increased presence of other languages in the top ranks. Figure 2: Visualization of language distribution in the top-20 retrieved documents for English queries. The heatmap intensity represents the density of retrieved doc- uments for each language, where darker shades indicate a higher concentration of documents. Distributions for additional models and data are provided in Appendix D. 5.3SHIFT in Multi-Source Languages To demonstrate the universality of SHIFT and verify that it functions effectively regardless of the source language, we conduct additional experiments using three different languages as new source queries. Experimental Setup We select Chinese (zh), Vietnamese (vi), and Hindi (hi) as non-English source (query) languages. These languages overlap with mMARCO and our evaluation benchmarks and cover distinct scripts and character distribu- tions. For each run, we treat the selected query lan- guage as the source languageâ src and re-estimate the relative language vectorsV â tgt using Eq. 1 for all target languagesâ tgt âL tgt . Subsequently, dur- ing indexing (as in Section 3.2), all documents are shifted into the embedding space of theâ src . For evaluation, we reprocess Belebele, MLQA, and XQuAD to construct MLIR test sets for eachâ src . We provide the constructed dataset statistics in Ap- pendix E. We report nDCG@20 and TLR@20 to measure overall retrieval effectiveness and target- languages coverage. Results Table 4 presents retrieval performance when non-English languages are set as the source query. The results show common performance im- provements across all source languages and models 6 QueryDataset multilingual-e5-largebge-m3llama-nemotron-embed-1b-v2gte-Qwen2-1.5B-instruct nDCG@20TLR@20nDCG@20TLR@20nDCG@20TLR@20nDCG@20TLR@20 BaseSHIFTBaseSHIFTBaseSHIFTBaseSHIFTBaseSHIFTBaseSHIFTBaseSHIFTBaseSHIFT ZH Belebele0.3210.8250.1950.8020.8240.8310.8130.8180.5970.6610.5720.6540.8750.8840.8670.878 MLQA0.3170.5490.0590.4830.5990.6070.6370.6530.4830.5190.4870.5520.5980.6230.6330.673 XQuAD0.4570.8800.3830.9070.8900.8980.9260.9310.8700.8890.9170.9300.9030.9160.9350.942 VI Belebele0.5960.8660.5510.8530.8360.8390.8280.8310.4610.6050.4210.5810.6580.8510.6290.842 MLQA0.4160.6300.2720.6740.6060.6100.6520.6630.4270.4940.4020.5350.3760.5430.2590.572 XQuAD0.6890.9110.7170.9410.8900.8960.9200.9260.7940.8700.8320.9050.6730.8660.7020.897 HI Belebele0.5340.8080.4820.7890.7750.7800.7640.7710.6290.6680.6110.6570.5120.6430.4660.613 MLQA0.4140.6140.2750.6460.5910.5950.6410.6490.4930.5280.5220.5880.3520.4580.2620.483 XQuAD0.6970.9020.7220.9270.8820.8870.9080.9130.8590.8770.9040.9210.6810.7990.7040.838 Table 4: Top-20 retrieval performance (nDCG@20, TLR@20) using non-English languages (Chinese, Vietnamese, Hindi) as the source query language. upon applying SHIFT, proving that our method is not limited to a specific source language. The most notable change is observed in the multilingual-e5- large model; it records significant increases in both nDCG@20 and TLR@20 compared to the base model, demonstrating the methodâs effectiveness. Another noteworthy observation is the bge-m3 model. As shown in Table 2, when the source language is English, the baseline performance is already strong and the gains from SHIFT are marginal. However, when non-English languages serve as source languages, applying SHIFT substan- tially improves TLR@20 on MLQA. This suggests that while state-of-the-art models like bge-m3 are strongly aligned around English, their alignment is relatively less consistent when the source language is non-English. Consequently, SHIFT effectively corrects these non-English-centric alignment dis- crepancies, demonstrating robustness across dif- ferent source languages. To reflect more realistic scenarios where users query in various languages, we provide additional experiments on Multilingual- to-Multilingual Information Retrieval (M2MIR) in Appendix J. 5.4 Analysis of Scale Factor α The scale factorαis a key hyperparameter in SHIFT that controls the strength of the language-vector subtraction. Varyingαadjusts the degree of debias- ing applied to the embedding space, and an overly aggressive subtraction may lead to over-correction for some models. To study this effect, we conduct a grid search overαfrom 0.1 to 1.0 and report the average nDCG@20 across all datasets. Figure 3 illustrates the impact ofαon retrieval performance. While the optimalαcan vary across models, multilingual-e5-large exhibits an inverted- U-shaped curve with a peak atα = 0.6. In con- 0.10.20.30.40.50.60.70.80.91.0 Alpha 0.60 0.62 0.64 0.66 0.68 0.70 0.72 0.74 0.76 Average NDCG@20 gte-Qwen2-1.5B-instruct Qwen3-Embedding-0.6B llama-nemotron-embed-1b-v2 bge-m3 multilingual-e5-large embeddinggemma-300m Base Figure 3: Impact of the scale factorαon the average nDCG@20 score. trast, all other models show a monotonic increase, reaching their maximum performance atα = 1.0. Crucially, SHIFT consistently outperforms the base- line across the entire range ofα. This demonstrates that our method effectively mitigates language bias without hinging on a narrowly tuned hyperparame- ter, although selecting a model-specificαcan fur- ther maximize the performance gains. 5.5 Sensitivity Analysis on SHIFT Magnitude While Section 5.4 explores the global impact ofα across all languages, this section investigates its local impact on individual target languages and the resulting trade-off with overall system stability. We conduct a fine-grained sensitivity analysis by varying theαvalue for a specific target language, while fixing theαvalues for other languages to their global optimal settings identified previously. Evaluation Setup To analyze target-language sensitivity to the scale factorα, we conduct ex- periments on two MLIR benchmarks, Belebele and XQuAD, using multilingual-e5-large with En- 7 0.20.40.60.81.0 Alpha 0.700 0.750 0.800 0.850 0.900 0.950 1.000 Target Recall (ZH) (a) Chinese (zh) Target Recall (ZH) Overall Recall 0.20.40.60.81.0 Alpha 0.825 0.850 0.875 0.900 0.925 0.950 0.975 1.000 Target Recall (HI) (b) Hindi (hi) Target Recall (HI) Overall Recall 0.20.40.60.81.0 Alpha 0.750 0.800 0.850 0.900 0.950 Target Recall (VI) (c) Vietnamese (vi) Target Recall (VI) Overall Recall 0.910 0.915 0.920 0.925 0.930 0.935 0.940 Overall Recall 0.920 0.922 0.925 0.927 0.930 0.932 0.935 0.938 0.940 Overall Recall 0.915 0.920 0.925 0.930 0.935 0.940 Overall Recall Figure 4: Sensitivity analysis of the scale factorαfor specific target languages (Chinese, Hindi, Vietnamese). The solid red line represents performance on the specific target language (Target Recall@20), while the dashed blue line represents the overall system performance (Overall Recall@20). glish as the fixed source (query) language. Based on Figure 3, we set the global optimal scale for multilingual-e5-large toα = 0.6. We experiment on three target languages (zh, hi, vi) and perform a per-language ablation: we varyαfor the target language while fixingα = 0.6for all other lan- guages. We report Overall Recall@20, computed over all relevant documents regardless of language, together with Target Recall@20 for each target language. Specifically, unlike TLR@20 (which aggregates over all non-query languages), Target Recall@20 recomputes recall by restricting the ground-truth set to relevant documents in a single target language and treating relevant documents in other languages as non-relevant. All metrics are averaged over the two benchmarks to assess generalization. Results Figure 4 illustrates the impact ofαon retrieval performance. Regardless of the target lan- guage, we observe two consistent trends. First, Target Recall@20 monotonically improves asα increases. This confirms the high precision of the directionality of the relative language vectors es- timated via mMARCO. Geometrically, if the rela- tive language vector fails to capture the languageâs characteristics properly, increasingαwould move document embeddings away from the query lan- guage space, causing a sharp performance drop. Therefore, the continuous rise in target language performance indicates that the relative language vectors extracted by our method form a correct tra- jectory from the target language space to the source language space. Second, Overall Recall@20 exhibits an inverted- U trend, decreasing beyond a certainα. This sug- gests over-shifting: pushing a single target lan- guage too aggressively can expose documents of that language in high ranks, but it simultaneously pushes down relevant documents in other non- target languages from the top ranks. In other words, optimizingαfor one language does not necessarily optimize overall retrieval across languages. Ad- ditional results for multilingual-e5-large on other datasets are reported in Appendix H, and results for additional retrievers are presented in Appendix I. Controllability and Robustness Furthermore, these results suggest thatαcan serve as a control- lable parameter in real-world environments. As observed, increasingαcan explicitly boost the ex- posure of documents in a specific target language, even if it slightly compromises the overall ranking balance. This offers a practical advantage when dy- namic adjustment of result composition is needed based on user language preferences or regional characteristics. In essence, SHIFT becomes an ef- fective tool for managing the trade-off between semantic alignment and language preference. 6 Conclusion In this work, we demonstrate that state-of-the-art multilingual dense retrieval models suffer from severe language bias, disproportionately favoring query-language documents. To address this, we propose SHIFT, a training-free, index-side trans- formation that aligns document embeddings by subtracting a relative language vector estimated from parallel data. Furthermore, we introduce Target-Languages Recall@k (TLR@k) to rigor- ously quantify this language bias. Extensive ex- periments across diverse benchmarks and multi- ple model architectures confirm that SHIFT signifi- cantly improves MLIR performance, successfully redirecting the focus from linguistic surface forms to underlying semantics. 8 Limitations Despite its effectiveness, our approach has certain limitations. First, the estimation of language vectors relies on the mMARCO dataset, which contains machine- translated texts. Consequently, the quality of the shift vector may be influenced by the translation quality of the parallel corpus. Second, SHIFT requires the target documents to be pre-labeled with their language to apply the correct vector subtraction during indexing. Future work will focus on relaxing this dependency on explicit language labels and exploring purely unsu- pervised methods for language vector estimation. Finally, SHIFT requires re-estimating the relative language vectors whenever the source language or the embedding model changes. In practice, how- ever, this cost is incurred only once per setting: once computed, the vectors can be cached and reused across indexing runs and datasets as long as the same model and source language are main- tained. Ethics Statement This study utilizes publicly available datasets and open-source libraries, specifically sentence- transformers, for all experiments. We adhered to the licenses and terms of use associated with these resources. regarding the use of AI tools, we utilized AI assistants (e.g., ChatGPT, Gemini) exclusively for grammatical error correction and polishing the text to enhance readability. The scientific ideas and contributions presented in this paper are entirely our own. References Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of mono- lingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguis- tics. Sunil Arya, David M Mount, Nathan S Netanyahu, Ruth Silverman, and Angela Y Wu. 1998. An optimal algorithm for approximate nearest neighbor search- ing fixed dimensions. Journal of the ACM (JACM), 45(6):891â923. Yauhen Babakhin, Radek Osmulski, Ronay Ak, Gabriel Moreira, Mengyao Xu, Benedikt Schifferer, Bo Liu, and Even Oldridge. 2025. Llama-embed- nemotron-8b: A universal text embedding model for multilingual and cross-lingual tasks. Preprint, arXiv:2511.07025. Lucas Bandarkar, Davis Liang, Benjamin Muller, Mikel Artetxe, Satya Narayan Shukla, Donald Husa, Naman Goyal, Abhinandan Krishnan, Luke Zettlemoyer, and Madian Khabsa. 2024. The belebele benchmark: a parallel reading comprehension dataset in 122 lan- guage variants. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 749â775, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics. Rexhina Blloshmi, Tommaso Pasini, NiccolĂČ Campol- ungo, Somnath Banerjee, Roberto Navigli, Gabriella Pasi, and 1 others. 2021. Ir like a sir: Sense-enhanced information retrieval for multiple languages. In Pro- ceedings of the 2021 Conference on Empirical Meth- ods in Natural Language, pages 1030â1041. Luiz Bonifacio, Vitor Jeronymo, Hugo Queiroz Abonizio, Israel Campiotti, Marzieh Fadaee, Roberto Lotufo, and Rodrigo Nogueira. 2022. mmarco: A multilingual version of the ms marco passage ranking dataset. Preprint, arXiv:2108.13897. Martin Braschler. 2002a. Clef 2001 â overview of re- sults. In Evaluation of Cross-Language Information Retrieval Systems, pages 9â26, Berlin, Heidelberg. Springer Berlin Heidelberg. Martin Braschler. 2002b. Clef 2002âoverview of re- sults. In Workshop of the Cross-Language Evalu- ation Forum for European Languages, pages 9â27. Springer. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. Preprint, arXiv:2402.03216. Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024b.M3- embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self- knowledge distillation. In Findings of the Asso- ciation for Computational Linguistics: ACL 2024, pages 2318â2335, Bangkok, Thailand. Association for Computational Linguistics. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco GuzmĂĄn, Edouard Grave, Myle Ott, Luke Zettle- moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Pro- ceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 8440â 8451, Online. Association for Computational Lin- guistics. Kareem Darwish and Douglas W Oard. 2003. Proba- bilistic structured query methods. In Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, pages 338â344. 9 Najim Dehak, Reda Dehak, James R Glass, Douglas A Reynolds, Patrick Kenny, and 1 others. 2010. Co- sine similarity scoring without score normalization techniques. In Odyssey, volume 15. Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, MĂĄrton Kardos, Ashwin Mathur, David Stap, Jay Gala, Wissam Siblini, Dominik Krzemi Ì nski, Genta Indra Winata, Saba Sturua, Saiteja Utpala, Mathieu Ciancone, Marion Schaeffer, Gabriel Se- queira, Diganta Misra, Shreeya Dhakal, Jonathan RystrĂžm, Roman Solomatin, and 67 others. 2025. Mmteb: Massive multilingual text embedding bench- mark. Preprint, arXiv:2502.13595. Koustava Goswami, Sourav Dutta, Haytham Assem, Theodorus Fransen, and John Philip McCrae. 2021. Cross-lingual sentence embedding using multi-task learning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9099â9113. Ximo Granell. 2014. Multilingual information man- agement: Information, technology and translators. Chandos Publishing. Seongtae Hong, Youngjoon Jang, Hyeonseok Moon, Jungseob Lee, and Heuiseok Lim. 2026. Improv- ing semantic proximity in english-centric informa- tion retrieval through cross-lingual alignment. In The Fourteenth International Conference on Learn- ing Representations. Xiyang Hu, Xinchi Chen, Peng Qi, Deguang Kong, Kun- lun Liu, William Yang Wang, and Zhiheng Huang. 2023. Language agnostic multilingual information retrieval with contrastive learning. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9133â9146. Zhiqi Huang, Hansi Zeng, Hamed Zamani, and James Allan. 2023. Soft prompt decoding for multilingual dense retrieval. In Proceedings of the 46th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1208â 1218. David A Hull and Gregory Grefenstette. 1996. Query- ing across languages: A dictionary-based approach to multilingual information retrieval. In Proceedings of the 19th annual international ACM SIGIR confer- ence on Research and development in information retrieval, pages 49â57. Mei Kobayashi and Koichi Takeda. 2000. Informa- tion retrieval on the web. ACM computing surveys (CSUR), 32(2):144â173. Wessel Kraaij, Jian-Yun Nie, and Michel Simard. 2003. Embedding web-based statistical translation models in cross-language information retrieval. Computa- tional Linguistics, 29(3):381â419. Dawn Lawrie, Sean MacAvaney, James Mayfield, Paul McNamee, Douglas W. Oard, Luca Soldaini, and Eugene Yang. 2025a. Overview of the trec 2024 neuclir track. Preprint, arXiv:2509.14355. Dawn Lawrie, James Mayfield, Douglas W Oard, and Eugene Yang. 2022. Hc4: A new suite of test col- lections for ad hoc clir. In European Conference on Information Retrieval, pages 351â366. Springer. Dawn Lawrie, James Mayfield, Eugene Yang, Andrew Yates, Sean MacAvaney, Ronak Pradeep, Scott Miller, Paul McNamee, and Luca Soldani. 2025b. Neuclir- bench: A modern evaluation collection for monolin- gual, cross-language, and multilingual information retrieval. Preprint, arXiv:2511.14758. Dawn Lawrie, Eugene Yang, Douglas W Oard, and James Mayfield. 2023. Neural approaches to multilin- gual information retrieval. In European Conference on Information Retrieval, pages 521â536. Springer. Jinhyuk Lee, Feiyang Chen, Sahil Dua, Daniel Cer, Madhuri Shanbhogue, Iftekhar Naim, Gus- tavo HernĂĄndez Ăbrego, Zhe Li, Kaifeng Chen, Hen- rique Schechter Vera, Xiaoqi Ren, Shanfeng Zhang, Daniel Salz, Michael Boratko, Jay Han, Blair Chen, Shuo Huang, Vikram Rao, Paul Suganthan, and 28 others. 2025. Gemini embedding: Generalizable em- beddings from gemini. Preprint, arXiv:2503.07891. Patrick Lewis, Barlas O Ì guz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2019. Mlqa: Eval- uating cross-lingual extractive question answering. arXiv preprint arXiv:1910.07475. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich KĂŒttler, Mike Lewis, Wen-tau Yih, Tim Rock- tĂ€schel, and 1 others. 2020. Retrieval-augmented gen- eration for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459â 9474. Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281. Jind Ë rich LibovickĂœ, Rudolf Rosa, and Alexander Fraser. 2020. On the language neutrality of pre-trained mul- tilingual representations. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2020, pages 1663â1674, Online. Association for Computa- tional Linguistics. Ting Liu, Andrew Moore, Ke Yang, and Alexander Gray. 2004. An investigation of practical approximate near- est neighbor algorithms. Advances in neural infor- mation processing systems, 17. Walid Magdy and Gareth JF Jones. 2011. Should mt systems be used as black boxes in clir? In European Conference on Information Retrieval, pages 683â686. Springer. Paul McNamee and James Mayfield. 2002. Comparing cross-language query expansion techniques by de- grading translation resources. In Proceedings of the 25th annual international ACM SIGIR conference on Research and development in information retrieval, pages 159â166. 10 Teruko Mitamura, Eric Nyberg, Hideki Shima, Tsuneaki Kato, Tatsunori Mori, Chin-Yew Lin, Ruihua Song, Chuan-Jie Lin, Tetsuya Sakai, Donghong Ji, and 1 others. 2008. Overview of the ntcir-7 aclia tasks: Ad- vanced cross-lingual information access. In NTCIR. Jeonghyun Park and Hwanhee Lee. 2025. Investigat- ing language preference of multilingual rag systems. Preprint, arXiv:2502.11175. Carol Peters and Martin Braschler. 2002a. The impor- tance of evaluation for cross-language system devel- opment: the clef experience. In LREC. Carol Peters and Martin Braschler. 2002b. The impor- tance of evaluation for cross-language system devel- opment: the clef experience. In LREC. Carol Peters, Martin Braschler, and Paul Clough. 2012. Multilingual information retrieval: From research to practice. Springer. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natu- ral Language Processing, pages 2383â2392, Austin, Texas. Association for Computational Linguistics. Uma Roy, Noah Constant, Rami Al-Rfou, Aditya Barua, Aaron Phillips, and Yinfei Yang. 2020. LAReQA: Language-agnostic answer retrieval from a multilin- gual pool. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5919â5930, Online. Association for Computational Linguistics. Nikhil Sharma, Kenton Murray, and Ziang Xiao. 2025. Faux polyglot: A study on information disparity in multilingual large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 1: Long Papers), pages 8090â8107, Albuquerque, New Mexico. Association for Computational Linguistics. Luo Si, Jamie Callan, Suleyman Cetintas, and Hao Yuan. 2008. An effective and efficient results merging strat- egy for multilingual information retrieval in federated search environments. Information Retrieval, 11(1):1â 24. Amit Singhal and 1 others. 2001. Modern information retrieval: A brief overview. IEEE Data Eng. Bull., 24(4):35â43. NLLB Team, Marta R. Costa-jussĂ , James Cross, Onur Ăelebi, Maha Elbayad, Kenneth Heafield, Kevin Hef- fernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scal- ing human-centered machine translation. Preprint, arXiv:2207.04672. Ming-Feng Tsai, Yu-Ting Wang, and Hsin-Hsi Chen. 2008. A study of learning a merge model for multi- lingual information retrieval. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 195â202. Henrique Schechter Vera, Sahil Dua, Biao Zhang, Daniel Salz, Ryan Mullins, Sindhu Raghuram Pa- nyam, Sara Smoot, Iftekhar Naim, Joe Zou, Feiyang Chen, and 1 others. 2025. Embeddinggemma: Pow- erful and lightweight text representations. arXiv preprint arXiv:2509.20354. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. Preprint, arXiv:2402.05672. Eugene Yang, Thomas JĂ€nich, James Mayfield, and Dawn Lawrie. 2024a. Language fairness in multi- lingual information retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval, pages 2487â2491. Eugene Yang, Dawn Lawrie, and James Mayfield. 2024b. Distillation for multilingual information re- trieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, page 2368â2373. ACM. Eugene Yang, Dawn Lawrie, Orion Weller, and James Mayfield. 2025. Hltcoe at trec 2024 neuclir track. Preprint, arXiv:2510.00143. Jinrui Yang, Fan Jiang, and Timothy Baldwin. 2024c. Language bias in multilingual information retrieval: The nature of the beast and mitigation methods. In Proceedings of the Fourth Workshop on Multilingual Representation Learning (MRL 2024), pages 280â 292. Ziyi Yang, Yinfei Yang, Daniel Cer, and Eric Darve. 2021. A simple and effective method to eliminate the self language bias in multilingual representations. In Proceedings of the 2021 Conference on Empiri- cal Methods in Natural Language Processing, pages 5825â5832, Online and Punta Cana, Dominican Re- public. Association for Computational Linguistics. Puxuan Yu, Luke Merrick, Gaurav Nuti, and Daniel Campos. 2024.Arctic-embed 2.0:Multilin- gual retrieval without compromise.Preprint, arXiv:2412.04506. Xinyu Zhang, Xueguang Ma, Peng Shi, and Jimmy Lin. 2021. Mr. TyDi: A multi-lingual benchmark for dense retrieval. In Proceedings of the 1st Workshop on Multilingual Representation Learning, pages 127â 137, Punta Cana, Dominican Republic. Association for Computational Linguistics. 11 Xinyu Zhang, Nandan Thakur, Odunayo Ogundepo, Ehsan Kamalloo, David Alfonso-Hermelo, Xi- aoguang Li, Qun Liu, Mehdi Rezagholizadeh, and Jimmy Lin. 2023. MIRACL: A multilingual retrieval dataset covering 18 diverse languages. Transactions of the Association for Computational Linguistics, 11:1114â1131. Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176. 12 A Evaluation Dataset Details In this work, we leverage multilingual Question An- swering (QA) datasets with parallel constructions, repurposed as retrieval tasks, to systematically as- sess MLIR performance. Since these datasets are originally designed for QA, the corresponding pas- sages serve as exact positive documents within the retrieval framework. Consequently, datasets devel- oped for QA are widely adopted for the evaluation of retrieval models in current literature (Enevoldsen et al., 2025; Lee et al., 2025; Zhang et al., 2025). BelebeleBelebele is a multilingual machine read- ing comprehension dataset spanning 122 language variants. Based on the FLORES-200 benchmark, each instance consists of a short passage and a corresponding multiple-choice question. For the retrieval task, we utilize the question stem as the query and the passage as the positive document. A defining feature of Belebele is its full parallelism: the entire set of questions and passages represents strictly parallel content across all 122 languages. For our evaluation, we select the 14 languages that overlap with the mMARCO dataset: Arabic (ar), German (de), Dutch (nl), English (en), Spanish (es), French (fr), Hindi (hi), Indonesian (id), Italian (it), Japanese (ja), Portuguese (pt), Russian (ru), Vietnamese (vi), and Chinese (zh). XQuAD XQuAD is a cross-lingual QA bench- mark consisting of a subset of the SQuAD v1.1 development set. The dataset was constructed by professionally translating both the questions and the paragraphs (contexts) from English into 10 tar- get languages. This rigorous human translation process guarantees that XQuAD is fully parallel; every question (query) and paragraph (document) has an exact, semantically aligned counterpart in all other languages. We utilize the 8 languages inter- secting with mMARCO for evaluation: Arabic (ar), German (de), English (en), Spanish (es), Hindi (hi), Russian (ru), Vietnamese (vi), and Chinese (zh). MLQA MLQA is a multi-way aligned extrac- tive QA benchmark covering seven languages. Un- like Belebele or XQuAD, MLQA is designed to be multi-way parallel; each QA instance is aligned across a subset of languages (typically four) rather than the entire set. This structure results from min- ing parallel sentences where overlaps exist across varying language combinations to maximize lin- guistic diversity. Consequently, queries are also not fully parallel across all seven languages for every instance. Instead, for a given document, queries ex- ist only in the specific subset of languages aligned with that document. Despite this partial parallelism, the dataset ensures that within each 4-way subset, the questions and contexts are semantically equiva- lent. Since all 7 languages overlap with mMARCO, we utilize the complete language set for evaluation: Arabic (ar), German (de), English (en), Spanish (es), Hindi (hi), Vietnamese (vi), and Simplified Chinese (zh). MultiEup-v2Following Yang et al. (2024c), we utilize MultiEup-v2 as an MLIR benchmark. The task is defined as retrieving the relevant parliamen- tary speech segment (document) given a subject descriptor (query). The dataset is constructed from European Parliament proceedings, where profes- sional archivists manually assign official subject descriptors (e.g., âInternational Human Rightsâ) to specific speech segments. Since the descriptors are professionally translated into 24 languages, the same information can be expressed as multilingual queries. In this framework, the ground truth is the specific speech text that was originally tagged with the descriptor by human experts. We use the 6 lan- guages that overlap with mMARCO: German (de), English (en), Spanish (es), French (fr), Italian (it), and Portuguese (pt). We note that Belebele, XQuAD, and MLQA are utilized in MMTEB (Enevoldsen et al., 2025), while MultiEup-v2 is adopted to evaluate MLIR tasks following Yang et al. (2024c). We do not include Mr. TyDi (Zhang et al., 2021) or MIR- ACL (Zhang et al., 2023) in our evaluation, since their relevance annotations are not explicitly la- beled across languages and the datasets are neither parallel nor aligned at the document level, mak- ing them unsuitable for measuring retrieval perfor- mance under our MLIR setup. B Implementation Details ModelDetails Weutilizethe sentence-transformers 1 libraryfor inference. For bge-m3, we exclusively employ the dense embedding component. The maximum sequence length is set to 512 tokens for all models. To ensure optimal performance, we strictly adhere to the specific prefixes and instruction prompts required by each retrieval model. Language vectors are computed from unnormalized embeddings, and 1 https://github.com/huggingface/ sentence-transformers 13 Modelquery prefixdocument prefix# of trained_langs embeddinggemma-300mtask: search result | query: querytitle: none | text: document100+ multilingual-e5-largequery: querypassage: document94 bge-m3--100+ Qwen3-Embedding-0.6B Instruct: Given a web search query, retrieve relevant passages that answer the query : query -100+ llama-nemotron-embed-1b-v2query: querypassage: document26 gte-Qwen2-1.5B-instruct Instruct: Given a web search query, retrieve relevant passages that answer the query : query -33 Table 5: Comparison of Multilingual Embedding Models during evaluation, we apply language shifts to the unnormalized embeddings before computing cosine similarity which internally normalizes the vectors. We provide the details in Table 5. Hardware Details We conducted our experi- ments using an Intel Xeon Gold 6230R @2.10GHz CPU, 376GB RAM, and 8 NVIDIA RTX A6000 48GB GPUs. The software environment included nvidia driver, CUDA, and PyTorch, running on Ubuntu 20.04.6 LTS. C Comparison with Baselines To situate SHIFT among prior training-free ap- proaches for mitigating language bias, we com- pare it with two representative post-hoc base- lines on three retrievers: embeddinggemma-300m, multilingual-e5-large, and bge-m3. Table 6 reports nDCG@20 and TLR@20 on four MLIR bench- marks. C.1 Baselines Language-wisecentering(Centering) Li- bovickĂœ et al. (2020) propose an unsupervised centering procedure motivated by the view that multilingual representations contain both language- specific and language-neutral components. Let x âR d denote thed-dimensional embedding produced by the retriever for an input text (a query or a document). Given a languageâ, the method estimates a language centroid as the mean embedding of texts in the language, ÎŒ â = 1 N â N â X i=1 x (â) i and applies language-wise re-centering by subtract- ing the centroid from each embedding: Ì x = xâÎŒ â We use this centering variant as a training-free base- line and apply the same transformation to both queries and documents based on their language. Language Information Removal (LIR) Yang et al. (2021) introduce Language Information Re- moval (LIR), a post-training method that sup- presses language identification information by re- moving dominant language directions in the rep- resentation space. For each languageâ, LIR con- structs an embedding matrixX â âR N â Ăd by stack- ing embeddings of texts written inâ. It then com- putes the Singular Value Decomposition (SVD), X â = U â ÎŁ â V †â and takes the top-rright-singular vectorsV â,r â R dĂr as principal directions associated with lan- guage identity. Each embedding is updated by sub- tracting its projection onto the subspace spanned by these directions: Ì x = xâ V â,r V †â,r x Equivalently, this removes components along the top directions, Ì x = xâ P r j=1 (v †j x)v j , wherev j is the j-th column of V â,r . C.2 Implementation For a controlled comparison, we compute all language-dependent statistics required by the base- lines from mMARCO using the same retriever un- der evaluation. Centering For each model and each languageâ, we embed mMARCO texts written inâand esti- mate the centroidÎŒ â as the mean of their embed- dings. At indexing time, each document embed- ding is transformed asx 7â xâÎŒ â . At retrieval time, we apply the same transformation to the query embedding before nearest-neighbor search. 14 Model BelebeleMLQAXQuADMultiEup-v2AVG nDCGTLRnDCGTLRnDCGTLRnDCGTLRnDCGTLR embeddinggemma-300m (base)0.9240.9220.6640.7060.9470.9660.4950.3180.7580.728 Centering (LibovickĂœ et al., 2020) 0.9240.9240.6690.7270.9470.9670.5080.3630.7620.745 LIR (Yang et al., 2021)0.9260.9250.6710.7280.9510.9700.5090.3570.7640.745 SHIFT0.9260.9260.6710.7360.9510.9710.5110.3710.7650.751 multilingual-e5-large (base)0.8160.7980.4940.3800.8550.8860.3670.0950.6330.540 Centering (LibovickĂœ et al., 2020) 0.8890.8770.6220.6240.9300.9420.4430.1960.7210.660 LIR (Yang et al., 2021)0.8910.8800.6230.6250.9310.9430.4440.2610.7220.677 SHIFT0.9130.8980.6490.6540.9440.9560.4420.2680.7370.694 bge-m3 (base)0.8740.8640.6060.6370.9230.9420.4550.2930.7150.684 Centering (LibovickĂœ et al., 2020)0.8490.8380.5980.6260.9120.9310.4540.2950.7030.673 LIR (Yang et al., 2021)0.8620.8530.6050.6370.9200.9390.4580.3000.7110.682 SHIFT0.8760.8680.6090.6430.9250.9450.4600.3000.7170.689 Table 6: Top-20 retrieval performance (nDCG@20, TLR@20) of base models, LibovickĂœ et al. (2020), Yang et al. (2021), and SHIFT across four MLIR benchmarks, evaluated with English as the source query language. LIR For each model and each languageâ, we construct the embedding matrixX â by stacking em- beddings of mMARCO texts inâ, and compute its SVD to obtainV â,r . Following Yang et al. (2021), which reports that removing only the top compo- nent performs best in their setting, we setr = 1 throughout. At indexing time, each document em- bedding is transformed asx 7â xâ V â,r V †â,r x , and we apply the same transformation to the query embedding at retrieval time. C.3 Results & Analysis Results Table 6 shows that SHIFT yields the strongest average performance across the three retrievers when considering both nDCG@20 and TLR@20. The difference is most pronounced for multilingual-e5-large, where the base model ex- hibits low TLR (AVG TLR@20 of 0.540). Both centering and LIR increase TLR@20 substantially, while SHIFT achieves the highest average TLR@20 (0.694) and also improves average nDCG@20. For embeddinggemma-300m, all methods yield modest but consistent improvements, with SHIFT attain- ing the best averages. For bge-m3, the improve- ments are smaller and both baselines are slightly worse than the base model on average, suggest- ing that removing global language components is not uniformly beneficial when the embedding space is already relatively well aligned. This be- havior is consistent with prior observations that removing dominant components may also remove semantic information (Huang et al., 2023). Over- all, the comparison indicates that SHIFT provides consistent gains across models, while maintaining strong target-language exposure as measured by TLR@20. Inference-time practicalityWhile all compared methods are training-free, the baselines require ad- ditional query-time transformations: centering sub- tracts a language-dependent centroid, and LIR re- moves language identification components via an additional projection step. In contrast, SHIFT can be applied to document embeddings offline dur- ing indexing, and the query-side operation is fixed given a source language. As a result, SHIFT re- duces inference-time overhead and is simpler to deploy in latency-sensitive retrieval systems. D Additional Language Distribution Analysis Figure 5 provides a comprehensive visualization of language distribution heatmaps for all evalu- ated models across the four benchmarks. Consis- tent with the findings in Section 5.2, the baseline results (left panels) universally exhibit a strong English-centric bias, regardless of the model archi- tecture. Whether using encoder-based models like embeddinggemma-300m or decoder-based models like gte-Qwen2-1.5B-instruct, the retrieval density is disproportionately concentrated in the âenâ col- umn, leaving other target languages marginalized. However, applying SHIFT (right panels) alleviates this skew across all architectures. The density, pre- viously locked on English, effectively disperses to the relevant target languages, resulting in a more semantically balanced distribution. This visual evi- dence further corroborates the quantitative improve- ments reported in Table 3, confirming that SHIFT 15 ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.030.040.050.410.050.050.060.050.050.030.050.040.040.04 Base ar de dt en es fr hi id it ja pt ru vi zh 0.050.060.070.120.070.070.070.060.070.060.070.060.060.12 SHIFT ar de en es hi ru vi zh XQuAD 0.050.090.480.100.110.060.070.05 ar de en es hi ru vi zh 0.070.120.190.140.140.090.130.12 ar de vi en es hi zh MLQA 0.010.020.020.910.020.020.00 ar de vi en es hi zh 0.040.060.100.510.100.100.10 de en es fr it pt MultiEup 0.010.940.010.020.010.01 de en es fr it pt 0.080.670.060.080.050.06 0.0 0.2 0.4 0.6 0.8 multilingual-e5-large ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.050.070.070.170.070.060.060.080.070.050.060.060.070.05 Base ar de dt en es fr hi id it ja pt ru vi zh 0.070.070.070.080.070.070.060.090.070.060.070.070.090.06 SHIFT ar de en es hi ru vi zh XQuAD 0.080.130.210.140.100.110.140.09 ar de en es hi ru vi zh 0.110.130.130.140.090.140.160.11 ar de vi en es hi zh MLQA 0.050.090.110.520.130.070.04 ar de vi en es hi zh 0.100.100.200.280.150.070.09 de en es fr it pt MultiEup 0.040.730.070.060.040.04 de en es fr it pt 0.080.530.110.110.080.08 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 gte-Qwen2-1.5B-instruct ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.080.060.080.120.060.070.080.060.070.060.070.070.060.06 Base ar de dt en es fr hi id it ja pt ru vi zh 0.080.070.070.060.070.070.070.070.070.070.070.070.080.07 SHIFT ar de en es hi ru vi zh XQuAD 0.110.130.190.120.120.110.110.11 ar de en es hi ru vi zh 0.110.130.130.140.120.120.130.12 ar de vi en es hi zh MLQA 0.100.080.080.450.090.090.09 ar de vi en es hi zh 0.130.110.150.250.130.110.12 de en es fr it pt MultiEup 0.070.710.060.080.040.04 de en es fr it pt 0.090.550.110.100.080.08 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 embeddinggemma-300m ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.060.070.070.090.070.070.070.080.070.060.070.060.070.08 Base ar de dt en es fr hi id it ja pt ru vi zh 0.060.070.060.080.080.080.070.080.070.060.080.070.070.07 SHIFT ar de en es hi ru vi zh XQuAD 0.080.120.160.130.130.100.150.13 ar de en es hi ru vi zh 0.090.130.150.140.120.110.140.12 ar de vi en es hi zh MLQA 0.060.080.140.340.170.100.11 ar de vi en es hi zh 0.080.080.130.320.200.090.10 de en es fr it pt MultiEup 0.070.630.080.100.060.06 de en es fr it pt 0.070.620.090.100.070.06 0.0 0.1 0.2 0.3 0.4 0.5 0.6 bge-m3 ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.090.070.060.100.070.050.090.040.080.070.060.050.070.09 Base ar de dt en es fr hi id it ja pt ru vi zh 0.080.070.070.100.070.070.070.060.070.070.070.060.070.08 SHIFT ar de en es hi ru vi zh XQuAD 0.110.130.160.140.110.090.130.14 ar de en es hi ru vi zh 0.100.130.160.140.090.110.130.13 ar de vi en es hi zh MLQA 0.120.090.120.320.120.110.13 ar de vi en es hi zh 0.110.100.120.320.130.090.12 de en es fr it pt MultiEup 0.100.650.090.060.060.05 de en es fr it pt 0.090.600.100.090.070.06 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Qwen3-Embedding-0.6B ar de dt en es fr hi id it ja pt ru vi zh Belebele 0.050.060.060.160.100.080.080.050.060.050.070.050.060.08 Base ar de dt en es fr hi id it ja pt ru vi zh 0.040.070.080.130.080.090.050.060.070.060.070.060.070.07 SHIFT ar de en es hi ru vi zh XQuAD 0.100.110.180.130.120.110.120.13 ar de en es hi ru vi zh 0.090.130.180.130.100.120.130.13 ar de vi en es hi zh MLQA 0.100.080.090.400.140.110.08 ar de vi en es hi zh 0.080.100.100.430.130.080.09 de en es fr it pt MultiEup 0.030.840.050.030.020.02 de en es fr it pt 0.050.750.070.060.030.04 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 llama-nemotron-embed-1b-v2 Figure 5: Extended visualization of language distribution heatmaps for all evaluated models. serves as a robust and universal solution for recti- fying language bias, independent of the underlying model type. E Multi-Source Language Evaluation Dataset Table 7 summarizes the detailed statistics of the datasets configured for the non-English source 16 Source LanguageZHVIHI DatasetsBelebeleMLQAXQuADBelebeleMLQAXQuADBelebeleMLQAXQuAD # of Languages147814781478 Fully ParallelOXOOXOOXO # of Queries9005,1361,1909005,4951,1909004,9161,190 # of Documents6,83217,1651,9206,83218,1501,9206,83215,6391,920 # Avg. d + /q143.798143.758143.778 Table 7: Detailed statistics of the datasets used for the multi-source language evaluation (Section 5.3), where Chinese (ZH), Vietnamese (VI), and Hindi (HI) serve as the source query languages. query evaluation, where Chinese (ZH), Vietnamese (VI), and Hindi (HI) serve as the source query languages. As shown in the table, fully parallel datasets such as Belebele and XQuAD maintain consistent statistics across all source languages, in- cluding English in Table 1. In contrast, MLQA exhibits variations in the number of queries and documents depending on the source language. This is due to its partially parallel nature, where the availability of aligned query-passage pairs differs across language subsets. F Additional Evaluation on NeuCLIR2023 To further evaluate SHIFT beyond parallel or translation-derived benchmarks, we conduct an additional experiment on a NeuCLIRBench dataset (Lawrie et al., 2025b), specifically Neu- CLIR2023RetrievalHardNegatives 2 . Unlike our main experiments where relevance is constructed from semantically aligned multilingual passages, this benchmark represents a more open multilin- gual retrieval setting. The original task uses En- glish queries and retrieves from a mixed multilin- gual document pool containing Chinese, Russian, and Persian documents. Since SHIFT requires rel- ative language vectors estimated for each target language, we evaluate the subset of Chinese and Russian documents, which overlap with the lan- guages supported by our vector-estimation corpus. We report nDCG@20 and apply the same model- specific scale factorαselected in the main experi- ments, without any dataset-specific tuning. As shown in Table 8, SHIFT improves the av- erage nDCG@20 from 0.4731 to 0.4839, yield- ing gains for 5 out of 6 evaluated retrievers. The most substantial improvement is observed for multilingual-e5-large (+0.039), which corresponds 2 https://huggingface.co/datasets/mteb/ NeuCLIR2023RetrievalHardNegatives ModelBaseSHIFTâ embeddinggemma-300m0.5140.520+0.006 multilingual-e5-large0.4090.448+0.039 bge-m30.4720.470-0.002 Qwen3-Embedding-0.6B0.4610.467+0.006 llama-nemotron-embed-1b-v20.5150.515+0.000 gte-Qwen2-1.5B-instruct0.4680.483+0.015 Average0.47310.4839+0.0108 Table 8: Retrieval performance (nDCG@20) on the NeuCLIR2023 benchmark. to the model exhibiting strong language bias in our previous analyses. Conversely, the performance of bge-m3 changes only marginally (-0.002). This aligns with our observation that bge-m3 already possesses relatively strong multilingual alignment, leaving less language-induced offset for SHIFT to correct. G Robustness to Imperfect Language Identification Because SHIFT applies language-specific transfor- mations during the indexing stage, its practical de- ployment depends on identifying the document lan- guages. In our primary experiments, we utilized ground-truth document language labels to strictly isolate the effect of the proposed transformation. To evaluate a more realistic deployment scenario, we conduct a robustness experiment where docu- ment languages are automatically predicted using an off-the-shelf FastText language identification model 3 , and SHIFT is applied based on these pre- dicted labels. We evaluate multilingual-e5-large using the same scale factor (α = 0.6) as in the main ex- periments. We compare three settings: Base, GT- SHIFT (using ground-truth document language la- 3 https://huggingface.co/facebook/ fasttext-language-identification 17 0.20.40.60.81.0 Alpha 0.650 0.700 0.750 0.800 0.850 0.900 0.950 Target Recall (AR) (a) Arabic (ar) Target Recall (AR) Overall Recall 0.20.40.60.81.0 Alpha 0.800 0.850 0.900 0.950 1.000 Target Recall (DE) (b) German (de) Target Recall (DE) Overall Recall 0.20.40.60.81.0 Alpha 0.850 0.900 0.950 1.000 Target Recall (ES) (c) Spanish (es) Target Recall (ES) Overall Recall 0.20.40.60.81.0 Alpha 0.750 0.800 0.850 0.900 0.950 Target Recall (RU) (d) Russian (ru) Target Recall (RU) Overall Recall 0.910 0.915 0.920 0.925 0.930 0.935 0.940 Overall Recall 0.920 0.925 0.930 0.935 0.940 Overall Recall 0.925 0.930 0.935 0.940 Overall Recall 0.915 0.920 0.925 0.930 0.935 0.940 Overall Recall Figure 6: Extended sensitivity analysis of the scale factor α across remaining 4 target languages. bels), and FT-SHIFT (using FastText-predicted document language labels). The evaluation spans the four MLIR benchmarks from the main experi- ments, as well as the NeuCLIR 2023 setting. We report Language Identification (LID) accuracy, de- fined as the percentage of corpus documents whose FastText-predicted language matches the ground- truth language. BenchmarkBaseGT-SHIFTFT-SHIFTâ (FTâGT)LID Acc. Belebele0.6230.8130.807-0.00699.4% MLQA0.4360.5480.534-0.01693.4% MultiEup0.3090.3890.389+0.00099.1% XQuAD0.7170.8640.846-0.01896.3% NeuCLIR 20230.4090.4480.421-0.02750.2% Average0.4990.6130.600-0.01387.7% Table 9: Robustness of SHIFT to imperfect language identification using FastText on multilingual-e5-large. Performance is measured by nDCG@20. The results are summarized in Table 9. FT- SHIFT consistently outperforms the Base model across all benchmarks, demonstrating that SHIFT remains effective even when relying on automati- cally predicted language labels. On average, FT- SHIFT improves nDCG@20 from 0.499 to 0.600, maintaining performance very close to GT-SHIFT with only a minimal average drop of 0.013. The most noticeable degradation occurs on Neu- CLIR 2023, where the FastText LID accuracy is only 50.2% (primarily due to a 7.4% accuracy drop on Chinese documents in this specific cor- pus). Nevertheless, even in this challenging setting, FT-SHIFT still surpasses the Base model (0.421 vs. 0.409). In contrast, on MultiEup, where all document languages are accurately identified, FT- SHIFT is virtually identical to GT-SHIFT. Ulti- mately, in deployment, SHIFT only requires pre- dicted document languages at indexing time. Be- cause both the language prediction and the corre- sponding vector transformation are applied offline, they introduce no additional overhead. H Additional Sensitivity Analysis for multilingual-e5-large We extend the sensitivity analysis presented in Sec- tion 5.5 to the remaining four languages available in the intersection of mMARCO, Belebele, and XQuAD: Arabic (ar), German (de), Spanish (es), and Russian (ru). The results are visualized in Fig- ure 6. As illustrated, the performance trajectories ex- hibit identical patterns to those reported in the main experiments. First, the Target Recall@20 (red line) monotonically increases asαgrows, reaffirming the precise directionality of the estimated relative language vectors across diverse language families. Second, the Overall Recall@20 (blue dashed line) displays a characteristic inverted-U or saturation trend. This consistent behavior across all evaluated 18 Figure 7: Sensitivity analysis with an Aggressive Default (α default = 1.0). We vary theαfor the target language while fixing α = 1.0 for all other languages. languages further validates that while increasingα effectively boosts target-specific retrieval, an opti- mal threshold exists to maintain the global stability of the multilingual ranking system. I Sensitivity Analysis on Diverse Architectures and Base Settings To verify the universality of our findings be- yond a single model architecture, we extend the sensitivity analysis to two distinct models: embeddinggemma-300m (encoder-based) and gte- Qwen2-1.5B-instruct (decoder-based). Further- more, to investigate how the global shift intensity influences the local sensitivity of a specific target language, we conduct experiments under two ex- treme baseline settings: âą Aggressive Default (α def ault = 1.0): All non- target languages are shifted maximally. This simulates a fully transformed, source-aligned embedding space. âąConservative Default (α def ault = 0.1): All non-target languages are shifted minimally. This simulates a scenario where the global embedding space remains close to the original, biased distribution. In both settings, we sweep theαfor the specific target language (Chinese, Hindi, Vietnamese) from 0.2 to 1.0 and observe the trade-off between Target Recall and Overall Recall. Aggressive DefaultFigure 7 presents the results when the base system is fully aligned to the source language. In this setting, the continuous improve- ment in Target Recall and the high stability of Overall Recall are expected outcomes. As demon- strated in Section 5.5, these models achieve their global optimum atα = 1.0. Therefore, increas- ing the specific targetαtowards 1.0 effectively moves the target language embeddings closer to the source languageâs semantic space, reinforcing the semantic signals that were previously diluted by language barriers. This confirms that the per- formance gains are a natural consequence of the embeddings converging to their optimal geomet- ric alignment, where both the target-specific and global retrieval capabilities are maximized. Conservative DefaultFigure 8 illustrates the re- sults when the base shift is minimal. In this setting, the base system retains a strong bias towards the source language, typically suppressing target lan- guage documents. However, as we increase theα for the specific target language, we observe a dra- matic surge in Target Recall (solid red line). This result is particularly significant from the perspec- tive of controllability discussed in Section 5.5. It demonstrates that even if the global system remains conservative, we can effectively "boost" the expo- sure of a specific language by selectively increasing its shift magnitude. Although this targeted boost- ing involves a trade-off with Overall Recall (dashed 19 Figure 8: Sensitivity analysis with a Conservative Default (α default = 0.1). We vary the α for the target language while fixing α = 0.1 for all other languages. blue line), it empirically confirms that SHIFT pro- vides a powerful mechanism to forcefully prioritize specific languages according to user requirements. J M2MIR: Multilingual-to-Multilingual Information Retrieval While standard MLIR tasks typically assume a monolingual query retrieving from a multilingual collection (one-to-many), real-world search scenar- ios often involve users issuing queries in diverse languages to find information across a globally mixed corpus. To evaluate SHIFT in this more prac- tical and challenging environment, we introduce the M2MIR (Multilingual-to-Multilingual Informa- tion Retrieval) scenario, where both queries and documents originate from multiple languages. J.1 Experimental Setup Data StatisticsBelebeleMLQAXQuADMultiEup-v2 # of Languages in queries 14786 # of Languages in collection 14786 Fully ParallelOXOX # of Queries12,60042,2459,52010,686 # of Documents6,83236,7991,92035,444 # of d + /q143.65819.9 # of d + en /q1117.91 Table 10: Statistics of the datasets configured for the M2MIR evaluation. We reconfigure the test sets of Belebele, MLQA, XQuAD, and MultiEuP-v2 to include queries from all available languages overlapping with mMARCO, rather than restricting them to a sin- gle source language. The detailed statistics of this constructed dataset are summarized in Table 10. In this setting, the retrieval pool contains documents in varying languages, while the query stream simul- taneously comprises a diverse mixture of languages. For all SHIFT operations in this scenario, we use English as the anchor (reference) language, i.e., we align embeddings toward the English embedding space. We set the scale factorαto 1.0 to apply the full shift toward the anchor space. Consistent with the main experiments, we employ nDCG@20 and Recall@20 as the primary evaluation metrics. We compare two strategies for applying SHIFT to align the embedding space toward English: âąDoc Only: We apply SHIFT exclusively to doc- ument embeddings. Specifically, we subtract the relative language vector corresponding to each documentâs language, aligning it with the English space. The query embeddings remain unmodified in their original language spaces. âąQuery and Doc: We apply SHIFT to both queries and documents.Queries in non- English languages are also shifted towards the English space using their respective rela- tive language vectors. This ensures that both the queries and the documents reside in the unified English-centric geometry. 20 Model BelebeleMLQAXQuADMultiEup-v2AVG nDCGRecallnDCGRecallnDCGRecallnDCGRecallnDCGRecall embeddinggemma-300m0.8750.8780.6220.6960.9130.9480.4570.4020.7170.731 + SHIFT (Doc Only)0.8920.8950.6300.7070.9280.9570.4900.4280.7350.747 + SHIFT (Query and Doc)0.8930.8960.6300.7070.9280.9570.4940.4310.7360.748 multilingual-e5-large0.6230.6010.4360.4520.7170.7710.3090.2560.5210.520 + SHIFT (Doc Only)0.8600.8560.5940.6640.9140.9470.4140.3550.6950.706 + SHIFT (Query and Doc)0.8620.8600.5970.6680.9160.9480.4090.3470.6960.706 llama-nemotron-embed-1b-v20.5410.5350.4490.5190.8510.8990.2180.1860.5150.535 + SHIFT (Doc Only)0.6040.6060.4650.5410.8740.9180.2580.2260.5500.573 + SHIFT (Query and Doc)0.6170.6180.4680.5430.8760.9180.2680.2350.5570.579 gte-Qwen2-1.5B-instruct0.7930.7900.5180.5770.8280.8750.4420.3880.6450.657 + SHIFT (Doc Only)0.8460.8450.5580.6330.8860.9230.4780.4190.6920.705 + SHIFT (Query and Doc) 0.8530.8520.5660.6420.8920.9270.4770.4170.6970.709 Table 11: Retrieval performance in the M2MIR scenario. We compare the effectiveness of shifting only the documents (Doc Only) versus shifting both queries and documents (Query and Doc) into the unified embedding space. J.2 Results Table 11 presents the retrieval performance in the M2MIR setting. We first observe that the Doc Only strategy significantly improves performance over the baseline across all models. For instance, embeddinggemma-300m increases from 0.717 to 0.735 on average nDCG@20, and llama-nemotron- embed-1b-v2 gains from 0.515 to 0.550. This in- dicates that simply shifting the multilingual docu- ment mixture into a single English-aligned space makes the documents more accessible, even if the queries are multilingual. However, the most robust performance is achieved with the Query and Doc strategy. By shift- ing the queries into the same English subspace as the documents, we maximize the geometric align- ment between queries and relevant documents. As shown in the results, applying SHIFT to both sides consistently yields the highest nDCG@20 and Re- call@20 scores. Notably, for embeddinggemma- 300m, the average nDCG@20 further improves to 0.736, and llama-nemotron-embed-1b-v2 reaches 0.557, outperforming the Doc Only approach. Even for strong baselines like multilingual-e5-large and gte-Qwen2-1.5B-instruct, the Query and Doc strat- egy maintains or slightly exceeds the high perfor- mance of Doc Only. This confirms that in M2MIR scenario, unifying both query and document repre- sentations into a common pivot space is the optimal strategy. K Qualitative Analysis To complement our quantitative results, we present qualitative retrieval examples that illustrate how SHIFT changes the ranking behavior in practice. Each table shows English queries and the top- ranked passages retrieved by the base model and by applying SHIFT. We highlight cases where the base retriever over-prefers documents written in the query language (English), while SHIFT promotes semantically relevant passages in other languages, improving target-language exposure without intro- ducing obvious topic drift. We report examples for four retrievers to demonstrate that these pat- terns are consistent across architectures and model families. 21 Q: What do some animals not have? Rank BaseSHIFT ContentLangRelevantContentLangRelevant 1 Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky. enTRUE ç©ç±ć€èæăç©ćäșéŁç©ććšäœć ¶æ¶ćă性 éšćç©éœèœç§»ăćȘæç©æïŒäœäčäžæŻææç© éœæïŒæŻćŠæ°ŽæŻć°±æïŒăć°çäžçćè§èœéœ æç©çèș«ćœ±ăćźæçćšć°äžææŽïŒæçćšæ”· æŽäžæžžïŒæçćšć€©ç©șäžçż±çżă [Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky.] zhTRUE 2 Out on the savanna, it is hard for a primate with a digestive system like that of humans to satisfy its amino-acid requirements from available plant resources. Moreover, failure to do so has serious consequences: growth depression, malnutrition, and ultimately death. The most readily accessi- ble plant resources would have been the proteins accessible in leaves and legumes, but these are hard for primates like us to digest unless they are cooked. In contrast, animal foods (ants, ter- mites, eggs) not only are easily digestible, but they provide high-quantity proteins that contain all the essential amino acids. All things consid- ered, we should not be surprised if our own ances- tors solved their âprotein problemâ in somewhat the same way that chimps on the savanna do to- day. enFALSE Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky. enTRUE 3 Insects were the first animals to take to the air. Their ability to fly helped them evade enemies more easily and find food and mates more effi- ciently. Most insects have the advantage of being able to fold their wings back along the body. This gives them a wider range of small places to hide from predators. Today, the only insects that can- not fold back their wings are dragon flies and mayflies. enFALSE Les animaux sont constituĂ©s de nombreuses cel- lules. Ils mangent toutes sortes de choses et les digĂšrent naturellement. La plupart des animaux peuvent se dĂ©placer. Seuls les animaux ont un cerveau (mĂȘme si tous les animaux nâen ont pas; les mĂ©duses, par exemple, nâont pas de cerveau). On trouve des animaux partout sur la Terre. Ils creusent le sol, nagent dans les ocĂ©ans et volent dans le ciel. [Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky.] frTRUE 4 Wildlife is among the most challenging motifs for a photographer, and needs a combination of good luck, patience, experience and good equipment. Wildlife photography is often taken for granted, but like photography in general, a picture is worth a thousand words. Wildlife photography often requires a long telephoto lens, though things like a flock of birds or a tiny creature need other lenses. Many exotic animals are hard to find, and parks sometimes have rules about taking photographs for commercial purposes. Wild animals might either be shy or aggressive. The environment might be cold, hot, or otherwise hostile. enFALSE Hewan tersusun dari banyak sel.Hewan memakan dan mencernanya. Sebagian besar hewan bisa bergerak. Hanya binatang memiliki otak (tetapi tidak semua binatang memilikinya; ubur-ubur, sebagai contoh, tidak memiliki otak). Hewan bisa ditemukan di berbagai tempat di Bumi. Hewan menggali tanah, berenang di laut, dan terbang di angkasa. [Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky.] idTRUE 5 Plants make oxygen which humans breathe, and they take in carbon-dioxide which humans exhale (that is, breathe out). Plants make their food from the sun by photosynthesis. They also provide shade. We make our houses from plants and make clothes from plants. Most foods that we eat are plants. Without plants, animals could not survive. enFALSE Os animais sĂŁo feitos de vĂĄrias cĂ©lulas. Eles comem coisas e as digerem em seu interior. A maior parte dos animais consegue se locomover. Apenas animais tĂȘm cĂ©rebros (embora nem todos os animais tenham; a ĂĄgua-viva, por exemplo, nĂŁo tem cĂ©rebro). Os animais sĂŁo encontrados por todo o planeta. Eles cavam o solo, nadam nos oceanos e voam pelo cĂ©u. [Animals are made of many cells. They eat things and digest them inside. Most animals can move. Only animals have brains (though not even all animals do; jellyfish, for example, do not have brains). Animals are found all over the earth. They dig in the ground, swim in the oceans, and fly in the sky.] ptTRUE Table 12: Qualitative analysis comparison on Belebele dataset. The left block shows the top retrieved documents by multilingual-e5-large, and the right block shows the results after applying our SHIFT method. 22 Q: Who said that the overheating incidents were not very serious? Rank BaseSHIFT ContentLangRelevantContentLangRelevant 1 âEs wurden ĂŒber 60 FĂ€lle von Ăberhitzung bei defekten iPods gemeldet. Insgesamt wurden dadurch sechs BrĂ€nde ausgelöst und vier Personen haben leichte Verbrennungen erlitten. Das japanische Ministerium fĂŒr Wirtschaft, Han- del und Industrie (METI) gab an, dass ihm 27 UnfĂ€lle im Zusammenhang mit den GerĂ€ten bekannt waren. â âLetzte Woche gab das METI bekannt, dass es von Apple ĂŒber 34 weitere VorfĂ€lle von Ăberhitzung informiert worden war, die das Unternehmen als ânicht schwerwiegendâ bezeichnete.â Das Ministerium reagierte, indem es die Verschiebung des Berichts durch Apple als âwirklich bedauerlich âbezeich- nete.â [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] deTRUE âEs wurden ĂŒber 60 FĂ€lle von Ăberhitzung bei defekten iPods gemeldet. Insgesamt wurden dadurch sechs BrĂ€nde ausgelöst und vier Personen haben leichte Verbrennungen erlitten. Das japanische Ministerium fĂŒr Wirtschaft, Han- del und Industrie (METI) gab an, dass ihm 27 UnfĂ€lle im Zusammenhang mit den GerĂ€ten bekannt waren. â âLetzte Woche gab das METI bekannt, dass es von Apple ĂŒber 34 weitere VorfĂ€lle von Ăberhitzung informiert worden war, die das Unternehmen als ânicht schwerwiegendâ bezeichnete.â Das Ministerium reagierte, indem es die Verschiebung des Berichts durch Apple als âwirklich bedauerlich âbezeich- nete.â [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] deTRUE 2 60 , , (METI) 27 . , METI Apple 34 , . [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] hiTRUE 60 , , (METI) 27 . , METI Apple 34 , . [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] hiTRUE 3 Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â enTRUE Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â enTRUE 4 Cerca de 60 casos de iPods com problemas de superaqueci- mento foram relatados, causando um total de seis incĂȘndios e deixando quatro pessoas com queimaduras leves. O Min- istĂ©rio da Economia, ComĂ©rcio e IndĂșstria do JapĂŁo (METI) disse estar ciente de 27 acidentes relacionados aos dispos- itivos. Na semana passada, o METI divulgou que a Apple o havia informado de mais 34 incidentes de superaqueci- mento, os quais a empresa classificou como ânĂŁo gravesâ. O ministĂ©rio respondeu chamando o adiamento do relatĂłrio da Apple de âverdadeiramente lamentĂĄvelâ. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] ptTRUE Cerca de 60 casos de iPods com problemas de superaqueci- mento foram relatados, causando um total de seis incĂȘndios e deixando quatro pessoas com queimaduras leves. O Min- istĂ©rio da Economia, ComĂ©rcio e IndĂșstria do JapĂŁo (METI) disse estar ciente de 27 acidentes relacionados aos dispos- itivos. Na semana passada, o METI divulgou que a Apple o havia informado de mais 34 incidentes de superaqueci- mento, os quais a empresa classificou como ânĂŁo gravesâ. O ministĂ©rio respondeu chamando o adiamento do relatĂłrio da Apple de âverdadeiramente lamentĂĄvelâ. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] ptTRUE 5 Passengers were given water as they waited in 90(F)-degree heat. Fire captain Scott Kouns said, âIt was a hot day in the Santa Clara with temperatures in the 90s. Any length of time trapped on a roller coaster would be uncomfortable, to say the least, and it took at least an hour to get the first person off the ride.â enFALSE Environ 60 cas de surchauffe dâiPods dĂ©fectueux ont Ă©tĂ© signalĂ©s, provoquant au total six incendies et laissant quatre personnes avec des brĂ»lures mineures. Le ministĂšre japon- ais de lâĂ©conomie, du commerce et de lâindustrie (METI) a dĂ©clarĂ© quâil avait eu connaissance de 27 accidents liĂ©s Ă ces appareils. La semaine derniĂšre, le METI a annoncĂ© quâApple lâavait informĂ© de 34 incidents de surchauffe sup- plĂ©mentaires, que la sociĂ©tĂ© a qualifiĂ©s de « non graves ». Le ministĂšre a rĂ©agi en qualifiant de « vraiment regrettable » lâajournement du rapport par Apple. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] frTRUE Table 13: Qualitative analysis comparison on Belebele dataset. The left block shows the top retrieved documents by bge-m3, and the right block shows the results after applying our SHIFT method. 23 Q: Who said that the overheating incidents were not very serious? Rank BaseSHIFT ContentLangRelevantContentLangRelevant 1 âEs wurden ĂŒber 60 FĂ€lle von Ăberhitzung bei defekten iPods gemeldet. Insgesamt wurden dadurch sechs BrĂ€nde ausgelöst und vier Personen haben leichte Verbrennungen erlitten. Das japanische Ministerium fĂŒr Wirtschaft, Han- del und Industrie (METI) gab an, dass ihm 27 UnfĂ€lle im Zusammenhang mit den GerĂ€ten bekannt waren. â âLetzte Woche gab das METI bekannt, dass es von Apple ĂŒber 34 weitere VorfĂ€lle von Ăberhitzung informiert worden war, die das Unternehmen als ânicht schwerwiegendâ bezeichnete.â Das Ministerium reagierte, indem es die Verschiebung des Berichts durch Apple als âwirklich bedauerlich âbezeich- nete.â [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] deTRUE âEs wurden ĂŒber 60 FĂ€lle von Ăberhitzung bei defekten iPods gemeldet. Insgesamt wurden dadurch sechs BrĂ€nde ausgelöst und vier Personen haben leichte Verbrennungen erlitten. Das japanische Ministerium fĂŒr Wirtschaft, Han- del und Industrie (METI) gab an, dass ihm 27 UnfĂ€lle im Zusammenhang mit den GerĂ€ten bekannt waren. â âLetzte Woche gab das METI bekannt, dass es von Apple ĂŒber 34 weitere VorfĂ€lle von Ăberhitzung informiert worden war, die das Unternehmen als ânicht schwerwiegendâ bezeichnete.â Das Ministerium reagierte, indem es die Verschiebung des Berichts durch Apple als âwirklich bedauerlich âbezeich- nete.â [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] deTRUE 2 60 , , (METI) 27 . , METI Apple 34 , . [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] hiTRUE 60 , , (METI) 27 . , METI Apple 34 , . [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] hiTRUE 3 Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â enTRUE Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI announced that Apple had informed it of 34 additional overheating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â enTRUE 4 Cerca de 60 casos de iPods com problemas de superaqueci- mento foram relatados, causando um total de seis incĂȘndios e deixando quatro pessoas com queimaduras leves. O Min- istĂ©rio da Economia, ComĂ©rcio e IndĂșstria do JapĂŁo (METI) disse estar ciente de 27 acidentes relacionados aos dispos- itivos. Na semana passada, o METI divulgou que a Apple o havia informado de mais 34 incidentes de superaqueci- mento, os quais a empresa classificou como ânĂŁo gravesâ. O ministĂ©rio respondeu chamando o adiamento do relatĂłrio da Apple de âverdadeiramente lamentĂĄvelâ. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] ptTRUE Cerca de 60 casos de iPods com problemas de superaqueci- mento foram relatados, causando um total de seis incĂȘndios e deixando quatro pessoas com queimaduras leves. O Min- istĂ©rio da Economia, ComĂ©rcio e IndĂșstria do JapĂŁo (METI) disse estar ciente de 27 acidentes relacionados aos dispos- itivos. Na semana passada, o METI divulgou que a Apple o havia informado de mais 34 incidentes de superaqueci- mento, os quais a empresa classificou como ânĂŁo gravesâ. O ministĂ©rio respondeu chamando o adiamento do relatĂłrio da Apple de âverdadeiramente lamentĂĄvelâ. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] ptTRUE 5 Passengers were given water as they waited in 90(F)-degree heat. Fire captain Scott Kouns said, âIt was a hot day in the Santa Clara with temperatures in the 90s. Any length of time trapped on a roller coaster would be uncomfortable, to say the least, and it took at least an hour to get the first person off the ride.â enFALSE Environ 60 cas de surchauffe dâiPods dĂ©fectueux ont Ă©tĂ© signalĂ©s, provoquant au total six incendies et laissant quatre personnes avec des brĂ»lures mineures. Le ministĂšre japon- ais de lâĂ©conomie, du commerce et de lâindustrie (METI) a dĂ©clarĂ© quâil avait eu connaissance de 27 accidents liĂ©s Ă ces appareils. La semaine derniĂšre, le METI a annoncĂ© quâApple lâavait informĂ© de 34 incidents de surchauffe sup- plĂ©mentaires, que la sociĂ©tĂ© a qualifiĂ©s de « non graves ». Le ministĂšre a rĂ©agi en qualifiant de « vraiment regrettable » lâajournement du rapport par Apple. [Around 60 cases of malfunctioning iPods overheating have been reported, causing a total of six fires and leaving four people with minor burns. Japanâs Ministry of Economy, Trade and Industry (METI) said that it had been aware of 27 accidents related to the devices. Last week, METI an- nounced that Apple had informed it of 34 additional over- heating incidents, which the company called ânon-serious.â The ministry responded by calling Appleâs postponement of the report âtruly regrettable.â] frTRUE Table 14: Qualitative analysis comparison on Belebele dataset. The left block shows the top retrieved documents by bge-m3, and the right block shows the results after applying our SHIFT method. 24 Q: Defending democracy from foreign interference Rank BaseSHIFT ContentLangRelevantContentLangRelevant 1 Madam President, defence of democracy means enhancing transparency and it also means public participation, protect- ing journalists, empowering civil society, investing in educa- tion, in critical thinking and informed decision-making. It means keeping a strong framework for the rule of law and sanctioning those who ignore it. It means free voices. While we are fully aware of the importance of such a package, we need to be extremely careful on which particular instruments we choose to defend and not... enFALSE Monsieur le PrĂ©sident, que dire aujourdâhui sur lâingĂ©rence Ă©trangĂšre mettant en danger la dĂ©mocratie? Nous nous con- centrons dĂ©sormais â Nous devons nous protĂ©ger et dĂ©tecter les failles dans notre systĂšme qui permettent ces prises de pouvoir, que je qualifierais de souterraines. Notre ouverture sur le monde, nos libertĂ©s dâexpression, de la presse et bien dâautres, ces valeurs qui nous sont si chĂšres contiennent en elles les risques qui, parfois, nous empĂȘchent de nous pro- tĂ©ger. Or, nous devon... [Mr. President, what can be said today about foreign inter- ference endangering democracy? We are now focusing â We must protect ourselves and detect the flaws in our system that allow these takeovers, which I would call underground. Our openness to the world, our freedoms of expression, of the press and many others, these values which are so dear to us contain within them the risks which, sometimes, prevent us from protecting ourselves. However, we must...] frTRUE 2 Madam President, Vice-Presidents, colleagues, this proposal arrives at an important time as we approach the European elections in a few months and Iâm very proud of our work in fighting foreign interference in the INGE and ING2 special committees in the European Parliament. Sadly, we are being proven right almost every day about how important this topic is. I believe that our special committeesâ commitment has been instrumental in shaping the discourse around foreign interference, and it is enco... enFALSE Madam President, defence of democracy means enhancing transparency and it also means public participation, protect- ing journalists, empowering civil society, investing in educa- tion, in critical thinking and informed decision-making. It means keeping a strong framework for the rule of law and sanctioning those who ignore it. It means free voices. While we are fully aware of the importance of such a package, we need to be extremely careful on which particular instruments we choose to defend and not... enFALSE 3 The work of the INGE Committee and this House has been a real source of inspiration for the Commission. I want to congratulate the rapporteur, Sandra Kalniete, for bringing forward this work, which seems to command broad support across this House. That is testament to the importance of the report in analysing the phenomena of foreign interference and reflecting the need for a truly whole-of-society approach. The report explores many dimensions. It looks at interfer- ence via elite capture, nationa... enFALSE Señor presidente, ÂżcĂłmo podemos defender nuestra democ- racia de injerencias extranjeras? Pues no permitiĂ©ndolo, no amparĂĄndolo, no legitimĂĄndolo. Porque durante demasiado tiempo hemos permitido, amparado y legitimado que Ru- sia e IrĂĄn, Venezuela y otros paĂses financien e impulsen movimientos polĂticos extremistas en diferentes paĂses eu- ropeos y tambiĂ©n en esta CĂĄmara. Porque hemos permitido, legitimado y amparado que paĂses como Rusia apoyaran a movimientos separatistas en Europa para debilitarla... [Mr. President, how can we defend our democracy from foreign interference? Well, by not allowing it, not sheltering it, not legitimizing it. Because for too long we have allowed, sheltered and legitimized Russia and Iran, Venezuela and other countries to finance and promote extremist political movements in different European countries and also in this Chamber. Because we have allowed, legitimized and shel- tered countries like Russia to support separatist movements in Europe to weaken it...] esTRUE 4 Madam President, dear Commissioners, colleagues, I have the tradition of reviewing voting results after this House votes on things the Chinese Communist Party really dislikes, and the Greek Social Democrats tend to consistently abstain or vote against China-critical texts. The underpinning problem is, of course, Greeceâs critical port of Piraeus now being China-owned. This is foreign interference at work. In my country, the Netherlands, we consistently supported Nord Stream 2, the pipeline, toge... enFALSE Comme vous le savez, la dĂ©fense des principes dĂ©mocra- tiques et la promotion dâĂ©lections libres et Ă©quitables comptent parmi les prioritĂ©s de la prĂ©sidence belge, et le renforcement de nos dĂ©mocraties va de pair avec une autre rĂ©sistance: la rĂ©sistance Ă la dĂ©sinformation, Ă la propagande et Ă lâingĂ©rence Ă©trangĂšre. Câest pourquoi nous avons pris lâinitiative de prĂ©parer des conclusions du Conseil sur la rĂ©silience Ă©lectorale et la sauvegarde des processus dĂ©mocra- tiques contre lâingĂ©rence Ă©trangĂš... [As you know, defending democratic principles and promot- ing free and fair elections are among the priorities of the Belgian presidency, and strengthening our democracies goes hand in hand with another resistance: resistance to disin- formation, propaganda and foreign interference. This is why we took the initiative to prepare Council conclusions on electoral resilience and safeguarding democratic processes against foreign interference...] frFALSE 5 Señor presidente, ÂżcĂłmo podemos defender nuestra democ- racia de injerencias extranjeras? Pues no permitiĂ©ndolo, no amparĂĄndolo, no legitimĂĄndolo. Porque durante demasiado tiempo hemos permitido, amparado y legitimado que Ru- sia e IrĂĄn, Venezuela y otros paĂses financien e impulsen movimientos polĂticos extremistas en diferentes paĂses eu- ropeos y tambiĂ©n en esta CĂĄmara. Porque hemos permitido, legitimado y amparado que paĂses como Rusia apoyaran a movimientos separatistas en Europa para debilitarla... [Mr. President, how can we defend our democracy from foreign interference? Well, by not allowing it, not sheltering it, not legitimizing it. Because for too long we have allowed, sheltered and legitimized Russia and Iran, Venezuela and other countries to finance and promote extremist political movements in different European countries and also in this Chamber. Because we have allowed, legitimized and shel- tered countries like Russia to support separatist movements in Europe to weaken it...] esTRUE Madam President, Vice-Presidents, colleagues, this proposal arrives at an important time as we approach the European elections in a few months and Iâm very proud of our work in fighting foreign interference in the INGE and ING2 special committees in the European Parliament. Sadly, we are being proven right almost every day about how important this topic is. I believe that our special committeesâ commitment has been instrumental in shaping the discourse around foreign interference, and it is enco... enFALSE Table 15: Qualitative analysis comparison on MultiEup-v2 dataset. The left block shows the top retrieved documents by Qwen3-Embedding-0.6B, and the right block shows the results after applying our SHIFT method. 25