PaperlayerView exact revision on arXiv

Source-linked AI summary

Bridging Language Gaps: Advances in Cross-Lingual Information Retrieval with Multilingual LLMs

Roksana Goworek, Olivia Macmillan-Scott, Eda B. Özyiğit

arXiv:2510.00908v1cs.IRcs.AIcs.CL

TL;DR

CLIR addresses unequal access to information across languages, while research has often treated it as monolingual retrieval plus translation. This survey reviews methods, evaluation, and resources across the field; its synthesis emphasizes that progress requires addressing linguistic mismatch, resource imbalance, and evaluation needs.

  • Problem

    CLIR research has often framed the task as monolingual retrieval plus translation, while linguistic differences, scarce parallel corpora, and domain mismatch complicate cross-lingual retrieval.

    Source: §1 · p. 2

  • Method

    The survey reviews the CLIR pipeline, including embedding alignment, multilingual pre-training, retrieval methods, generative models, datasets, and evaluation.

    Source: §1 · p. 2

  • Results

    The survey concludes that effective CLIR requires more than adapting monolingual approaches or reusing multilingual models, with progress depending on addressing linguistic mismatch, resource imbalance, and evaluation.

    Source: §7 · p. 26

  • Takeaways & Limitations

    Inclusive multilingual data, stronger evaluation benchmarks, fairness, bias mitigation, and attention to linguistic diversity are important directions for equitable CLIR.

    Source: §7 · p. 26

    Short-query ambiguity, OOV terms, and typological differences hinder retrieval and alignment, while low-resource languages remain unsupported.

    Source: §6.1 · p. 24

Abstract

from arXiv · show

Cross-lingual information retrieval (CLIR) addresses the challenge of retrieving relevant documents written in languages different from that of the original query. Research in this area has typically framed the task as monolingual retrieval augmented by translation, treating retrieval methods and cross-lingual capabilities in isolation. Both monolingual and cross-lingual retrieval usually follow a pipeline of query expansion, ranking, re-ranking and, increasingly, question answering. Recent advances, however, have shifted from translation-based methods toward embedding-based approaches and leverage multilingual large language models (LLMs), for which aligning representations across languages remains a central challenge. The emergence of cross-lingual embeddings and multilingual LLMs has introduced a new paradigm, offering improved retrieval performance and enabling answer generation. This survey provides a comprehensive overview of developments from early translation-based methods to state-of-the-art embedding-driven and generative techniques. It presents a structured account of core CLIR components, evaluation practices, and available resources. Persistent challenges such as data imbalance and linguistic variation are identified, while promising directions are suggested for advancing equitable and effective cross-lingual information retrieval. By situating CLIR within the broader landscape of information retrieval and multilingual language processing, this work not only reviews current capabilities but also outlines future directions for building retrieval systems that are robust, inclusive, and adaptable.

1 Introduction

CLIR bridges languages so queries can retrieve documents written in other languages, against a backdrop of unequal digital and model resources. The survey reviews retrieval methods, multilingual advances, evaluation, and persistent challenges across the CLIR pipeline.

1 Introduction

  • CLIR enables queries in one language to retrieve documents in another, adding cross-lingual methods to the query-expansion, ranking, re-ranking, and generation pipeline.Methods include translating queries or documents, aligning multilingual embeddings, or using generative multilingual models.Source: §1 · pp. 1–2
  • English-heavy web content and model training leave many low-resource languages with less digital content and NLP support.The passage reports English at almost 80% of web content in the 1990s and around 90% of training data in most popular models.Source: §1 · pp. 1–2
  • Embedding and multilingual pre-training advances improve CLIR, but alignment, scarce parallel corpora, domain mismatch, and noisy judgments remain obstacles.The survey notes that translation or alignment can also cause semantic drift.Source: §1 · p. 2
  • The survey covers the full pipeline from query reformulation to re-ranking and answer generation, emphasizing embedding alignment, contrastive learning, multilingual pre-training, and generative models.Source: §1 · p. 2
  • Its contributions include reviewing CLIR techniques and retrieval architectures alongside multilingual NLP developments relevant to cross-lingual retrieval.The methods review addresses strengths and trade-offs; the multilingual review considers how advances in cross-lingual models support retrieval.Source: §1 · p. 2
  • The survey examines datasets, evaluation protocols, metrics, and limitations, and analyzes challenges including linguistic divergence, data scarcity, generalization, and fairness.Source: §1 · p. 2

2 CLIR Architecture

The section treats retrieval architecture as a foundation for CLIR, focusing on core retrieval stages that can strengthen cross-lingual systems. Its diagram traces an example query through retrieval and answer generation.

2 CLIR Architecture

  • The section focuses on foundational retrieval stages, complementing the separate challenge of bridging language gaps.Source: §2 · p. 2
  • Figure 1 provides an overview of the survey's cross-lingual and information retrieval approaches.Source: Figure 1 · p. 3
  • The architecture covers query expansion, ranking, re-ranking, and optional question answering, drawing on monolingual IR advances to strengthen CLIR.Source: §2 · p. 3

2.1 Query Expansion

Query expansion adds terms to address short, ambiguous, or underspecified queries and mismatches between query and document vocabulary. The section reviews expansion pipelines, user feedback, LLM-based approaches, and cross-lingual variants.

2.1 Query Expansion

  • Expansion adds relevant terms or synonyms to queries, addressing ambiguity, low-quality queries, and vocabulary mismatch.It is especially relevant to broad informational queries.Source: §2.1 · p. 4

2.1.1 Query Expansion Pipeline

  • The pipeline processes a data source, weights and ranks candidate terms, selects terms for augmentation, then reformulates the query.Sources include thesauri, logs, embeddings, and relevance feedback; proposed term counts range from 5–10 to hundreds or a third of all candidates.Source: §2.1.1 · p. 4

2.1.2 User Interaction & Relevance Feedback

  • Expansion can be automatic, manual, or interactive; relevance feedback uses relevant documents, while pseudo-relevance feedback assumes top-ranked documents are relevant.Local feedback is reported as more effective than global feedback, whereas pseudo-relevance feedback is generally less robust than explicit feedback.Source: §2.1.2 · p. 5

2.1.3 Query Expansion with LLMs

  • LLM-based expansion includes prompting, supervised fine-tuning, and reinforcement learning, with prompting the most common approach.Prompting can generate reformulated queries or pseudo-documents; one comparative study found chain-of-thought prompting most effective.Source: §2.1.3 · p. 5
  • LLM expansions can be irrelevant or erroneous, while conversation search can clarify intent; knowledge gaps and ambiguous prompts are reported sources of hallucinations.Consistency verification is one proposed way to filter meaningless outputs.Source: §2.1.3 · p. 5
  • Multilingual query expansion research remains limited; hybrid pre- and post-translation expansion often yields the best performance.Methods include sequential or joint translation and expansion, as well as implicit expansion through adjusted vector representations.Source: §2.1.3 · p. 5

2.2 Ranking

Ranking retrieves and scores an initial candidate list, prioritizing efficiency and recall before re-ranking. The section contrasts translation-dependent statistical methods with multilingual embeddings and other ranking signals.

2.2 Ranking

  • Ranking retrieves and scores an initial document list for efficiency and recall, using statistical, embedding-based, or hybrid strategies.The initial ordering can later be refined during re-ranking.Source: §2.2 · p. 5

2.2.1 Traditional Statistical Ranking

  • TF-IDF and BM25 are fast, interpretable statistical foundations, but require translation when query and document languages differ.With translation, these methods remain competitive in low-resource or high-latency settings; BM25 is a common baseline.Source: §2.2.1 · pp. 5–6

2.2.2 Neural Embedding-based Ranking

  • Multilingual embedding rankers map queries and documents into a shared vector space, enabling similarity-based retrieval beyond lexical overlap.Bi-encoders encode queries and documents independently, with precomputed document embeddings supporting efficient retrieval over large corpora.Source: §2.2.2 · p. 6
  • Language metadata can reduce irrelevant matches, while domain adaptation uses in-domain pairs, continued pre-training, or instruction tuning.General-purpose multilingual models can underperform in specialized domains such as law, medicine, and technology.Source: §2.2.2 · p. 6
  • Topic models remain useful in low-resource or hybrid settings, while learning-to-rank can incorporate structural and behavioral signals alongside text.Signals include freshness, categorical tags, clicks, ratings, and purchases; normalization strategies help reduce bias.Source: §2.2.2 · p. 6

2.3 Re-ranking

Re-ranking refines a candidate list with more detailed relevance signals, trading computation against precision, efficiency, and other ranking objectives. The section surveys encoder, LLM, QA-oriented, and complementary methods alongside architectural trade-offs.

Re-ranking objectives

  • Pointwise, pairwise, and listwise re-rankers differ in complexity and optimization goals, with listwise methods capturing ranking interdependencies for long-context or diversity-sensitive tasks.Re-ranking can apply computationally intensive models to a limited candidate set, such as the top 100.Source: §2.3 · pp. 6–7

Encoder-based Re-ranking

  • Cross-encoders model fine-grained query-document interactions and typically offer higher precision, while bi-encoders support efficient initial ranking; distillation transfers teacher relevance scores to a student bi-encoder.Cross-encoder inference costs limit scalability to large candidate sets, while distilled bi-encoders can approach cross-encoder accuracy.Source: §2.3 · p. 7

LLM-based Re-ranking

  • LLM re-rankers can score or reorder candidate lists in zero-shot or few-shot settings, but inference cost, latency, and prompt sensitivity remain challenges.Prompting may be pointwise, pairwise, or listwise; reported examples include multilingual nDCG gains and improved performance from few-shot prompting.Source: §2.3 · p. 7

Additional Approaches

  • QA-oriented re-ranking can prioritize answerability, while feature-enhanced ranking combines textual similarity with structural and behavioral signals.The MIA shared task's query-generation approach outperformed BM25 by 6–18% in top-20 accuracy; feature examples include freshness, authority, and engagement.Source: §2.3 · pp. 7–8
  • Score fusion and calibration address mismatched score scales and language or model biases; proposed mitigations include normalization, balanced training, and adaptive thresholds.Reciprocal Rank Fusion avoids score-scale mismatches when integrating retrieval components.Source: §2.3 · p. 8
  • Cascades reduce re-ranking cost, while adaptive methods, feedback-driven ranking, and explainability approaches address efficiency, changing interactions, and transparency.Bi-encoder cascades cut compute by up to sixfold with minimal performance loss; high-stakes settings motivate interpretable ranking.Source: §2.3 · p. 8

Retrieval architectures

  • Bi-encoders favor scale, cross-encoders detailed interactions, late interaction token-level matching, and sparse retrieval inverted-index efficiency.Table 1 compares representation and retrieval approaches across embedding granularities; architecture choice trades off scalability, semantic depth, precision, and efficiency.Source: §2.3 · pp. 8–9, Table 1 · p. 9
  • Hybrid retrieval combines sparse and dense signals, while multi-vector, hierarchical, and memory-augmented methods preserve finer-grained or global-local document information at added storage or computational cost.Dense Passage Retrieval performance depends on segmentation, and hybrid retrieval with rank fusion or concatenation often outperforms either signal alone.Source: §2.3 · p. 9

2.4 Question Answering

Question answering moves information access from ranked documents toward direct, contextually appropriate answers, often through a reader component. In cross-lingual settings, retrieved content can be answered or summarized in the query language, while retrieval-augmented generation grounds responses in documents.

2.4 Question Answering

  • Question answering provides direct responses ranging from factoids to extracted passages and abstractive summaries, often using a reader with retrieved documents.This shifts the user's task from sifting through ranked results to receiving an answer.Source: §2.4 · p. 9
  • Retrieval-augmented generation combines retrieval with generation so retrieved documents ground answers, with some systems adding references and citations.The paper describes RAG either as an integrated IR-and-QA system or as a retriever plus reader, particularly for open-book scenarios.Source: §2.4 · p. 10
  • Cross-lingual QA commonly pairs English with another language, using translation for sparse retrieval or multilingual embeddings for dense retrieval.A query-language answer or summary can bridge the gap when users cannot understand retrieved documents; QA systems also face hallucinations, overconfidence, and limited access to recent or proprietary data.Source: §2.4 · p. 10

2.5 Current CLIR Approaches

Contemporary CLIR combines retrieval architectures ranging from sparse and dense models to hybrid, generative, and multimodal systems. Despite broader capabilities, gaps remain in domain adaptation, handling morphologically rich and code-switched languages, and using LLMs as multilingual rankers.

2.5 Current CLIR Approaches

  • CLIR methods span sparse, dense, late-interaction, hybrid, cross-encoder, generative, and multimodal approaches.Examples include SPLADE-family sparse models, LaBSE and mDPR dense models, ColBERT late interaction, and BM25-neural hybrids.Source: §2.5 · p. 10
  • Cross-modal pre-training, LECCR, and M-SpeechCLIP align text with image or speech embeddings for multilingual retrieval across modalities.Source: §2.5 · p. 10
  • Open gaps include domain adaptation, morphologically rich and code-switched languages, and incorporating LLMs as multilingual rankers.The survey identifies these as remaining challenges for accurate, scalable, and equitable CLIR.Source: §2.5 · p. 10

3 Dealing with Cross-Linguality

CLIR has moved from translation-centered pipelines toward shared cross-lingual representations and multilingual LLMs, while translation remains useful for some settings. The survey reviews translation choices and techniques, multilingual model training and architectures, and the trade-off between language coverage and per-language capability.

3 Dealing with Cross-Linguality

  • Cross-lingual embeddings map texts into a shared semantic space, while contrastive, adversarial, and self-supervised alignment strategies address linguistic and resource disparities.The survey frames cross-lingual retrieval as benefiting from methods designed for cross-lingual use, rather than treating it solely as monolingual retrieval augmented by translation.Source: §3 · pp. 10–11

3.1 Translation

  • Query translation is cheaper and more flexible than document translation, while document translation adds context, reduces ambiguity, and suits collections with queries in one language.Query translation can still require later document translation for user access; document translation permits direct examination of retrieved texts.Source: §3.1 · p. 11

3.1.1 Translation Techniques

  • Dictionary translation can retain alternatives to improve recall but struggles with proper nouns and OOV terms; parallel corpora are costly, while comparable corpora typically yield lower translation quality.SMT was effective with large parallel corpora but typically produced one translation; NMT offers fluent, context-aware translation and handles OOV terms through subword segmentation.Source: §3.1.1 · pp. 11–12
  • NMT systems such as Translate-Train, Translate-Distill, mDPR, and ColBERT-X integrate translation with retrieval, but short queries, ambiguity, domain adaptation, data needs, compute, and scalability remain challenges.The survey describes NMT as improving on SMT in fluency, adaptability, and context handling, while retaining low-resource constraints.Source: §3.1.1 · p. 12

3.1.2 Indirect Translation

  • Indirect translation uses a high-resource pivot language or translates both texts into a third language, including a semantic space.Triangulation through three pivots outperformed pairwise merging in reported findings, with benefits most evident for unstructured queries; ESA uses weighted vectors and human-readable labels.Source: §3.1.2 · p. 12

3.2 Multilingual LLMs for CLIR

  • Multilingual LLMs use multilingual corpora and generally follow pre-training, fine-tuning, and RLHF, with encoder-only, encoder-decoder, and decoder-only architectures.The survey notes that encoder-decoder models may suit CLIR question answering, while decoder-only models are preferable for some specialised tasks.Source: §3.2 · p. 12

3.2.1 Training Stages of Multilingual LLMs

  • Training spans multilingual pre-training, task-specific supervised fine-tuning, and preference alignment; continual pre-training is cheaper but risks catastrophic forgetting.RLHF relies on preference data and reward-model fine-tuning, and is resource-intensive because it requires substantial human annotation.Source: §3.2.1 · pp. 12–13

3.2.2 Model Architectures of Multilingual LLMs

  • Encoder-only models produce contextual representations, encoder-decoder models support sequence-to-sequence generation, and decoder-only models generate text autoregressively.Source: §3.2.2 · p. 14

3.3 Embeddings

Multilingual embeddings enable direct cross-language comparison, but effective CLIR depends on how representations align and preserve task-relevant meaning. The survey reviews alignment granularity and methods, query-document asymmetry, and why strict geometric alignment may not be necessary or sufficient for transfer.

3.3 Embeddings

  • Contextual multilingual embeddings enable cross-language comparison without explicit translation, and have been reported to surpass translation-based methods in cross-lingual retrieval.Joint pre-training and alignment data at word, sentence, or domain level help map equivalent concepts to nearby positions.Source: §3.3 · p. 15

3.4 Alignment in CLIR

  • Multilingual pre-training alone may leave language-specific subspaces, so additional mechanisms are needed to align semantically equivalent content across languages.Alignment can support transfer from monolingual fine-tuning and reduce the need for language-specific supervision, especially in low-resource settings.Source: §3.4 · p. 17

3.4.1 Levels and Methods of Alignment

  • Word-level alignment is limited by lexical gaps and cultural specificity, while sentence-level methods capture sentence or phrase meaning; contrastive learning pulls equivalent pairs together and separates dissimilar ones.GAN-based mapping can be unstable and struggle with polysemy and compositional semantics, and its approximate-isomorphism assumption may fail for distant languages or mismatched domains.Source: §3.4.1 · p. 18

3.4.2 Query-Document Alignment

  • Query-document alignment must account for short, focused queries versus longer, diverse documents, motivating separate encoders trained for semantic similarity rather than exact pairwise alignment.A common design maps queries into document space; LLMs can also generate pseudo-documents from queries or pseudo-queries from documents.Source: §3.4.2 · p. 19

3.4.3 Is Alignment Necessary?

  • Cross-lingual alignment is neither always necessary nor sufficient for capability transfer; retrieval may benefit more from context-sensitive representations that preserve semantic distinctions and query-relevant information.The survey favors approximate, task-sensitive alignment with contextual understanding over strict geometric uniformity.Source: §3.4.3 · p. 19

4 Evaluation

CLIR evaluation spans component-level and end-to-end assessment, using retrieval measures alongside translation, embedding, and generation metrics. In practice, direct system evaluation remains centered on retrieval performance.

4 Evaluation

  • Full-system evaluation assesses cross-lingual retrieval across settings such as RAG, QA, and domain-specific retrieval, alongside component-wise assessment.Datasets and metrics are drawn from machine translation, monolingual IR, multilingual generation, and embedding alignment; some datasets use proxy or reference-free metrics.Source: §4 · p. 19

4.1 Translation Evaluation

  • Translation evaluation prioritises retrieval effectiveness over fluency, with BLEU, METEOR, and TER complemented by neural metrics BLEURT and COMET.Recall is generally prioritised over precision because the goal is to retrieve relevant documents and minimise false negatives.Source: §4.1 · pp. 19–20

4.2 Datasets for CLIR

  • CLIR datasets range from broad multilingual collections to language- and region-specific resources, but many are Wikipedia-derived and English-centred.Some newer resources use LLM-aided generation or machine translation to produce query-document triplets, while human-generated content is less common.Source: §4.2 · p. 20, Table 5 · p. 21

4.3 CLIR Evaluation Metrics

  • CLIR evaluation uses fixed corpora, query topics, and relevance judgments; recall often takes precedence, while MRR and nDCG reward early relevant results.Resource imbalance and translation or embedding quality complicate evaluation, and auxiliary measures are generally not final end-to-end criteria.Source: §4.3 · p. 20

4.3.1 Retrieval Metrics

  • Hit@K tests whether any relevant document appears in the top K, while Recall@K measures completeness and MRR and MAP reward earlier retrieval.nDCG normalises discounted gain to compare queries with different numbers of relevant documents; it is popular for multilingual and graded-relevance settings.Source: §4.3.1 · p. 20

4.3.2 Cross-lingual Performance Metrics

  • Cross-lingual assessment spans multilingual accuracy benchmarks, translation and generation scores, embedding-alignment measures, and learned semantic metrics, but direct CLIR evaluation remains retrieval-centred.Standalone cross-lingual metrics are mainly used for development, diagnostics, alignment analysis, or auxiliary evaluation.Source: §4.3.2 · pp. 21–22

4.3.3 Generation Evaluation Metrics

  • QA evaluation relies on human judgments because automatic metrics capture narrow qualities and often fail to align with human assessments.Crowdworkers may lack expertise to judge factuality, while reference-based metrics are limited for long-form QA because valid answers vary.Source: §4.3.3 · p. 22

5 Applications

CLIR addresses barriers created by the uneven distribution of online content and supports access across languages in search, professional resources, QA, media, research, and commerce.

Applications

  • English makes up 49.1% of websites as of February 2025, and CLIR helps users access information beyond their native language.The section identifies low-resource-language users as facing significant barriers.Source: §5 · p. 22

Search Engines

  • Cross-lingual search engines can retrieve higher-resource-language content, but Google cross-lingual search underperforms monolingual search; syntactic analysis improves results.Translation and AI-generated search summaries illustrate potential applications in translation and summarisation.Source: §5 · p. 22

Specialised Databases

  • CLIR gives legal and medical professionals access to technical resources without requiring duplicate content in multiple languages.Examples include English–Persian retrieval and Hindi–English medical retrieval using morphology and query expansion.Source: §5 · p. 23

LLMs and Question Answering

  • Retrieval-augmented generation can retrieve multilingual evidence and generate answers independently of the source language, though cross-lingual QA and reasoning remain challenging.The cited benchmarks include XOR-TyDi QA and XRAG.Source: §5 · p. 23

News, Media and Security

  • News and security applications use multilingual retrieval for global monitoring, fact-checking, crisis response, and awareness of threats and local protocols.NeuCLIR, xMIND, and CLSD evaluate news retrieval; MMTweets and MultiClaim assess fact-checking.Source: §5 · p. 23

Scientific Research

  • CLIR can help researchers access international literature without English translation, supporting collaboration and reducing duplication.CLIRudit and OPTICAL address retrieval over academic documents.Source: §5 · p. 23

E-commerce

  • E-commerce uses CLIR to let customers search product catalogues and reviews in their native language, with cross-lingual product QA also supported by datasets.The passage identifies CLPR-9M, multilingual ranking systems, and xPQA.Source: §5 · p. 23

Applications

  • Across domains, CLIR is presented as a means of reducing language barriers and promoting equitable access, while timeliness, accuracy, and coverage remain central to credibility.The applications discussed include medicine, law, research, media, security, and commerce.Source: §5 · p. 23

6 Challenges and Future Directions

CLIR faces interacting linguistic, representational, data, evaluation, and engineering challenges, especially for low-resource languages. Future directions target more robust and equitable systems through improved representations, resources, evaluation, and retrieval settings.

6.1 CLIR Challenges

  • Short-query ambiguity, polysemy, OOV terms, inconsistent transliteration, and typological variation complicate cross-lingual retrieval and alignment.Translation from a morphologically simple language to a richer one tends to perform poorly compared with the reverse direction; culturally embedded terms can also defy literal translation.Source: §6.1 · p. 24, Table 6 · p. 24
  • Semantic differences, catastrophic forgetting, and inadequate tokenisation hinder language-agnostic representations, while scarce and imbalanced resources weaken specialised-domain performance.Most datasets rely heavily on Wikipedia, and high-resource languages dominate pre-training corpora.Source: §6.1 · p. 24
  • Uneven performance across languages, demographic bias, and skewed evaluation data or metric choices can distort perceived system effectiveness.Translation pipelines may use pivot languages, whereas embedding-based CLIR requires uniform representations across languages.Source: §6.1 · p. 24
  • CLIR metrics may miss cross-lingual challenges, generation metrics remain insufficient, and hallucinations or fabricated references risk undermining user trust.High multilingual-LLM inference costs also challenge interactive systems, while independently designed pipeline components complicate integration.Source: §6.1 · p. 25

6.2 Future Directions in CLIR

  • Future work proposes semantic-consistency objectives, expanded training and pseudo-parallel corpora, and methods to reduce bias and improve representation of low-resource languages.Suggested approaches include LLM-generated training data, OPTICAL, adversarial training, universal tokenisers, and cross-lingual vocabulary transfer.Source: §6.2 · p. 25
  • Multimodal CLIR would incorporate images, captions, and speech queries, while graph-based retrieval could extend multilingual RAG to knowledge graphs for grounding and disambiguation.These directions require new pre-training strategies, benchmarks, and evaluation metrics for multimodal inputs.Source: §6.2 · p. 25
  • Future systems could address misinformation with fact-checking and transparent references, and reduce ambiguity through conversational clarification of user intent.The passage also calls for assessing information-source reliability across languages and building on query-expansion methods.Source: §6.2 · p. 25
  • Multilingual, multi-domain benchmarks and metrics for semantic drift, retrieval robustness, and QA performance are needed to advance robust and equitable CLIR.The survey identifies resource imbalance, bias, misinformation, and disambiguation as persistent challenges.Source: §6.2 · p. 25

7 Conclusion

The survey reviews CLIR across multilingual representation learning and information retrieval, arguing that progress requires more than adapting monolingual methods or reusing multilingual models. It identifies scarce high-quality multilingual data and weak evaluation benchmarks as major bottlenecks to equitable access.

Conclusion

  • The survey covers representation learning, embedding alignment, retrieval techniques, evaluation, and system architectures at the intersection of multilingual modelling and information retrieval.It argues that effective CLIR requires innovations addressing linguistic mismatch, resource imbalance, and reliable evaluation rather than direct adaptation or reuse alone.Source: §7 · p. 26
  • The survey frames equitable access to knowledge across science, education, healthcare, commerce, and other domains as a potential of robust, accurate, language-agnostic CLIR.Continued effort is needed to ensure these technologies serve languages equitably.Source: §7 · p. 26
  • Scarcity of high-quality multilingual data, especially for low-resource languages, remains a major bottleneck, making inclusive corpora and stronger benchmarks essential.The conclusion also calls for collaboration and attention to fairness, bias mitigation, and linguistic diversity in system design.Source: §7 · p. 26
Loading 2510.00908v1…