CLIR: Cross-Language Information Retrieval

• CLIR is a retrieval framework that processes queries in one language to find relevant documents in another, integrating methods like query translation, transliteration, and cross-lingual embeddings.
• The field employs diverse methodologies including dictionary-based translations, structured probabilistic models, and neural ranking systems that leverage multilingual transformers and representation learning.
• CLIR research focuses on overcoming challenges such as translation ambiguity, sparse training data, and operational constraints, while being evaluated through metrics like MAP, nDCG, and mAQWV.

Cross-language information retrieval (CLIR) denotes retrieval systems that accept a query in one language and retrieve documents written in another. In the standard formulation, a query $q$ is expressed in language $L_q$, documents $d_j$ are expressed in language $L_d$, and the system ranks documents by relevance even when $L_q \ne L_d$. CLIR differs from monolingual retrieval in two foundational ways: the searcher cannot reliably choose terms that appear in relevant documents, and ranking alone does not suffice because the searcher may not be able to read the retrieved documents. Across the literature, CLIR has been implemented through query translation, document translation, pivot or interlingual representations, structured probabilistic translation models, cross-lingual embeddings, multilingual neural rankers, and multilingual interaction models (Galuščáková et al., 2021).

1. Definition, scope, and task variants

CLIR is usually studied in the ad-hoc retrieval setting, where a user issues a query in a source language and the system retrieves relevant documents in a target language. The canonical example is English queries over non-English collections, but the field includes English–Hindi search in FIRE, English-to-Dutch/Italian/Finnish/Spanish search in CLEF, English-to-Chinese/Persian/Russian search in HC4 and NeuCLIR, and non-English-query to English-passage retrieval in XOR-TyDi (Varshney et al., 2014, Yu et al., 2020, Lawrie et al., 2022, Li et al., 2021).

The retrieval unit varies. In document-level CLIR, the system ranks full documents, as in CLEF, FIRE, HC4, and NeuCLIR. In sentence- or passage-level CLIR, the ranking target is a sentence or passage, as in Europarl sentence retrieval, MS MARCO-derived dense retrieval, and XOR-TyDi passage retrieval. Some evaluations depart from ranked-list assumptions entirely. In MATERIAL, systems are evaluated on their ability to return a query-relevant document set rather than only a ranked document list, and the metric mAQWV explicitly penalizes both misses and false alarms (Litschko et al., 2021, Yang et al., 2024, Agarwal et al., 5 Nov 2025).

The field also spans several operational regimes. Some systems assume that the query language is the user’s native language and the document language is unknown to the searcher. Others assume multilingual organizations or web platforms in which users search in one language but consume content in several. The chapter-length review of the field situates CLIR across “three waves”: controlled vocabularies, full-text ranked retrieval, and multilingual transformer-based retrieval (Galuščáková et al., 2021).

2. Translation-based and lexical formulations

Historically, CLIR was organized around three questions: what to translate, which translations are possible, and how translations should be used by the retrieval model. Query translation translates $q$ into the document language and then applies monolingual retrieval. Document translation translates each document into the query language and indexes the translated collection. Pivoting translates both sides into a third language. None of these architectures dominates universally; empirical findings summarized in the literature show that outcomes depend on language direction, translation quality, and domain (Galuščáková et al., 2021).

Dictionary-based CLIR remains important, especially where parallel data are limited. Its central problems are translation ambiguity, out-of-vocabulary terms, and missing lexical formations. Structured Query formulations treat multiple translations of a query term as synonyms, while Probabilistic Structured Queries estimate expected query-language term counts in a document via translation probabilities: $$e_{i,j} = \sum_{k \in j} p(e_i \mid f_k)\, f_{k,j}.$$ This lets BM25, vector-space, and language-model rankers operate as if the document had been written in the query language (Galuščáková et al., 2021).

A substantial line of work improved dictionary-based CLIR by learning which candidate translations to trust. In English–Persian CLIR, Learning to Rank has been used to combine bilingual dictionaries, parallel corpora, and comparable corpora through translation-relation and context-based features, significantly outperforming single-resource baselines (Azarbonyad et al., 2014). For agglutinative or morphologically rich languages, the Minimum Edit Support Candidates method combines a bilingual dictionary with a monolingual corpus to generate plausible target-language formations, score them probabilistically, and resolve ambiguity in English–Persian retrieval (Dadashkarimi et al., 2014).

Transliteration is a distinct but often crucial component. In English–Hindi CLIR, transliteration is used alongside dictionary translation because many Hindi documents contain English-origin terms written in Devanagari, and named entities or technical terms may not appear in bilingual dictionaries. A query such as “Democracy in India” can therefore generate alternatives such as भारत में लोकतंत्र, इंडिया में लोकतंत्र, भारत में डेमोक्रेसी, and इंडिया में डेमोक्रेसी; offline evaluation on FIRE 2010 showed that selecting the best translation–transliteration combination improved MAP over translation-only baselines (Varshney et al., 2014). In e-commerce CLIR, the query translation module itself has been modernized from SMT toward Transformer-based NMT, with clickthrough-derived in-domain query pairs and an asynchronous SMT+NMT architecture to meet latency constraints in production (Yao et al., 2020).

3. Cross-lingual representations and neural ranking

A major transition in CLIR has been the move from explicit translation tables toward shared multilingual representations. Cross-lingual word embeddings (CLWEs) map words from different languages into a common space, enabling both term-by-term translation and direct semantic matching. In the fully unsupervised setting, CLWEs can be induced from monolingual corpora alone through adversarial alignment and Procrustes refinement, after which query and document vectors can be formed by embedding aggregation or nearest-neighbor query translation. On CLEF EN→NL, EN→IT, and EN→FI, such unsupervised CLIR models outperformed alternatives that relied on word-level or document-level bilingual alignments, and simple rank-based ensembles of term-by-term translation and embedding aggregation further improved MAP (Litschko et al., 2018).

Neural CLIR then diversified into representation-based and interaction-based models. Representation-based methods compare a single query vector with a single document vector, typically by cosine similarity. Interaction-based models preserve word-level cross-lingual similarities and learn relevance patterns over interaction matrices, histograms, or kernel features. On four CLEF language pairs, interaction-based neural models built over fixed CLWEs—MatchPyramid, DRMM, and K-NRM—outperformed CLWE aggregation and term-by-term translation baselines, with DRMM-Cosine performing best across all tested pairs (Yu et al., 2020).

A notable empirical correction emerged from later work: pretrained multilingual encoders did not significantly outperform strong CLWE baselines for unsupervised document-level CLIR, even though sentence-level cross-lingual retrieval benefited greatly from specialized multilingual sentence encoders such as LASER, LaBSE, and DISTIL variants. This directly challenged the view that multilingual transformers had rendered CLWEs obsolete for all CLIR settings (Litschko et al., 2021).

Subsequent neural work focused on supervised dense retrieval and reranking. DR.DECR showed that a strong English-only retriever could be converted into an efficient end-to-end CLIR retriever through multi-stage knowledge distillation: relevance distillation from an MT+English ColBERT teacher and representation distillation from parallel corpora via cross-lingual token alignment. On XOR-TyDi, this substantially closed the gap between direct multilingual dense retrieval and MT+English IR while removing MT at inference time (Li et al., 2021). Translate-Distill extended this logic to English-query, non-English-document passage retrieval, using cross-encoder teachers and translation only where needed for offline supervision; the resulting ColBERT-X students achieved state-of-the-art single-stage CLIR on NeuCLIR and HC3 (Yang et al., 2024).

Cross-encoder rerankers form another branch. On TREC NeuCLIR 2022, mT5-XXL rerankers fine-tuned only on same-language query–document pairs nevertheless proved viable for CLIR, delivering top leaderboard performance across Chinese, Persian, and Russian. The study is notable because zero-shot cross-lingual reranking worked even when the first-stage retrieval was suboptimal (Jeronymo et al., 2023). Beyond text-only representations, HIKE introduced multilingual knowledge graphs into CLIR: multilingual BERT encodes queries, documents, and Wikidata entity descriptions, while hierarchical fusion integrates entity neighborhoods and source- and target-language knowledge to bridge the language gap explicitly (Zhang et al., 2021).

4. Supervision, datasets, and evaluation

CLIR has always been constrained by training data and evaluation design. Traditional resources include bilingual dictionaries, parallel corpora, comparable corpora, mixed-language web data, transliteration rules, and human interaction. Modern neural systems further rely on multilingual pretrained encoders, translated MS MARCO variants such as mMARCO and NeuMARCO, parallel corpora for distillation, and clickthrough logs for domain adaptation (Galuščáková et al., 2021, Yao et al., 2020).

Evaluation collections have evolved accordingly. Earlier CLIR benchmarks came from TREC, CLEF, NTCIR, and FIRE. FIRE 2010, for example, supports English–Hindi retrieval over 149,481 Hindi documents and was used to evaluate translation–transliteration query refinement (Varshney et al., 2014). CLEF collections remain central for English-to-European-language CLIR, including both document-level and sentence-level experiments (Yu et al., 2020, Litschko et al., 2021). XOR-TyDi evaluates non-English question retrieval over English passages, and NeuCLIR evaluates English queries over large Chinese, Persian, and Russian news collections (Li et al., 2021, Jeronymo et al., 2023).

HC4 was explicitly created because older CLIR collections, built by pooling traditional CLIR runs, have systematic gaps in their relevance judgments when used to evaluate neural CLIR methods. HC4 uses Common Crawl News in Chinese, Persian, and Russian, graded three-level relevance judgments, and active learning rather than conventional pooling. It contains 60 topics each for Chinese and Persian, 54 topics for Russian, and baseline results with machine and human translation, BM25+RM3, and multilingual neural rerankers (Lawrie et al., 2022).

The scarcity of large-scale CLIR supervision has also motivated synthetic data generation. JH-POLO starts from naturally occurring non-English passage pairs, uses a generative LLM to produce English queries for which one passage is relevant and the other is not, and validates triples with a multilingual cross-encoder. On NeuCLIR-like news collections this synthetic data was on par with translate-train data derived from machine-translated MS MARCO, and on HC3 tweet collections it substantially outperformed translated MS MARCO as CLIR training data (Mayfield et al., 2023).

The metrics used in CLIR reflect both ranked retrieval and operational constraints. MAP, nDCG, P@k, Recall@k, and R-Precision dominate benchmark evaluation (Varshney et al., 2014, Lawrie et al., 2022). In document-set retrieval settings such as MATERIAL, mAQWV replaces ranked-list metrics because the task is to choose which subset of the collection to return: $\text{mAQWV} = \frac{1}{N_q} \sum_q QV(q_e).$ This emphasizes that evaluation in CLIR is not always reducible to the ranking-only view inherited from monolingual IR (Agarwal et al., 5 Nov 2025).

5. Applications and system-level designs

The application space for CLIR includes web search, multilingual regions and institutions, medical and consumer health search, patent retrieval, security and intelligence analysis, speech search, and cross-language question answering (Galuščáková et al., 2021). In India, English–Hindi CLIR has the pragmatic motivation that users may find it easier to formulate queries in English while preferring to read Hindi content (Varshney et al., 2014). In e-commerce, CLIR supports multilingual search over a shared English inventory, and query translation quality has direct commercial impact on result clicks and purchases (Yao et al., 2020).

System design choices differ sharply by application. In latency-sensitive web or commerce settings, translate-query architectures remain common because the document index can stay monolingual and query translation is cheap enough to run online. The Aliexpress system exemplifies this: a cache serves NMT translations for repeated queries, SMT handles cache misses synchronously, and asynchronous NMT updates the cache in the background (Yao et al., 2020). In high-effectiveness settings, sparse first-stage retrieval followed by multilingual reranking has become a standard architecture, as illustrated by BM25/SPLADE plus mT5-XXL reranking in NeuCLIR (Jeronymo et al., 2023).

Low-resource and speech settings require additional mechanisms. SARAL for MATERIAL combines three evidence generators—SEARCHER shared embeddings, translation tables, and multiple MT systems—into a sentence-level relevance probability $p(\text{rel}\mid s_f,w_e)$, aggregates these hierarchically to document-level relevance, and then chooses an optimal document set by maximizing expected query value under the mAQWV objective. This design is explicitly intended for Farsi, Kazakh, and Georgian text and speech retrieval (Agarwal et al., 5 Nov 2025).

A plausible implication is that CLIR system design is increasingly bifurcated between architectures optimized for online efficiency and architectures optimized for cross-lingual semantic precision. The recent literature supports both directions: fast translation-based query pipelines remain viable in production, while dense retrieval, reranking, distillation, and multilingual LMs increasingly dominate benchmark effectiveness (Yao et al., 2020, Yang et al., 2024, Jeronymo et al., 2023).

6. Limitations, misconceptions, and open directions

A common misconception is that CLIR is simply monolingual IR preceded by machine translation. Translation remains central, but the literature shows that CLIR also depends on transliteration, lexical ambiguity resolution, morphology, evaluation methodology, and user-facing relevance interpretation. Another common misconception is that multilingual encoders have superseded older methods across the board; for unsupervised document-level CLIR, pretrained multilingual encoders did not significantly outperform strong CLWE baselines, even though specialized multilingual sentence encoders did excel at sentence-level retrieval (Galuščáková et al., 2021, Litschko et al., 2021).

Several limitations recur across paradigms. Query translation suffers from short, telegraphic inputs and domain mismatch. Dictionary-based methods suffer from ambiguity, OOV terms, and incomplete morphological coverage. Translation+transliteration query selection can improve effectiveness offline, but selecting the best query by MAP requires relevance judgments and is therefore not directly deployable at query time (Varshney et al., 2014, Dadashkarimi et al., 2014). Neural systems depend on large-scale supervision, yet CLIR relevance data remain scarce; translated MS MARCO introduces translationese and fixed-domain bias, while synthetic data depend on LLM quality and validation heuristics (Mayfield et al., 2023).

Open research directions in the literature include low-resource CLIR, domain adaptation, interaction design, conversational CLIR, microtext and social media, fairness across languages, and multimodal retrieval (Galuščáková et al., 2021). More technically, several lines are active: tailoring histogram bins and kernel placements to cross-lingual similarity distributions, learning end-to-end CLIR with multilingual interaction models, combining knowledge graphs with multilingual encoders, distilling from stronger teachers into efficient CLIR retrievers, and improving calibration for document-set retrieval under detection-style objectives (Yu et al., 2020, Zhang et al., 2021, Li et al., 2021, Yang et al., 2024, Agarwal et al., 5 Nov 2025).

Taken together, the literature depicts CLIR as a mature but still unsettled field. Its central problem—ranking relevance across languages—has not collapsed into a single dominant formulation. Instead, CLIR remains a layered discipline in which lexical translation, transliteration, probabilistic modeling, representation learning, knowledge integration, large-scale supervision, and evaluation design all continue to shape the state of the art.