mLLeQA: Translation-Augmented Dense Retrieval for Multilingual Legal Question Answering

mLLeQA: Translation-Augmented Dense Retrieval for Multilingual Legal Question Answering | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article mLLeQA: Translation-Augmented Dense Retrieval for Multilingual Legal Question Answering Rachit Verma, Alluri Lakshman Narendra, Abhinav Shankar, Riya Chitnis, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9448649/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Multilingual legal question answering remains underexplored, despite its practical importance in jurisdictions like the European Union, where citizens have a constitutional right to access legal information in their native language. However, constructing multilingual legal benchmarks is prohibitively expensive, requiring annotators with both multilingual proficiency and legal expertise. This paper investigates whether high-quality neural machine translation can produce synthetic multilingual legal corpora that preserve retrieval utility for downstream dense retrieval systems. We present mLLeQA, a synthetic multilingual benchmark derived by translating the French LLeQA legal retrieval dataset into five additional European languages (Dutch, English, Finnish, Italian,and Spanish) using Seamless-M4T-v2-large. We train language-specific BERT encoders (MDRL) and multilingual BERT encoders (MDRM) as dense retrievers using contrastive learning with hard negatives, and evaluate retrieval performance against BM25 lexical baselines across 195 test queries per language. Language-specific BERT retrievers show no statistically significant performance degradation from the French baseline (p > 0.05), with average cross-language deviation from the French baseline of 5.14 pp. Dense retrievers consistently outperform BM25 across all languages, achieving 29.49 – 62.47 pp absolute improvement on Recall@500. Notably, Finnish (typologically distant) shows minimal deviation (1.61 pp) while Italian exhibits maximal deviation (10.12 pp), suggesting that pre-training corpus characteristics dominate linguistic similarity effects. BLEU scores prove poor predictors of dense retrieval performance but correlate with lexical retrieval, revealing disconnect between surface-level translation quality metrics and semantic task utility. These findings suggest translation-augmented dense retrieval as a potentially viable and cost-effective pathway toward multilingual legal information access in the EU, with immediate applications in retrieval-augmented generation pipelines for citizen-facing legal question-answering systems. All codes, datasets and output files are available for review at https://github.com/rachitprojects/mLLeQAAllCodeResultsUpload legal multilingual retrieval neural machine translation dense passage retrieval contrastive learning multilingual corpora synthesis natural language processing Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9448649","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":624936844,"identity":"f928cd76-9509-48fb-9bd1-b559af86aba7","order_by":0,"name":"Rachit Verma","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIie3QsQrCMBCA4StCXE67RoT0FVL6Qi2Cs5M4xqWOXV18D8eUG7qUunZwUAQ3QXB1MLWDiJB2dMg/JSEfXALgcv1lAzohCGTDtdLA9PtM2wmbS4RIjDFXWvcjKLkhkeCxuco6Ljf5mZbytpdJOrms6b48gr/RHi0shNdxfN6VhkwTM1h1BV7GQFsLkRx0NEqfLclTAqgBCG3Ez9V0lDaD5S0JOgnMoCER415LZBfh9ZyF5i2CoRmsrAjDMlFW4meHx8n8GAab4nJfLUmIguhhI+889Vnj97YPcblcLtdPL4epVzqC99UPAAAAAElFTkSuQmCC","orcid":"","institution":"IIT Dharwad","correspondingAuthor":true,"prefix":"","firstName":"Rachit","middleName":"","lastName":"Verma","suffix":""},{"id":624936845,"identity":"e9321e39-8703-4193-a468-dd3db2e46926","order_by":1,"name":"Alluri Lakshman Narendra","email":"","orcid":"","institution":"IIT Dharwad","correspondingAuthor":false,"prefix":"","firstName":"Alluri","middleName":"Lakshman","lastName":"Narendra","suffix":""},{"id":624936846,"identity":"85e2b1a5-4ce1-4b58-9638-366dfa386cc4","order_by":2,"name":"Abhinav Shankar","email":"","orcid":"","institution":"IIT Dharwad","correspondingAuthor":false,"prefix":"","firstName":"Abhinav","middleName":"","lastName":"Shankar","suffix":""},{"id":624936847,"identity":"56c2f0cc-7a18-47ad-ad07-26a9e9ec4a61","order_by":3,"name":"Riya Chitnis","email":"","orcid":"","institution":"IIT Dharwad","correspondingAuthor":false,"prefix":"","firstName":"Riya","middleName":"","lastName":"Chitnis","suffix":""},{"id":624936848,"identity":"97d81be1-a32f-43bf-90f6-82b79b74146c","order_by":4,"name":"Achyut Mani Tripathi","email":"","orcid":"","institution":"IIT Dharwad","correspondingAuthor":false,"prefix":"","firstName":"Achyut","middleName":"Mani","lastName":"Tripathi","suffix":""},{"id":624936850,"identity":"a6dfe954-016a-426c-ba13-f692675dcae0","order_by":5,"name":"Konjengbam Anand","email":"","orcid":"","institution":"IIT Dharwad","correspondingAuthor":false,"prefix":"","firstName":"Konjengbam","middleName":"","lastName":"Anand","suffix":""}],"badges":[],"createdAt":"2026-04-17 11:40:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9448649/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9448649/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107706819,"identity":"1efc7e59-835f-4bae-8847-3ea99c7bf67f","added_by":"auto","created_at":"2026-04-24 09:18:49","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":999659,"visible":true,"origin":"","legend":"","description":"","filename":"MultilingualLLeQAJIIS16042026.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9448649/v1_covered_886ace72-5593-44da-b2b3-4334cbba2bb4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"mLLeQA: Translation-Augmented Dense Retrieval for Multilingual Legal Question Answering","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"legal multilingual retrieval, neural machine translation, dense passage retrieval, contrastive learning, multilingual corpora synthesis, natural language processing","lastPublishedDoi":"10.21203/rs.3.rs-9448649/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9448649/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Multilingual legal question answering remains underexplored, despite its practical importance in jurisdictions like the European Union, where citizens have a constitutional right to access legal information in their native language. However, constructing multilingual legal benchmarks is prohibitively expensive, requiring annotators with both multilingual proficiency and legal expertise. This paper investigates whether high-quality neural machine translation can produce synthetic multilingual legal corpora that preserve retrieval utility for downstream dense retrieval systems. We present mLLeQA, a synthetic multilingual benchmark derived by translating the French LLeQA legal retrieval dataset into five additional European languages (Dutch, English, Finnish, Italian,and Spanish) using Seamless-M4T-v2-large. We train language-specific BERT encoders (MDRL) and multilingual BERT encoders (MDRM) as dense retrievers using contrastive learning with hard negatives, and evaluate retrieval performance against BM25 lexical baselines across 195 test queries per language. Language-specific BERT retrievers show no statistically significant performance degradation from the French baseline (p \u003e 0.05), with average cross-language deviation from the French baseline of 5.14 pp. Dense retrievers consistently outperform BM25 across all languages, achieving 29.49 – 62.47 pp absolute improvement on Recall@500. Notably, Finnish (typologically distant) shows minimal deviation (1.61 pp) while Italian exhibits maximal deviation (10.12 pp), suggesting that pre-training corpus characteristics dominate linguistic similarity effects. BLEU scores prove poor predictors of dense retrieval performance but correlate with lexical retrieval, revealing disconnect between surface-level translation quality metrics and semantic task utility. These findings suggest translation-augmented dense retrieval as a potentially viable and cost-effective pathway toward multilingual legal information access in the EU, with immediate applications in retrieval-augmented generation pipelines for citizen-facing legal question-answering systems. All codes, datasets and output files are available for review at https://github.com/rachitprojects/mLLeQAAllCodeResultsUpload","manuscriptTitle":"mLLeQA: Translation-Augmented Dense Retrieval for Multilingual Legal Question Answering","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-23 02:44:23","doi":"10.21203/rs.3.rs-9448649/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"120f4ee3-2364-40d7-8da6-39bd52e5dbf1","owner":[],"postedDate":"April 23rd, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-04-30T04:04:05+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-30T04:01:12+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-30T03:43:46+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-30T04:09:28+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-23 02:44:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9448649","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9448649","identity":"rs-9448649","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}