Fact-checking systems for non-English social media mostly do one thing well. They take a viral post and find an existing fact-check article that already debunks it. When a claim is new, or when checking it requires joining two facts that live in different documents (often in different languages, and often only in local news), these systems have nothing to find. Multi-hop retrieval addresses exactly this. It finds a first document, uses what it learned there to find a second, and so on. But the retrievers that do this well learned the skill from Wikipedia's hyperlink graph. That graph is English-heavy, and it covers local politicians, colloquial bill nicknames and regional events poorly, which are the things vernacular misinformation is about.
We propose to teach a multilingual retriever to "hop" using structure that exists in the target ecosystems themselves. The first source of structure is social posts that link to news articles. The second is named entities shared across nearby local-language news stories. The retriever takes the claim plus the document it already found and directly retrieves the next document. It does not ask a general-purpose language model to rewrite the intermediate step as a new search query. Our bet is that this rewriting step is where current iterative pipelines break on non-English content, because general models mangle local names and idioms. A retriever pre-trained on local structure should skip that failure.
Core Claim
A dense retriever trained with a self-supervised "next document prediction" objective has two properties we care about. It is dense in that it retrieves by comparing learned vector representations. It is trained on hop pairs mined from regional news and link-sharing social posts, with no hyperlinks and no labels. We claim such a retriever will recover two-document evidence chains for multilingual claims substantially better than the two obvious alternatives. The first alternative is an iterative loop in which a language model writes each follow-up query. The second is the same hop objective trained on multilingual Wikipedia.
We further predict that the advantage will be concentrated in a specific place. It should appear on claims whose bridge entity, meaning the intermediate person, place or event linking the two documents, is weakly covered in English Wikipedia. If the advantage is instead uniform, or absent, the "local structure" story is wrong. That prediction is what makes the claim falsifiable.
Where This Sits
The closest work is CONCRETE. It pre-trains a cross-lingual claim retriever with a Cross-lingual Inverse Cloze Task: news titles are translated to act as proxy claims, and the retriever must find the source passage in about 347k BBC passages. CONCRETE reached 19.83% macro-F1 on zero-shot X-FACT. Its retrieval is single-hop, however. The authors name multi-step evidence chains and local-news coverage as open problems, and our proposal targets both.
The method we borrow comes from LOUVRE. LOUVRE builds synthetic two-hop chains from Wikipedia hyperlinks by re-phrasing the bridge entity, then trains next document prediction. It reaches PathR@1 of 67.0 on HotpotQA and matches full-data baselines with 1% of the labels. It is monolingual English and depends entirely on hyperlinks. We keep its objective and replace its supervision source.
Two findings constrain the design and push us toward domain-matched hop data rather than more generic data. Multi-Hop Fact Checking of Political Claims found that transfer from LIAR-PLUS helped on its PolitiHop benchmark while transfer from large, generic FEVER did not. X-Fact found that adding 12,000 English PolitiFact claims hurt multilingual performance. It also found that only 45% of top search snippets sufficed to verify a claim.
Three further results motivate the gap. MMTweets formulates cross-lingual fact-checking as matching a tweet to one existing debunk, and it reports a sharp drop on Hindi queries. MultiHop-RAG shows that retrieval, not generation, is the bottleneck even on clean English news: its best Hits@10 is 0.7467, and GPT-4 accuracy falls from 0.89 with gold evidence to 0.56 with retrieved evidence. The social-science findings supply the reason we expect English-centric bridges to fail on local claims. These include vernacular misogyny missed by English analysis (Dehingia et al.), caste prejudice carried by regional idiom (Vijayaraghavan et al.), and the advantage of Hebrew-tuned models on political Facebook text (HebID).
The iterative loop we compete with is IRCoT. IRCoT interleaves chain-of-thought sentences with retrieval, using each generated sentence as the next query. On English benchmarks it improved recall by 11 to 21 points. We do not dispute that iteration helps. We dispute that a language-model-written query is the right bridge when the model does not know the local entity.
Research Program
The first component is hop-pair mining, done without hyperlinks, in four pilot languages: Hindi, Arabic, Portuguese and Spanish. We chose them because they overlap with MMTweets and X-FACT and cover two non-Latin scripts. The corpus is local-language news crawled from regional outlets, plus the BBC multilingual collection, plus public social posts that contain URLs to news articles. Two kinds of pairs come from this corpus.
The first kind is post-to-article pairs. A post that links an article gives a natural first hop, from claim-like text to its source.
The second kind is entity-bridged pairs. We take two articles published within 30 days of each other that share a named entity $e$, where $e$ is found by multilingual entity recognition and matched across scripts by transliteration and embedding similarity. We keep a pair only if $e$ is rare, measured by corpus inverse document frequency, and is central to article $B$ but only described in article $A$. Following LOUVRE's re-phrasing, we build a query from a sentence of $B$ in which $e$ is replaced by the description of $e$ taken from $A$. For example, "the minister who pushed the farm ordinance" replaces a name. Retrieving $B$ therefore requires first finding $A$. A fraction of pairs are made cross-lingual by pairing $A$ and $B$ in different languages, or by translating the query as in CONCRETE's X-ICT.
The second component is the retriever and its objective. We start from a multilingual encoder (LaBSE or XLM-R base) and train it with contrastive next document prediction. Given query $q$, the model must score $A$ above in-batch negatives. Given the concatenation $q \oplus A$, it must score $B$ above negatives, including hard negatives that share surface words with $B$ but not the entity $e$. This is LOUVRE's loss with a different data source.
The third component is the verification pipeline. We use beam search over two to three hops of the trained retriever, followed by a fixed reader that predicts veracity. The reader is a multilingual verification model; as a secondary factor, we also try a regionally pre-trained small language model, motivated by the Arabic-centric results in Rahman et al.. The pipeline makes no language-model calls between hops, which also addresses the latency cost of IRCoT-style loops.
The fourth component is an evaluation set, because none currently exists for multilingual multi-hop claim verification. We sample about 800 claims across the four languages from X-FACT and MMTweets fact-check articles. Annotators, who are native speakers, trace the cited sources in each fact-check article and mark whether verification needs at least two documents, which documents, and what the bridge entity is. They also record whether that bridge entity has an English Wikipedia page, which becomes our locality stratum. The corpus to retrieve from is the pooled news collection with the held-out months removed from pre-training. Labels are gated by agreement between two annotators.
Baselines
The naive reference point is a claim-only classifier with no retrieval. On X-FACT, X-Fact reports 38.2% macro-F1 in-domain for claim-only and 41.9% for Attn-EA+Meta, with both falling to about 15 to 16% on out-of-domain and zero-shot splits. Any retrieval system must clear the zero-shot figure by a wide margin to matter.
The strongest single-hop competitor is CONCRETE. Its reported number is 19.83% zero-shot macro-F1 on X-FACT (CONCRETE). We will rerun it on our corpus with the same reader and report chain recall on our multi-hop set. Beating it means exceeding 19.83 on X-FACT zero-shot, and showing a clear gain on the two-document subset without loss on the single-document subset. On MMTweets single-hop retrieval, the reference is LaBSE bi-encoder plus cross-encoder at 0.795 average MRR (MMTweets). We require our retriever not to fall below it, since hop training should not break direct matching.
The strongest multi-hop competitor is IRCoT, run with an off-the-shelf multilingual retriever (mContriever or LaBSE) and an instruction-tuned open model writing the queries. There is no reported result found on Lacuna for IRCoT on X-FACT or MMTweets, so we produce it ourselves. We give it the same corpus, reader, hop limit and a tuned prompt with an equal development budget. We report its language-model calls and latency alongside accuracy. Beating it means higher two-document chain recall at the same or lower cost.
We also run a translate-to-English variant, which translates claims and runs IRCoT over English plus translated news.
Two ablations isolate our key ingredient. The first trains the same encoder with the same token budget on the same local corpus, but with CONCRETE's single-hop X-ICT objective. This separates the hop objective from the domain data. The second trains LOUVRE's hop objective on multilingual Wikipedia hyperlinks with a matched budget. This separates the domain data from the hop objective. Every retriever shares its initialization, pre-training token count, fine-tuning data and reader.
Why not just…?
Why not just run IRCoT with a good multilingual retriever? This is the serious threat. IRCoT's English gains came from iteration alone, and a strong language model may produce usable follow-up queries. Our counter-hypothesis is narrow. The model's rewritten query degrades when the bridge entity is local or transliterated, and a hop-trained retriever conditioned on the retrieved document sidesteps that step. The locality-stratified comparison settles it. If IRCoT matches us on bridges with poor Wikipedia coverage, the pre-training is unnecessary and we will report that. If IRCoT wins only when its model is much larger, the result becomes a cost trade-off rather than a win.
Why not just translate everything to English? MMTweets found translation plus BM25 consistently beaten by native cross-lingual encoders, with particular trouble on Devanagari. CONCRETE found that evidence for local claims often never appears in English outlets, which translating the query cannot fix. The translate-to-English baseline measures this directly on our evaluation set.
Why not just run LOUVRE on Wikipedia in each language? Non-English Wikipedias do have hyperlinks, so this is cheap and plausible. Against it stands the transfer evidence: FEVER-style generic data did not help political claims in PolitiHop, and off-domain English claims hurt X-FACT. Wikipedia's coverage of local actors is also exactly what our locality stratum measures. The matched-budget Wikipedia ablation decides this, and a tie would demote our data pipeline to an engineering detail.
Why not just prompt a frontier model with web search? X-Fact's snippet audit (45% sufficient) and CONCRETE's analysis of search-engine noise both argue against single search calls. MMTweets also found GPT-3.5 no better than a cross-encoder at about seven times the latency. We include one frontier-model-with-search run as a reference ceiling on the evaluation set, reported but not tuned.
Decisive Checks
The first check is two-document chain recall at 10 on our evaluation set, split by whether the bridge entity has an English Wikipedia page. The central prediction is that the gap between our retriever and both IRCoT and the Wikipedia-hop ablation is at least twice as large on poorly covered bridges as on well-covered ones. The second check is X-FACT zero-shot macro-F1 against CONCRETE's 19.83. The third checks that nothing broke: MMTweets MRR must stay at or above 0.795. As an English sanity check, we fine-tune on PolitiHop and compare against Transformer-XH's 47.2 evidence F1 and 24.5 FEVER score, which shows whether local hop pre-training transfers or at least does no harm. The fourth is a noise audit. We sample 200 mined entity-bridged pairs per language and have them hand-checked for a true bridge relation, and we check whether chain recall degrades as we lower the rarity filter. This tells us whether the method depends on clean mining.
If our retriever beats the single-hop ablation but not the Wikipedia-hop ablation, the hop objective matters and local data does not. If the reverse holds, domain data matters and hopping does not. Either result is informative. Only a win over both, concentrated on local bridges, supports the full claim.
References
- CONCRETE: Improving Cross-lingual Fact-checking with Cross-lingual Retrieval
- Weakly Supervised Pre-Training for Multi-Hop Retriever
- Multi-Hop Fact Checking of Political Claims
- X-Fact: A New Benchmark Dataset for Multilingual Fact Checking
- Breaking Language Barriers with MMTweets: Advancing Cross-Lingual Debunked Narrative Retrieval for Fact-Checking
- MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries
- Violence against women on Twitter in India: Testing a taxonomy for online misogyny and measuring its prevalence during COVID-19
- Modeling Social Media Narratives about Caste-related News Stories
- HebID: Detecting Social Identities in Hebrew-language Political Text
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Do LLMs Distinguish Between Halal and Haram? Benchmarking Islamic Cultural Alignment in General vs. Arabic-Centric SLMs