When a false claim goes viral in one language, it has often already been debunked in another. One study found that about 18% of fact-checks address claims that were already checked in a different language (Comparative Analysis of Engagement, Themes, and Causality of Ukraine-Related Debunks and Disinformation). Finding that earlier debunk is a retrieval problem. A messy social media post written in Polish, Hindi or Thai must be matched against a large database of clean journalistic fact-checks written mostly in other languages. Two things make this hard. The posts are full of slang, hashtags and tangents. And the facts that make a match decisive are easy to lose, because a summarizer drops them or a translator mangles them. Those facts are the person's name, the number, the date, the place.
We propose a small model that rewrites a noisy post, in any language, into a short English claim. While writing, the model can copy those "anchor" tokens (names, numbers, dates, URLs) straight from the post instead of regenerating them. The copying uses a pointer-generator: a decoder that, at each output word, mixes "write a word from the vocabulary" with "copy a word from the input" through a learned switch. We train it on synthetic data made in the reverse direction, starting from real fact-check claims and having a modest open model write fake noisy posts about them. Because the training targets are real human-written claims, a teacher model's mistakes can only corrupt the inputs, never the labels. The claim to test is that this cheap, anchor-preserving condenser, placed in front of an off-the-shelf retriever, closes part of the large gap between monolingual and cross-lingual retrieval. Off-the-shelf retrievers score about 0.83 versus 0.56 Success@10 in Multilingual Previously Fact-Checked Claim Retrieval. The condenser should also cost far less than prompting a large model per post.
Core Claim
The pieces of this idea exist separately, and each has a specific gap.
Condensing noisy tweets before retrieval is known to help. Harnessing Abstractive Summarization for Fact-Checked Claim Detection uses Distilled PEGASUS to turn raw tweets into short claims and raises Recall@5 from 9.52% to 34.91% against Google Fact Check Explorer. That work motivates ours, but it is English-only, uses 567 pairs, and the authors note that the summarizer drops named entities.
Pointer layers are known to keep proper nouns and rare tokens intact. A Pilot Study of Domain Adaptation Effect for Neural Abstractive Summarization shows this under domain shift. A Copy-Augmented Generative Model for Open-Domain Question Answering supplies our mechanism: it reuses decoder cross-attention as the copy distribution, adds only 1,537 parameters, and gains 9.6 Exact Match points on answers unseen in training. Both are monolingual, so copying there means copying the same surface string into the same language.
Our observation is that the tokens worth copying across languages are mostly the ones that do not change across languages: numerals, Latin-script names, dates, handles and URLs. So a pointer used inside a translate-and-condense decoder does not hit the usual cross-lingual objection that verbatim copying cannot translate. We deliberately copy only what should not be translated.
The synthetic-data angle is the cost constraint, not decoration. Promptagator: Few-shot Dense Retrieval From 8 Examples needs a 137B teacher plus round-trip filtering that discards many samples. TrueTeacher: Learning Factual Consistency Evaluation with Large Language Models needs a 540B teacher, and teacher errors flow into its labels.
Reversing the direction of generation changes both costs:
- Filtering. We know the anchors of each source claim before generating, so we can filter a synthetic post with a string or transliteration check. No auxiliary retriever is needed.
- Copy supervision. The same anchor alignment tells us exactly which target tokens should be copied.
Research Program
Data generation. We start from the roughly 205,000 fact-check claims in MultiClaim, which span 39 languages. For each claim we extract anchors with a multilingual tagger plus number and date rules. An open 7–8B instruction model is then prompted with a few real post–claim pairs. It writes a social-media-style post about the claim in a target language, with injected noise: hashtags, slang, emoji, irrelevant asides, and code-switching. It is instructed to keep anchors verbatim or transliterated.
Filtering. A post is kept if its anchors are recoverable. A cheap multilingual similarity floor rejects posts that drift off topic. We log the teacher tokens spent per retained pair, so cost is a measured outcome rather than an assumption.
Model. The condenser is an mT5-base-sized encoder-decoder with a pointer layer built as in FiD-PGN. During training, target tokens that are aligned anchors receive an auxiliary loss that pushes the copy switch toward copying. All other tokens are left to the generator.
Inference. At test time, the condensed English claim is used as the query. It is used alone and also concatenated with the raw post, and is fed to whichever retriever is being compared. This leaves the fact-check index untouched.
Analyses.
- Anchor recall: the fraction of gold-claim anchors present in the condensed query.
- Retrieval by script and by seen versus unseen language: covering Latin, Cyrillic, Devanagari, Thai and Burmese, plus the unannounced Polish and Turkish in SemEval.
- Behaviour under topic density: many near-duplicate COVID-style debunks, where MultiClaim reports retrievers confusing similar claims.
We expect the pointer to help least for non-Latin-script names that need transliteration. That is a limitation we want to measure, not hide.
Baselines
Do nothing. The naive reference point is the raw post fed directly to each retriever. Two versions are run:
- An English-centric pipeline that machine-translates the post and retrieves with GTR-T5-Large. Multilingual Previously Fact-Checked Claim Retrieval reports 0.56 Success@10 cross-lingually for this pipeline.
- Native multilingual encoders with no translation.
Strongest existing competitor. The main competitor is a strong multilingual embedder on raw posts. We use the models CAISA and ClaimCatchers used in SemEval-2025 Task 7: Jina embeddings v2, e5-large and bge-multilingual-gemma2. CAISA at SemEval-2025 Task 7 reports 55.2% Success@10 for Jina standalone and 57.7% when LLaMA-3 8B reranks with a task prompt. The task overview reports top cross-lingual systems around 0.81–0.85 Success@10 (SemEval-2025 Task 7). The file does not describe those systems' components, so we treat that range as a ceiling reference, not a reproducible baseline.
Fair comparison. For every retriever we compare raw post versus raw post plus condensed claim. The retriever weights, index and candidate depth are the same in both arms.
What counts as success. The proposal succeeds if the condenser raises cross-lingual Success@10 over the same retriever on raw posts. That gain must be larger than seed variance across three condenser training runs, and it must hold on the unseen-language split.
Prompted condensation. The same 7–8B model used as teacher is prompted at inference time to rewrite each post into an English claim. It is run once with ordinary decoding and once with LargePiG decoding. LargePiG is a training-free way to make a decoder LLM copy from its input, using its attention plus a layer-divergence copy switch, at about 6% extra latency on a 7B model. This pair separates "a big model rewriting" from "a big model rewriting with copying." We compare on quality and on per-post compute.
Ablations of the key ingredient.
- An identical condenser trained on identical synthetic data with the pointer removed. This isolates copying.
- The pointer kept but the anchor-supervised copy loss removed. This isolates anchor supervision.
- Anchor filtering replaced by Promptagator-style round-trip filtering, holding teacher compute fixed. This isolates the cheaper filter and lets us report discard rates.
Why not just…?
Why not just use a strong multilingual embedder on the raw post? This is the most serious threat. The 9.52% verbatim-query figure in the PEGASUS study comes from a keyword search tool, not a modern dense encoder, so dense encoders may already ignore much of the social noise. Our first experiment is exactly this comparison. If condensing adds nothing on top of bge-multilingual-gemma2 or Jina, the noise-robustness motivation is wrong and we will say so. The remaining hope would then rest on anchor-heavy and topic-dense subsets, which we report separately.
Why not just prompt an LLM to translate and summarize each post? It may work nearly as well. The file has no reported number for prompted rewriting on MultiClaim. CAISA's LLaMA-3 8B result concerns reranking, not rewriting, so it does not settle the question. We will run the prompted-rewrite baseline with and without LargePiG copying and report three things for each: Success@10, anchor recall, and compute per post. If the prompted LLM matches the condenser, the condenser's case reduces to a large drop in per-post inference cost for streaming monitoring. That is still useful, but it is a weaker claim, and we would state it that way.
Why not just machine-translate to English and use an English retriever? MultiClaim already shows this leaves the cross-lingual score at 0.56 against 0.83 monolingual. It also struggles on scripts such as Thai and Burmese. Matching Tweets with Applicable Fact-Checks Across Languages finds LaBSE (55.48% MAP@1) beating translation plus BM25 on Hindi–English. Our condenser is itself a translator, so the relevant test is whether anchor copying beats plain translation on anchor recall. The no-pointer ablation and the machine-translation baseline answer that directly.
Why not just fine-tune the retriever on the same synthetic post–claim pairs? This is a real alternative. It is Promptagator's recipe applied cross-lingually, and it could absorb most of the benefit. We run it on the identical synthetic data with the same tuning budget, and we also test the combination. If retriever fine-tuning alone matches condenser plus retriever, the contribution narrows to the reverse-synthesis data recipe and its cost accounting. The condenser would then be justified only by not needing to re-index and by producing a human-readable claim. If the two are complementary, both stay in the pipeline.
Proposed Experiments
Datasets. The primary evaluation uses the MultiClaim cross-lingual pairs and the SemEval-2025 Task 7 cross-lingual test split, including Polish and Turkish as languages unseen during training. Secondary checks use:
- The Hindi–English tweet set from Matching Tweets, with LaBSE at 55.48% MAP@1 as reference.
- MMTweets, for which no reported result was found on Lacuna.
- The English tweet–claim set from the PEGASUS study, where fine-tuned PEGASUS reports 34.91% Recall@5 and MRR 0.30.
Metrics. Retrieval is scored with Success@10 and MRR. Condensation is scored with anchor recall. Cost is scored as teacher tokens per retained training pair and GPU-seconds per test post.
Decisive checks. Four results decide the proposal's fate:
- Does condensing help at all? Condenser plus the strongest embedder must beat that embedder on raw posts cross-lingually. If not, the idea fails.
- Is copying the source of the gain? The pointer must beat the no-pointer ablation on anchor-rich posts. If it does not, the copy mechanism is not what helps.
- Is the filter cheaper without losing quality? Anchor filtering must match round-trip filtering at lower teacher cost and lower discard rate. If not, the data-cost claim fails.
- Is the small model worth it? The small condenser must come within a few points of the prompted 7–8B rewriter at a fraction of per-post compute. If not, the case for a small model fails.
Diagnostic, not decisive. A per-script breakdown will show where copying stops helping because names need transliteration rather than copying.
References
- Comparative Analysis of Engagement, Themes, and Causality of Ukraine-Related Debunks and Disinformation
- Multilingual Previously Fact-Checked Claim Retrieval
- Harnessing Abstractive Summarization for Fact-Checked Claim Detection
- A Pilot Study of Domain Adaptation Effect for Neural Abstractive Summarization
- A Copy-Augmented Generative Model for Open-Domain Question Answering
- Promptagator: Few-shot Dense Retrieval From 8 Examples
- TrueTeacher: Learning Factual Consistency Evaluation with Large Language Models
- CAISA at SemEval-2025 Task 7: Multilingual and Cross-lingual Fact-Checked Claim Retrieval
- SemEval-2025 Task 7: Multilingual and Crosslingual Fact-Checked Claim Retrieval
- LargePiG: Your Large Language Model is Secretly a Pointer Generator
- Matching Tweets with Applicable Fact-Checks Across Languages