Skip to content

Cross-lingual Biography Enrichment via Claim Extraction and Alignment

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/MeloS7/cross_lingual_biography_enrichment
Area: Multilingual & Machine Translation
Keywords: Cross-lingual Biography Enrichment, Claim Extraction and Alignment, Wikipedia Information Asymmetry, Factuality and Hallucination Control, Structured Evidence Rewriting

TL;DR

Addressing the pronounced knowledge asymmetry where non-English Wikipedia articles document long-tail figures with far richer detail than English Wikipedia, this paper presents a claim-based cross-lingual biography enrichment framework that extracts atomic claims into a shared English space, filters already covered information via fine-grained semantic alignment, and rewrites the English article with selected novel claims, achieving significantly higher supported additions with substantially lower hallucination rates.

Background & Motivation

Wikipedia is an indispensable foundational resource for natural language processing, providing vital supervision and factual grounding for language model pretraining, multilingual representation learning, retrieval-augmented generation (RAG), and open-domain verification. Nonetheless, encyclopedic coverage across Wikipedia language editions remains deeply asymmetrical: the same historical or contemporary entity is frequently described with starkly divergent granularities and breadths. While English Wikipedia is almost universally treated as the default, authoritative encyclopedic repository, it is rarely the most comprehensive source for long-tail entities whose cultural relevance and primary documentation reside in non-English communities. This information gap is particularly acute for notable women from non-English-speaking contextsโ€”such as French, Chinese, and Azerbaijani scientists, artists, and leadersโ€”who enjoy exhaustive, community-curated biographies on their local Wikipedia editions while being relegated to brief stubs in English.

Previous efforts to improve biographical coverage either frame the task as data-to-text generation from Wikidata infobox triplesโ€”typically generating single-sentence leads or brief summariesโ€”or synthesize full-length articles using open-web search and retrieval-augmented generation. However, structured data-to-text remains fundamentally limited in discourse depth and nuance, whereas web-retrieval pipelines suffer severely from the scarcity of high-quality English web evidence for regional figures. Exploiting existing, human-curated non-English Wikipedia articles represents a compelling alternative, but direct end-to-end cross-lingual generation presents formidable modeling hurdles. Feeding a short English biography alongside a lengthy raw non-English counterpart (which reaches up to 28,420 tokens in real-world cases) forces the language model to concurrently decode foreign-language discourse, perform implicit cross-lingual alignment, discard redundant facts, and draft coherent English prose. This cognitive overload induces severe attention dispersion and factual hallucinations. Alternatively, translating the non-English biography before generation reduces the cross-lingual decoding burden but introduces cascading translation noise, syntactic clunkiness, and redundant passages, leaving the average hallucination rate unacceptably high (~29.6%).

The pivotal insight of this work is that long-context cross-lingual enrichment must be decoupled into explicit evidence selection in an atomic claim space prior to generation, followed by controlled text rewriting. By projecting both biographies into a shared space of self-contained English factual propositions, the model can perform rigorous set-difference filtering to isolate genuinely novel, non-redundant information. Core idea: decompose cross-lingual biography enrichment into "cross-lingual atomic claim extraction into a shared English space โ†’ fine-grained semantic alignment and novel evidence selection โ†’ controlled biographical rewriting conditioned on selected claims", which removes redundant context before generation and uses atomic claims as structured factual anchors to simultaneously maximize supported knowledge gains and suppress hallucination.

Method

Overall Architecture

Given a paired biography \((B_{en}, B_X)\) for the same person, where \(B_{en}\) is an existing English Wikipedia article and \(B_X\) is a corresponding non-English Wikipedia article in French, Chinese, or Azerbaijani, the goal is to synthesize an enriched English biography \(\tilde{B}_{en}\) that preserves the integrity of \(B_{en}\) while integrating novel factual information supported by \(B_X\). The framework operates across five sequential stages: (1) inputting paired cleaned biographies; (2) performing cross-lingual claim extraction to project both \(B_{en}\) and \(B_X\) into English atomic claim sets \(C_{en}\) and \(C_X\); (3) executing coarse-to-fine claim-pair alignment to classify semantic relationships; (4) selecting non-English-side enrichment claims \(C_{add}\) by discarding redundant facts already expressed in \(B_{en}\); and (5) applying a section-level incremental rewriting generator conditioned on \(B_{en}\) and the structured evidence \(C_{add}\).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Paired Biographies<br/>(English B_en and Non-English B_X)"] --> B["Cross-Lingual Claim Extraction & Decontextualization<br/>(X-Claimify multi-turn atomic English projection)"]
    B --> C["Coarse-to-Fine Claim Alignment & Classification<br/>(MPNet retrieval filter + Qwen3.5-9B 6-way relation verifier)"]
    C --> D["Enrichment Evidence Selection & Conflict-Aware Filtering<br/>(Filter covered facts, retain novel & conflict claims C_add)"]
    D --> E["Output: Controlled Enriched English Biography"]

Key Designs

1. Cross-Lingual Claim Extraction & Decontextualization: Cross-lingual syntactic divergence and pronoun ambiguity in isolated propositions -> Multi-turn verification, disambiguation, and decomposition pipeline -> Canonical, self-contained atomic English claims
Extracting claims directly in source languages complicates subsequent cross-lingual matching and risks carrying context-dependent pronouns (e.g., "she", "this work") that cannot be verified out of context. To overcome this, the authors adapt the Claimify framework into X-Claimify, a multi-turn prompting pipeline that projects both English and non-English texts directly into self-contained English atomic claims. The pipeline executes three disciplined stages: first, Verifiable Selection filters out non-factual discourse such as subjective praise, evaluations, and transitional meta-text, retaining only checkable factual sentences; second, Disambiguation leverages the full paragraph context and biographical subject to resolve pronouns into full entity names and ground implicit temporal/spatial references; third, Decomposition decomposes compound clauses into independent atomic propositions expressed in English. To ensure rigorous evaluation against human annotations without inflation from one-to-many matches, the authors enforce degree-normalized precision and recall credits: $\(p_k(i) = \frac{1}{|\{j' : (i, j') \in E_k\}|}, \quad r_k(j) = \frac{1}{|\{i' : (i', j) \in E_k\}|}\)$ X-Claimify establishes a state-of-the-art benchmark on the human-annotated CLAW-4L-CX dataset, reaching 71.18% Exact-Aligned-F1 across four languages and outperforming prior extraction frameworks by over 20 points.

2. Coarse-to-Fine Claim Alignment & Classification: Intractable Cartesian claim pairing and limited granularity of standard NLI labels -> Two-stage MPNet embedding retrieval coupled with a fine-grained 6-way LLM classifier -> Rapid candidate pruning and asymmetric semantic enrichment modeling
With non-English biographies yielding up to 1,363 extracted claims, full Cartesian verification across all claim pairs is computationally prohibitive (\(O(|C_{en}| \times |C_X|)\)), while conventional Natural Language Inference (NLI) collapses rich relational dynamics into mere entailment, neutral, and contradiction. The authors introduce a coarse-to-fine alignment pipeline. In the coarse stage, candidate English claims for each non-English claim are retrieved using All-MPNet-Base-v2 sentence embeddings with a cosine similarity threshold of 0.7 and a top-5 nearest-neighbor cutoff, safely eliminating 96% of irrelevant pairings while maintaining near-perfect candidate recall (mean cosine similarity of aligned pairs spans 0.843 to 0.960). In the fine stage, a specialized Qwen3.5-9B classifier operates on the retained pairs under a two-tiered, 6-way schema: top-level relations comprise Aligned, Contradicted (\(A \perp B\)), and Not Relevant (\(A \dashv B\)); Aligned is further decomposed into Exact Equivalence (\(A = B\)), English-More-Specific (\(A > B\)), Non-English-More-Specific (\(B > A\)), and Mutual Enrichment (\(A \leftrightarrow B\)). On CLAW-4L-RC, Qwen3.5-9B attains 95.5% macro ARC and 89.5% Align-FG accuracy, rivaling proprietary GPT-5.1 (95.9% / 90.8%) while bypassing brittle structured infobox parsing.

3. Enrichment Evidence Selection & Conflict-Aware Filtering: Factual redundancy, Wikipedia narrative discordance, and generator context overflow -> Rule-based asymmetric filtering and section-level incremental rewriting -> High-density factual enrichment with strict hallucination control
Equipped with predicted claim relationships, the framework executes a deterministic set-theoretic filtering protocol to curate the enrichment evidence set \(C_{add}\). Non-English claims categorized as Exact Matches (\(A = B\)) or English-More-Specific (\(A > B\)) are deemed already covered and discarded (accounting for 13.9% of non-English claims). Conversely, Non-English-Additive claims (\(B > A\) and \(A \leftrightarrow B\), totaling 26.2%) and completely Unmatched claims (\(A \dashv B\), comprising 48.6%) are retained as novel evidence. Crucially, classifier-labeled Contradictions (\(A \perp B\), 11.4%) are not automatically pruned: because both sources are human-edited encyclopedias, discrepancies often reflect complementary perspectives or date shifts that require contextual reconciliation by the generator. Overall, 86.2% of non-English claims are selected. To rewrite biographies without exceeding LLM context windows or losing local cohesion, the system processes evidence incrementally by section: the original English biography is bounded at 4,096 tokens, while non-English claims are fed in section-aligned chunks of up to 2,048 tokens, rolling forward intermediate rewrites to produce the final comprehensive biography.

Cross-Lingual Alignment and Generation Strategy

During the generation step, the generator LLM receives the source English biography \(B_{en}\) together with the filtered enrichment claims \(C_{add}\). The rewriting instruction mandates that the model preserve the factual propositions of \(B_{en}\) while organically weaving \(C_{add}\) into the chronologically and thematically appropriate sections (e.g., placing novel early-education claims into the early life section). For claims flagged as conflicting, the model is prompted to synthesize a coherent, neutral encyclopedic account without speculative extrapolation.

To evaluate biographical enrichment objectively in the shared semantic space, an evaluation pipeline extracts claims \(\hat{C}_{gen}\) from the generated text using X-Claimify (with Gemma-4-31B-it) and compares them against the combined reference pool \(C_{ref} = C_{en} \cup C_X\): - Claim Gain: \(\#Gain = |\hat{C}_{gen}| - |C_{en}|\) - Supported Additions: \(\#Valid = Valid\% \times \#Gain\), where \(Valid\%\) is the percentage of generated claims verified as supported by \(C_{ref}\) (\(A=B\) or \(A>B\)) - Hallucination Rate: \(Halluc.\% = Unsup.\% + Contr.\%\) - Balanced Trade-off Score (Bal.): $\(Bal_s = 50 \times \left( \frac{\#Valid_s - \min(\#Valid)}{\max(\#Valid) - \min(\#Valid)} + \frac{Valid\%_s - \min(Valid\%)}{\max(Valid\%) - \min(Valid\%)} \right)\)$ balancing normalized factual growth and reliability on a 0โ€“100 scale.

Key Experimental Results

Main Results

The primary evaluation on CLAW-4L compares the three evidence representationsโ€”Raw (unprocessed non-English text), Translation (machine-translated non-English text), and Claims (our claim-based evidence)โ€”across three representative open-weight instruction-tuned generators: Gemma-4-31B-IT, Qwen3.6-27B, and Mistral-Small-3.2-24B-IT.

Evidence Representation Generator Model #Gain Coverage% Valid% #Valid Halluc.% Bal. Score
Raw (Direct Input) Gemma-4-31B-IT +19.71 25.99 71.43 14.09 28.57 21.05
Raw Qwen3.6-27B +25.14 33.19 67.41 16.96 32.59 20.22
Raw Mistral-3.2-24B-IT +22.85 30.32 72.68 16.62 27.32 33.67
Translation (Full MT) Gemma-4-31B-IT +22.01 29.07 70.01 15.43 29.99 21.91
Translation Qwen3.6-27B +26.54 35.16 68.15 18.11 31.85 26.40
Translation Mistral-3.2-24B-IT +32.29 42.58 73.08 23.61 26.92 59.99
Claims (Ours) Gemma-4-31B-IT +23.78 31.45 74.98 17.85 25.02 44.49
Claims Qwen3.6-27B +29.88 39.31 73.65 22.01 26.35 55.79
Claims (Best Setting) Mistral-3.2-24B-IT +26.67 34.89 80.77 21.56 19.23 74.02

Note: Table aggregates language splits from paper Table 28 across Azerbaijani (AZ), French (FR), and Chinese (ZH), aligning with Table 22 averages. Mistral-3.2 under Claims achieves peak individual Bal. scores of 89.57 on French and 81.42 on Chinese.

Ablation Study: Claim Extraction Frameworks on CLAW-4L-CX

To validate the upstream extraction backbone, five claim extraction frameworks were evaluated on the 600-sentence human-annotated CLAW-4L-CX benchmark with GPT-5.1 as the shared engine.

Extraction Framework Pipeline Architecture EN E-F1 FR E-F1 ZH E-F1 AZ E-F1 Macro Avg. E-F1 Macro Avg. A-F1
X-DnDScore Decompose first, decontextualize later 11.61 16.08 16.02 22.87 16.65 44.40
X-FactScore Direct joint single-step decomposition 27.83 35.23 26.30 39.51 32.22 54.33
X-FactCheck-GPT Two-stage decomposition + checkability 43.70 49.87 49.82 44.38 46.94 57.34
X-VeriScore Single-prompt joint extraction & coref 46.92 52.91 42.94 55.11 49.47 68.31
X-Claimify (Ours) Multi-turn selection, disambig. & decomp. 72.44 73.56 71.91 66.80 71.18 73.72

Note: Data from Table 2 and Table 17. A-F1 denotes soft Aligned-F1; E-F1 denotes strict Exact-Aligned-F1.

Key Findings

  • Decisive Hallucination Suppression: Replacing raw text or full translation with structured claims reduces overall hallucination from 29.59% to 23.54% on average across all generators. In Mistral outputs, hallucination drops to an impressive 15.95% on Chinese and 19.69% on French, confirming that pre-filtering redundancy and presenting self-contained facts prevents LLM confabulation.
  • Lower-Resource Error Propagation: Azerbaijani represents the sole setting where Translation achieves a higher Bal. score than Claims with Mistral (56.54 vs. 51.07). Upstream extraction on Azerbaijani is noticeably more error-prone (AZ Exact-F1 of 66.80% vs. 73.56% on FR), inadvertently filtering out valid non-English evidence before generation. However, for hallucination-prone generators like Gemma and Qwen, Claims remains superior even on Azerbaijani.
  • Preserved Encyclopedic Prose Quality: A human evaluation of 45 Mistral-generated biographies across three expert annotators demonstrates that providing atomic claims does not fragment text fluency. In fact, Claims achieved the highest Readability/Fluency (3.933 vs. Raw 3.844 vs. Translation 3.756) and Wikipedia Style scores (3.822 vs. Raw 3.778 vs. Translation 3.689), with Coherence remaining competitive (3.867 vs. 3.911).

Highlights & Insights

  • Formulating Cross-Lingual Enrichment as Set Difference: The paper ingeniously maps unstructured, non-parallel multilingual documents into a shared English claim space, transforming a messy text-generation challenge into a clean set-difference operation (\(C_{add} = C_X \setminus C_{en}\)) that eliminates hallucinations at the source.
  • Calibration against Multilingual Tokenization Bias: When curating candidate biography pairs, the authors introduce a tokenizer inflation factor \(b_L\) computed on parallel FLORES+ corpora (\(b_{FR}=1.41\), \(b_{AZ}=1.36\), \(b_{ZH}=0.91\)), preventing surface-level subword segmentation artifacts from skewing content-richness estimation.
  • Generalizable Structured Generation Blueprint: The pipeline of "extract atomic claims \(\rightarrow\) coarse-to-fine semantic graph alignment \(\rightarrow\) condition-filtered rewriting" provides a robust, deployable blueprint for enterprise document consolidation, multi-source intelligence synthesis, and clinical report reconciliation.

Limitations & Future Work

  • Constrained Demographic and Linguistic Scope: The empirical benchmark is restricted to biographies of women across three non-English languages (French, Chinese, Azerbaijani). Extending the methodology to low-resource African and Indigenous languages, non-Latin scripts, and non-biographical encyclopedic domains remains to be demonstrated.
  • Wikipedia-as-Ground-Truth Assumption: The system presumes that human-curated Wikipedia entries are factually sound. When language editions conflict, the framework preserves the discrepancy for LLM reconciliation rather than grounding against external primary archives or trusted knowledge graphs.
  • Future Direction: Generate-Verify-Revise Loops and Bidirectional Enrichment: Future iterations could integrate a post-generation self-checking loop using claim-level verifiers (e.g., MiniCheck) to prune unsupported statements. Additionally, the paradigm should be inverted to enable bidirectional and many-to-many enrichment, utilizing English Wikipedia to expand underrepresented local editions.
  • vs Data-to-Text Biography Generation (Lebret et al., 2016; Mille et al., 2024): Structured data generation relies on rigid tabular infoboxes and produces short summaries; this work operates over unaligned, full-length prose to enrich complex narratives.
  • vs Retrieval-Augmented Biography Generation (Fan and Gardent, 2022; Zhang et al., 2025): Web-retrieval methods fail on long-tail figures due to the absence of English web pages; this work taps into native-language Wikipedia communities as high-quality, curated knowledge hubs.
  • vs Monolingual Claim Extraction Tools (Min et al., 2023 FActScore; Metropolitansky and Larson, 2025 Claimify): Prior claim extraction models were strictly engineered for monolingual English verification; this research extends claim decomposition and decontextualization to cross-lingual projection.

Rating

  • Novelty: โญโญโญโญโญ Formulates the cross-lingual biography enrichment task and introduces an elegant shared-claim-space alignment framework.
  • Experimental Thoroughness: โญโญโญโญโญ Comprehensive evaluation on 300 full biography pairs, 600 annotated extraction sentences, and 600 relation pairs across 3 languages, 5 extraction backbones, 6 classifiers, 3 generation architectures, and expert human audits.
  • Writing Quality: โญโญโญโญโญ Outstanding clarity, thorough mathematical definitions, and meticulous documentation of hyper-parameters and prompts.
  • Value: โญโญโญโญโญ Highly actionable and impactful methodology for mitigating multilingual and gender knowledge disparities across global knowledge bases.