Cross-lingual Biography Enrichment via Claim Extraction and Alignment
Proposes a claim-based cross-lingual biography enrichment framework leveraging non-English Wikipedia, achieving 75.17 F1 in claim extraction and 86.2% claim coverage.
Key Findings
Methodology
This paper introduces a claim-centric framework for cross-lingual biography enrichment, involving four main steps: claim extraction from English and non-English biographies using models like FactScore and Claimify; semantic claim alignment via LaBSE to identify potential evidence; fine-grained relation classification with GPT-5.1 to determine whether claims are aligned, contradictory, or supplementary; and finally, content generation with models such as Qwen3.6-27B to incorporate validated claims into the English biography. The system effectively filters out redundant or conflicting claims, ensuring factual consistency. The approach is evaluated on the CLAW-4L benchmark, which includes 300 paired biographies across four languages, with detailed claim annotations and relation labels. The methodology emphasizes explicit evidence selection before generation, reducing hallucination and improving factual accuracy.
Key Results
- The claim extraction model (X-Claimify) achieved an F1 score of 75.17 on CLAW-4L-CX, with claim coverage reaching 86.2%, significantly enhancing the richness of enriched biographies. The filtering process effectively removed 13.9% of claims already covered by English, focusing on novel information.
- In experiments comparing raw, translated, and claim-driven generation, the claim-based approach supported 20% more factual additions while reducing hallucination rates to below 10%. Multilingual evaluations demonstrated consistent improvements across French, Chinese, and Azerbaijani, especially in low-resource settings.
- Ablation studies confirmed that fine-grained claim relation classification and semantic filtering are crucial for balancing content richness and factual correctness, with claim filtering notably improving the precision of enriched biographies.
Significance
This work advances the state-of-the-art in multilingual knowledge integration by systematically leveraging non-English Wikipedia as a rich knowledge source. Its claim-centric approach offers a controllable, interpretable, and scalable solution for automatic biography enrichment, addressing long-standing issues of content gaps and inaccuracies in low-resource languages. The framework's ability to reduce hallucination and improve factual fidelity has broad implications for knowledge base construction, multilingual NLP, and AI-powered content creation, fostering more inclusive and accurate information ecosystems.
Technical Contribution
The paper introduces a novel claim-based pipeline combining semantic claim alignment, relation classification, and controlled content generation. It innovates with fine-grained claim relation labels (exact, enrichment, contradiction) and a filtering mechanism that explicitly selects non-redundant, relevant claims before generation. The integration of large-scale multilingual models (LaBSE, GPT-5.1, Qwen3.6-27B) enables robust cross-lingual transfer, while the newly constructed CLAW-4L benchmark provides a comprehensive dataset for future research. The approach demonstrates significant improvements over baseline methods, notably in reducing hallucination and enhancing factual coverage.
Novelty
This research is the first to systematically combine claim extraction, fine-grained relation classification, and content generation for cross-lingual biography enrichment. Unlike prior works that rely solely on translation or structured data, this framework explicitly filters and incorporates non-English claims, ensuring factual relevance and reducing hallucination. Its claim-centric design and multi-model integration set a new standard for multilingual knowledge augmentation.
Limitations
- Despite high accuracy, claim extraction and alignment still face challenges in low-resource languages due to limited training data and linguistic variability, leading to potential missed facts or misalignments.
- The relation classifier may produce errors in complex or ambiguous contexts, affecting the quality of claim filtering and subsequent content generation.
- While hallucination rates are reduced, the generated biographies still occasionally contain inaccuracies, especially when dealing with conflicting claims or sparse evidence. Further refinement of relation modeling and multi-modal integration is needed.
Future Work
Future directions include integrating multi-modal data such as images and structured databases to enhance claim verification and enrichment robustness. Developing more sophisticated relation models capable of handling nuanced or conflicting claims will further improve factual fidelity. Additionally, expanding the benchmark to include more languages and domain-specific biographies will facilitate broader applicability and推动跨语种知识图谱的自动构建与推理能力。
AI Executive Summary
In the era of digital knowledge, Wikipedia stands as a cornerstone resource, yet its coverage across languages remains uneven. While English Wikipedia is extensive, it often lacks detailed information about long-tail entities, especially those rooted in non-English-speaking communities. Conversely, non-English editions frequently contain richer, locally grounded facts that are absent from the English version. This disparity hampers the development of comprehensive multilingual knowledge bases and limits the effectiveness of NLP applications relying on factual data.
Traditional approaches to enriching biographies have focused on structured data extraction or web retrieval, but these methods struggle to leverage the nuanced, narrative-rich content found in non-English Wikipedia articles. Recognizing this gap, the authors propose a novel claim-based framework that systematically extracts, aligns, and filters factual claims from non-English biographies to enrich English biographies. The core innovation lies in treating claims as atomic factual units, which are semantically aligned across languages using models like LaBSE, and then classified into detailed relation categories with GPT-5.1. This fine-grained classification distinguishes between claims that are exact matches, partial overlaps, or contradictory, enabling precise selection of evidence.
The framework employs large-scale language models such as Qwen3.6-27B for content generation, integrating only those claims verified as relevant and novel. Extensive experiments on the CLAW-4L benchmark, which includes 300 paired biographies in French, Chinese, and Azerbaijani, demonstrate that the claim-driven approach significantly outperforms raw translation or naive retrieval methods. The enriched biographies exhibit a 20% increase in factual content support and a reduction of hallucination rates below 10%, confirming the effectiveness of explicit evidence filtering.
This research marks a substantial step forward in multilingual NLP, offering a scalable, interpretable, and accurate method for knowledge augmentation. Its implications extend to automatic knowledge base construction, AI-powered content creation, and cross-lingual information retrieval, fostering a more inclusive and factually reliable digital ecosystem. Future work aims to incorporate multi-modal data, enhance relation modeling, and expand to more languages and domains, pushing the boundaries of automated, trustworthy knowledge integration.
Deep Dive
Abstract
English Wikipedia is often treated as the default encyclopedic source, yet non-English Wikipedia editions can contain richer locally grounded information for long-tail figures. We study cross-lingual biography enrichment: enriching an existing English biography with facts supported by a non-English biography about the same person. Focusing on women from non-English-speaking contexts, we introduce \textsc{CLAW-4L}, a benchmark consisting of 300 Wikipedia biography pairs linking an English biography with its French, Chinese or Azerbaijani counterpart, along with claim annotations and a fine-grained claim-pair relation corpus. We propose a claim-based enrichment framework that extracts English claims from both biographies, aligns them to identify enrichment evidence from the non-English biography, and rewrites the English biography using the selected claims. Our results show that non-English Wikipedia biographies provide valuable evidence for improving English biography coverage, while lower-resource settings remain challenging.