💬 NLP

Cross-lingual Biography Enrichment via Claim Extraction and Alignment

This paper introduces the CLAW-4L benchmark and a claim-based framework to enrich English Wikipedia biographies of women with locally grounded facts from non-English editions (French, Chinese, and Azerbaijani), demonstrating that cross-lingual alignment can improve coverage despite challenges in lower-resource settings.

Original authors: Yifei Song, Ziyang Chen, Emil Sayilov, Claire Gardent

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yifei Song, Ziyang Chen, Emil Sayilov, Claire Gardent

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Wikipedia is the world's most popular encyclopedia, a vast library built by volunteers in nearly every language. Yet, this library is not perfectly balanced. While English speakers have access to millions of detailed articles, many people from non-English speaking backgrounds are described with far less detail in English, even though they are well-documented in their local languages. This gap creates a blind spot for artificial intelligence. Modern computer programs that read and write text are trained heavily on English data, meaning they often know less about the rich lives of people documented primarily in French, Chinese, or Azerbaijani. When these programs try to write a biography for such a person, they often end up with a thin, incomplete story, missing the very details that make that person's life unique and important.

A team of researchers set out to fix this imbalance by teaching computers how to borrow the best parts of a biography from one language and weave them into another. They focused specifically on women, a group that has historically been underrepresented in encyclopedias, to see if they could use the detailed information found in French, Chinese, or Azerbaijani Wikipedia articles to flesh out the corresponding English versions. The challenge was not just to translate the missing words, but to understand which facts were new, which were already known, and how to combine them without inventing false details. The researchers developed a new method that acts like a careful editor, first breaking down the foreign text into small, verifiable facts, matching them against what is already known, and then using only the new, confirmed information to expand the English story.

To test their idea, the team created a specialized dataset called CLAW-4L, which pairs 300 English biographies of women with their counterparts in French, Chinese, or Azerbaijani. They found that the non-English versions were often much richer, containing thousands more words and hundreds more specific facts about the subjects' lives, careers, and achievements. The researchers then built a system to extract these facts. Instead of feeding the entire foreign biography into a computer program to read and rewrite, which can confuse the machine with too much information, they first broke the foreign text down into individual claims. For example, instead of a long paragraph about a scientist's career, the system identified specific statements like "She directed the Radium Institute in Paris" or "She received the Nobel Prize in Chemistry in 1911."

Once the facts were isolated, the system compared them to the existing English biography. It acted as a strict gatekeeper, discarding any fact that was already present in the English version, but retaining any that contradicted it for potential reconciliation. The goal was to find only the "enrichment claims"—the pieces of information that existed in the foreign language but were missing from the English one. This process was crucial because it prevented the computer from simply repeating what it already knew or getting lost in the noise of a long, unstructured text. The researchers tested three different ways to do this: feeding the raw foreign text directly to the computer, translating the foreign text into English first and then feeding it in, and using their new method of feeding in only the selected, verified facts.

The results showed that the new method was the most effective. When the computer tried to read the raw foreign text or a direct translation, it often added new information but also made up facts that were not true, a problem known as hallucination. By contrast, the method that used the selected, verified facts allowed the computer to add significantly more accurate details while making far fewer mistakes. In their tests, the system successfully added new, supported facts to the English biographies while keeping the rate of invented information very low. This suggests that breaking a complex text down into small, manageable facts and checking them before using them is a much safer and more reliable way to improve computer-generated writing than simply translating a whole document.

The study also highlighted that this approach works best when the source language is well-resourced, like French or Chinese, but remains challenging for languages with fewer digital resources, such as Azerbaijani. In those cases, the system sometimes missed useful information because the initial step of extracting facts was less accurate. However, even with these limitations, the research demonstrates a clear path forward. By treating biographies not as blocks of text to be translated, but as collections of facts to be verified and aligned, computers can learn to tell more complete and accurate stories about people from all over the world. This work suggests that the future of digital encyclopedias lies not just in translating more words, but in intelligently connecting the specific, verified details that exist in every language to create a fuller picture of human history.

Technical Summary: Cross-lingual Biography Enrichment via Claim Extraction and Alignment

Problem Statement
While English Wikipedia is often treated as the default encyclopedic source in Natural Language Processing (NLP), its coverage is uneven. Non-English Wikipedia editions frequently contain richer, locally grounded information for "long-tail" entities, particularly women from non-English-speaking contexts. Existing approaches to biography generation typically focus on creating short summaries from structured data (e.g., Wikidata) or generating full biographies from scratch using web retrieval. However, there is a lack of research on enriching existing English biographies using curated, non-English Wikipedia evidence. The challenge lies in effectively identifying, aligning, and integrating missing factual information from a non-English source into an English article without introducing hallucinations or requiring the model to implicitly handle complex cross-lingual alignment and long-context processing.

Methodology
The authors propose a framework for Cross-lingual Biography Enrichment, where an existing English biography ( $B_{en}$ ) is rewritten to incorporate additional facts supported by a non-English biography ( $B_X$ ) of the same person. The approach is structured around three core components:

Benchmark Construction (CLAW-4L):
The authors introduce CLAW-4L (Cross-Lingual Alignment on Women biographies in 4 Languages), a benchmark comprising 300 paired biographies linking English articles with their French, Chinese, or Azerbaijani counterparts. The dataset is constructed to favor pairs where the non-English biography is richer, utilizing a language-calibrated token ratio to normalize for tokenizer biases. The benchmark includes three sub-tasks:
- CLAW-4L-CX: 600 sentences manually annotated with English claims to evaluate cross-lingual claim extraction.
- CLAW-4L-RC: 600 claim pairs annotated with fine-grained relation labels (Exact Alignment, One-sided Enrichment, Mutual Enrichment, Contradiction, Not Relevant) to evaluate claim alignment.
Claim-Based Enrichment Framework:
The proposed pipeline decomposes the enrichment task into explicit steps to reduce the burden on the generative model:
- Claim Extraction: Both the English and non-English biographies are processed to extract atomic, verifiable factual claims. These are normalized into a shared English space. The authors evaluate several extraction frameworks (FactScore, FactCheck-GPT, DnDScore, VeriScore, Claimify) and find that X-Claimify (adapted from Claimify) performs best, particularly in preserving exact claim formulation.
- Claim Alignment: Extracted claims from the non-English side are aligned against English claims using a classifier (Qwen3.5-9B). The classifier determines if a non-English claim is already covered, contradicts, or provides new information (enrichment).
- Evidence Selection: The system selects "enrichment claims" ( $C_{add}$ ) which are non-English claims that are not fully covered by the English biography (including additive details, mutual enrichments, and unresolved conflicts).
- Controlled Rewriting: The final English biography is generated by an LLM conditioned on the original English text and the selected enrichment claims ( $C_{add}$ ), rather than the raw non-English text.
Comparative Evaluation:
The framework is evaluated against two baselines:
- Raw Cross-lingual Generation: The generator receives the original English text and the raw non-English text.
- Translation-based Generation: The non-English text is first translated to English, then used as evidence for rewriting.

Key Results
Experiments conducted on CLAW-4L using three open-weight multilingual generators (Qwen3.6-27B, Gemma-4-31B-it, Mistral-3.2-24B-it) yield the following findings:

Trade-off between Support and Hallucination: The Claim-based setting achieves the best trade-off. It produces the highest number of supported additions (average #Valid of 20.47) while significantly reducing the hallucination rate (23.54%) compared to Raw (29.49%) and Translation (29.59%) settings.
Effectiveness of Explicit Selection: By filtering out information already present in the English biography and normalizing evidence into English claims, the claim-based approach reduces the long-context burden and implicit alignment requirements, leading to more reliable generation.
Language and Resource Disparities: Performance is generally stronger for French and Chinese than for Azerbaijani. The authors note that in lower-resource settings (specifically Azerbaijani with the Mistral model), translation-based methods occasionally outperform claim-based methods. This is hypothesized to be due to error propagation in the upstream claim extraction pipeline for low-resource languages, where useful evidence might be missed before the generation step.
Human Evaluation: A human evaluation of writing quality (readability, coherence, and Wikipedia style) indicates that using selected claims as structured evidence does not incur a penalty in writing quality compared to raw or translation-based pipelines.

Significance and Contributions
The paper claims three primary contributions:

Task Definition: It formally defines and studies the task of cross-lingual biography enrichment, moving beyond generation from scratch to the refinement of existing articles using curated non-English evidence.
Benchmarking: It provides the CLAW-4L suite of benchmarks, which are essential for evaluating cross-lingual enrichment, claim extraction, and fine-grained claim-pair relation classification.
Methodological Framework: It demonstrates that a claim-based framework, which explicitly selects and normalizes enrichment evidence before generation, offers a superior balance between factual support and hallucination reduction compared to raw cross-lingual or translation-based approaches.

The authors conclude that while non-English Wikipedia biographies provide valuable evidence for improving English coverage, the reliability of this enrichment depends heavily on robust claim extraction and alignment, particularly for lower-resource languages. They suggest that future work should explore bidirectional enrichment and the integration of post-generation verification loops to further enhance factual reliability.

Technical Summary: Cross-lingual Biography Enrichment via Claim Extraction and Alignment

More like this