A preprint comparing three ways to enrich English biographies found that outputs using selected English claims had the highest average number of supported additions and the lowest average hallucination rate. The comparison covered raw non-English biographies, machine-translated biographies and English claims selected from paired non-English material.
Across the tested languages and generators, outputs using Claims averaged 20.47 supported additions per biography and a 23.54% hallucination rate. Raw outputs averaged 15.89 supported additions with 29.49% hallucination, while Translation averaged 19.05 additions with 29.59% hallucination. Here, hallucination refers to generated claims counted as unsupported or contradicted by the evaluation reference.
A pipeline for selecting usable evidence
The task starts with an existing English biography and a paired non-English biography. The second biography is used as additional evidence for enrichment, with factual claims expressed in English before the rewriting stage. The claim pipeline adapts FactScore, FactCheck-GPT, DnDScore, VeriScore and Claimify to accept English or non-English biographies and produce English factual claims.
For alignment, All-MPNet-Base-v2 retrieves likely English matches for each claim. The system keeps candidates whose cosine similarity is at least 0.7, considers the five nearest neighbours and sends the retained claim pairs to a relation classifier.
The main CLAW-4L benchmark contains 300 biography pairs: 100 English-to-French cases, 100 English-to-Chinese cases and 100 English-to-Azerbaijani cases. The candidates were required to be women with the relevant citizenship and both English and non-English sitelinks, with coarse occupation-aware balancing used in selection.
Two supporting resources were used for development and assessment. CLAW-4L-CX contains 600 manually annotated sentences, while CLAW-4L-RC contains 600 annotated claim pairs. The annotation recruited 13 English-fluent annotators.
Most source claims were not already covered
The analysis found that 13.9% of non-English claims were already covered by the English biographies. A further 26.2% added detail to partially aligned English claims, while 48.6% had no relevant English counterpart. Overall, 86.2% were selected as enrichment evidence.
The extraction comparison found that X-Claimify performed best across the tested languages, particularly on exact alignment. The study used it to extract the reference claims for the full benchmark.
For claim-pair relation classification, the study selected Qwen3.5-9B as its default open-source classifier. Its reported ARC accuracy was 95.5%, close to the 95.9% reported for GPT-5.1.
The average did not hold in every setting
The overall averages conceal a language-and-generator exception. Mistral had the highest Bal. score in every language-evidence setting tested, but Translation scored better than Claims for Azerbaijani with Mistral on that measure.
Bal. is the study's combined measure of how much supported information was added and how reliable it was. The generation comparison also assessed claim growth, factual support and hallucination, so the headline averages describe several related outcomes rather than a single score.
A small writing check found mixed results
In the writing assessment, Claims received the highest average ratings for readability and fluency, at 3.933 on a five-point scale, and for Wikipedia-style writing, at 3.822. Raw outputs scored slightly higher for coherence and integration, at 3.911 compared with 3.867 for Claims.
The check covered 45 Mistral-generated biographies across three languages. It used five length-stratified biography pairs per language and tested three enrichment methods; three computer-science Ph.D. researchers rated every output on a 1-to-5 Likert scale. The paper describes the exercise as a descriptive, non-significance-tested sanity check.
What the result does not settle
The findings are averages across the tested languages and generators, and the reported Azerbaijani result shows that Claims did not lead in every combination. The main benchmark covers 300 biography pairs from French, Chinese and Azerbaijani cases, with English as the target, so the averages describe those tested settings.
The manuscript is an arXiv preprint identified as arXiv:2608.23390v2 and dated 26 August 2026.
Paper data and sources
Original title: Cross-lingual Biography Enrichment via Claim Extraction and Alignment
Authors: Yifei Song, Ziyang Chen, Emil Sayilov, Claire Gardent
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text