Malaga researchers replace word alignment with LLM label insertion for cross-lingual clinical NER
Researchers at the University of Malaga propose projecting annotations across languages by having an LLM insert entity tags directly into the target-language text, with a deterministic checker rejecting any output where the original text was altered and exact character offsets rebuilt from passing outputs. A Juejin AI article reporting the paper (arXiv 2609.11450) says the method was state of the art across all 18 settings of six target languages and three clinical entity types, with strict F1 gains of 0.0564 to 0.1512, an average strict F1 of 0.9201 for GLM 5.2 and 0.9133 for locally deployed Gemma4:31B, and 55,416 grounded annotations with precise offsets. The article cautions that the evaluation covers only three entity types and six languages, that nested or overlapping annotation schemes remain untested, and that inference cost is higher than dedicated alignment approaches.
Replacing the translate-align-map pipeline with constrained generation plus deterministic validation could make LLM annotation auditable and runnable on local 31B-class open-weight models, which the report says is useful for any domain where one language has labels and another has none.
💡 一句话总结:把 A 语言的标注搬到 B 语言(标注投影)过去要靠脆弱的词对齐流水线。这篇论文换了条路:让 LLM 直接在原文里「插标签」,用确定性校验保证原文一个字符都没被动过。六个语言全 SOTA(最高 F1 0.92),本地开源模型就能跑。「约束生成 + 确定性校验」这对组合,值得所有用 LLM 做标注的团队抄走。
导语:标注贵,跨语言的标注更是贵上加贵
医疗 NLP 的老难题:标注一份临床语料要专业医生按小时计费,英语语料尚且稀缺,其他语言基本空白。经典解法「标注投影」很聪明——英语有标注,机器翻译过去,再把标注「投影」到目标语言上。