Cross-Lingual UMLS Named Entity Linking using UMLS Dictionary Fine-Tuning

Findings of the Association for Computational Linguistics: ACL 2022, pages 3380 - 3390 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Cross-Lingual UMLS Named Entity Linking using UMLS Dictionary Fine-Tuning Rina Galperin∗ rinag@post.bgu.ac.il Shachar Schnapp† schnapp@post.bgu.ac.il Michael Elhadad‡ elhadad@cs.bgu.ac.il Abstract We study cross-lingual UMLS named entity linking, where mentions in a given source lan- guage are mapped to UMLS concepts, most of which are labeled in English. Our cross- lingual framework includes an offline unsuper- vised construction of a translated UMLS dictio- nary and a per-document pipeline which identi- fies UMLS candidate mentions and uses a fine- tuned pretrained transformer language model to filter candidates according to context. Our method exploits a small dataset of manually an- notated UMLS mentions in the source language and uses this supervised data in two ways: to extend the unsupervised UMLS dictionary and to fine-tune the contextual filtering of candidate mentions in full documents. We demonstrate results of our approach on both Hebrew and En- glish. We achieve new state-of-the-art (SOTA) results on the Hebrew Camoni corpus, +8.9 F1 on average across three communities in the dataset. We also achieve new SOTA on the English dataset MedMentions with +7.3 F1. 1 Introduction Public health practices are becoming increasingly digital, with tools to explore scientific sources of information such as medical literature and online health communities rising in popularity. Such tools are essential in offering insights to researchers, pro- viding information to patients and to their care- givers. Reliable identification of mentions of bio- medical concepts in free text is a key technique to enable robust mining of such textual resources. Named-Entity Recognition (NER) is the task of classifying entities in text to high level classes (Per- son, Organization, Gene, Disease, Treatment, etc.). ∗Ben-Gurion University. Supported by the Israeli Min- istry of Science and Technology (grant 8777221). †Ben-Gurion University. Partially supported by Israel Science Foundation (grant 1871/19) and by the Cyber Security Research Center at Ben-Gurion University of the Negev. ‡Ben-Gurion University. Named-Entity Linking (NEL) seeks to addition- ally classify entity mentions in text into specific concepts according to an existing reference list or knowledge base. We focus in this work on biomed- ical NEL, i.e., identifying mentions referring to biomedical concepts such as disorders and drugs and linking them to normalized concepts, for ex- ample, concept unique identifiers (CUIs) listed in the Unified Medical Language System (UMLS) controlled vocabulary. Biomedical NEL has been mostly studied in English. Other languages present additional challenges because terms in the ontol- ogy are described in English. We address cross- lingual NEL which consists of mapping mentions in a source language to concepts labeled and de- scribed in a different target language. We focus on cross-lingual UMLS NEL, where mentions in the source language (we specifically test Hebrew, see Figure 1 for a Hebrew tagging example) are mapped to UMLS concepts. We aim for a general solution that can be adapted to any source language. We operate in a low resource setting, where the on- tology is large, text …

Content truncated to 500 words