End to end approach for i2b2 2012 challenge based on Cross-lingual models

Abstract

BACKGROUND - We propose a Cross-lingual approach to i2b2 2012 challenge for Clinical Records focused on the temporal relations in clinical narratives. Corpus of discharge summaries annotated with temporal information was provided for automatically extracting : (1) clinically significant events, including both clinical concepts such as problems, tests, treatments, and clinical departments, and events relevant to the patient’s clinical timeline, such as admissions, transfers between departments, etc; (2) temporal expressions, referring to the dates, times, duration, or frequencies in the clinical text. The values of the extracted temporal expressions had to be normalized to an ISO specification standard; and (3) temporal relations, among the clinical events and temporal expressions. GOALS - The objectives involved in the current work consists on outperforming previous State of the Art for the i2b2 2012 challenge and adapting Cross-lingual model into clinical specific domain with low Data resources available. METHODS - The task has been conceived as a pipeline of different modules, an event and temporal expression token-classifier and a text-classifier for relation extraction, each of them independently developed from the other. We used XLM-RoBERTa Cross-lingual model. RESULTS - For event detection, the proposed token-classifier obtains a 0.91 Span F1. For temporal expressions, our sentence-classifier achieves a 0.91 Span F1. For temporal relation, we propose sentence classifier based on sequential-taggers that performs at 0.29 F1 measure.

DESKRIBAPENA - Narratiba klinikoen domeinuan i2b2 2012 erronkarako hizkuntzarteko ikuspegia jorratzen duen soluzioa proposatzen dugu. Erronka honek txosten medikuetan islatzen diren gertaeren arteko denbora-erlazioak iragartzea du helburu. Horretarako, lan hau alde batetik (1) klinikoki esanguratsuak diren gertaerak, adibidez, kontzeptu klinikoak, probak, tratamenduak, sail klinikoak eta bestetik, (2) denbora-adierazpenak, adibidez, txostenak esleituta duen data, denbora, iraupen edo maiztasuna adierazten duten espresioak antzeman eta bukatzeko gertaera klinikoen eta (3) denbora-adierazpenen arteako erlazioak anotatuta duen corpus batetik abiatzen da. HELBURUAK - Lanaren helburuak i2b2 2012 artearen egoera hobetzea eta Cross-lingual modeloa Data baliabide baxuak dituen domeinu kliniko espezifikora egokitzea dira. METODOAK - Lana modulu desberdinetako hobi gisa ulertu da, gertaera eta denbora-adierazpenetarako sekuentzia-markatzaileak, eta denbora-erlaziorako perpaus-sailkatzailea, independenteki garatu dira. XLM-RoBERTa Cross-lingual modeloa erabili izan da lan honetan. EMAITZAK - Gertaerak atzemateko, 0.91 Span F1 exekutatzen duen sekuentzia-markatzailea proposatzen dugu. Denbora-adierazpenetarako, 0.91 Span F1 egiten duen sekuentzia-markatzailea bat proposatzen dugu. Denbora-erlaziorako, 0.29 F1 neurria egiten duten sekuentzia-markatzaileetan oinarritutako perpaus-sailkatzailea proposatzen dugu.