Santana , SergioPerez Ramirez, AliciaCasillas Rubio, ArantzaOronoz Antxordoki, Maite2024-11-272024-11-27production.44722https://dx.doi.org/10.26876/ikergazte.iv.01.07https://gordailua.ueu.eus/handle/123456789/2531Testu klinikoetan informazio aberatsa dago, besteak beste, medikuntzako entitate izendunak ditugu (botika-izenak,gaixotasun-izenak, etab.) eta hauen arteko erlazioak. Informazio hori erauzteko, ikasketa sakoneko algoritmoakhartu ditugu oinarri eta antzeko patroiak, ez derrigor berdinak, topatzeko gai den sistema bat eraiki dugu. Sistemakbi entitate (adb. botika bat eta gaixotasun bat) erlazionatuta dauden ala ez detektatu eta erlazio mota (adb. kausa)esleitu behar du. Horretarako, Joint AB-LSTM deritzon algoritmoaren ahuleziak aztertu eta desoreka estatistikoariaurre egiteko bi modu proposatu ditugu. Hurbilpena nazioarte mailan ezaguna den ingelesez idatzitako BioNLP2011 datu-sortan ebaluatu dugu.This work deals with relation extraction in clinical text mining. To this end we turn to algorithms based on deeplearning: given examples from a set of texts, the aim is to build a system able to find similar though not necessarilyequal patterns. In this work we analyse systems that automatically extract information from clinical texts. Giventhat the context is crucial in this task, we resorted to Joint AB-LSTM algorithm. The system has to detect if twoentities (a medicine and a disease) are related, and if they are, assign a class (for example cause). In this work weexplore the weak points of the algorithm and we propose two ways to tackle statistical inbalance. Our proposalwas assessed in a well-known English medical data-set, i.e. BioNLP 2011.sare neuronal sakonakjoint AB-LSTMhizkuntzaren prozesamenduadeep neural networksjoint AB-LSTMnatural language processingHizkuntzalaritzaInformatikaMedikuntzaintroduction