Medical Event Extraction with Question Answerability Judgment

Abstract:
Medical event extraction is an important task in medical information extraction, which aims to extract useful information from the unstructured medical text. Existing work in event extraction mostly formats it as a classification task, which requires a large amount of labeled training data to ensure good performance, but it is difficult to obtain enough labeled data in the medical field. In this paper, a new medical event extraction method based on Machine Reading Comprehension (MRC) is proposed. Our method relies on MRC framework for event argument extraction, which includes two subtasks: the Question Answerability Judgment (QAJ) and the Span Selection (SS), to overcome the labeled data scarcity problem. In QAJ, a question answerability discriminator—Judger is designed to predict whether a question can be answered or not, avoiding argument extraction from unanswerable questions. In SS, a new span selection model is designed to obtain effective contextual representations, avoiding the difficulty of extracting answers to questions with contextual representations due to the poor logic of context in medical reports. Additionally, inspired by multi-task learning with auxiliary tasks, a new loss function with the result of QAJ is designed to improve the training efficiency. A large number of experiments have shown that our method achieves noticeable performance comparing with existing medical event extraction methods.