Bi-directional LSTM Model with Symptoms-Frequency Position Attention for Question Answering System in Medical Domain

Abstract

Online medical intelligent question answering system plays an increasingly important role as a supplement of the traditional medical service systems. The purpose is to provide quick and concise feedback on users’ questions through natural language. The technical challenges mainly lie in symptom semantic understanding and representation of users’ description. Although the performance of phrase-level and numerous attention models have been improved, the lexical gap and position information are not emphasized enough. This paper combines word2vec and the Chinese Ci-Lin [it is a dictionary that plays an auxiliary role in word2vec where processing Chinese (https://www.ltp-cloud.com/download)] to propose synonyms-subject replacement mechanism (i.e., map common words as kernel words) and realize the normalization of the semantic representation; Meanwhile, based on the bi-directional LSTM model, this paper introduces a method of the combination of adaptive weight assignment techniques and positional context, enhancing attention to the typical symptoms of the disease. More attention weight is given to the neighboring words and propose the Bi-directional Long Short Term Memory Model with Symptoms-Frequency Position Attention (BLSTM-SFPA). The good performance of the BLSTM-SFPA model has been demonstrated in comparative experiments on the medical field dataset (MED-QA and GD-QA).