BRCNN: Chinese Medical Question Answer Selection Based on BERT.

Ruilin Qi,Pengyue Sun,Qingchen Zhang

DOI: https://doi.org/10.1109/smartworld-uic-atc-scalcom-digitaltwin-pricomp-metaverse56740.2022.00094

2022-01-01

Abstract:Nowadays, automatic question answer systems are beneficial to improve the efficiency of computer-aided diagnosis. It is difficult for current methods to select best matched answers for patients’ questions since they are hard to dynamically generate word vectors or to provide rich sentence representations. This paper presents an BRCNN model that combines BERT, bidirectional GRU(BiGRU) and multi-scale CNNs for the selection of Chinese medical question answer. In detail, the proposed model works in four steps. First, we input the question or answer into BERT to generate dynamic word vectors. Next, we input word vectors into BiGRU to obtain bidirectional sequential information. And then, we input the bidirectional information into multi-scale CNNs to capture features carried by different n-grams. Finally, we measure the similarity between the question and the answers for selecting a best matched answer by using cosine similarity. In particular, we combine the advantages of different models to obtain more accurate and rich representations. We conduct some experiments to verify the proposed model by comparing with a series of models on cMedQA V1.0 dataset. The results demonstrate that our proposed model achieves 10.6% higher accuracy than other models.