Question Classification Method in Disease Question Answering System Based on MCDPLSTM
Xiaosheng Yu, Ruxin Gong, Peng Chen · 2021 IEEE 21st International Conference on Software Quality, Reliability and Security Companion (QRS-C) · 2021
In disease question answering system, question text has some characteristics of diverse expression forms, close semantic connection and uneven distribution of key features. These characteristics make it difficult to extract information and lead to low information utilization. In order to solve the problems, a question classification method based on Merge-Convolutional Neural Networks - Deep Pyramid Convolutional Neural Networks-Long-Short Term Memory (MCDPLSTM) model is proposed. Firstly, question text is vectorized by BERT language model. Secondly, Convolutional Neural Networks (CNN), Deep Pyramid Convolutional Neural Networks (DPCNN) and Bidirectional Long-Short Term Memory (BiLSTM) constitute a multi-channel model to process feature vectors concurrently. And a method based on attention mechanism is used to fuse processed vectors of three channels. Finally, Softmax classifier is used for classification. The experimental results show that MCDPLSTM model has less time cost. Compare with the other models, precision of MCDPLSTM rose by 0.18%-5.43%, recall of MCDPLSTM rose by 0.42%-7.42% and F1-Score of MCDPLSTM rose by 0.47%-7.23%.