Answer Selection Method Based on BERT and Parallel Multi-Channel Convolution

Aiming the problem that insufficient attention to the feature information of existing answer selection algorithms, we proposed a model B-PC which based on BERT and parallel multi-channel convolutional neural network Firstly, using BERT to get the global feature information of question-answer pair. Then, using parallel multi-channel CNN to obtain the local feature information. Finally, fusing the global and local feature information, and using fully connected neural network to calculate the score. Experimental results show that B-PC can effectively improve the effect of answer selection. On the Wiki-QA data set, the MAP is 4.2% higher than the BERT-LSTM-Attention and 9.3% higher than RE2.