Investigations on Answer Selection Based on Co-attention Guided Stacked BiLSTM with BERT

With the rapid development of the Internet, massive amounts of data accumulate on the Internet. Users need to quickly obtain effective text information from these large amounts of documents. The automatic question answering technology can quickly analyze the input questions and find the answers. In this paper, we focus on answer selection, one kind of automatic question answering. Different from the traditional word2vector embedding representation, we use the output representation of the BERT pre-training model. we propose a multi-layer LSTM network with a co-attention mechanism. The common attention mechanism allows questions to directly affect the representation of answers. On the contrary, answers also affect questions, It strengthens the connection between the questions and its correct answers. Our model was evaluated on the NLPCC 2016 DBQA dataset, the MRR score of baseline model is 78.53%. Experimental results show that co-attention guided stacked BiLSTM(Bi-directional Long Short-Term Memory) are obvious effective. Compared with the word2vector model, the BERT pre-training model has greatly improved in terms of word embedding. The highest MRR of the Co-attention Guided Stacked BiLSTM with BERT model proposed in this paper is 83.73%, which is an increase of 5.2%.