关键词: 神经网络, 答案选择, 动态注意力机制, 多角度匹配, 预训练模型

Abstract: Abstract: Since the current mainstream neural network can not satisfy the full expression of sentences and the full interaction of information between sentences at the same time when processing sentence pairs, a dynamic attention and multi-perspective matching method (DAMPM) is proposed. First, the word embedding method of language model (ELMo) is called to obtain the word vector containing simple semantic information, and then the filtering mechanism is used in the attention layer to effectively remove the noise in the sentence, and the sentence representation of question and answer sentences is obtained. Secondly, a variety of matching strategies are introduced in the matching layer to complete the information interaction between sentence vectors. Finally, the sentence vectors obtained from the matching layer are spliced by BiLSTM and calculate the sentence similarity. The experimental results on the text retrieval conference question answering (TRECQA) data set show that compared with the method of Dynamic-clip Attention Network (DCAN), which is the best algorithm in the comparison aggregation framework of the baseline model, DAMPM improves 1.6% on the mean average precision (MAP) and mean reciprocal rank (MRR). Secondly, the experimental results on the wiki question answering (WIKIQA) data set show that DAMPM improves the performance of DCAN by 0.7% and 0.3% respectively. Experimental results show that DAMPM has better performance than baseline model overall.

Key words: neural network, answer selection, dynamic attention mechanism, multi-perspective matching, pre-trained model