Document-Based Auto Question Answering Model Based On Convolutional Neural Network And Attention Mechanisms

In recent years,with the rapid development of social informatization and Internet industry,the demand for automated Q&A technology has been increasing day by day.In particular,the importance of automatic question answering system has gradually emerged in the e-commerce system and social areas.Document-based automatic question answering(DBQA)system with adaptability,rapid adjustment features,it's ideal for intelligent customer service assistants and chat robots and other applications.Benefited from the rapid progress of deep learning technology and the emergence of large-scale network knowledge base,DBQA already has some theoretical research and realization foundation.The essence of DBQA is the semantic matching of sentences,in order to improve the accuracy of semantic matching,this paper proposes a Chinese Q&A matching method based on deep learning.This paper mainly completed the following work:1.propose a new method of attention concentration which can be used to weight the semantic features learned by convolutional neural networks and highlight the semantic features that can better represent the whole sentence.2.Through the discovery of new words,a user-defined dictionary was created to avoid the ambiguity of the word segmentation algorithm.And a high-performance Chinese word vector model was trained using the user-defined dictionary and word2vec tools.3.A mixed semantic model of attention mechanism,LSTM and CNN was implemented and a comparative experiment was carried out.The experimental results show that the mixed model with attention mechanism can further improve the accuracy of the matching,with the improvement of 2.2%and 2.1%respectively in the Top-1 accuracy and the Mean reciprocal rank(MRR).