Research And Application Of Key Technologies Of Question Answering System Based On Deep Neural Network

As an important way of human-computer interaction,question answering system directly and accurately answers user's questions in the form of natural language,and gradually shows more effective performance in information retrieval,intelligent services and other applications.At the same time,due to the development of deep learning and the improvement of computing power,and more deep learning technologies are also introduced into the task of question answering system,prompting the field to gradually shift from information retrieval based on feature engineering or shallow semantic analysis to intelligent question answering based on deep understanding.Among them,the use of natural language processing technology to represent human language,semantic understanding and knowledge reasoning are still the focus and difficulty of research on question answering systems.In this thesis,a machine reading comprehension model,which can capture the deep semantic relationship between the user's questions and related articles,is selected as the deep neural network of the question answering system,a reading comprehension question answering system is designed and implemented in an end-to-end way,and conduct in-depth research on some key issues in this QA system,pretrained language models,attention mechanisms and other technologies in deep learning are used to improve the question answering system.Finally,the knowledge source,context modeling and knowledge representation,semantic understanding and other aspects of the question answering system are improved considerably,which effectively improved the performance of the QA model.The following is the main contents of this thesis:(1)In the aspect of text representation,a text representation method based on multistage feature fusion is proposed.Firstly,the characteristics of Chinese characters and the limitations of existing Chinese representation models are analyzed in detail,and a morphological n-gram Chinese representation method suitable for Chinese word representation learning is proposed;then,feature vectors of different granularity,which are extracted from grammatical categories and pretrained language models,are used as the supplement to the grammar and external knowledge of the original text,and multi-stage feature fusion method is adopted to fuse these feature vectors.Finally,the superiority of the Chinese representation method based on morphological n-grams and the effectiveness of the multi-stage feature fusion method are verified by experiments.(2)In the aspect of semantic understanding,a machine reading comprehension model(HANet)based on hierarchical attention mechanism is proposed,which vividly simulates the human understanding process from shallow to deep.In this model,different kinds of attention mechanisms are applied to multiple network layers to capture the relationship between questions and articles at different granularity levels,and gradually focus on the boundary part of the best answer,and finally,the correct answer is predicted by elaborating the details.The validity of the model is verified by experiments on different datasets.(3)In the aspect of application,a question answering system based on deep neural network is designed and implemented in an end-to-end way to provide real-time question answering service for alumni of UESTC.This question answering system is based on the methods and models proposed in(1)(2),and interacts with users through wechat and web pages.The user submits a query request on any client,then the QA system will return short and accurate query results to the user.