Research On Key Techniques Of Question Answering System Based On Neural Network

With the rapid development of the Internet,the information that people can access is growing explosively,which greatly exceeds the human capacity.The search engine is emerging under this background.It can do a rudimentary screening of these large amounts of information.However,people are more likely to get answers they need directly.The question answering system is such a system that can quickly give the answer according to the description of the user's question.This paper mainly studies the answer selection part of the question answering system,that is,given a question and a set of candidate answers,choose the best one or several answers.The essence of the answer selection task is to calculate the semantic similarity between the question and the answer sentences.Different from traditional schemes such as WorldNet,dependency grammar analysis tree and so on,this paper uses a modular framework of answer selection model based on neural network,and designs different implementations for some of these modules.For each answer,the model calculates the semantic relevancy score with the question and selects the result.This paper tests the traditional method and several neural network-based methods on the QA dataset,and the experimental results show that the model based on neural network has better effect and relies on fewer external linguistic tools.In view of the answer selection especially the non-factoid answer selection,considering that the previous methods are generally processing question and answer sentences individually until the last step,this paper proposes a semantic calculation method that uses the correlation information between question and answer sentences.It can calculate the semantic similarity between the question and the answer at the text level.These correlation information represent the importance of each part of the question and answer sentence,and also participates in the later modeling step,providing additional interactive features.Experiment results show that the proposed method can combine with neural network to achieve higher accuracy than the original model.