Research On Key Technologies Of Question Answering System Based On Deep Learning

With the rapid development of the modern Internet,the amount of information that people are exposed to and process every day have exploded.But in the vast information space,what are the resources we really need? In order to deal with this problem,search engines came into being.Its appearance greatly facilitated people’s preliminary screening of information resources,so as to further find the effective information they need.However,in decades of search engine development,it starts to expose some problems:the search results fed back by search engines do not match people’s query intent very well.And the query results are too redundant,causing users to spend extra time to select the useful information.Therefore,people expect that there will be a new way of obtaining information that is more efficient,concise,and different from traditional search engines.The question answering system is the one that meets people’s such needs.It can accept input in natural language and provide users with more accurate answers than traditional search engines.This novel information retrieval model is favored by the public and gradually becomes a research hotspot in the field of natural language processing.This paper mainly studies the answer extraction part of the question answering system based on deep neural network.That is,the input question through natural language is matched with the set of candidate answers,and the most suitable one or more answers are selected.The core of this method is the matching of questions and answers.Different from the Boolean model and syntactic analysis model used by traditional search engines,we choose a set of answer extraction models based on deep neural networks,and use this model to vectorize the input question text and candidate answer sets.The matching calculation is performed at the vector space level to obtain final results.We compared the accuracy of related models on open source data sets,including traditional language models represented by the bag-of-words model and models based on neural networks.Our experiments show that when the external language tools are the same,the neural network-based language models have better results.In terms of sentence modeling in the answer extraction framework,although the existing neural network language model can handle the vector set from the question text and the answer text well,it does not process the vector results of both text accordingly and loses some features.This paper proposes an algorithmic method based on the semantic level,which aims to calculate the lexical dependency correlation between the input and the answer,that is,the semantic similarity.We also combine it with other technologies in deep learning to enrich the modeling features.The experimental results also show that the combination of the method proposed in this paper and the neural network will have a higher accuracy than the traditional scheme.