Research On Question Answering System Algorithm Based On Deep Learning And Topic Model

With the intelligentization of our lives,human-machine dialogues can interact with various hardware devices,thereby providing humans with language information services,intelligent voice control devices,and providing entertainment chat.Nowadays,it has become a hot topic.This paper mainly studies the question-answering system based on deep learning and topic model.The traditional neural network question-answering model has many problems to be solved.One of the key issues is how to introduce exogenous background knowledge for opendomain questions and answers.In order to promote the richness and diversity of question and answer,this paper starts the research work from the topic model and the Seq2Seq(sequenceto-sequence)framework of the neural network:1)Twitter-LDA(Latent Dirichlet Allocation)applies to short text probabilistic topic models.The hypothesis is that each short text is categorized into a topic,and the topic model is introduced into deep neural network.The seq2 seq question and answer model introduces exogenous knowledge,so the question answering model not only can learn dialogue patterns from question and answer corpora,the model can also use topic words extracted from the topic model to make up for the lack of exogenous background knowledge in the seq2 seq generative model and to promote the generation of rich and diverse answer content.2)Integrate topics into a neural network question answering model and use large-scale corpus to train.In the decoding stage of seq2 seq,the joint attention mechanism,ie topic attention mechanism and message attention mechanism,is used to make the Q&A model dynamically use the semantic vector of the question and the semantic vector of the topic words when decoding the generated words.At the same time,increase the topic word bias probability and promote the generation of topic-related answers.3)Aiming at the problem of the topic word noise and question semantics understanding in the neural network topic question answering model,this paper proposes a new topic-based QA model for attention enhancement,and further optimizes the question answering model.Using Seq2Seq-encoded global semantic vectors and dynamically-weighted local semantic vectors,both of the mixed semantic vectors input to the attention mechanism can better mine the semantic information of the words in the question-and-answer corpus and reduce the influence of unrelated topic words,while using topic attention.The weight coefficient of the topic attention is used to adjust the bias probability of the topic words,further enhances the probability of occurrence of the topic word with strong relevance to the question,and reduces the influence of the noise data of the unrelated topic words.Finally,based on the large-scale open domain question and answer corpus training topic question answering model,the experimental results verify the validity of the elucidation method.