Research On Intelligent Chat Method Based On Text Analysis Of Community Topic

With the development of large-scale deep learning algorithms,the field of natural language processing has revived the upsurge of research on dialogue systems in the open domain.At the same time,the influence of social media is rapidly spreading,the vast amount of data available on the network has also greatly assisted researchers in building data-driven smart chat systems.However,few people focus on analyzing contextual topics to guide users to more meaningful conversations.At present,when researchers study the generative dialogue model in the open field,they generally improve under the sequence-to-sequence framework of deep learning technology.Under the technical principle,considering the topic information of the dialogue into the basic framework,a smart chat robot model(RS2S-TA)combining the topic model and the dialogue generation model under the open domain is proposed.The model uses the keywords acquired by the topic model to simulate the user's prior knowledge,and uses the joint attention mechanism and the generation probability of the biased topic information to guide the dialogue system to generate informative and interesting responses.The joint attention mechanism compresses the implicit state of the input message into a semantic vector through the contextual attention layer,and synthesizes the topic vector from the related topic keywords obtained from the pre-trained i BTM topic model through the topic attention layer.Both together affect the choice of utterance in the decoding process.In order to increase the possibility that the topic keyword appears in the reply,the model gives an additional bias term to the word generation probability,so that the overall distribution is biased toward the topic keyword.Considering that the generative dialogue model only takes a short question as input,there may be a problem that the amount of information is insufficient and the feature cannot be captured by the deep learning network,and the retrieval model is integrated at the input end.The retrieval model uses cosine similarity as a measure.The retrieved best response is sent to the neural network of the generative model as additional information of the original question.The Chinese corpus is built by crawling the Weibo user's homepage and the hot topic list to train the model.The results show that the model improves the fluency of the response statement and the diversity of the utterances compared to the traditional dialogue generation model when conducting a continuous and meaningful conversation.