Research On Short Text Conversation In Combination Of Retrieval And Generation

Short text conversation refers to a single round of open domain brief conversation,which can be regarded as a variant of the question answering system.It is generally a form of question and answer.Users ask questions and the machine generates the text of the corresponding answer.A single round of conversations can be rule-based or data-based.Data-based methods can be further divided into search-based methods and generation-based methods.In the dialogue task,the search-based method can get a smooth reply.The disadvantage is that a new reply cannot be generated and sometimes the retrieved reply is not related to the topic.The advantage of a method based on generation is that a new reply can be generated.The disadvantage is that the reply contains very limited information.The development of deep learning technology has provided new research directions for short text conversations.After continuous exploration,deep learning has achieved a wide range of applications in various fields such as image processing,computer vision,natural language processing,and machine translation.Deep learning performs better than traditional methods in semantic representation and can better solve problems in natural language tasks such as semantic understanding of sentences and understanding of contextual information.These deep learning methods are very useful for dialogue tasks and provide new ideas for dialogue modeling.The main research direction of this article is from the following three directions:Research on Short Text Conversation Based on Retrieval.A candidate reply ranking model based on pattern is implemented.A reply-matching model using a recurrent neural network is implemented.By encoding questions and replies,the semantic representation vector of the sentence is obtained,and then the degree of matching between the problem vector and the candidate reply vector is obtained by using the matching degree calculation and the like so that the candidate reply can be get the best reply.The reply matching model based on long short term memory networks is implemented,which improves the sorting effect of the model and improves the overall result.Based on the generated short text conversation research.The circulatory neural network excels in the processing of sequence problems,and can utilize historical information in the sequence,and can obtain the semantic impact of each word in the sentence and its corresponding position on the entire s entence.In this paper,a Seq2Seq-based model is implemented.At the same time,LSTM and Bi LSTM are used to improve the model's ability in sentence coding.In this paper,the Attention mechanism is introduced,and combined with the response generation mode l based on Seq2 Seq,the experimental results are improved.Finally,this paper uses improved methods such as Beam Search to improve the diversity of generated responses.Search and generation combined short text conversation research.In the fusion model,the extracted candidate responses and questions are input into the encode-decode model together,providing more information for generating models.The usual reply of the query dialogue will contain information related to the topic,so the retrieved topic word will participate in the process of reply generation.This method combines the advantages of the retrieval and generation model and has a great advantage in performance.