Text Summarization Research On Dialogue Unique Feature Representation

With the rapid development of artificial intelligence,as an important research direction,natural language processing(NLP)is developing vigorously.Text summarization is an important method to extract and summarize text information in natural language processing.In recent years,the dialogue summarization task has raised attention,such as online customer service summary,consultation summary,etc.,which has a broad application prospect.Dialogue,however,is different from news,literature of science and technology,the text has unique structure and information distribution features,such as a dialogue with the speaker structure,temporal structure and adjacent utterance dependency structure.Besides,dialogue information distribution has features of loose distribution,high redundancy and small granularity.The traditional summarization model cannot make full use of these dialogue features in the research,which makes the representation of the dialogue structure features and the information distribution features incomplete and makes the text summary poor in readability and information relevance.In order to solve these problems,this thesis makes full use of the dialogue structure features and the information distribution features to study then dialogue abstractive summarization.A dialogue abstractive summarization model based on dialogue feature representation was constructed to provide an effective solution for dialogue abstractive summarization.Specifically,the main research contents of this paper are as follows:1.According to the unique dialogue structural features,a Dialogue Structure Focused Hierarchical Model is proposed.This model analyzes the unique structures of dialogue,puts forward the hierarchical model of dialogue and the Dialogue-Structure-Focused Attention mechanism.The model is on the basis of presenting the unique speaker structure,temporal structure and adjacent discourse dependency structure of dialogue,and makes full use of the unique structural information of dialogue.2.According to the distribution characteristics of dialogue information,a key information perception model based on graph is proposed.This model obtains the representation of fusing key information of utterances by establishing the key information perception graph In the process of generation,the model can effectively perceive the key information of the dialogue text,which can be used for summary generation.3.A dialogue abstractive summarization model based on the unique dialogue features of the dialogue is designed and the text summarization system is implemented.The model fully integrates the feature representation of the dialogue structure and the perception of the dialogue information to generate summaries with good relevance and readability.In this thesis,experiments are carried out on SAMSUM dataset.The experiment results show that the proposed model is effective,and can generate dialogue summaries with good readability and relevance compared with the baseline models.