Research On Unsupervised Summarization For Dialogue Text

In the field of natural language processing,the task of automatic text summarization has always been a popular research.At present,the main research object of text summary is news texts,and it is relatively mature.However,in addition to news texts,there are other formats of texts in real life,such as scripts,conferences,customer service dialogues,legal texts,etc.In recent years,dialogue summarization is one of the emerging research tasks,its purpose is to extract dialogues to obtain short texts that can summarize dialogue information.Compared with news texts,dialogue texts have the characteristics of richer content,more complex structure,and many topics,which make the process of extracting dialogue summaries more challenging.Therefore,the model needs to grasp the accuracy of content extraction from a global perspective.This paper studies unsupervised dialogue summarization from three different aspects,and experiments on typical representative script dialogue texts.Specifically include the following:First of all,this paper proposes a dialogue summarization method based on a pre-trained model,it can effectively avoid extracting sentences with non-key information on dialogue texts.Firstly,By adding two pre-training tasks for text sequence processing during the pre-training process,the model can fully consider the scene and character information,Secondly,this pre-trained model predicts the similarity between two sentences,Finally,the sentences are scored and sorted by TextRank,and the top-ranked sentences are extracted as summaries.The experimental results show that this method achieves better results than the benchmark model method,and the system performance is significantly improved in ROUGE evaluation.Secondly,this paper proposes a dialogue summarization method based on the combination of Siamese network and attention mechanism,it can fully consider the sentence semantic information of the dialogue texts.Firstly,the method uses BERT to obtain the text vector,and then uses the cross-attention mechanism to fuse the current text vector with the key information vector.Secondly,the result is calculated through the Siamese network for semantic similarity.Finally,the sentences are scored and sorted by TextRank,and the top-ranked sentences are extracted as summaries.The experimental results show that this method achieves better results than the benchmark model method,and the system performance is significantly improved in ROUGE evaluation.Finally,this paper proposes a dialogue summarization method based on association graph,it can fully consider the scene and character structure information in the dialogue texts.Firstly,it constructs structure graphs at the input end.Secondly,in order to enrich text features,it uses Graph Convolution Neural Networks to understand the relationship between sentences.Finally,the sentences are scored and sorted by TextRank,and the top-ranked sentences are extracted as summaries.The experimental results show that this method achieves better results than the benchmark model method,and the system performance is significantly improved in ROUGE evaluation.