| The emergence of automatic text summarization technology has effectively relieved the pressure of facing a large amount of information in daily life and helped people to obtain the target information efficiently and accurately.Text summarization methods can be divided into two categories: extractive and abstractive.Extractive abstracts extract key words and sentences directly from the original text,while abstractive abstracts summarize information from the original text and allow the generation of words and sentences that do not exist in the original text.The thesis mainly focuses on the abstractive summarization that is more in line with the way human beings summarize information.The rise of deep learning in recent years has led to the rapid development of abstractive summarization,but there are still some issues in the existing generative summary models,such as inadequate access to semantic information,information redundancy,loss of key information,exposure bias,and poor readability of the generated summary.To address the above issues,the main research of this thesis is as follows.(1)The encoder is lack of coding ability,which easily leads to insufficient Semantic information of the generative summary model and loss of key information.To address the above issues,this thesis proposes a global and local summary generation model with fusion selection mechanism.The model investigates the encoding side,combines local and global information to enrich semantic information,and fuses the selection mechanism to extract key information.The Bidirectional Gated Recurrent Unit(Bi-GRU)extracts information from a global perspective,and the Dilated Convolutional Network(DCN)extracts information from a local perspective.The two modules provide semantically richer source-side text information for the decoder part.For the effect of redundant information on decoding,the model adds a selection mechanism after the global encoder.Using information from all time steps for selection and not ignoring information from each time step ensures the quality of the information that goes to the decoder.The model is experimentally verified to effectively avoid the generation of redundant information and enhance the conciseness of the abstract.(2)To address the problems of exposure bias and poor readability of generated summaries,this thesis investigates from the decoding side to solve these problems and proposes a summary generation model based on joint attention mechanism and reinforcement learning.The first step is to add a weighted summation process that allows the encoder to determine how well the vectors generated by the neural network match the original text and to assign higher weights to important parts of the original text.Second,a joint attention mechanism is established and a word-sentence level joint attention mechanism is used to improve the pointer network,forcing the decoder to pay attention to both the keywords in the source text and the sentences in which the keywords are located.Then,combined with the a priori knowledge,on the one hand,it enables the decoder to get the original information of the source text to avoid the information loss brought by layers of passing in the abstract generation process;on the other hand,it transmits higher-level abstract semantic information to the decoder to enable the generation of abstracts that are more in line with human reading habits.Finally,reinforcement learning is introduced to optimize for the problems of exposure bias and inconsistency of evaluation metrics with the loss function.The experimental results show that the proposed model outperforms existing baseline models,and the readability and fluency of generating abstracts are improved. |
