Research And Application Of Automatic Text Summarization Method Based On Encoder-decoder Network

As Internet technology becomes more powerful and the amount of network data grows exponentially,it becomes cumbersome for people to find useful information in numerous texts.The emergence of automatic text summarization technology helps people to quickly extract the effective information in the text.Automatic text summarization technology is mainly divided into two categories:extraction and generative,which have different principles and usage scenarios.At present,the extractive summarization model based on the Encoder-Decoder network has the problems of weak ability to capture key semantics and excessive semantic loss during decoding;the generative summarization model based on the Encoder-Decoder network has the problems of low global correlation and one-sided calculation of memory node information.This paper proposes solutions to the above problems and completes the following work:(1)A text summarization model ESMTF that fuses multi-level topic features is proposed.Aiming at the problem that the model’s ability to capture key semantics is weak,this paper designs a multi-level topic feature extraction method,which extracts and fuses global information features and local topic features during encoding to improve the model’s ability to capture key semantics.Aiming at the problem of excessive semantic loss during decoding,this paper designs a confidence calculation method Ext Conf,which adds a buffer layer in the process of confidence calculation,which alleviates the problem of excessive semantic loss during decoding,thereby improving the accuracy of confidence calculation.Combined with the above methods,a text summarization model ESMTF is proposed.The experimental results on the Pubmed dataset show that ESMTF has a significant improvement in the Rouge index compared with other models,which verifies the effectiveness of the model improvement.(2)A text summarization model Sum CNN-PCM based on semantic enhancement is proposed.Aiming at the problem that the global relevance of the generated abstract is low,this paper designs a multisemantic feature extraction method Sum CNN,which extracts and fuses the multi-semantic features before encoding.By enhancing the semantic information,the global relevance of the generated abstract is improved.Aiming at the one-sided problem of memory node information generation,this paper integrates full-text compression semantics when calculating memory nodes,so that the generated memory node information is more abundant.And introduce point-network and Coverage mechanism to solve the problem of OOV and repeated words in the generated summary.Combined with the above methods,a text summarization model Sum CNN-PCM based on semantic enhancement is proposed.The experimental results on the CNN/Daily dataset show that Sum CNN-PCM has a significant improvement in the Rouge index compared with other models,which verifies the effectiveness of the model improvement.(3)Design and implement an automatic text summarization system.The system uses the text summarization models ESMTF and Sum CNN-PCM proposed in this paper as the algorithm model support,and completes the demand analysis,design and implementation of the text automatic summarization system.