Research On Automatic Summary Generation Of Short Text Based On Deep Learning

With the rapid development of mobile devices,the ways of information dissemination are gradually increasing,and the importance of timeliness is getting stronger and stronger.To face on these huge amounts of information,how to quickly and comprehensively grasp it is very important;automatic sunmmarization technology is a good solution.With automatic summarization technology,most of the information can be covered by a short text description.Today,with the rapid development of society,grasping the main information quickly can help people to improve the speed of information acquisition and work efficiency,thus creating more social value.The main work of this paper is divided into three parts:(1)Use the fundamental called Seq2seq +attention(sequence to sequence with attention)to generate abstracts.Seq2seq + attention uses encoding and decoding methods.Firstly,it learns the text content,adds attention attention vector as an intermediate semantic vector to the decoding part,and jointly decides the generation words at a certain time in the decoding module.The model consists of two parts:the coding language model encodes the input sequence and decodes the decoding language model;the intermediate semantic vector C is generated dynamically at each time of decoding,and the generated words at T-1 are determined by the output words at T-1 and the intermediate semantic vector C.generated at the current time.(2)Optimization of seq2seq + attention model.The model is improved by using modified probability and coverage mechanism to solve most of the duplication problems and the phenomenon of OOV(out of vocabulary)in abstract generation.(3)In the experimental part,two methods,ROUGE automatic evaluation and manual evaluation,are used to generate summary evaluation.The experimental results show that the generated summary algorithm proposed in this paper is relatively higher than the traditional extracted summary evaluation in ROUGE-1,ROUGE-2 and manual evaluation.The experimental results show that the improved generative summary based on seq2seq + attention improves the integrity and coherence of the document summary to a great extent.