| Nowadays,the rapid development of information technology has spawned many channels for information generation and dissemination and information interaction methods,especially the widespread use of mobile devices,making the generation and dissemination of information more convenient.At the same time,as the generated network information becomes more and more large and redundant,querying effective information often wastes more time and energy.The emergence of automatic summary technology provides a good solution to this type of problem.By extracting concise and concise abstracts from massive amounts of information,quickly conveying the central idea and content of the text,it can increase the speed at which people can obtain information and save time in querying information to improve work efficiency and reading experience.The main work of this article is as follows:In terms of automatic abstract data set selection,the open source data set,LCSTS(Document Understanding Conference)constructed by the Intelligent Computing Research Center of Shenzhen Graduate School of Harbin Institute of Technology,was selected.This is a data set that uses Weibo content as the corpus source.The object is short Chinese text.For each Chinese text posted on Weibo,there is a corresponding Chinese abstract.In terms of automatic abstract generation technology,this paper focuses on the TextRank algorithm in the extractive automatic technology and the Seq2Seq model in the generated automatic abstract technology.Aiming at the extractive automatic summary technology,this paper focuses on a graph model-based method—TextRank algorithm,and improves the calculation method of its edge weights.The TF-IWF value is used instead of the TF-IDF value as the word frequency feature to solve the problem of the algorithm assigning too high weight to words with low word frequency.Fusion of text feature extraction methods based on TF-IWF and Word2Vec to achieve the purpose of considering both text statistical and semantic features.The topic keywords of the text are extracted using the LD A topic model,and the edge weights are adjusted according to the similarity between the topic keywords and the sentence feature words and sentence vectors.The experimental results prove that the abstract obtained by the improved algorithm has higher accuracy.Aiming at the generative automatic summary technology,this paper selects the Seq2Seq model as the key research object,and introduces the RNN-based Seq2Seq model,RNN+Context model,and the improvement process for its existing defects.The LSTM unit is used instead of the RNN unit to enhance the memory of the neural network,and solve the problem of gradient disappearance.Introduce the double-layer attention mechanism and construct a Seq2Seq model of the double-layer attention mechanism to solve the problem that the Seq2Seq model loses too much important information in the process of processing long sentences.Combined with the LDA topic model,a Seq2Seq model with a two-level attention mechanism incorporating topic keywords.The experimental results prove that the accuracy of the text is improved and the readability of the abstract is increased. |
