| Automatic text summarization technology is the process of inducing input text by computer to generate a concise and generalized text.The automatic text summary technology can play a "dimensionality reduction" effect on the amount of text data on the network,and effectively alleviate the pressure caused by information overload.In recent years,the development of deep learning has greatly promoted the progress of automatic text summary technology,but the abstract model still has problems such as poor controllability,incomplete generation of summaries,and long-term dependence.To solve these problems,this paper uses deep learning algorithms to study automatic text abstracts as follows:First,aiming at the problem that the generative summary algorithm ignores the important clues provided by keywords and the final summary is incomplete during the generation process,a summary generation method is proposed to fuse the BERT pre-trained model and keyword information with the pointer generation network as the skeleton.First of all,the BERT is embedded in the model,and the sequence-to-sequence model based on the attention mechanism and the TextRank algorithm extract keywords from the original text.The extracted keyword information is then added to the attention mechanism of the pointer generation network to guide the generation of the summary,so that the generation process pays more attention to the main message of the original text.In addition,the two-pointer copy mechanism is used instead of the copy mechanism of the pointer generation network to improve the coverage of the model copy mechanism.The results on the LCSTS dataset show that the model designed improves the accuracy and readability of the summary.Second,abstract generation is generally divided into two methods: extraction and generation.When the input text is too long,the generative algorithm will have a long-term dependency problem in the generation process,and the generative model may be uncontrollable,resulting in the final summary effect being unsatisfactory.In this article,in order to solve the problem of long text dependencies and explore the impact of multiple guide signals on digest generation,Generate summaries by combining extractive and generative.First of all,the improved BERTSUM model is used to compress the original text and reduce the redundancy of the original text.Secondly,the ROUGE value between the sentence and the abstract is calculated to extract the sentence closest to the abstract,and the keywords of the original text are extracted using the TextRank algorithm,and the extracted sentences and keywords are used as the guide information.Finally,the two kinds of guidance information are integrated into the Transformer model,and the influence of guidance information on summary generation is explored from multiple dimensions,which increases the controllability of the model to a certain extent.The results on the NLPCC2017 Chinese dataset show that our model improves the relevance and accuracy of the abstract to the original text. |
