| The automatic text summarizing uses algorithms to extract important information from the original document to form a summary,which is easy to read and understand.Aiming at the shortcomings of existing automatic text summarization and keyword extraction algorithms,and reducing the degree of information overload,this paper integrates automatic text summarization and keyword extraction techniques to try high-quality text summarization methods.The main research contents are as follows:(1)In view of the traditional keyword extraction methods that do not effectively use the structural characteristics of Chinese,and the lack of text feature information,this paper proposes an At-Bi-LSTM-CRF network model to transform the keyword extraction into an entity labeling problem.Firstly,the vectorized text sequence is put into the two-way long and short-term memory neural network layer to extract the two-way long-distance dependent features of the text.Secondly,the attention mechanism is used to obtain the correlation between the input and the output,and the new output feature value is obtained after the new weight is calculated.The context information is further extracted,and the text content feature can be extracted more accurately.Finally,the linear conditional random field is used to process the state relationship between the tags to obtain the global optimal tag sequence to achieve a better keyword extraction effect.The findings show that compared with the traditional TF-1DF,CRF,LSTM and other algorithms,the proposed method has a better extraction effect,and compared with the algorithm without attention mechanism,the F1 value still has a good improvement.(2)When calculating text similarity,the traditional TextRank algorithm only compares the number of overlapping words in a sentence,without considering semantic and structural information,resulting in the problem of reduced accuracy.This paper proposes a text abstract extraction algorithm based on keywords and graph models.First of all,the feature information such as keywords,part of speech,word and sentence position is integrated and used in the word shift distance algorithm,which overcomes the defects of the traditional word shift distance algorithm,implements a text similarity calculation algorithm based on multi-feature weighted fusion,obtains semantic correlation level information and fundamentally improves the accuracy of similarity.Secondly,the weight of each sentence is calculated iteratively,according to the similarity result until it is converged.Finally,the abstract sentences are sequenced based on the weight value to filter the abstract sentences from high to low.From the experimental results,the extraction effect of the algorithm in this paper is better than the traditional MMR,TextRank algorithm and the optimized ImpTextRank algorithm.The extracted abstract is more similar to the artificial abstract,which improves the quality of article abstract extraction.In summary,with an optimized deep learning method,this article processes keywords,uses the characteristics of keywords as special abstract sentences,and combines the optimized TextRank method to extract text summaries.It overcomes the defects of the traditional TextRank method,such as ignoring the coverage of keywords and the lack of semantic information and structural information. |
