Optimization Selection-based Extractive Automatic Text Summarization Research

With the advent of the big data era,all kinds of textual information in daily life has increased exponentially.With the limitation of technical processing capabilities today,the problem of information overload caused by the problem gradually haunts people's daily life.Therefore,how to obtain as much effective information as possible in the shortest time has become the main problem to be solved urgently.For text information,the appearance of automatic text summaries can greatly alleviate information overload.In recent years,many experts and scholars have proposed many relatively successful technical methods for automatic text summaries.However,the generated text summaries are poorly readable,and the source document center Practical application problems such as large deviations in thoughts and high repetition frequency of repetitive words or sentences in generating abstracts are still the main bottlenecks in current research on text summaries.Extractable automatic text summaries form abstracts by extracting key words or sentences that already exist in the document.The traditional extractive text summarization method is highly dependent on the artificial features of the text,and has the disadvantages of poor readability of the generated results,large deviation from the central idea of the source document,and repeated occurrence of repeated words or sentences in the generated summary.In view of the above problems,this paper studies the text representation based on deep learning,the influence degree of words and sentences,and the selection and scoring method of key sentences,and proposes an extractive automatic text summary algorithm based on optimized selection.The specific contents are as follows:Using the BiGRU network to process the corpus sequentially from the sentence level to the document level,a joint selection and scoring strategy based on deep learning is constructed to generate candidate text summaries.By fine-tuning the input parameters and adding labels,the corpus is represented on the text based on the Bert model,combined with the positional relationship and other information,and the sentence influence in the candidate text summary is calculated based on the Bert text representation.The combined influence degree of the extracted sentences in the candidate text summary is selected,and the first three sentences in the comprehensive result ranking are selected and output as the final summary result.After verification on the DUC2004 and CNN / Daily Mail datasets,compared with the classic algorithm,the method in this article is significantly higher in readability and fluency in generating abstracts,and the degree of fit with the original content.Degree of performance improvement.