Research On Key Issues Of Abstractive Automatic Text Summarization

There are two types of automatic text summarization: extractive methods and abstractive methods.With the continuous advancement of deep learning technology,abstractive text summarization has gradually become a popular research direction.The core problem of the current automatic text summarization task is to extract or learn the semantics in the original document to generate summaries.This paper takes abstractive automatic text summarization as the main line,and divides it into multiple sub-problems according to different goals,including: redundancy,lack of key information,factual consistency,and controllability.First,in order to solve the redundancy and lack of key information,this paper combines the advantages of extractive and abstractive methods to quickly obtain important semantic information of the original text and use the model to generate smooth abstracts.This joint method is then trained in an end-to-end manner using a reinforcement learning framework.On this basis,a word-level extraction method is introduced in the extraction module,and the extracted keyword set is used to enrich the semantic information of the abstract,so as to avoid the redundancy problem caused by the sentence granularity being too large.In addition,the graph-informed attention mechanism is introduced to learn more deeply the semantic connection between incoherent key sentence sets,and to pay attention to all the key semantic information related to the current lemma in the process of summarization generation,so as to avoid the problem of missing key points in the summary.Second,aiming at factual consistency and controllability,this paper uses a Transformer-based encoder-decoder architecture model and pre-trains the input text with the BERT language model.On this basis,a variety of external guidance signals are extended as input to help the model pay attention to various types of external guidance signals while generating summaries,constrain the output of the summarization model,and control the semantic direction of summarization generation.Then extract the fact entity from the original document,construct the graph structure,increase the fact perception of the decoder by means of knowledge map enhancement,and improve the fact consistency of the abstractive model.Third,on the basis of the above two research contents,a abstractive text summarization system oriented to the field of public health was designed and built.The system automatically generates a complete,concise and smooth text summary according to the input original document,and provides an experimental evaluation function to help Users can more intuitively judge the accuracy of the generated summaries.