Research On Automatic Summarization Based On Pointer Generator Networks Model

Automatic summarization is one of the research topics in the direction of natural language processing.Because it can automatically extract the main information of the article to form an abstract and save people's reading cost,it becomes a research hotspot.However,the problems such as incoherent summary statements and poor readability,caused by the traditional extractive automatic summarization methods,have plagued the researchers and brought great challenges to the commercial use of the automatic summarization.In the past five years,with the rapid development of deep learning,the abstractive automatic summarization method has gradually begun to attract attention.It has provided new ideas for solving automatic summarization.This article starts with the basic model commonly used in the current ive automatic summarization method,the Sequence-to-Sequence model based on the Attention mechanism.And introduces in detail the main framework of the basic model,the main structure of the encoder-decoder framework,and the Attention mechanism.At the same time,the article gives the common training and evaluation corpus datasets and evaluation methods in the field of automatic summarization.We designed experiments on the basic model and discovered several problems with the basic model,including factual errors caused by Out-Of-Vocabulary words,repetition of summary sentences,and inefficient model training.To solve the problems of the basic model,we proceed from the principle and make some improvements to the basic model.1)By introducing the pointer mechanism of the pointer network and combining it with the generating mechanism of the basic model to guide the generation of the summary,the accuracy of the summary information is effectively improved.2)By introducing a "coverage vector",the Attention mechanism is reused and improved,and the repeated sentences of the abstract text are greatly reduced.3)For the problem of low training efficiency of the basic model,we also propose the idea of using convolutional neural networks instead of the recurrent neural network in the basic model for parallel computation.We trained and tested the improved model.Through qualitative and quantitative comparison and analysis of the generated summaries,we can see that our improved model outperformed the basic model in terms of actual results and evaluation indicators,and basically solved the problem of the details of summary facts errors and the repeated statements,proves the effectiveness of the improved model.