Text Summarization Based On Neural Network Joint Learning

With the development of the Internet,the information that people are exposed to is exploding,and how to make knowledge efficiently becomes especially important.To better deal with the problem of information expansion,the researchers proposed the automatic generation of text summaries.It helps to get knowledge faster and more accurately by compressing and refining articles,which saves a lot of reading time.After more than ten years of research,the text abstracts can be divided into two types:extractive summaries and generative abstracts.In recent years,due to the development of deep learning and the attention-based end-to-end model,the automatic generation of text abstracts has achieved good results.In addition,these models lack the ability to identify key information,making the generated summary lack critical information,and the key information is the most indispensable part of the abstract.In response to this problem,we propose a key information guidance generation network that combines the text summary generation model and the decimation model.The main work and contributions are as follows:1.A new Key Information Guidance Network(KIGN)has been proposed that successfully combines extractive summaries with generative s.First,we get the keywords from the input text through a decimation model.Then,we introduce the Key Information Guidance Network(KIGN),which encodes keywords and guides the generation process.Since the above model is actually two models,we further adopt the method of joint learning,so that the acquisition of keywords and the generation of abstracts are simultaneously implemented in one model.2.Influenced by the reinforcement learning thought,we use the predictive guidance mechanism to further guide the generation of the,to ensure that the choice of the generated words at the current moment is not only the current optimal,but also optimal in the long run.Specifically,we have added a feedforward neural network to estimate the impact of the generated words selected at the current decoding time on the extent of the critical information contained in the final summary.3.We evaluated the model on the CNN/Daily Mail dataset and verified the validity of the model.The accuracy of the automatic summary improved to 39.64%in the ROUGE-1,which is significantly better than other existing methods.In addition,we also tested the model on the reading comprehension dataset MARCO proposed by Microsoft,and also achieved a high ROUGE index,which verified the universal adaptability.