Research And Implementation Of Multi-entity Relationship Recognition And Automatic Text Summarization Based On Deep Learning

Relation extraction and automatic text summarization are two typical applications in the field of information extraction,which have attracted more and more scholars' attention.Relation extraction task recognizes the relationship between entities from original texts.However,most of the current researches focus on simple situations,such as a sentence containing only one relationship,and there are few in-depth discussions on complex situations of relation extraction.In addition,the automatic text summarization task performs semantic analysis on input texts,which can improve users' reading efficiency by extracting abstracts from original long texts.But current researches on automatic text summarization still face some challenges,such as poor readability of generated summaries,inconsistency in the core content of generated summaries and original texts,and some repeated tuples in generated summaries.Aiming at the above problems,this paper proposes a multi-entity relation extraction method which is based on multi-label technology to deal with the relation extraction of complex situations,and an automatic text summarization method which is based on cumulative attention mechanism to improve the readability of the summary and the relevance of the core content of the original text.The work of this paper mainly includes the following parts:(1)This paper proposed a new multi-entity relation extraction method based on multi-label technology which can improve the performance of entity relation extraction of complex situations.This method combines the new multi-tag loss function and traditional sigmoid loss function to train the network together,which introduces more tag association information in the multi-label classification process,and balances the loss of tags in the training process as the denominator of the loss function.This paper has carried out experiments on public data set NYT.The results show that our proposed model has improved in the relation extraction compared with the existing model.And there is a significant optimization on F1 indicator.(2)This paper proposed a new automatic text summarization method based on cumulative attention mechanism which can improve the performance of indicators such as the readability of summary and the relevance of the original content.The method is based on the encoder-decoder framework,combined with pointer network and cumulative attention mechanism proposed in this paper.The cumulative attention mechanism refers to the fact that the decoder performs the internal attention calculation to obtain the context vector of the decoder,and then adds this vector to the encoder to participate in the attention calculation of the input text,which can optimize the context vector of the input text.We think that the attention to the already output text can be achieved to reduce the duplicates of the output content.At the same time,the attention to the input text can be optimized,so that the model input and output can be more related.In addition,we carried out experiments on public data set CNN/DailyMail.The results show that the proposed model has improved in every indicators compared with the existing supervised learning model.(3)Based on the above two models,we builds an online demonstration platform based on Django Web framework,which implements receiving user input,calling the model online,obtaining the output and returning the result to the user.Furthermore,the output information by proposed models will be analyzed.