Research On Social Media Summarization Based On Context Awareness And Relation Denoising

Automatic text summarization is an effective technology to solve the problem of information overload in social media.Its main purpose is to mine important content from the original large-scale short texts and form a concise summary,so as to improve the efficiency of user information acquisition.Different from traditional long documents,posts on social media often have short length and low writing quality,which lead to incomplete information and sparse features.Latest researches propose to mine social relations in social networks to assist the selection of summaries,while these methods still have shortcomings:(1)The representation of posts only considers the independent text content of posts,which does not essentially solve the problems of sparse features and lack of information.Also,the rule-based methods can not adapt to more complex networks and have poor generalization ability.(2)There are often noise relations in real social networks that do not conform to sociological theories.However,existing methods cannot eliminate the interference of noise relations,which leads to additional errors and reduces the accuracy of the summary.To solve the above problems,this paper studies unsupervised social media summarization method based on context awareness and relationship denoising.The main contents are as follows:(1)Aiming at the problems of sparse features and lack of information caused by the short and noisy nature of posts,this paper proposes to integrate social context and multi-granularity relations for social media summarization.The model collects relevant background information in the social context and relation clues at different level to alleviate the problem of insufficient information of a single post.Meanwhile,the feature encoder based on graph neural network is used to avoid manually designing rules and improve the flexibility and generalization ability of the model.(2)In view of the problem of noise relationships in real networks,this paper further proposes a social media summarization method based on denoising graph auto-encoder.It synthesizes training data through data augmentation,which avoids the dependence on labeled data and enables the model to remove unreliable relations in an unsupervised manner,so as to produce more salient and reliable summaries.The experimental results on data collected from two real-world social media platforms(Twitter and Weibo)show that considering the social context and the topological information of posts on social media effectively improve the understanding of short texts,thereby improving the quality of summaries.Also,The results show that the proposed denoising graph auto-encoder model can learn to eliminate noise relations in social networks without labeled data and alleviate the influence of noise relationships on social media summarization,thus improving the accuracy of summaries.