Research On Stance Detection And Opinion Summarization Towards Microblog Topics

At present,more and more users choose to participate in the discussion of current hot topics and express their emotional experience on Microblogs.The hot topics on Microblogs are often a hot event,a controversial figure,a social policy that has attracted wide attention,and so on.Each hot topic involves a wide range of people and things.Different users express their opinions from different angles and stance.The details of the opinions are also different.Extracting the opinions expressed by the public on these hot topics on Microblogs,studying the public's stance and summarizing the public's opinions can help Microblogs users to have a more objective and comprehensive understanding of hot topics in a very short period of time,and help governments and enterprises to carry out more timely public opinion supervision.Therefore,this thesis will use Natural language processing technology to carry on stance detection towards the target topic and extracts opinion summary,in order to provide some new ideas for the related research work of Microblog.Stance detection focuses on whether the text holds support(FAVOR),opposition(AGAINST)stance or no stance(NONE)for a given target topic.However,a given target topic does not necessarily appear directly in the text,and the author of the text may indirectly present the author's stance on a given target topic by expressing his or her views on other events,so the positive sentiment does not necessarily represent a support stance,and a negative sentiment does not necessarily represent an opposition stance.In this paper,a shared convolution plus conditional-lstm model based on attention mechanism is designed to automatically extract the fusion features of semantic information of target topic text and micro-blog text.Shared convolution layer and LSTM conditional encoding layer are designed to combine the information of target topic and micro-blog text,and then adopt attention mechanism to make the model enhance the influence degree of target topic in the process of processing micro-blog text.The final obtained(topic-microblog)features are fed into the output layer for stance classification.Higher experimental results have been achieved on official datasets.The innovation of our stance detection model in this paper is that it can extract the features from the text essential semantic level to integrate target topic and micro-blog text from the beginning to the end,and it can fully and effectively take into account the semantic factors of target topic in the process of processing micro-blog text.A shared convolution mechanism is proposed.By sharing the same convolution between the target topic and the micro-blog text,the n-gram semantic features of the two can be extracted by the same standard,and the semantic association and indirect fusion of the two can be carried out.Shared convolution is successfully connected with LSTM conditional coding model and attention mechanism.Opinion summarization refers to automatically summarize a large number of viewpoints and generate a concise and all-inclusive summary without redundancy.Aiming at a given number of micro-blog topics and a collection of multiple Microblog texts related to each micro-blog topic,this paper designs an opinion summarization model suitable for micro-blog text.Firstly,keywords corresponding to the target topic are extracted based on BTM model.Based on the co-occurrence relationship between each micro-blog text and keywords set of the target topic,the correlation between micro-blog text and the target topic is analyzed in fine granularity.Then,on the basis of weighted averaging of word vectors,PCA mechanism is added to vectorize the microblog text semantically,and then clustering is used to identify the opinion categories of a given target topic.Finally,considering the importance of micro-blog,the similarity of microblog,the length of microblog,the most representative microblog is extracted from each opinion category.Representative microblogs of all categories are combined together to form a non-redundant extractive opinion summarization for a given target topic.The innovation of this model lies in: on the one hand,the importance and information comprehensiveness of micro-blog text are analyzed from the fine-grained level by extracting viewpoint keywords;on the other hand,viewpoint categories are further identified from the overall semantic level of text sentences.The combination of local analysis and global analysis improves the accuracy of viewpoint Abstract extraction for short text.