Research On Stance Detection Based On Graph Model Information Enhancement

Stance detection aims to automatically judge the stance and attitude(such as support,opposition,etc.)of users’ opinion text under a given target topic,which is of great significance to understand people’s will,government decision-making and intelligent recommendation.At present,deep neural network models have been widely used because of its excellent performance,but a large amount of information outside the text has not been well considered and used,especially the user information interaction in social media.The research content of this paper mainly includes the following three aspects:(1)Research on stance detection based on context relevanceAiming at the problem that the interaction information is easy to be ignored due to the single exchange of context information,this paper expands the context interaction information based on the principle of logic,and uses the graph convolution network to construct the corresponding graph structure.Firstly,the maximum pool and two-way long-term and short-term memory network are used to improve the ability of the model to extract semantic features from text.Secondly,the feature weight is dynamically adjusted through the self attention mechanism.Finally,the original context reply relation is expanded to enrich the context association relation,and the graph convolution network is used to mine the text interaction information to improve the utilization of context information.Experimental results show that the proposed method is more effective than the baseline model.(2)Research on stance detection based on user associationAiming at the problem of insufficient utilization of user information and failure to effectively combine the existing text,this paper explores the user information of the text and establishes the semantic features of text and user information.This paper improves the performance of the model from two aspects:enhancing the information of the text itself and the information of the associated text.This paper uses attention mechanism to enhance the perceptual representation of text and users,constructs a user-based graph network by mining user information,uses convolution operation to mine the stance information of similar texts under the same user,and improves the similar association between different texts from the perspective of users.In this paper,experiments are carried out on the debate data set,and the results show the effectiveness of the user association model.(3)Research on stance detection based on dependency syntaxFor the problem that the relationship between different components in a sentence is weak and the model is difficult to completely identify the syntactic relationship of different components in a sentence,this paper employs dependency to enhance the information sharing among multiple components in a sentence,and designs a graph structure based on syntactic dependency.At the same time,in order to solve the importance of different component information in sentences,this paper introduces graph attention network,which further adaptively allocates the weights of different neighbors by aggregating neighbor nodes,and improves the performance of stance detection model combined with user-based graph structure.Finally,considering the universality of data sets and the practicability of dependency model,this paper also proposes a dependency based system,which makes the system applied to a wider range of data sets.To sum up,from the perspective of multi-level information,this paper studies the graph based convolution network from the perspective of constructing and expanding context relevance,user relationship and dependency relationship,so as to solve the problems of different information utilization and graph model construction.Finally,the information enhancement method based on graph model proposed in this paper has achieved some preliminary results.It is hoped that these studies can bring some help to the stance detection task.