Research On Stance Detection Of Users In Social Media

With the rapid development of social media and internet,more and more people express their stance towards a certain target through message on social media.The target may be a person,an organization,a hot issue,a movement,or a policy.Formally,there are three stance categories:support,opposition and non-stance.Stance detection aims to automatically determine whether the author is in favor of or against a given target.Automatic detection of user's stance is conducive for the improvement of products,services,policies,etc.and the government's monitoring of public opinion,and the customers to make decisions on whether or not to buy products or servicesThis dissertation launches the research of the theory and application of stance detection based on text information extracted from message itself and external messages.First,to overcome the problem that the polarity of the text and the stance of the user's are not always consistent,we propose a joint neural network model,which can jointly learn the stance and sentiment information.Then,we propose a hierarchical attention network,which can effectively integrate the influence of lexical,sentiment,syntactic and argument information extracted from messages on stance detection.Finally,we explore the impact of multi-target data on stance detection to overcome the problem of fewer labeled samples in a single target.The main contributions of this thesis lie in;(1)Stance detection via sentiment information.In principle,the sentiment information of a message highly influences the stance and the stance may be opposite to the opinion So,we propose a joint neural network model to predict the stance and sentiment of a message.This model can learn both representation and interaction between the stance and sentiment collectively.Empirical studies demonstrate that our proposed joint neural model can effectively leverage the sentiment information of a message to improve the performance of stance detection(2)Stance detection model based on multiple linguistic information.With the traditional machine learning model,we explore and analyse the influence of four linguistic information including lexical,morphology,semantic and syntax features for stance detection.Based on this,we propose a hierarchical attention neural model to explore the influence of various linguistic representations for stance detection.We use two types of attention mechanisms:linguistic attention and hyper attention,while the former helps learn the mutual attention between the document and the linguistic information,the latter helps adjust weight of different linguistic features.Detailed evaluation on two benchmark datasets demonstrates that the proposed hierarchical attention neural model can effectively integrate multiple linguistic features for stance detection.(3)Stance detection model based on multi-target data.There is some common information among different targets that can be used to transfer knowledge between targets to make up for the lack of labeled samples in one target.Therefore,we explore the mutual influence and assistance of stance detection among multi-target data,and propose an adversarial multitask learning stance detection model on multi-target data,which takes stance detection as the main task,and target classification and sentiment classification as the auxiliary task.At the same time,the joint learning among three tasks and the adversarial learning between different targets,could obtain target-independent features to help stance detection.The experimental results in four targets show that multi-target adversarial multi-task learning is beneficial for stance detection,and demonstrate the effectiveness of the proposed method.At present,the research of stance detection of social media texts is still in the initial stage and this thesis is an explorative work.Focusing on text information,we propose effective methods to improve the performance of stance classification.The research has achieved some preliminary success.We hope it can not only help researchers in this area but also promote the development of deep natural language understanding.