Stance Detection-based Group Decision Prediction

Nowadays,more and more people are beginning to share their opinions on social media online.Obtaining the user's stance from these data and predicting future group decision-making based on the stance is a powerful way for the government to understand the people's sentiment and for the company to improve its products.At present,a variety of solutions have been proposed in academia,but there are still the following limitations:1)Stance detection problem: Existing models only consider text dimensional information without introducing user characteristics;2)Group decision prediction problem: Existing models only consider the influence of the interaction of opinions,but not the influence of the environment on the individual.This thesis mainly studies the above-mentioned problems.On the one hand,it introduces user characteristics and designs a cross-modal algorithm that combines user and text characteristics to achieve stance detection;on the other hand,it combines stance detection model,attention mechanism and recurrent neural network to model the environmental impact of opinion evolution,the stance of each user are predicted,and the stances of all users are accumulated through a integration algorithm to obtain the prediction of group decision-making.Specifically,the main research contents of this article are as follows:Firstly,a position detection algorithm based on user and text representation learning is proposed.The algorithm converts the text data into low-dimensional vectors through the pre-trained BERT model,and designs a multi-layer neural network model to capture the non-linear mapping relationship between users,text vectors and stance: The model firstly embeds the user's social relationship into a low-dimensional vector representation through GCN,and then uses a multi-modal fusion algorithm to fuse the obtained user vector and text vector to obtain a joint vector which fully integrated the text dimension information and user dimension information semantically.And finally the stance of the user is classified based on the joint vector.Besides,a group decision prediction algorithm based on position detection is proposed.The algorithm obtains the position vector representation of the text through the position detection model,and designs a multi-layer neural network model to capture the transfer relationship of the user's stance changes: First,the attention mechanism is used to capture the influence of the background on the user's text generation.Then the transfer relationship of the stance change in the text sequence is captured based on the recurrent neural network.Finally,based on the transfer relationship,predict the user's stance at the next epoch.The algorithm also designs a stance integration module to integrate each user's stance predicted by the above-mentioned multi-layer network model to realize the prediction of group decision-making.The experimental results show that the accuracy and stability of this algorithm in group decision-making tasks are better than other comparison methods,which proves the feasibility of the group decision prediction algorithm based on neural network.In general,the model designed in this thesis proposes different degrees of improvement for stance detection and group decision prediction,and experiments on real datasets have proved that the algorithms in this thesis have improved performance in their respective tasks.