Research On Deep Learning Based Stance Detection In Social Media Text

With the rapid popularization of intelligent terminals and rapid development of Internet technology,more and more users express their views and stances towards emergencies on social media platform.The attitudes of users are also valuable in business decisions of commercial organizations and policy decisions of government.Traditional sentiment analysis technique classifies the texts in to positive or negative,but it is difficult to mine the stance of users to specific topics.Therefore,the study on stance detection to specific topic shows important academic value and broad application prospect.The existing stance detection research is mainly based on machine learning with feature engineering approach and deep learning based approach.The machine learning based approach in stance detection requires well organized linguistic knowledge and a large number of manually selected features,which is costly.The deep learning based approach normally regards the stance detection as text classification,but rarely uses the background knowledge in social media text.Meanwhile,the information of specific topic is seldom used in stance detection.Considering the above problems,in this study,the word embedding learned from social media text adopted to provide the background knowledge,and the deep learning based stance detection approach adopting the attention mechanism in deep memory network is investigated.Based on using large scale social media text to train word embedding as background knowledge,a stance detection method based convolutional neural networks(CNN)is firstly investigated.The experimental results on Sem Eval English stance detection dataset show that this method achieves average F-measure as 0.6752 in English dataset while the experimental results on NLPCC Chinese stance detection dataset achieves average F-measure as 0.7036.The overall performance are ranked the second in both Semeval and NLPCC evaluation.The observation show that the word embe dding pre-trained on social media text improves the stance detection performance effectively which is better than the random assignment of word embedding.Target to the problem that most existing stance detection methods ignore the information related to specific topic,in this study,a stance detection method which uses the attention mechanism in deep memory network to estimate the association between text sentiment expression and target topic.This method firstly loads the word embedding of text and topic,and then utilizes multiple computation hops to learn the representation of texts following the memory mechanism and attention mechanism in deep memory network.Finally,the stance to specific topic is determined based o n the learned representations.The experimental results on Sem Eval dataset shows that this method achieved 0.6821 average F-measure which outperforms the top system in Semeval based on transfer learning for 0.39%.Meanwhile,the experimental results on NLP CC dataset shows that 0.7140 average F-measure is achieved which outperforms the top system in NLPCC for 0.34%.These results indicate that the proposed method is effective in stance detection.