Research On Stance Detection In Social Media Text By Incorporating Topic Target Information

With the rapid development of the Internet and mobile Internet and the popularity of social media,more and more people browse the hot news report and post their viewpoints,stance and emotions in the network any time and any place.Therefor,the research on text sentiment analysis attracts much research interests.The majority of existing sentiment analysis only focus on the sentiment polarity determination for input text,but ignores the detection of stance to specific target topic in text.Actually,in many practical application scenarios,users desire more on the stance in the text rather than the sentiment in text.Therefore,the research of social media text stance detection towards specific target has great social and commercial value.The existing research on stance detection are mainly based on machine learning with traditional semantic features and based on deep learning.The construction cost of traditional semantic features is costly.Meanwhile,the migration of model is complicated.As for the end to end deep learning based methods,they usually ignore the topic target information.Aiming at the shortcomings of the existing methods,this study focuses on stance detection techniques from the following two aspects.Aiming at the problem that the existing stance detection research lacks the consideration of target information,we investigate a stance detection method which combines the target topic information and text information through conditional encoding.The target topic information is employed as priori knowledge to guide the encoding of textual information in stance detection.By considering the characteristics of text with stance,the conditional encoding model is further improved.The evaluations on Sem Eval 2016 English stance detection dataset and NLPCC 2016 Chinese stance detection dataset show that our proposed stance detection method based on conditional encoding achieves the micro-average F1 values of 0.671 and 0.698,respectively.It is shown that this method improve the performance of the stance detection effectively.Due to the characteristics that different target topics have different priorities in text content,we propose a novel stance detection method which utilizes the target topic as the guide of attention mechanism to assign different attention weights to text content for extracting the pattern of stance detection.Considering that attention mechanism and conditional encoding introduce topic target information from encoding perspective and decoding perspective,respectively,we propose a stance detection method by incorporating neural network with attentional mechanism and conditional encoding.In the“encoding”process,conditional encoding is applied to guide the encoding of text information,while in the “decoding”process,attention mechanism is applied to extract classification patterns from encoded information for stance detection.The experimental results on Sem Eval2016 English dataset and NLPCC 2016 Chinese dataset show that the proposed method achieves the micro-average F1 values of 0.689 and 0.716,respectively.Comparing with the top performance system in the two corresponding evaluation tasks,our method improves the micro-average F1 value for 1.08% and 0.61%,respectively.It has shown the effectiveness of incorporation of attention mechanism and conditional encoding neural network model in social media stancedetection.