Textstance Detection Based On Attention Mechanism And Adversarial Multitask Learning

With the rapid development of social media platform,network users may browse news reports and express and exchange their views and feelings at any time.Thus,sentiment analysis has attracted worldwide research attention.However,traditional sentiment analysis methods only determine the polarity the text,but lack of the capability to detect the stance corresponding to specific topics in the text,while the stance is always the interest of users in many practical applications.Therefore,the stance detection of social media text has shown great scientific and commercial value.Stancedetection aims to analyze and determine the stance contained in the text on specific topics,namely support,opposition and neutral.Most existing research may be camped into machine learning based approach and deep learning based approach.Machine learning based approach usually relies on artificial feature construction and filtering,but the process of feature construction is often time consumes.The deep learning based approach avoids tedious feature engineering through end-to-end encoding and decoding,but often lacks of the deep modeling capability during text interaction process.In addition,due to the small size of available data set,the learning ability of a single model is often limited.Therefore,this study focuses on the text stance detection method by incorporating attention mechanism and adversarial multi-task learning.In order to learn the stance-related semantic representation in the interactive text,this paper first studies the stance detection method based on bidirectional long-term and short-term memory network.On this basis,attention mechanism and match mechanism are further introduced to reduce the influence of the excessive length differences between texts and enhance the weight of key matching to stance detection.Experimental results on the Fake News Challenge 2017 dataset show that the proposed method achieved the same F1 Averageperformance as the best system,while the Score value increases by 0.2.Considering that the small size of available training dataset puzzles the representation learning capability of single model,this study further investigates the stance detection method based on multi-task adversarial learning.This method firstly established a multi-task learning framework by taking stance detection as the as the main task and reading comprehension as the second task.In order to solve the noise problem during parameter sharing process in the multi-task learning,the adversarial training mechanism is introduced.It is expected to improve the stance detection performance through reduce the overlap between feature spaces fordifferent tasks.The experimental results show that by introducing multi-task adversarial learning mechanism,the F1 Average is increased by 0.05 while Score value is further increased by 0.7,which achieves the best-known performance.This shows the effectiveness of our proposed stance detection method incorporating attention mechanism and multi-task adversarial learning mechanism.