Research On Deep Learning Based Stance Detection Of Interactive Text

With the rapid development of Mobile Internet and Social Media,more and more people are able to participate in the discussion of Hot Spots and express their opinions at anytime anywhere.For some specific topics,interactive texts between users are the main carrier of their emotion and stance.Detecting users' stance of interactive text is of great value in both research and application.The existing researches of stance detection consist of machine learning based approach and deep learning based approach.The former relies on the construction and selection of features,which stays in the stage of shallow semantic representation.As for the latter,it can extract deep semantic features for both texts in interaction.However,it ignores the interactive context of discussion.Another equally important aspect is,due to limited information conveyed by interactive texts,the deep learning approaches cannot leverage the relevant background knowledge for stance detection.In view of the above problems,this paper proposes a deep learning model incorporating attention mechanism and background knowledge for stance detection of interactive text.In this paper,the bidirectional Long Short-Term Memory is used for jointly modeling the sequence of interactive text,Quote-Response pair(Q-R pair).The Self Attention and Cross Attention are incorporated to locate salient part from textual context of Q/R and capture the interactive argumentations between Q-R pair,which acquire a better representation for stance detection of interactive text.Experimental results on three online debate datasets,including Internet Argument Corpus(IAC),Debatepedia(DP)and Agreement by Create Debaters(ABCD),show that the model achieves significant performance in stance detection of interactive text,outperforming the baseline by 8.8% and 19.6%,measured by accuracy on IAC and DP,26.1% measured by average 1-score on ABCD.The End-to-End deep learning based model for stance detection is short of capability to incorporate background knowledge.Target to the problem,this paper further studies a stance detection model with background knowledge embedding for interactive text.The model first constructs the notional words of interactive text as corresponding query to retrieve relevant background knowledge from Wikipedia.Then,with the help of attention mechanism,the feature of background knowledge is extracted by Deep Memory Network from the retrieved results,and embedded into the representation of interactive text.Experimental results indicate that,with suitable settings for embedding layers,including the number of layers and connection ways,the model has a higher improvement of 1.7% measured by accuracy on DP and 0.5% measured by average 1-score on ABCD.These results prove that the background knowledge embedding can further enhance the performance of stance detection of interactive text.