BLCU_NLP at SemEval-2019 Task 7: An Inference Chain-based GPT Model for Rumour Evaluation

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1090–1096 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1090 BLCU NLP at SemEval-2019 Task 7: An Inference Chain-based GPT Model for Rumour Evaluation Ruoyao Yang, Wanying Xie, Chunhua Liu, Dong Yu (�) Beijing Language and Culture University, Beijing, China {yangruoyao97, xiewanying07, chunhualiu596}@gmail.com yudong@blcu.edu.cn Abstract Researchers have been paying increasing at- tention to rumour evaluation due to the rapid spread of unsubstantiated rumours on social 7.media platforms, including SemEval 2019 task However, labelled data for learning ru- mour veracity is scarce, and labels in rumour stance data are highly disproportionate, mak- ing it challenging for a model to perform supervised-learning adequately. We propose an inference chain-based system, which fully utilizes conversation structure-based knowl- edge in the limited data and expand the train- ing data in minority categories to alleviate class imbalance. Our approach obtains 12.6% improvement upon the baseline system for subtask A, ranks 1st among 21 systems in sub- task A, and ranks 4th among 12 systems in subtask B. 1 Introduction With the universality of the Internet, social media has become the main channel for acquiring and ex- changing information. However, the free flow of information has given rise to the prevalence of ru- mours, among which fake ones are harmful since they are generally convincing and hard to distin- guish. To address this problem, we need automatic rumour veracity classification on social media. A large amount of rumour stance instances on social media have been employed to assist the model in making better predictions regards the rumour’s ve- racity. Rumour stance classification and rumour ve- racity classification are two subtasks of SemEval 2017 Task 8 (Derczynski et al., 2017) and Se- mEval 2019 Task 7 (Gorrell et al., 2018). Sub- task A predicts the stance of a post replying to a rumourous post, in terms of supporting, denying, querying and commenting the rumour. Subtask B anticipates the veracity of a rumour as true or false given the rumourous post and a set of additional resources. Apart from variations in models, research in this area mainly focuses on the special characteristics of data coming from social media: conversation structure, rich intrinsic features, skewed distribu- tion toward the comment class in rumour stance data and scarcity of available data for rumour ve- racity classification. While most pioneering works treated rumour evaluation as a single-tweet task, attempts to utilize the conversation structure in- cluded pairing source and replies together to make up input (Singh et al., 2017), and adopting the full conversation thread as input in the time se- quence (Kochkina et al., 2017). With the real- ization of rich features hidden in tweet contexts, Qazvinian et al. (2011) was one of the first who extracted them and combined them with model in- put. The feature sets were augmented during the following work. In trying to acquire more com- prehensive information, not only features of the tweet for prediction were taken into considera- tion, but also …

Content truncated to 500 words