ECNU at SemEval-2017 Task 8: Rumour Evaluation Using Effective Features and Supervised Ensemble Models

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 491–496, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics ECNU at SemEval-2017 Task 8: Rumour Evaluation Using Effective Features and Supervised Ensemble Models Feixiang Wang1, Man Lan1,2∗, Yuanbin Wu1,2 1Department of Computer Science and Technology, East China Normal University, Shanghai, P.R.China 2Shanghai Key Laboratory of Multidimensional Information Processing 51151201049@stu.ecnu.edu.cn, {mlan, ybwu}@cs.ecnu.edu.cn Abstract This paper describes our submissions to task 8 in SemEval 2017, i.e., Determin- ing rumour veracity and support for ru- mours. Given a rumoured tweet and a plethora of replied tweets, subtask A is to label whether these tweets are support, de- ny, query or comment, and subtask B aim- s to predict the veracity (i.e., true, false, and unverified) with a confidence (in range of 0-1) of the given rumoured tweet. For both subtasks, we adopted supervised ma- chine learning methods incorporating rich features. Since the training data is imbal- anced, we specifically designed a two-step classifier to address subtask A . 1 Introduction With the rapid development of social media in re- cent years, people cannot only stay abreast of on- going events and breaking news, but also express their own views freely. News can spread quickly in social media platforms through a large amount of users, whilst those pieces of unverified infor- mation often spawn rumours. The RumourEval (Derczynski et al., 2017) task aims to identify how users in social media networks regard the originat- ing rumours and reply to them, as well as analy- s.sis and determining veracity of rumoured tweet- The organizer provides tree-structured con- versations that are associated with breaking news and consisting of originating rumoured tweets and tweets replying to them. There are two subtasks in RumourEval. The propose of subtask A is, given the related break- ing news, to predict the class (i.e., support, de- ny, query, and comment) of the originating ru- moured tweet (i.e., source tweet) and reactions (i.e., replied tweets). The goal of subtask B is to determine the veracity and confidence of the given rumoured tweet, participants are required to return a label of rumour as true, false or unverified, with a confidence value in the range of 0-1. We treated the two subtasks as multi- classification problems, and designed multiple effective natural language processing (NLP) features to build classifiers to make predictions. Besides, rumour detection is relevant to sentiment analysis, for example, support and deny can be viewed as positive and negative sentiment respectively. Therefore, we solved the problem with the aid of a number of sentiment-related features. Due to the imbalanced characteristic of the training data, we specifically adopted a two-step classifier to deal with subtask A. Firstly, tweets would be separated into two categories: comment and non-comment, then the tweets labeled as non-comment would be classified as support, deny or query. On the other hand, we directly adopted a three-classification system for subtask B to label rumoured tweets as true, false or unverified along with confidence. 2 System …

Content truncated to 500 words