Abstract

Join Zoom Meeting https://cl-cam-ac-uk.zoom.us/j/91334942020?pwd=eDJXd09wdW1FUDMvWFBoenovMDdNUT09

Meeting ID: 913 3494 2020 Passcode: 565263

In this talk I am going to present our recent work on rumour stance classification. This task consists of automatically classifying the stance of replies to a given rumour, which has been proven to help in the verification of the rumour itself. Traditionally framed as a four-class classification problem (support, deny, query, comment), the available datasets are highly imbalanced, with a large majority of examples in the comment class. However, the two most important classes for this task are deny and support, since they express the crowd perception towards the rumour. Aiming to improve the performance of classification models in these two most important classes, we experiment with traditional feature-based as well as BERT -based approaches, using well-known techniques for dealing with data imbalance problems. In addition, we re-evaluate two widely known shared tasks on rumour stance classification, highlighting that reliable and detailed evaluation needs to be performed in order to select systems for this task.