Ic-Bait: An Inferential Commonsense-Driven Model for Predicting Political Polarity in News Headlines

Abstract

The identification of political polarity in news headlines is a critical task since it influences the selection and distribution of news articles. In particular, when it comes to politics, many outlets prefer to publish their stories with misleading headlines that support their political ideology. Predicting this ideology using conventional methods is challenging due to the absence of sufficient syntactic and semantic information in short headline text. To compensate for this lack of information, Inferential Commonsense Knowledge (IC_Knwl) can be employed in a way similar to how people use commonsense inferences to perform a variety of tasks. However, without proper emphasis, the additional inferential context is prone to introduce unnecessary noise, preventing models from fully exploiting the acquired knowledge. To address the aforementioned bottlenecks, in this work we present our model IC-BAIT (Inferential Commonsense aware BiAs IdenTifier), which makes use of IC_Knwl to aid the task of political bias prediction in news headlines. Furthermore, we present two datasets annotated with bias labels, MediaBias and GoodNews. We conduct experiments on both of them to evaluate the reliability of our model. The experimental results demonstrate that our proposed model outperforms baseline models, regardless of whether they are trained with or without IC_Knwl. We also present a detailed analysis of when and why IC_Knwl is beneficial or counterproductive. Our data and scripts are publicly available at:  https://drive.google.com/file/d/1faCayrWTvs--EGVp0lCSvrPqUWwHr3x3/view?usp=sharing

Keywords: News bias, Media bias, Machine learning, NLP (Natural Language Processing), Inferential Commonsense Knowledge, Short Text Classification