PANther: Party-specific Attention-based Networks for Accurate Political Perspective Detection

초록

The rise of online news platforms has worsened echo chambers and political polarization, prompting research on detecting the political perspective of news articles. News articles often contain complex information, making it challenging to understand them solely through text analysis. To address this, researchers have turned to external knowledge graphs (KGs) to enrich textual information. Existing KG-based approaches incorporate common-sense knowledge about political entities, but fail to capture different opinions and sentiments associated with the same entity across parties. In our work, we propose party-specific attention-based networks named PANther, that aims to construct two political KGs (i.e., KG-liberal, KG-conservative) that leverage distinct opinion information for different parties. We then apply attention-based networks that effectively integrate the external knowledge from these KGs to detect the political perspective of news articles. By conducting extensive experiments, we demonstrate the enhanced performance of PANther in terms of both accuracy and efficiency.

키워드