PARS: A Position-Based Attention for Rumor Detection Using Feedback From Source News

Rumor detection focuses on verifying unconfirmed information. Using user feedback within a timeframe enhances the quality of detection. Given the rapid spread and societal impact of rumors, both accurate validation and early detection are critical. Despite numerous studies in this field, designing effective feature representations still remains a central challenge, especially for models aiming to capture the dynamics of rumor propagation. This paper introduces PARS (Position-based Attention for Rumor detection using feedback from Source news), a novel method that treats a rumor and its associated feedback as a unified unit. Each tweet is considered like a word, and the sequence of tweets forms a sentence that conveys the collective meaning. This sentence-level-like structure captures the semantic and structural aspects of rumor spread. PARS incorporates a position-aware multi-head attention layer, which attends to tweets based on their roles in the propagation tree. A tweet reordering mechanism is applied to reflect the actual dissemination pattern, ensuring that position-sensitive attention captures the most relevant signals. Experimental results show that PARS achieves superior performance compared to existing methods, particularly in terms of accuracy, F1-score, and early detection, demonstrating its effectiveness in real-world rumor classification scenarios.