Abstract
Currently, early rumor detection has garnered significant attention due to the rapid dissemination of information on social media. Many existing studies approach rumor detection as a traditional classification task, disregarding the existence of unverified rumors that require future verification. To address this limitation, we present T3-ERD, a novel framework that employs an improved BigBird transformer-based model for detecting rumors of specific types while concurrently leveraging reinforcement learning to determine early time checkpoints and categorize unverified rumors. And we introduce a novel bucket strategy, Online Dynamic, which dynamically clusters comments and processes them in batches, demonstrating remarkable effectiveness. The T3-ERD achieves F1 score, accuracy, and early rate of 0.830, 0.837, and 0.220 on the Twitter15 dataset, and 0.837, 0.844, and 0.205 on the Twitter16 dataset, respectively.

This publication has 41 references indexed in Scilit: