Abstract: We develop a double-channel classifier to detect the veracity of social media rumors, relying only on the most basic textual information. Our model first assigns each thread into a “certain” or “uncertain” category. Since authors with a proprietary source of information are likely to post threads with a certain textual tone, we apply lie detection algorithms to certain texts. In contrast, as uncertain threads are arbitrary, we examine whether the replies are in accordance with the threads instead of applying the lie detection algorithms. This approach yields a macro-F1 score of 0.4027, outperforming all the baseline models and the second-place winner of SemEval 2019 Task 7. Further, we show that dividing the sample into two subgroups significantly improves the classification accuracy, reinforcing our claim that applying appropriate classifiers is crucial in rumor veracity detection.
0 Replies
Loading