DAAL: Domain Adversarial Active Learning Based on Dual Features for Rumor Detection

Abstract

The widespread of rumors has a negative impact on society, and rumor detection has attracted significant attention. When a new event appears, the scarcity of corresponding labeled data causes a severe challenge to rumor detection. It is necessary to query high-quality unlabeled data and annotate it for detection. Previous studies on active learning for rumor detection can query the unlabeled samples on the decision boundary based on the textual features of posts. These studies, when selecting the optimal samples, have not sufficiently considered that new event posts are usually far from old events and differ in terms of sentiment, which often plays a key role in rumor detection. Therefore, overlooking these characteristics could potentially lead to sub-optimal performance in rumor detection based on these active learning methods. To this end, we propose domain adversarial active learning based on dual features (DAAL) for rumor detection, considering these characteristics. Specifically, we first extract dual features, including affective and textual features, to obtain representations of the posts. We then propose a new active learning method that selects the samples furthest from all labeled samples based on their dual features. This method helps our rumor detection model gain more labels from new, distant event posts for training. Finally, we introduce adversarial domain training, a method designed to extract transferable features across different events (or domains), to enhance the adaptability of our rumor detection model to new events. Experimental results demonstrate DAAL can select high-quality candidates and achieve superior performance compared to existing methods.