Abstract:
Sentiment analysis plays a pivotal role in the detection of rumors on social media. Existing rumor detection methods predominantly rely on the emotional features of articles, often neglecting the combined sentiment emanating from user comments and the articles themselves. Single-task deep learning models for rumor detection face challenges due to a shortage of labeled data, hindering substantial improvements in accuracy. To address these challenges, this paper introduces a multi-task joint learning approach for rumor detection, leveraging the inherent correlation between rumor detection and joint sentiment detection tasks.This innovative method integrates joint sentiment detection and rumor detection tasks within the same model. Initially, an encoder is constructed to extract semantic features from articles and their corresponding comments, mapping them into a unified semantic space. An attention mechanism is employed to meticulously blend and weigh these semantic features. Subsequently, an emotion extraction network, founded on a common emotion classifier and emotion lexicon, extracts emotional features from both articles and corresponding comments. Finally, the semantic features and joint emotional features are fused and fed into two shared loss function classifiers, yielding classification results for both rumor detection and joint sentiment detection.Experimental results demonstrate the superior performance of the multi-task model over its single-task counterpart. When evaluated on the publicly available Chinese Weibo-16 dataset and English Twitter-15 dataset, the proposed model exhibits a notable improvement of 3.5 and 2.9 percentage points in accuracy, along with a commendable increase of 3.1 and 3.9 percentage points in F1 score, respectively.