Abstract:

[Objective] With the rapid development of the internet and the accelerated advancement of generative artificial intelligence technologies, the spread of fake news has become increasingly severe. Detecting fake news efficiently and accurately has emerged as a critical research focus in academia. Traditional fake news detection methods primarily rely on semantic features of news texts, while neglecting in-depth exploration of features from other dimensions of the text as well as information such as user comments. [Methods] This paper proposes a fake news detection method based on BERT-SentiMHCA (Bidirectional Encoder Representations from Transformer-Sentiment Features with Multi-Head Cross Attention) model. Building on semantic features of news content, the method comprehensively integrates sentiment features of both news content and user comments, as well as their sentimen similarity. Through multi-head self-attention and multi-head cross-attention mechanisms, it achieves effective fusion of multi-level features, combining textual and emotional characteristics to enhance the accuracy and robustness of fake news detection. The specific methodology includes: (1) utilizing the pre-trained language model BERT to extract deep semantic representations of news texts; (2) constructing sentiment classifiers to extract sentiment features from both news content and comments; (3) designing a fusion module that integrates features via attention mechanisms and employs a neural network classifier to build the fake news detection model. [Conclusions] Experimental results on four public fake news datasets demonstrate that the proposed BERT-SentiMHCA model achieves notable performance improvements in terms of accuracy, precision, recall, and F1-score compared to several mainstream baseline models. These results validate the capability of the proposed method in multi-level textual feature extraction, significantly enhancing the performance of fake news detection.

Key words: fake news detection, sentiment features, BERT, feature fusion