Advanced Fake News Detection in Chinese Using ECL (ERNIE-CNN-LSTM) Model

With the proliferation of Internet technology and social media platforms, information can be easily exchanged across regions at unprecedented speeds, but this can also reduce the quality of information - especially in an era when false and misleading content spreads like wildfire on social media. False news spreads among people, misleads individuals and causes great harm in politics, economy, public health, and other fields. Despite the challenges of fake news detection and intervention, excellent technical support is provided in this regard thanks to recent advances in big data, natural language processing (NLP), deep learning paradigms, and more. Currently, however, the technology that can be used to determine whether news is true or false has been developed primarily in English - and there is an urgent need to optimize doing this in China at a non-minimal scale and complexity. The motivation and purpose of this study is to achieve the same high accuracy of fake news detection in the Chinese context. This article suggests a high-precision Chinese fake news detection system that innovatively combines the ERNIE 3.0, TextCNN, and BiLSTM models. Each model has a distinct role in our system: ERNIE 3.0 collects sophisticated semantic information using its robust pre-trained language model, which is particularly useful for comprehending Chinese text; TextCNN is good at extracting local features from text and identifying essential patterns. BiLSTM excels in identifying long-distance dependencies and contextual information in text. We employ a progressive training technique, initially training and optimizing the performance of each model independently, and then in the integrated model by weighting the outputs of these independent models so that they learn from one another, hence enhancing overall predictive power. As a result, our model can effectively integrate the benefits of each, improve the complete analysis capability of the complex characteristics of fake news, and increase robustness. It is worth noting that our ECL model (Ernie-TextCNN-BILSTM integration) achieved an accuracy of 0.7608 in the Chinese fake news identification assignment, outperforming typical machine learning methods. This will be an important step in contributing to research in the broader field of information verification and fake news detection while paving the way for future work across other languages and other areas.