Fake Information Detection Method Based on BERT Pre-Trained Model

With the rapid evolution and ubiquitous penetration of social media, the proliferation of false information-characterized by rapid diffusion and contextual complexity-has become a critical societal challenge, distorting public cognition in domains like public health and electoral politics, and threatening social cohesion or even triggering unrest in extreme cases. Traditional detection methods, such as keyword matching or shallow machine learning models, often struggle to capture semantic ambiguity and contextual dependencies. To address this, this study proposes a refined false information detection framework based on BERT, leveraging its strengths in deep semantic parsing and dynamic contextual representation. The framework integrates multi-dimensional text preprocessing (e.g., noise reduction and entity recognition) with a pre-trained fine-tuning paradigm to capture nuanced semantic associations critical for distinguishing false information, avoiding information loss from manual feature engineering. Experimental results on the Weibo dataset demonstrate that the proposed method achieves relatively advanced performance across core evaluation metrics. Moreover, it maintains good stability under data distribution shifts and adversarial perturbations, highlighting its potential for deployment in real-world applications.