摘要
With the surge in online information, unverified claims and fake news, especially in politics and economics, have become rampant. This rise makes distinguishing true from false news crucial, as fake news can manipulate public perception and cause socio-economic harm. Previous research indicates that human-written fake news often employs propaganda techniques to seem credible. However, the smaller volume of fake news data compared to real news leads to imbalances in training datasets. This imbalance can cause models to learn incorrect correlations instead of accurately identifying the semantic features of deceptive content. This paper introduces a counterpropaganda semantic learning framework for detecting fake news. Utilizing existing fake news datasets, we automatically generate propaganda-related summaries, reasons, and scores using a large language model. We then create counterpropaganda semantics by enhancing the potential propaganda elements in news content. The model learns the relationship between news and propaganda through a multi-task learning strategy. Experimental results demonstrate that this framework improves small language models’ accuracy and semantic understanding in detecting fake news while reducing the need for intermediary pre-training and manual review of generated training data.