Filling the Blanks: Context-driven Detection and Correction of Cherry-picking in News Reporting

Abstract

Cherry-picking involves suppressing (censoring) or distorting evidence that supports the counter argument. Cherry-picking facts in news reports by mainstream media distorts public perception, undermines trust, and fuels misinformation by presenting a biased or incomplete narrative. Manually identifying suppressed statements in news stories is challenging and time consuming. In this study, we introduce a novel importance-based approach to automatically spotting and correcting cherry-picking by identifying then substituting missing important statements in a target news story with the help of contextual information from other news sources with different biases. Additionally, we showcase the flexibility of our approach by utilizing different methods to estimate a statement's importance including fine-tuned embedding models, zero and few-shot generative models, in addition to unsupervised methods. Furthermore, this research introduces a novel dataset specifically designed for training and evaluating cherry-picking detection methods. Our best performing method achieves an F-1 score of above 90% in estimating a statement's importance. Moreover, results show the effectiveness of the proposed approach in correcting cherry-picking and bringing the biased narrative closer to its neutral alternative by nearly 12%. Finally, through thorough experimentation, we provide answers to a set of important research questions related to cherry-picking detection and correction.