Improving Rumor Detection Performance by Using Bias Attributes

With the rapid advancement of social media, the barriers to sharing information have significantly decreased, leading to the rampant spread of rumors across various platforms. Current natural language processing (NLP) techniques use models trained on datasets to detect rumors. However, these datasets often contain inherent biases, which, as has been demonstrated in other NLP tasks, can negatively impact the accuracy of the results. This study confirms the presence of significant bias in rumor detection datasets. Given the intertwined and complex nature of bias and rumors, both of which play a crucial role in assessing the trustworthiness of online content, the article argues that improving detection models should not only focus on performance but also prioritize detecting biased rumors. Addressing this dual challenge is essential to prevent rumor creators from exploiting biases to enhance the spread of false information. To tackle this issue at the data level, this study proposes integrating bias attributes into the training datasets, thereby improving the ability of models to identify biased rumors. Through experiments conducted across multiple rumor detection models, the approach has been shown to enhance the detection of biased rumors in both Chinese and English datasets without compromising overall detection accuracy. This method might be helpful in reducing the spread of biased rumors online, contributing to a healthier information ecosystem.