Ruddit: Norms of Offensiveness for English Reddit Comments
This addresses the need for more nuanced offensive language detection to improve user well-being and inclusivity on platforms like Reddit, though it is incremental as it builds on existing datasets with a new annotation approach.
The authors tackled the problem of detecting offensive language on social media by creating the first fine-grained, real-valued dataset of English Reddit comments with offensiveness scores from -1 to 1, annotated using Best-Worst Scaling to reduce biases, and showed that this method produces highly reliable scores while evaluating neural models on it.
On social media platforms, hateful and offensive language negatively impact the mental well-being of users and the participation of people from diverse backgrounds. Automatic methods to detect offensive language have largely relied on datasets with categorical labels. However, comments can vary in their degree of offensiveness. We create the first dataset of English language Reddit comments that has fine-grained, real-valued scores between -1 (maximally supportive) and 1 (maximally offensive). The dataset was annotated using Best--Worst Scaling, a form of comparative annotation that has been shown to alleviate known biases of using rating scales. We show that the method produces highly reliable offensiveness scores. Finally, we evaluate the ability of widely-used neural models to predict offensiveness scores on this new dataset.