Ruddit: Norms of Offensiveness for English Reddit Comments

TL;DR

This paper introduces Ruddit, a novel dataset of Reddit comments with fine-grained offensiveness scores, and evaluates neural models' ability to predict these scores, aiming to improve nuanced offensive language detection.

Contribution

The creation of the first dataset with real-valued, fine-grained offensiveness scores for Reddit comments and an assessment of neural models' predictive performance.

Findings

High reliability of offensiveness scores from Best--Worst Scaling

Neural models show promising but imperfect prediction accuracy

Dataset enables more nuanced offensive language detection

Abstract

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.