Hidden behind the obvious: misleading keywords and implicitly abusive language on social media

TL;DR

This paper investigates how abusive language detection models rely heavily on keywords, leading to false negatives and positives, and offers insights to improve model robustness against implicit abuse.

Contribution

It provides a detailed analysis of keyword influence on model errors and interactions with unseen data, highlighting gaps in current abusive language detection research.

Findings

Models often miss abuse without keywords

Models falsely flag non-abuse with keywords

Performance drops on unseen data

Abstract

While social media offers freedom of self-expression, abusive language carry significant negative social impact. Driven by the importance of the issue, research in the automated detection of abusive language has witnessed growth and improvement. However, these detection models display a reliance on strongly indicative keywords, such as slurs and profanity. This means that they can falsely (1a) miss abuse without such keywords or (1b) flag non-abuse with such keywords, and that (2) they perform poorly on unseen data. Despite the recognition of these problems, gaps and inconsistencies remain in the literature. In this study, we analyse the impact of keywords from dataset construction to model behaviour in detail, with a focus on how models make mistakes on (1a) and (1b), and how (1a) and (1b) interact with (2). Through the analysis, we provide suggestions for future research to address all three problems.