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

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