Check out this new paper titled “Subjective Isms? On the Danger of Conflating Hate and Offence in Abusive Language Detection” by Amanda Cercas Curry, Gavin Abercrombie, and Zeerak Talat. This study delves into the critical nuances of detecting abusive language through natural language processing (NLP) and the implications of annotator subjectivity on research findings.
The methodology employed in this research involves examining how annotators’ personal views influence the labelling of hate speech and offensive content. The authors argue that while subjectivity can be valid in tasks like sentiment analysis, it poses significant challenges in hate speech detection. They highlight that conflating hate and offence can lead to invalid conclusions about the nature of hate speech.
Data for the study was drawn from various sources, including previous research on racism, sexism, xenophobia, homophobia, and transphobia. The researchers analysed the extent of annotator disagreement and its impact on the classification of abusive language. They propose that annotators’ backgrounds and lived experiences significantly influence their labelling decisions, leading to inconsistencies in data interpretation.
One of the key findings is the distinction between hate and offence. The paper posits that hate speech should be considered a culturally defined concept, whereas offence is subjective and varies from person to person. This differentiation is crucial for developing more accurate and reliable NLP models for detecting hate speech. The authors recommend that future work should incorporate theoretical frameworks to disentangle hate from offence and ensure that labelling practices reflect these distinctions.
The implications of this research are far-reaching, suggesting that NLP models need to be more context-aware and culturally sensitive. By recognising the orthogonality of hate and offence, researchers can improve the accuracy of hate speech detection and contribute to more effective policy-making in combating online abuse.