See the words through my eyes: The role of personal traits in abusive language detection
Tsungcheng Yao, Sebastian Binnewies, Ernest Foo, Masoumeh Alavi · Expert Systems with Applications · 2025
Abusive language detection systems play a significant role in addressing cyberbullying. However, conventional detection systems primarily focus on the textual pattern of messages for their prediction, overlooking the reality that the same message can usually provoke different consequences for different users. In other words, personalised predictions are essential but currently absent. To address this limitation, this paper considers the rationality of introducing a set of psychological features to personalise abusive language detection tasks. The foundation of our work lies in the Antecedents, Behaviours, Consequences model (ABC model), asserting that an individual’s response to a trigger message is impacted not only by the trigger itself but also by their personal beliefs and attitudes. We develop a novel data preparation framework and construct a new abusive language dataset, incorporating psychological features from 505 online users. Logistic regression analysis illustrates that Irrationality and Self-down features positively correlate with the abusive class, while Rationality features exhibit a negative correlation. These results are supported by established psychological findings. Furthermore, a preliminary evaluation showed that our proposed psychological features improve a CNN-based detection system’s Macro and Weighted F1 scores by 4%–5% points when making personalised predictions. These results collectively make for an empirical case that underscores the importance of considering users’ psychological features in abusive language detection. Crucially, these findings also pave the way for developing personalised prediction systems. • Psychological features improve the detection system’s F1 scores by 4%–5% points. • Psychological theory and features are transferable to detection systems studies. • Due to a self-selection bias, online and ordinary users are substantially different. • A new abusive language dataset incorporates psychological features from 505 users.