Predicting social media users' indirect aggression through pre-trained models.

Area of Science:

• Social Psychology
• Computational Linguistics
• Cyberpsychology

Background:

• Indirect aggression is a growing problem on social media, difficult to detect with traditional methods.
• Existing self-reporting questionnaires are impractical for online environments due to cost and objectivity challenges.
• Understanding and mitigating online indirect aggression is crucial for social media platform management.

Purpose of the Study:

• To develop and evaluate a novel model for predicting indirect aggression in online social media environments.
• To classify indirect aggression into subtypes: social exclusion, malicious humour, and guilt induction.
• To leverage pre-trained models for enhanced prediction accuracy.

Main Methods:

• Constructed features from Weibo user data, including basic, dynamic, and content-based attributes.
• Developed a prediction model by integrating these features with large-scale pre-trained models.
• Classified indirect aggression into three distinct subtypes for granular analysis.

Main Results:

• The proposed prediction model (ERNIE) demonstrated superior performance compared to baseline pre-trained models.
• The model significantly improved the accuracy of predicting indirect aggression online.
• Empirical evidence supports the model's effectiveness in identifying various forms of indirect aggression.

Conclusions:

• The study presents a practical and effective model for predicting users' indirect aggression on social media.
• This research enhances the understanding of online indirect aggression behaviors.
• The findings can aid social media platforms in content moderation and user management strategies.