Abusive language detection in youtube comments leveraging replies as conversational context.

Area of Science:

• Computational Linguistics
• Social Media Analysis
• Natural Language Processing

Background:

• Social media platforms experience increasing hostility, leading to widespread online abusive behavior and harassment.
• Existing datasets for abusive language detection often lack contextual information, limiting their effectiveness.

Purpose of the Study:

• To introduce a novel, publicly available annotated dataset for abusive language detection in short texts from YouTube.
• To provide refined annotation guidelines for classifying abusive comments by topic (politics, religion, other).
• To establish strong baseline results for abusive language and topic classification using various models.

Main Methods:

• Development of a new annotated dataset comprising YouTube comments with contextual information (replies, video details).
• Labeling comments as abusive or not, and classifying them by topic.
• Implementation and evaluation of baseline classifiers (e.g., linguistic features, conversational context) for detection and classification tasks.

Main Results:

• The dataset includes labeled abusive comments from YouTube, with contextual data like replies and video information.
• Baseline models demonstrate strong performance on both abusive language detection and topic classification.
• Utilizing conversational context, specifically replies, significantly improves classification accuracy compared to using only linguistic features.

Conclusions:

• The newly introduced dataset is valuable for advancing research in abusive language detection and online harassment.
• Conversational context is a critical factor in accurately identifying abusive language in social media.
• The dataset enables further investigation into topic-specific nuances of online hostility.