Leveraging deep learning for toxic comment detection in cursive languages.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Social media facilitates public expression, including hate speech and harmful content.
• Detecting toxic online content is vital for community safety and well-being.
• Urdu, a widely spoken low-resource language, lacks standard tools for toxic comment detection.

Purpose of the Study:

• To develop and evaluate a novel model for identifying toxic comments in Urdu text.
• To address the challenges of tokenizing and classifying complex, often ungrammatical Urdu comments.
• To enhance content moderation capabilities for diverse linguistic platforms.

Main Methods:

• A novel deep learning model utilizing transformers to identify salient features in Urdu sentences.
• Binary classification of text to flag toxic comments.
• Evaluation of models including bidirectional encoder representations from transformers (BERT) and GPT-2.

Main Results:

• The proposed fine-tuned model achieved a precision of 88.38%, outperforming existing methods.
• Bidirectional encoder representations from transformers (BERT) showed strong performance with 85.45% accuracy and 85.71% precision.
• GPT-2 was identified as the lowest-performing model in the study.

Conclusions:

• The developed model represents a significant advancement in Urdu toxic comment detection.
• The findings contribute to optimizing content moderation across various languages and platforms.
• This research addresses a critical need for effective tools in low-resource language environments.