Tackling toxicity in Arabic social media through advanced detection techniques

Area of Science:

• Natural Language Processing
• Social Media Analysis
• Computational Linguistics

Background:

• Online social networks facilitate communication but also host toxic content like hate speech and cyberbullying.
• Existing toxicity detection efforts primarily focus on English, neglecting other languages.
• There is a significant need for resources to address online abuse in Arabic.

Purpose of the Study:

• To create a standard Arabic dataset for toxicity and abuse detection on online social networks (OSNs).
• To evaluate the performance of various machine learning and transfer learning models on this new dataset.
• To advance the field of Arabic natural language processing for content moderation.

Main Methods:

• Construction and expert annotation of a novel Arabic dataset for toxicity detection.
• Experimentation with sixteen machine learning algorithms, FastText, and seven transfer learning architectures.
• Utilized four word embedding techniques: bag of words (BOW), term frequency-inverse document frequency (TF-IDF), FASTTEXT, and bidirectional encoder representations from transformers (BERT).

Main Results:

• The fine-tuned MARBERTv2 model combined with BERT embedding achieved the highest performance.
• Achieved an F1-score of 92.43% and an accuracy of 92.21% in classifying toxic Arabic content.
• Demonstrated the effectiveness of the developed dataset and advanced models for Arabic toxicity detection.

Conclusions:

• The study presents a significant breakthrough in classifying toxic tweets in Arabic.
• Highlights the importance of developing multilingual resources for addressing online toxicity.
• The proposed dataset and MARBERTv2 model offer a robust solution for Arabic content moderation.