Multilingual Hate Speech Detection: A Semi-Supervised Generative Adversarial Approach.

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)

Background:

• Social media's global reach presents challenges in detecting hate speech across diverse languages.
• Scarcity of labeled data hinders the development of effective hate speech detection models.
• Semisupervised and Generative AI approaches are crucial for overcoming data limitations.

Purpose of the Study:

• To develop an innovative multilingual semisupervised model for hate speech and offensive language detection.
• To combine Generative Adversarial Networks (GANs) with Pretrained Language Models (PLMs) for improved performance.
• To address the challenge of limited labeled data in cross-lingual hate speech detection.

Main Methods:

• Implementation of a multilingual semisupervised model integrating GANs with mBERT and XLM-RoBERTa.
• Utilizing only 20% annotated data from the HASOC2019 dataset for training.
• Evaluation across multilingual, zero-shot cross-lingual, and monolingual training scenarios.

Main Results:

• The proposed SS-GAN-mBERT model demonstrated significant effectiveness in detecting hate speech and offensive language in Indo-European languages.
• The model achieved high performance even with limited annotated data (20%).
• The mBERT-based model (SS-GAN-mBERT) outperformed the XLM-RoBERTa-based model (SS-GAN-XLM), showing an average F1 score increase of 9.23% and accuracy increase of 5.75% over the baseline.

Conclusions:

• The developed semisupervised GAN model offers a robust solution for multilingual hate speech detection.
• The approach effectively mitigates the challenge of limited labeled data in NLP tasks.
• The findings highlight the potential of combining GANs and PLMs for advancing cross-lingual content moderation.