Multilingual hope speech detection from tweets using transfer learning models.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Social Media Analysis

Background:

• Social media significantly influences public discourse and community emotional states.
• Hateful speech is prevalent, but so is 'hope speech'—supportive and encouraging online content.
• Automatic detection of hope speech has been explored in several languages, but not Urdu and English using translation-based methods.

Purpose of the Study:

• To address the gap in Urdu and English hope speech detection.
• To create a novel multilingual dataset for English and Urdu hope speech.
• To benchmark the dataset using state-of-the-art machine learning, deep learning, and transfer learning models.

Main Methods:

• Development of a multilingual dataset in English and Urdu.
• Application of a translation-based approach to manage multilingual challenges.
• Utilized state-of-the-art machine learning, deep learning, and transfer learning methods, including the Bert transformer model, for benchmarking.

Main Results:

• The Bert transformer model achieved benchmark performance, with 87% accuracy for English and 79% for Urdu.
• This represents an improvement of 8.75% in English and 1.87% in Urdu over baseline Support Vector Machine (SVM) models.
• Rigorous annotator selection and detailed guidelines significantly enhanced dataset quality.

Conclusions:

• The proposed translation-based methodology using the Bert transformer model is effective for multilingual hope speech detection.
• The created English-Urdu dataset and benchmarking provide a foundation for future research in this area.
• High-quality annotation processes are crucial for developing robust datasets for social media content analysis.