Pashto poetry generation: deep learning with pre-trained transformers for low-resource languages

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Artificial Intelligence

Background:

• Poetry generation using machine learning is a complex research area.
• Pashto poetry generation presents unique linguistic challenges.
• Transformer models offer advanced capabilities for text generation.

Purpose of the Study:

• To develop and evaluate a novel approach for generating high-quality Pashto poetry.
• To leverage pre-trained transformer models for Pashto language tasks.
• To assess the quality of machine-generated Pashto poetry against human standards.

Main Methods:

• Utilized two pre-trained transformer models: LaMini-Cerebras-590M and bloomz-560m.
• Trained models on a new, extensive dataset of Pashto poetry.
• Employed seed text prompting for poetry generation.
• Conducted subjective and objective evaluations, including human assessments.

Main Results:

• The proposed approach successfully generated Pashto poetry.
• Experimental results indicated that generated poetry quality is comparable to human-generated poetry.
• Demonstrated the effectiveness of transformer models for Pashto poetry generation.

Conclusions:

• The study presents a significant contribution to Pashto language and poetry generation.
• The developed method shows potential for applications in natural language processing and computational linguistics.
• Machine learning models can produce Pashto poetry of remarkable quality.