Semantic textual similarity for modern standard and dialectal Arabic using transfer learning

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Artificial Intelligence

Background:

• Semantic Textual Similarity (STS) is crucial for NLP tasks like information retrieval and machine translation.
• High-performance STS models exist for English due to abundant data, but Arabic lacks sufficient resources.
• This gap hinders the development of advanced NLP applications for Arabic speakers.

Purpose of the Study:

• To propose and evaluate three distinct approaches for generating effective Arabic STS models.
• To address the scarcity of Arabic STS training and evaluation data.
• To extend STS capabilities to Egyptian and Saudi Arabian Arabic dialects.

Main Methods:

• Approach 1: Fine-tuning using English STS data translated automatically to Arabic.
• Approach 2: Interleaving Arabic models with existing English data resources.
• Approach 3: Fine-tuning knowledge distillation-based models with a proposed translated dataset.

Main Results:

• Achieved 81% correlation with human judgment on the STS 2017 Arabic evaluation set using limited Arabic STS pairs.
• Successfully extended models to Egyptian Arabic (77.5% correlation) and Saudi Arabian Arabic (76% correlation).
• Demonstrated the effectiveness of proposed methods in low-resource STS model development.

Conclusions:

• The proposed methods effectively generate high-performance Arabic STS models despite data limitations.
• The developed models show significant potential for various NLP applications in Arabic and its dialects.
• This research contributes valuable resources and methodologies for low-resource NLP tasks.