A transformer fine-tuning strategy for text dialect identification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Healthcare Informatics

Background:

• Online medical consultations enhance primary care efficiency.
• Current systems connect patients to consultants based on questions.
• Linguistic variations in patient queries necessitate improved referral systems.

Purpose of the Study:

• To propose a novel fine-tuning strategy for pre-trained transformers.
• To identify the social origin of text authors for better patient-doctor matching.
• To improve the efficiency of online medical consultation referral systems.

Main Methods:

• Developed a novel fine-tuning strategy for pre-trained transformer models.
• Integrated the proposed strategy with an existing adapter model.
• Evaluated performance on the Nuanced Arabic Dialect Identification (NADI) dataset.

Main Results:

• Achieved an overall accuracy of 53.96% for Arabic dialect identification.
• Exceeded the previous best accuracy by 0.54% on the NADI dataset.
• Demonstrated the utility of custom fine-tuning for transformer models.

Conclusions:

• Custom fine-tuning strategies are effective for pre-trained transformer models.
• Social origin identification can enhance online medical consultation referral systems.
• The proposed method shows promise for improving cross-cultural communication in healthcare.