Advancing arabic dialect detection with hybrid stacked transformer models.

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Natural Language Processing

Background:

• The proliferation of Arabic dialects online necessitates accurate classification for effective Natural Language Processing (NLP) applications.
• Deep learning (DL) models offer potential solutions for the challenges in identifying diverse Arabic dialects.

Purpose of the Study:

• To propose a novel stacking model for enhanced Arabic dialect classification.
• To improve the accuracy and efficacy of NLP applications dealing with dialectal Arabic.

Main Methods:

• A two-level stacking model combining two transformer models (Bert-Base-Arabertv02 and Dialectal-Arabic-XLM-R-Base) was developed.
• Base models generated class probabilities, which were used to train a meta-learner in the second level.
• The stacking model was compared against LSTM, GRU, CNN, and individual transformer models.

Main Results:

• The proposed stacking model significantly outperformed single-model approaches in classifying Arabic dialects.
• The model achieved high performance metrics, including 89.73% accuracy for Shami and 93.06% accuracy for IADD.
• The stacking approach effectively captured a broader range of linguistic features, leading to better generalization.

Conclusions:

• The novel stacking model provides a reliable solution for precise Arabic dialect recognition.
• This advancement enhances the efficacy of NLP applications by improving dialect classification accuracy.
• The model's ability to capture diverse linguistic variables is key to its superior performance.