A Transformer-Based Neural Machine Translation Model for Arabic Dialects That Utilizes Subword Units

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Translation

Background:

• Neural machine translation (NMT) struggles with free word order languages like Arabic dialects due to scarce and out-of-vocabulary words.
• Unknown Word (UNK) tokens are a limitation in fixed-size vocabulary NMT systems, particularly for morphologically rich languages.

Purpose of the Study:

• To develop the first Transformer-based NMT model for Arabic vernaculars utilizing subword units.
• To address the challenge of translating scarce and unknown words in Arabic dialects to Modern Standard Arabic (MSA).

Main Methods:

• The proposed model is based on the Transformer architecture, incorporating subword units and a shared vocabulary between Arabic dialects and MSA.
• Multi-head attention sublayers in the encoder are utilized to capture dependencies within Arabic vernacular input sentences.

Main Results:

• The model successfully addresses the issue of unknown words (out-of-vocabulary) in Arabic dialect translation.
• Experiments demonstrated a significant boost in translation quality for various Arabic vernaculars (Levantine, Maghrebi, Gulf, Nile, Iraqi) to MSA.

Conclusions:

• The proposed Transformer-based NMT model with subword units is effective for translating Arabic vernaculars to MSA.
• This approach enhances translation quality by better handling linguistic variations and unknown words inherent in Arabic dialects.