Context-Aware Dialectal Arabic Machine Translation with Interactive Region and Register Selection

1. Problem Statement

The paper addresses a critical limitation in current Arabic Machine Translation (MT) systems: "Dialect Erasure."

• Diglossia Challenge: Arabic is characterized by diglossia, where Modern Standard Arabic (MSA) is used for formal contexts, while diverse regional dialects (e.g., Egyptian, Levantine, Gulf, Maghrebi) are used for daily communication.
• Current Limitations: Existing MT systems (including large models like NLLB) typically normalize dialectal inputs into MSA or produce a single "average" output. They lack mechanisms for users to explicitly control the target dialect or social register (formal vs. informal).
• Evaluation Gap: Standard automatic metrics (BLEU, METEOR) are biased toward MSA. Consequently, models that generate authentic dialectal variations often receive lower scores than models that default to MSA, creating an "Accuracy Paradox" where high metric scores correlate with low dialectal fidelity.

2. Methodology

The authors propose a context-aware, steerable framework that treats dialect and register as explicit conditioning variables rather than latent noise.

A. Data Engineering: Rule-Based Data Augmentation (RBDA)

To overcome the scarcity of parallel dialectal data, the authors developed a novel pipeline:

• Seed Corpus: Started with a small, high-quality seed of ~3,000 dialectal sentences extracted from noisy web-scale data (OPUS) and the Tatoeba Project.
• RBDA Pipeline: Applied a rule-based augmentation strategy to expand the seed into a balanced 57,600-sentence parallel dataset.
  • Lexical Injection: Used manually verified synonym maps to transform MSA samples into specific dialectal variants (e.g., converting MSA 'urīdu to Egyptian 'āyez or Levantine biddī).
  • Balancing: Upsampled minority dialect classes (e.g., Moroccan, Algerian) to match the MSA baseline, ensuring uniform representation across eight regional varieties.
  • Schema: Structured data with metadata tags: {input, target, region, context, style}.
• Domains: Covered five contexts: Medical, Travel, Education, Restaurant, and General Life.

B. Model Architecture & Training

• Base Model: Utilized mT5-base (Multilingual T5), a 580M parameter encoder-decoder model.
• Control Mechanism: Introduced a Multi-Tag Prompt Structure ([Region] [Context] Input).
  • Example: [EGYPTIAN][MEDICAL] I want to see a doctor
  • This allows the model to condition generation on specific geo-lexical choices and social registers (Formal/Informal) without requiring separate models for each dialect.
• Training: Fine-tuned on a single NVIDIA GPU using the Hugging Face transformers library with FP16 precision, AdaFactor optimizer, and early stopping.

C. Interactive Interface

Developed a Gradio-based web interface allowing real-time selection of dialect region, formality level, and context, demonstrating the system's steerability.

3. Key Contributions

1. RBDA Framework: A novel data augmentation pipeline that transforms a small seed corpus into a large, balanced, multi-dialect parallel dataset (57.6k sentences) covering eight Arabic varieties.
2. Steerable Generation: A unified multilingual model capable of explicit, user-controlled generation of specific dialects and registers via metadata tagging, avoiding the need for multiple specialized models.
3. The "Accuracy Paradox" Identification: Empirical evidence showing that standard MT metrics (BLEU) penalize dialectal specificity. Models defaulting to MSA achieve higher scores, while models generating authentic dialects score lower despite higher cultural fidelity.
4. New Resource: Release of a publicly available, balanced Arabic MT dataset and a fine-tuned model on Hugging Face, designed specifically for controllable dialectal translation.

4. Results

Quantitative Analysis

• Automatic Metrics: The proposed model achieved lower aggregate scores compared to baselines (NLLB-200 and Helsinki-NLP):
  • BLEU: Proposed (8.19) vs. NLLB (13.75).
  • Interpretation: The lower scores are attributed to the model generating region-specific lexical forms that diverge from MSA-heavy reference translations, not to translation errors.

Qualitative & Authenticity Analysis

• LLM-Assisted Evaluation: An LLM (GPT-5) evaluated dialectal alignment on a 1–5 scale.
  • Proposed Model: Achieved an average score of 4.80/5, successfully generating region-appropriate markers (e.g., l-makla bnina for Moroccan).
  • Baselines: Scored 1.0/5, consistently defaulting to MSA forms (e.g., al-ta'ām ladhīdh) regardless of the requested dialect.
• Performance Distribution: While baselines showed a "spike-and-drop" pattern (high performance on MSA, near-zero on low-resource dialects), the proposed model maintained a uniform performance profile across all eight dialects.

5. Significance and Conclusion

• Paradigm Shift: The paper argues for shifting Arabic MT research from optimizing for dialect-agnostic aggregate metrics to prioritizing sociolinguistic appropriateness and user intent.
• Evaluation Reform: It highlights the urgent need for new evaluation frameworks that account for linguistic diversity, as current metrics fail to capture the value of dialectal authenticity.
• Practical Impact: By enabling "steerable" translation, the system empowers users to receive culturally authentic outputs tailored to specific regions (e.g., Gulf vs. Maghrebi) and contexts (e.g., Medical vs. Casual), addressing the real-world needs of Arabic speakers that current systems ignore.

Limitations: The study acknowledges that the RBDA pipeline focuses primarily on lexical substitution rather than deep morphological/syntactic changes, and the reliance on synthetic data may introduce "translationese." Future work aims to incorporate human evaluation and deeper syntactic modeling.