Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation

AI summary

Overview

Research area: Natural Language Processing, specifically machine translation, dialectal Arabic, dialogue context modeling, and parameter-efficient fine-tuning. Technical level: Intermediate. Scope: The paper presents Rosetta, a LoRA-adapted NileChat-3B system for context-aware English-to-Dialectal Arabic dialogue translation in the AlexandriaX-2026 shared task, comparing constrained and unconstrained tracks.

What This Paper Is About

Translating English dialogue into Arabic dialects is difficult because Arabic has many regional dialects and each conversational turn depends on prior dialogue context. This paper builds Rosetta to generate dialect-specific Arabic translations while conditioning on the target dialect and surrounding dialogue. It evaluates whether adding external dialect data improves performance across thirteen dialects.

Key Contributions

1. Developed the Rosetta system for Subtask 1 of AlexandriaX-2026, covering both constrained and unconstrained tracks for context-aware English-to-Dialectal Arabic dialogue translation.
2. Fine-tuned a LoRA adapter on NileChat-3B using structured system and user prompts that condition generation on both target dialect and dialogue context.
3. Created an unconstrained variant by additionally pretraining the adapter on MADAR and PADIC, two external Arabic dialect resources.
4. Provided an empirical analysis of external pretraining across thirteen dialects, showing limited benefit and evidence of negative transfer.

Main Findings

• Constrained track performance: Rosetta ranked 4th with an spBLEU score of 26.10.
• Unconstrained track performance: Rosetta ranked 5th with an spBLEU score of 25.09.
• Limited gains from external data: Additional pretraining on MADAR and PADIC helped only two of thirteen dialects.
• Negative transfer: External pretraining slightly hurt overall performance, suggesting that adding dialect data is not automatically beneficial.
• Effective lightweight adaptation: Prompt-conditioned LoRA fine-tuning on NileChat-3B produced competitive context-aware dialectal Arabic dialogue translation without full-model retraining.

Methodology in Plain English

The researchers started with NileChat-3B, a 3-billion-parameter chat model. Instead of retraining all model weights, they trained a small LoRA adapter, which updates only a compact set of extra parameters. Each example was formatted as system and user prompts that tell the model the target Arabic dialect and include the surrounding dialogue context. For the unconstrained track, the same adapter was additionally pretrained on MADAR and PADIC before evaluation. The final systems were evaluated using spBLEU and ranked within the shared task.

Why This Matters

Impact on research: The paper provides evidence that parameter-efficient adaptation can be effective for low-resource dialectal Arabic dialogue translation, while also cautioning that external dialect corpora can cause negative transfer. It contributes a reproducible prompt-conditioning setup for context-aware translation and highlights the importance of dialect-level evaluation rather than aggregate scores alone.

Real-world applications:

• Customer support chatbots that respond in a user’s specific Arabic dialect.
• Localization of chat, messaging, and social media platforms for MENA users.
• Subtitle, dubbing, and conversational media translation for Arabic-speaking audiences.
• Assistive agents in healthcare, government services, tourism, or e-commerce that need regional Arabic dialogue.

Industry relevance: The work is directly useful for companies serving Arabic-speaking markets, where dialect variation and conversational context affect user experience. LoRA reduces training and deployment costs compared with full fine-tuning, and the negative-transfer finding shows that data curation and dialect-aware evaluation matter more than simply adding more Arabic dialect data.

Future Directions

• Diagnose the negative transfer: identify whether dialect imbalance, domain mismatch, data noise, or forgetting causes external pretraining to hurt overall performance.
• Develop dialect-specific adapters, routing strategies, or data selection and curriculum methods to improve all thirteen dialects.
• Extend context modeling to longer multi-turn dialogues, speaker roles, code-switching, and informal conversational text.
• Add human evaluation and test larger base models, additional Arabic dialects, and broader low-resource translation settings.

Target Audience

NLP researchers and machine translation practitioners working on Arabic dialects, low-resource translation, dialogue systems, and parameter-efficient fine-tuning. The paper also benefits industry ML engineers building multilingual or MENA-focused conversational products, as well as shared-task participants interested in LoRA adaptation and context-aware dialectal Arabic translation.

Authors’ abstract

This paper describes the Rosetta system for Subtask 1 (Context-Aware English-to-Dialectal Arabic Dialogue Translation) of the AlexandriaX shared task, participating in both constrained and unconstrained tracks. The approach fine-tunes a LoRA adapter on NileChat-3B using structured system/user prompts that condition generation on dialect and dialogue context. For the unconstrained track, the adapter is additionally pretrained on MADAR and PADIC. Rosetta ranked 4th in the constrained track (spBLEU 26.10) and 5th in the unconstrained track (spBLEU 25.09). The experimental results demonstrate that external pretraining helps only two of thirteen dialects while slightly hurting overall performance, suggesting negative transfer.