Source-linked AI summary
Rosetta at AlexandriaX-2026: LoRA-Adapted NileChat for Context-Aware Dialectal Arabic Dialogue Translation
Nada Esmaeil, Fathima Rena, Sibi Subhash, Osama Elgendy, Mina Naguib, Salma Omar, Muhammad Arif
TL;DR
Dialectal Arabic dialogue translation is difficult because dialects differ substantially from formal MSA and from one another, with limited parallel resources. Rosetta uses LoRA adaptation of NileChat-3B with structured prompts containing dialect and dialogue context, adding MADAR and PADIC pretraining in the unconstrained track. It ranked 4th with spBLEU 26.10 in the constrained track and 5th with spBLEU 25.09 in the unconstrained track, while external pretraining improved only Libyan and Moroccan and lowered overall performance.
Problem
Arabic dialects differ from MSA and one another, while most lack standardized orthography or large parallel resources, limiting dialectal translation.
Method
Rosetta fine-tunes a LoRA adapter on NileChat-3B using structured prompts with dialect, metadata, and dialogue history, adding MADAR and PADIC pretraining unconstrained.
Results
Rosetta ranked 4th with spBLEU 26.10 in the constrained track and 5th with spBLEU 25.09 in the unconstrained track.
Takeaways & Limitations
External pretraining improved only Libyan and Moroccan and slightly degraded overall performance across the unconstrained track.
Takeaways & Limitations
The negative-transfer pattern is only partially understood, and the paper leaves controlled study of pretraining-data quality per dialect to future work.
Abstract
from arXiv · showhide
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.
1 Introduction
Arabic’s diglossia and dialect diversity make dialectal translation difficult, motivating AlexandriaX’s context-aware English-to-dialectal Arabic task. Rosetta addresses it with LoRA adaptation and participates under both constrained and unconstrained settings.
- Motivation: Arabic’s formal MSA variety differs from regional dialects that dominate everyday and online communication.Dialects also diverge lexically, morphologically, and pragmatically, while often lacking standardized orthography or large parallel resources.
- Task: AlexandriaX Subtask 1 translates English dialogue turns into a specified Arabic dialect using dialogue history and speaker metadata.Inputs include the target country/dialect, domain, gender, and persona information when available.
- Tracks: The study compares constrained training on Alexandria data with an unconstrained setting permitting external data and larger models.The constrained track limits models to 5B parameters and provided Alexandria data, whereas the unconstrained track allows external data and larger models.
- Approach: Rosetta fine-tunes a LoRA adapter on NileChat-3B with structured prompts encoding dialect and conversational context.The unconstrained configuration additionally pretrains the adapter on MADAR and PADIC.
- Results: 4th place and spBLEU 26.10 were achieved in the constrained track, while the unconstrained track placed 5th with spBLEU 25.09.External pretraining improved only Libyan and Moroccan and slightly reduced the overall average.
2 Background
AlexandriaX evaluates translation of individual English dialogue turns into specified Arabic dialects while preserving conversational context. The task builds on limited, domain-narrow dialectal resources and Arabic-focused language-model adaptation.
- Task setup: Subtask 1 supplies an English dialogue turn with dialect, country, domain, gender, history, and optional persona metadata, requiring translation of that turn only.Previous turns provide context rather than being retranslated.
- Dataset: The Alexandria dataset contains dialectal Arabic dialogues spanning multiple countries and dialects, with train, development, and test splits.Table 1 summarizes turn counts by dialect.
- Dialect resources: Arabic dialect resources remain limited, with PADIC covering a handful of dialects and MADAR providing 25 city-level dialects mainly in the travel domain.These resources generally lack dialogue-level context.
- Arabic language models: NileChat continues pretraining Qwen2.5-3B on synthetic and retrieval-based Arabic dialectal data for targeted dialect communities.It is reported to outperform comparable Arabic-aware models while targeting Egyptian and Moroccan Arabic rather than treating Arabic as monolithic.
3 System Overview
Rosetta freezes NileChat-3B and attaches a LoRA adapter, using structured prompts to condition translation on target-dialect metadata and dialogue history. The two tracks differ only in whether MADAR and PADIC pretraining precedes Alexandria fine-tuning.
- Architecture: Rosetta freezes NileChat-3B and attaches a LoRA adapter targeting attention and MLP projections.The adapter uses r = 32, α = 32, and dropout = 0.
- Prompting Strategy: The task prompt instructs translation-only generation and supplies target dialect, metadata, and preceding conversation history.At inference, the model’s previous outputs are fed back as history.
- Pretraining Prompt: External pretraining prompts specify the target dialect, country, and English sentence without dialogue history or detailed conversational metadata.This differs from the richer Alexandria task prompts.
- Tracks: The constrained track uses only Alexandria training data, whereas the unconstrained track first uses MADAR and PADIC before Alexandria fine-tuning.The external datasets are mapped to shared country-level dialect labels and converted into English–dialect pairs via NLLB-200.
- External data: Table 2 summarizes MADAR and PADIC training and development examples by dialect.These counts describe the external data used for the unconstrained pretraining stage.
4 Experimental Setup
Rosetta is trained and evaluated with official Alexandria data, optional MADAR/PADIC adapter pretraining, and deterministic beam-search decoding. The setup also constructs external English–dialect pairs and applies history noising during task fine-tuning.
- Data: The official Alexandria splits are used for task training and evaluation, with MADAR and PADIC added during unconstrained adapter pretraining.The number of external training and development examples is summarized by dialect.
- External-data construction: MADAR and PADIC English–dialect pairs are constructed by translating their MSA side to English with NLLB-200.MADAR city labels are mapped to country-level categories, and unaligned PADIC rows are filtered.
- Training: Training uses one epoch on a single Tesla T4 with maximum sequence length 1,024 and 4-bit NF4 quantization.The setup uses batch size 8, gradient accumulation 4, learning rate 2 × 10^-4, cosine scheduling, and AdamW 8-bit.
- Training: History noising truncates dialogue history with probability 0.15 and applies word-level drop, swap, or duplication with probability 0.25.The clean reference remains the training target, and this augmentation is used only during task fine-tuning.
- Decoding: Inference uses deterministic five-beam search with length penalty 0.7, disabled sampling, and early stopping for both tracks.The same decoding configuration is applied to constrained and unconstrained systems.
- Inference context: Previously generated translations are inserted into the prompt as conversation history for subsequent turns within the same dialogue.This preserves the dialogue-conditioned inference setup across turns.
5 Results
Rosetta performs unevenly across dialects and tracks: constrained decoding ranks higher overall, while additional MADAR/PADIC pretraining helps only selected dialects and can produce negative transfer. Qualitative analysis further identifies literal translation and suppressed Latin-script code-switching as recurring errors.
- Overall performance: 26.10 spBLEU / 41.79 chrF++ in the constrained track versus 25.09 / 41.02 in the unconstrained track, placing Rosetta 4th and 5th respectively.These are official test-set results.
- Dialect variation: 12.13 constrained and 10.83 unconstrained spBLEU make Mauritanian the clear low-end outlier, roughly one-third of the best dialect scores.Syrian, Jordanian, Palestinian, Saudi, Lebanese, and Egyptian exceed 29 spBLEU in the constrained track.
- Constrained vs. unconstrained: 25.09 versus 26.10 average spBLEU shows that unconstrained pretraining underperforms constrained training on 11 of 13 dialects.Libyan and Moroccan are the only exceptions, improving by +0.32 and +0.42.
- Constrained vs. unconstrained: Sudanese declines by -1.10 despite receiving 1,600 MADAR examples, so data availability alone does not explain the track divergence.The authors treat the pattern as partially understood negative transfer and suggest uneven back-translation quality as one possible contributor.
- Qualitative error analysis: Manual analysis finds overly literal translations, with the model translating every source word even when Egyptian Arabic naturally condenses routine exchanges.The analysis inspected the 20 lowest-scoring Egyptian dev outputs.
- Qualitative error analysis: 319 Moroccan, 148 Lebanese, 123 Tunisian, and 70 Egyptian examples omit Latin-script tokens present in the reference, indicating widespread code-switching under-generation.The reverse mismatch is comparatively rare except for Tunisian, with 212 cases.
6 Conclusion
Rosetta applies structured prompting and history noising to a LoRA-adapted NileChat system for context-aware English-to-dialectal Arabic translation. It achieves competitive official rankings, while external dialectal pretraining improves only Libyan and Moroccan and slightly harms overall performance.
- Rosetta covers thirteen regional varieties with a LoRA-adapted NileChat system for context-aware English-to-dialectal Arabic dialogue translation.
- Structured prompting incorporates dialogue history and speaker metadata, while history noising supports performance under tight parameter and data constraints.
- 4th place (spBLEU 26.10) was achieved by the constrained system, while the unconstrained configuration placed 5th (spBLEU 25.09).
- External sentence-level dialectal pretraining produced no overall downstream gains and induced mild negative transfer across most dialects.
- Improvements from external pretraining were restricted to Libyan and Moroccan, with no single explanatory factor fully accounting for the pattern.
A Prompt Templates
The appendix presents two prompt templates: sentence-level prompts for MADAR and PADIC pretraining, and multi-turn prompts for Alexandria translation that encode dialect, conversational context, and participant metadata.
- A Prompt Templates: The appendix distinguishes a single-turn MADAR/PADIC pretraining template from a full multi-turn Alexandria conversational template.
- A.1 Pretraining Prompt Template (MADAR / PADIC): The pretraining prompt contains sentence-level translation instructions and target-country metadata.
- A.1 Pretraining Prompt Template (MADAR / PADIC): Its system prompt requires translation-only output without code, explanations, comments, or other extra text.
- A.1 Pretraining Prompt Template (MADAR / PADIC): The user prompt specifies Egyptian Arabic as the target dialect and supplies the English sentence to translate.
- A.2 Alexandria Conversational Prompt Template: The Alexandria prompt includes conversational history, participants, domain, and gender direction.
- A.2 Alexandria Conversational Prompt Template: The conversational system prompt preserves meaning and tone, respects gender direction, and considers country, domain, participants, and speaker.
- A.2 Alexandria Conversational Prompt Template: The Alexandria user prompt identifies Egyptian Cairene as the target dialect and provides country, agriculture domain, participants, speaker, and female-to-male direction.
- A.2 Alexandria Conversational Prompt Template: The example supplies a wholesale-buyer history and a wholesale-seller turn about artichokes, followed by its dialectal translation.