PaperlayerView exact revision on arXiv

Source-linked AI summary

alkhder at AlexandriaX-2026: Context-Aware English-to-Dialectal Arabic Dialogue Translation with QLoRA-Fine-Tuned NileChat-3B

Hasan Alkhder, Maher Pasa, Maria Hanini, Amro Najjar, Ahmet Zengin

arXiv:2610.22525v1cs.CL

TL;DR

Dialectal Arabic translation is challenged by differences among varieties and limited, domain-specific data, alongside spelling variation and the need for dialogue and speaker context. The paper fine-tunes NileChat-3B-Base with QLoRA using shared adapters and metadata-rich prompts, then reports sixth-place results in both tracks. The system is presented as a practical constrained-track approach, though self-generated dialogue history did not improve its reported results.

  • Problem

    Dialectal Arabic MT faces linguistic differences among varieties, scarce domain-specific parallel data, non-standardized spelling, and the need to use dialogue context and speaker information.

    Source: §1 · p. 1

  • Method

    The system fine-tunes NileChat-3B-Base with QLoRA on turn-level Alexandria prompts, using one adapter across dialects and conditioning on target-dialect and dialogue metadata.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    20.36 country-macro-averaged spBLEU and 36.64 chrF++ on both tracks, ranking sixth in each.

    Source: §5.1 · p. 4, §6 · p. 6

  • Takeaways & Limitations

    The compact dialect-focused system is presented as a practical constrained-track approach, training in about one GPU-hour and covering thirteen dialects.

    Source: §6 · p. 6

    The submitted self-generated history condition performs worse than gold or empty history, and its previous-translation errors can outweigh context benefits.

    Source: §5.2 · p. 5

Abstract

from arXiv · show

This paper describes team Alkhder's submission to Subtask 1 of the AlexandriaX-2026 shared task, which asks systems to translate English dialogue turns into thirteen Arabic dialects. We fine-tune NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, training one set of adapter weights for all dialects. Each turn is translated conditioned on the target country and dialect, the domain, the participant roles, the speaker-addressee gender direction, and the preceding dialogue context. At inference, conversations are translated autoregressively and the model is fed its own previous outputs back as context without using any gold target-side test information. Under the constrained track, using only Alexandria data, it trains in about one hour and translates the private test set in under two hours. The system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64 on both tracks.

1 Introduction

Dialectal Arabic translation is difficult because varieties differ from MSA and one another, while data scarcity, spelling variation, and the need for dialogue context and speaker information complicate progress. Alkhder’s system uses one QLoRA adapter across dialects and reports its efficiency and shared-task results.

1 Introduction

  • Dialectal Arabic MT remains difficult because varieties differ in linguistic and sociolinguistic features, parallel data are scarce and domain-specific, spelling is non-standardized, and translation needs dialogue context and speaker information.The shared task evaluates meaning preservation and dialect-specific adaptation.Source: §1 · p. 1
  • The system fine-tunes NileChat-3B-Base with QLoRA on turn-level Alexandria prompts containing target-dialect and dialogue metadata, using one adapter-weight set for all dialects.It is submitted to both tracks.Source: §1 · p. 1
  • The system achieves country-macro-averaged scores of 20.36 spBLEU and 36.64 chrF++ on both tracks, while ranking sixth in each.It trains in about one GPU-hour and covers all dialects.Source: §1 · p. 1

2 Background

AlexandriaX-2026 Subtask 1 evaluates English-to-dialectal Arabic dialogue translation across thirteen varieties using conversational and speaker metadata. The paper situates the task amid data scarcity, orthographic variation, and prior work on Arabic MT and context-aware translation.

2.1 Task setup

  • Subtask 1 translates selected English turns into a specified dialect using dialogue context, country and dialect labels, domain, personas, speaker–addressee gender direction, and turn position.The constrained track permits only Alexandria data and models up to 5B parameters; the unconstrained track permits external data and models of any size.Source: §2.1 · p. 2

2.2 Data

  • Training and development data cover eleven countries, with Libya and Sudan evaluated zero-shot; the training pool is highly imbalanced, from 14,933 Palestinian turns to 2,034 Tunisian turns.The private test set contains 14,459 turns across 4,673 conversations, distributed near-uniformly over thirteen countries.Source: §2.2 · p. 2

2.3 Related Work

  • Prior Arabic MT expanded corpus coverage but remained constrained by domain and data scarcity, while pretrained models and instruction-tuned LLMs show uneven performance across dialects.Work on orthographic normalization and context-aware MT addresses spelling variation, coherence, pronoun and ambiguity resolution, and gender agreement; AlexandriaX-2026 supplies dialogue history and speaker–addressee gender metadata.Source: §2.3 · p. 2
  • LoRA and QLoRA are presented as parameter-efficient methods that make pretrained-model adaptation feasible with limited computational resources.Source: §2.3 · p. 2

3 System Overview

The system formats each turn as a metadata-rich prompt, fine-tunes one dialect-focused model jointly across training-data dialects, and translates conversations sequentially using its own prior outputs as context.

3.1 Prompt construction

  • Each prompt instructs translation of the final English turn and supplies dialect, country, domain, participant and gender metadata, up to four prior turns, and the current English query.Previous turns contain English sources and, when available, Arabic translations; training uses gold Arabic context, while inference uses the model’s own outputs.Source: §3.1 · pp. 2–3

3.2 Base model and fine-tuning

  • NileChat-3B-Base is fine-tuned with QLoRA using one adapter jointly trained on eleven data-supported dialects, supporting transfer and zero-shot evaluation for Libyan and Sudanese.At inference, natural-language dialect labels and pretraining exposure to related varieties are used for the two dialects without fine-tuning data.Source: §3.2 · p. 3

3.3 Context-aware inference

  • At test time, turns are translated in dialogue order, and each generated Arabic turn becomes context for the next without target-side test information.Errors may remain visible for up to four subsequent turns; decoding is greedy, and prompts from 16 conversations are batched at each turn position.Source: §3.3 · p. 3

4 Experimental Setup

The experimental setup creates turn-level supervised examples, limits per-country training data, and evaluates private-test predictions with country-macro-averaged spBLEU as the primary metric.

4.1 Data preparation

  • The prepared corpus contains 78,730 turn-level examples, then caps each of the eleven data-supported varieties at 2,500 randomly sampled examples to equalize country counts at 27,500 total.Each example includes metadata, preceding turns with English sources and gold Arabic translations, and the current turn’s gold translation.Source: §4.1 · p. 3

4.2 Preprocessing and training

  • Prompts are limited to four preceding turns and 896 tokens, with older context removed first; training runs for one epoch on an A100 and takes about one hour.No orthographic normalization, dediacritization, or transliteration is applied; private-test decoding takes approximately 1.5 hours.Source: §4.2 · pp. 3–4

4.3 Submission and evaluation

  • Predictions for every private-test turn are validated for complete, nonempty, nonduplicate coverage, with country-macro-averaged spBLEU as the primary metric and chrF++ as secondary.The system submits in both tracks using the official JSON format.Source: §4.3 · p. 4

5 Results

The system ranks sixth in both tracks, with country-macro-averaged scores of 20.36 spBLEU and 36.64 chrF++. Post-submission analyses find that self-generated history underperforms empty history, evaluation scores are sensitive to matched orthographic normalization, and several dialect-specific behaviors remain.

5.1 Official results

  • 20.36 spBLEU and 36.64 chrF++ place the system sixth in both tracks, with the same predictions evaluated in each track.The system retains 80.7% of the leading systems’ chrF++ score and 66.9% of their spBLEU score.Source: Table 2 · p. 4, §5.1 · p. 4

5.2 Post-submission analysis

  • Empty history matches or outperforms self-generated history in all thirteen varieties; gold history scores 21.62 macro spBLEU and 37.66 chrF++, versus 20.56 spBLEU and 36.73 chrF++ for the submitted condition.The comparison used 200 conversations per country; errors in previous translations can outweigh the benefits of context.Source: §5.2 · p. 5, Table 4 · p. 5, Table 5 · p. 5
  • 20.80 to 22.36 macro spBLEU and 37.01 to 38.28 chrF++ follow normalization of both system outputs and references, while normalizing outputs alone lowers spBLEU to 15.49.The authors conclude that normalization is useful for analysis only when outputs and references are normalized alike.Source: §5.2 · p. 5, Table 6 · p. 5
  • The apparent MSA signal largely reflects reference texts rather than systematic model fallback, with only small positive function-word-rate residuals in three dialects.The reported residuals are +0.34 for Sudanese, +0.35 for Tunisian, and +0.18 for Mauritanian.Source: §5.2 · p. 5, Table 7 · p. 5
  • In 60 manually annotated private-test turns, content errors were most common (19), while six understandable turns sounded non-native in register and were missed by the LLM annotator.The analysis also found 11 dialect–MSA confusion cases, eight morphological agreement errors, and two orthographic errors; the authors note the value of human evaluation for dialectal naturalness.Source: §5.2 · p. 5

6 Conclusion

The conclusion presents a compact QLoRA-fine-tuned 3B model for context-aware dialectal Arabic dialogue translation, while the prompt template specifies the metadata and dialogue context supplied for each turn.

6 Conclusion

  • The system achieves 20.36 spBLEU and 36.64 chrF++ on both tracks while covering thirteen dialects and training in about one GPU-hour.At inference, it feeds its own previous outputs back as context without gold leakage.Source: §6 · p. 6

A Prompt Template

  • During training, prior-turn Arabic fields contain gold translations; at inference, they contain the model’s own previous outputs.The prompt template marks the target English turn separately for translation.Source: A Prompt Template · pp. 7–8
  • The prompt instructs the model to translate only the final English turn, preserve meaning, tone, register, persona, and speaker gender agreement, and output only the translation.Generation ends at the end-of-sequence token or after 200 new tokens.Source: A Prompt Template · pp. 7–8
  • Prompts provide country, dialect, domain, participant roles, speaker role, and gender direction alongside the conversation history.The history includes the last four turns, with English turns and their Arabic translations.Source: A Prompt Template · pp. 7–8

B Full Configuration

Table 8 is identified as presenting the system configuration.

B Full Configuration

C Orthographic Normalization Rules

The appendix describes orthographic normalization mappings applied to hypotheses and references before rescoring.

C Orthographic Normalization Rules

D Per-Country Results

Table 9 is identified as reporting per-country results on the public test split, with normalization gains and zero-shot varieties noted.

D Per-Country Results

  • Table 9 reports per-country results on the public test split; “Norm.” denotes spBLEU gain from orthographic normalization, and an asterisk marks zero-shot varieties.Source: Table 9 · p. 9

E Annotation Protocol

The team manually reconciled native-speaker and LLM labels for a stratified sample of 60 private-test predictions, using inputs without reference translations.

E Annotation Protocol

  • 60 private-test predictions were sampled, approximately five per country, and shown with their English source, preceding context, country and dialect, and speaker–addressee gender direction, but no reference translation.Private-test references were not released.Source: E Annotation Protocol · p. 9
  • A native-speaker annotator and an LLM independently assigned labels using the same category definitions, and the human annotator reconciled their labels.Source: E Annotation Protocol · p. 9
  • The labels covered five error types adapted from the Subtask 3 LQM-inspired typology and a sixth human-added category for unnatural but content-acceptable turns.The added category captured recoverable content in turns whose register sounded non-native; Table 10 reports annotation counts over the 60 sampled turns.Source: E Annotation Protocol · p. 9, Table 10 · p. 9
Loading 2610.22525v1…