PaperlayerView exact revision on arXiv

Source-linked AI summary

A Parallel Multi-Dialect Arabic Dependency Treebank

Noor Abo Mokh, Mohamed Adel, Jamila El-Gizuli, Go Inoue, Ahmed Elshabrawy, Nizar Habash

arXiv:2610.28823v1cs.CL

TL;DR

Arabic dependency treebanks have mainly focused on MSA, leaving dialectal coverage limited. MADARTB adds parallel dependency annotations for five city dialects alongside MSA and evaluates syntactic patterns and parsing across the varieties. The analysis finds shared global construction profiles alongside local divergence, and supervised parsing remains strongest for dialectal Arabic.

  • Problem

    Arabic dependency parsing resources have primarily focused on MSA, while dialectal treebank coverage remains uneven and lacks large-scale, consistently annotated parallel dependency trees.

    Source: §1 · p. 1

  • Method

    The paper constructs a parallel CATiB treebank for five city dialects and MSA, then analyzes syntactic variation and benchmarks supervised and few-shot LLM parsing.

    Source: §1 · p. 1

  • Results

    Global construction profiles show high similarity but local syntactic divergence, and supervised parsers remain strongest for dialectal Arabic.

    Source: §1 · p. 1, §8 · p. 8

  • Takeaways & Limitations

    The parallel resource supports controlled comparison of syntactic similarity and divergence and study of cross-dialect transfer in syntactic modeling.

    Source: §1 · p. 1

    The treebank covers only five city varieties and 1,000 translated travel-domain sentences per variety, so it may not capture spontaneous speech, online writing, code-switching, other domains, or broader variation.

    Source: §9 · p. 9

Abstract

from arXiv · show

Dependency parsing resource development for Arabic has largely focused on Modern Standard Arabic, leaving dialectal Arabic comparatively under-resourced. We introduce MADARTB, the first large-scale parallel dependency treebank spanning five city dialects: Beirut, Cairo, Doha, Rabat and Tunis, and extending an existing parallel MSA version. In total MADARTB consists of 5,000 sentences annotated using the CATiB dependency scheme. We analyze syntactic similarity across all city dialects and MSA. Although global construction profiles exhibit high similarity across varieties, local syntactic structures frequently diverge, revealing substantial structural variation across dialects. We further establish a cross-dialect dependency parsing comparison using a supervised parser and four LLMs. The supervised parser achieves 87.5% LAS, while the best LLM setup trails at 85.7% LAS across all six varieties on the test set. We release our annotations and models publicly.

1 Introduction

Arabic dependency parsing resources have concentrated on MSA and Classical Arabic, while dialectal coverage is uneven and no large-scale resource provides consistently annotated parallel dependency trees across Arabic dialects. MADARTB brings five city dialects together with parallel MSA in a unified representation, supporting controlled syntactic comparison and cross-dialect parsing research.

Motivation

  • Arabic treebank coverage has focused mainly on MSA and Classical Arabic, while dialectal resources remain uneven and mostly Egyptian.No large-scale resource previously provided full, consistently annotated dependency trees for parallel sentences across Arabic dialects.Source: §1 · p. 1

Contributions

  • Parallel sentences in a unified dependency representation enable controlled comparison of syntactic similarity and divergence across MSA and five dialects.The same resource also supports study of cross-dialect transfer in syntactic modeling.Source: §1 · p. 1
  • MADARTB adds 5,000 manually annotated parallel sentences across five city dialects and extends an existing MSA version for 6,000 sentences total.The five dialects are Beirut, Cairo, Doha, Rabat, and Tunis.Source: §1 · p. 1
  • The paper benchmarks supervised and few-shot LLM parsing, finding pooled supervised training best overall and establishing new state-of-the-art results on several parsing datasets.Figure 1 illustrates CATiB-style parallel dependency annotations across MSA and the five city dialects.Source: §1 · p. 1, Figure 1 · p. 2

2 Linguistic Background

Arabic varieties share substantial linguistic structure, alongside systematic lexical, morphological, and syntactic differences. The examples illustrate how local forms and constructions vary even when the broad construction inventory is shared.

Overview

  • Arabic varieties share a linguistic core, but systematic differences occur at lexical, morphological, and syntactic levels.MSA is used in formal communication, while regional dialects dominate daily communication and increasingly appear in informal and online writing.Source: §2 · p. 2

Lexical variation

  • Lexical similarity averages 25.8% pairwise across the full dataset, with varieties differing in words such as ‘bus’ and ‘city’.Beirut and Doha share a word for ‘bus’, while Tunis, Rabat, and Cairo use different forms.Source: §2 · p. 2

Morphological variation

  • Verbal morphology overlaps across varieties, but affix realizations and distributions can differ, as with the first-person prefix n- marking singular or plural.The n- example marks first-person singular in Tunis and Rabat, but first-person plural in MSA, Beirut, Cairo, and Doha.Source: §2 · p. 2
  • Pre-verbal particles mark mood and tense, with different forms used across varieties for habitual present and future marking.Beirut uses b+ for habitual present, while Rabat uses k+; Doha uses b+ for future marking.Source: §2 · p. 2

Syntactic variation

  • The broad inventory of syntactic constructions is highly shared, including idafa possession, subject–verb–object configurations, and noun–adjective order.All six example trees use an idafa relation for ‘center of city’ despite lexical differences.Source: §2 · pp. 2–3
  • Destinations take different syntactic forms: MSA and some dialects use prepositions, Doha uses a direct nominal argument, and Cairo uses a different argument structure.Beirut uses ‘at’, Rabat and Tunis use ‘to/for’, and Cairo uses a verb meaning ‘take’ with an object clitic.Source: §2 · p. 3

3 Related Work

Arabic syntactic treebanks have chiefly covered MSA and Classical Arabic, with dialectal resources often limited to individual varieties or narrow settings. MADARTB differs through its scale and consistent parallel coverage across five dialects.

Arabic dependency parsing

  • Earlier Arabic treebanks and parsing work have primarily addressed MSA and Classical Arabic, with dialectal parsing receiving comparatively less attention.Dialectal approaches have included MSA-resource adaptation and multilingual transfer.Source: §3 · p. 3

Arabic Treebanks

  • Most dialect treebanks cover individual varieties, while multi-dialect efforts are fewer and often smaller or focused on specific constructions.Abo Mokh et al. (2024) annotated 200 parallel sentences in four dialect groups using CATiB-style dependencies; their data is not publicly available.Source: §3 · p. 3
  • MADARTB provides uniform CATiB dependency annotations for 5,000 parallel sentences across five dialects, expanding the scale and coverage of earlier work.Source: §3 · p. 3

Arabic dependency parsing

  • This work benchmarks supervised and few-shot LLM parsing and reports new state-of-the-art results on several parsing datasets.Source: §3 · p. 3

4 The MADARTB Treebank

MADARTB aligns five city dialects with MSA, uses CATiB annotations, and combines automatic annotation with expert correction and validation. The resulting corpus contains parallel sentence sets split consistently across training, development, and test data.

4.1 Source Data

  • The corpus draws 1,000 aligned sentences per city dialect from MADAR-CODA and pairs them with corresponding MSA sentences from CamelTB.This parallel structure supports comparison across the five dialects and MSA.Source: §4.1 · p. 3

4.2 Annotation Scheme

  • Annotations use CATiB, an Arabic-specific functional-head dependency representation that captures language-specific structures, unlike UD’s shared multilingual scheme.CATiB is similar to SUD in being a functional-head representation.Source: §4.2 · pp. 3–4

5 Dialect Syntax Similarity

MADARTB enables controlled syntactic comparison across MSA and five dialects, showing highly similar global construction profiles alongside substantial local differences. Alignment-level agreement and construction analyses identify where those differences appear.

Alignment Pipeline and Manual Evaluation

5.1 Global Construction Profiles

  • 93.3–98.3% pairwise cosine similarity across global construction profiles indicates a shared constructional core.Rabat–Tunis is the closest pair, followed by Doha–Tunis and Cairo–Beirut, suggesting similarity is not determined by geography alone.Source: §5.1 · p. 5, Table 2 · p. 6

5.3 Syntactic Structure Analysis

  • MSA strongly favors verb–subject order, whereas all five dialects favor subject–verb order; verbless constructions occur 63–91 times per dialect versus 26 in MSA.The Rabat example has a noun-headed predicate.Source: §5.3 · p. 6
  • Analytical Idafa is realized differently across dialects, with relational nouns more frequent in Doha, Rabat, and Tunis, while Cairo and Beirut also use preposition + l+ constructions.These alternatives result in different dependency structures.Source: §5.3 · p. 6
  • Verbs are the most frequent ROOT dependents in MSA, Rabat, and Tunis; nominals predominate in Cairo and Beirut, while Doha is nearly balanced.Nominally annotated “desire” expressions may partly account for the higher nominal ROOT frequency in Cairo and Beirut.Source: §5.3 · p. 6

Alignment Pipeline and Manual Evaluation

6 Parsing Benchmark

The benchmark evaluates supervised parsers and prompted LLMs across individual varieties and pooled data, using joint parallel splits to prevent translated evaluation sentences from appearing in training.

Data

  • Sentence IDs are split jointly across varieties, preventing translations of evaluation sentences from appearing in training.Evaluation covers each variety, the five dialects together, and all six varieties together, on development and test sets.Source: §6 · p. 6

Metric

  • LAS measures the percentage of tokens assigned both the correct head and dependency relation; development LAS selects models, while test results provide the final comparison.System differences are assessed with paired bootstrap 95% confidence intervals resampled by sentence ID.Source: §6 · p. 6

Supervised Parsers

  • Supervised parsers use a CAMeLBERT-Mix encoder and biaffine parser, with projective multi-root decoding.The eight MADARTB training configurations cover individual varieties, pooled dialects, and all six varieties; external-treebank combinations are also evaluated.Source: §6 · pp. 6–7

LLM Parsers

  • Four LLMs are evaluated with zero-shot prompting and retrieval, with prompts requesting one CATiB head–relation prediction per token.Retrieval adds the ten MADARTB training trees most lexically similar to each evaluation sentence by chrF++ score.Source: §6 · p. 7

8 Conclusions and Future Work

The paper introduces MADARTB, a parallel multi-dialect Arabic dependency treebank, and reports syntactic comparisons and parsing benchmarks across five cities. Future work will extend its variety coverage and annotation uses.

8 Conclusions and Future Work

  • MADARTB is introduced as the first parallel multi-dialect Arabic dependency treebank.The paper reports quantitative syntactic comparisons across five cities and a dialect-parsing benchmark using a supervised parser and four LLMs; supervised parsers remain strongest for dialectal Arabic.Source: §8 · p. 8
  • Future work will extend MADARTB to additional MADAR varieties through parallel annotation projection and controlled transfer.Source: §8 · p. 8
  • Planned extensions include multidialect tokenization for end-to-end parsing and automatic mapping to UD/SUD for broader cross-linguistic comparison.Source: §8 · p. 8

Related papers

Loading 2610.28823v1…