UD-Quran: A Universal Dependencies Conversion of the Extended Quranic Treebank for Interoperable Quranic/Classical Arabic Parsing

Technical Summary: UD-Quran

Problem Statement
Quranic Arabic possesses a stable, canonical text, yet its computational encoding lacks uniformity. Divergent practices in orthography, diacritization, and, most critically, clitic segmentation create incompatible token inventories that hinder cross-resource comparison and reuse in modern Natural Language Processing (NLP). While resources like the Quranic Arabic Corpus and the Extended Quranic Treebank (EQTB) offer high-quality, tradition-aligned annotations, their specific segmentation conventions and syntactic formalisms often diverge from the standards required by contemporary NLP ecosystems, particularly the Universal Dependencies (UD) framework. This lack of interoperability limits the ability to train standard parsers, utilize shared evaluation scripts, and integrate Quranic data into broader multilingual toolchains.

Methodology
The paper presents UD-Quran, a Universal Dependencies (UD) v2 conversion of the EQTB dataset. The conversion process is designed to balance strict adherence to UD conventions with the preservation of Quran-specific analytic distinctions.

• Dual-Variant Strategy: The conversion produces two distinct releases from the same source file:
  1. Surface Variant: Excludes all analytically inserted nodes (e.g., implied pronouns, ellipsis placeholders) to align strictly with the observable Quranic string. This variant is optimized for downstream NLP tasks on the raw text.
  2. Augmented Variant: Retains analytically inserted nodes (marked with loc=_ in the source) to preserve EQTB's explicit modeling of ellipsis and implied arguments. This variant supports syntactic research where ellipsis resolution is a primary object of inquiry.
• Tokenization and Mapping: EQTB morpho-syntactic segments are treated as UD tokens. Part-of-speech (POS) categories are mapped to 12 UD Universal POS (UPOS) tags using a conservative mapping table (e.g., mapping specific demonstratives and relatives to PRON), while original EQTB tags are preserved in the XPOS column for traceability. Morphological features are derived directly from explicit EQTB columns rather than inferred.
• Dependency Normalization: The conversion enforces a single-root structure per sentence, prioritizing lexical categories (VERB > NOUN/PROPN) for the root. It collapses 129 fine-grained EQTB dependency labels into a compact inventory of 18 UD relations. Key structural normalizations include:
  • Adposition Inversion: Adpositions (ADP) are converted to case dependents (case), with the nominal complement becoming the head, aligning with UD content-head conventions.
  • Negation Inversion: Negative particles governing a single verbal predicate are inverted so the predicate becomes the head.
  • Cycle Resolution: Deterministic heuristics resolve rare cycles by reattaching nodes to the sentence root.
• Traceability: The conversion script preserves provenance by storing EQTB identifiers and location data in the MISC field. Constituency labels from EQTB are maintained as a separate stand-off layer, preventing the overloading of core UD fields while allowing for reproducible reconstruction.
• Benchmarking: The authors established fixed train/dev/test splits (80/10/10, seed 42) and trained lightweight Stanza baselines (BiLSTM-based tagger and biaffine parser) from scratch without pretrained vectors or character encoders to provide a reproducible lower bound for performance.

Key Contributions

1. UD-Quran Resource: A fully compliant UD v2 treebank of the entire Quran, available in both surface and augmented variants, containing 11,693 sentences and over 128,000 tokens (surface) or 139,000 tokens (augmented).
2. Interoperability Layer: A conversion framework that enables the use of standard UD tooling (validators, parsers, evaluation scripts) on Quranic data without erasing the analytic richness of the source EQTB.
3. Controlled Comparison: The dual-release design allows for controlled investigation into the impact of explicit ellipsis modeling on parsing performance and structural statistics.
4. Reproducible Baselines: The paper provides fixed data splits and baseline results using the CoNLL 2018 evaluation script, establishing a reference point for future work on Quranic/Classical Arabic parsing.

Results

• Corpus Statistics: The surface variant contains 128,219 tokens with a mean of 1.656 tokens per orthographic word, reflecting systematic clitic segmentation. The augmented variant adds 11,157 inserted nodes. The distribution is dominated by pronouns (22.86%), nouns (19.60%), and verbs (15.10%).
• Parsing Performance:
  • Gold-Tag Parsing: Using gold UPOS and FEATS, the Stanza parser achieved a Labeled Attachment Score (LAS) of 80.66 (surface) and 82.47 (augmented).
  • End-to-End Pipeline: With predicted tags, the LAS dropped to 64.52 (surface) and 68.54 (augmented).
  • Error Analysis: The results indicate that the primary bottleneck is morphosyntactic prediction (tagging) rather than the parser architecture itself. The augmented variant consistently outperformed the surface variant, suggesting that explicit modeling of ellipsis aids parsing accuracy.
• Coverage: The conversion is highly specific, restricting the generic dep relation to only 0.881% of tokens (1,129 tokens), ensuring that most dependencies are mapped to specific, meaningful UD relations.

Significance
The paper positions UD-Quran not as a replacement for traditional Quranic annotation but as a pragmatic interoperability layer. Its significance lies in bridging the gap between philologically grounded, tradition-specific analyses and the standardized requirements of the modern UD ecosystem. By maintaining traceability to the original EQTB annotations, the resource allows researchers to utilize standard NLP pipelines while retaining the ability to inspect and revert to the original hybrid analysis. The authors emphasize that this approach facilitates reproducible benchmarking and integration into widely used toolchains without requiring Quranic resources to abandon their analytic commitments. The paper concludes that while the current baselines are lightweight, the resource provides a necessary foundation for future, higher-capacity experiments and the iterative refinement of mapping decisions.