Datasets:

commul
/

universal_dependencies

Blocked Treebanks Support
- Added blocked_treebanks.yaml configuration for license-restricted treebanks
- Metadata generation now includes "blocked" property for each treebank
- Dataset card (README.md) automatically filters blocked treebanks from configs
- See tools/BLOCKED_TREEBANKS.md for details
Documentation Improvements
- Added RELEASE.md - Framework version release guide
- Updated MIGRATION.md - Documents external ud-hf-parquet-tools library
- Rewrote tools/README.md - Concise quick reference for developers
- Updated ADDING_NEW_UD_VERSION.md - Comprehensive guide with current architecture
- Added comprehensive Python patterns to .gitignore

Script Renaming for Clarity
- 02_traverse_ud_repos.py → 02_generate_metadata.py (better reflects purpose)
- 03_fill_universal_dependencies_tamplate.py → 03_generate_README.py (fixed typo, clarified scope)
- Updated all documentation and workflows to reference new names
Tooling Simplification
- Removed generation of universal_dependencies-{VER} files (obsolete with Parquet-only architecture)
- Removed templates/universal_dependencies.tmpl (no longer needed)
- Moved tests to external ud-hf-parquet-tools library
- 03_generate_README.py now only generates README, not Python loader script
Metadata Generation
- 02_generate_metadata.py now reads blocked_treebanks.yaml and adds blocked status to metadata
- Enhanced logging for blocked treebanks
Template Updates
- templates/README.tmpl now filters blocked treebanks from configs section
- Updated CONLLU_PARSING documentation link to point to ud-hf-parquet-tools repository

License Compliance
- Properly excludes treebanks with restrictive licenses (e.g., CC BY-NC-SA) that prohibit redistribution in derived formats
- Currently excludes: pt_cintil (Portuguese-CINTIL)

Obsolete Files
- Removed all universal_dependencies-2.{7-17} generated files (11 files)
- Removed templates/universal_dependencies.tmpl template
- Removed local tests/ directory (functionality moved to ud-hf-parquet-tools)

All documentation now reflects the current architecture:

[2.0.0] - 2026-01-09

Requires datasets>=4.0.0 for Parquet format support
- Python dataset scripts are deprecated in datasets 4.0.0 (July 2025)
- Parquet format is now the primary distribution method
Token sequences now correctly exclude MWT forms
- Previous versions incorrectly included Multi-Word Token (MWT) surface forms in token sequences
- Token counts now match num_words from stats.xml, not num_tokens
- Affects ~50+ treebanks with contractions (French, Italian, Portuguese, Spanish, Arabic, etc.)
- Example: French "des" (MWT 18-19) → "de" + "les" (syntactic words 18, 19)
  - v1.x (buggy): tokens = [..., "des", "de", "les", ...]
  - v2.0 (correct): tokens = [..., "de", "les", ...], mwt = [{"id": "18-19", "form": "des", "misc": ""}]

Multi-Word Token (MWT) Support
- New mwt field in dataset schema containing structured MWT information
- Schema: [{"id": "1-2", "form": "surface_form", "misc": "metadata"}]
- Enables research on contractions, clitics, and word segmentation across languages
- MWT statistics now collected from stats.xml (num_fused field)
Parquet Format Distribution
- All 339 treebanks available as Parquet files for efficient loading
- 5-10x faster loading compared to on-the-fly CoNLL-U parsing
- Reduced memory footprint and better compression
- Compatible with HuggingFace datasets >=4.0.0
Separate Semantic Versioning
- Loader version (2.0.0) now tracked independently from UD data version (2.17)
- Version in pyproject.toml reflects codebase changes
- UD data version remains in dataset configuration
- Enables clearer communication of breaking changes in tooling
Parquet Generation Pipeline
- New tools/04_generate_parquet.py script for CoNLL-U → Parquet conversion
- Supports batch processing and validation
- Test mode for quick iteration (--test flag)
- Progress tracking and error reporting

Critical Bug: MWT lines incorrectly included in token sequences
- Fixed filtering in _generate_examples() to exclude MWT lines (tuple IDs) and empty nodes
- Now uses sent.filter(id=lambda x: type(x) is int) to extract only syntactic words
- Impact: Token counts were inflated by ~0.1-2% in affected treebanks
- All token-level annotations (lemmas, UPOS, XPOS, features, heads, deprels) now correctly aligned

Dataset schema extended (non-breaking addition)
- Added mwt field to all treebanks (empty list for treebanks without MWTs)
Dependencies updated
- Added pyarrow>=14.0.0 for Parquet support
- Added datasets>=4.0.0 for compatibility
- Added conllu>=5.0.0 for reliable CoNLL-U parsing
Metadata collection enhanced
- tools/02_traverse_ud_repos.py now extracts <fused> counts from stats.xml
- Enables validation of MWT extraction accuracy

Dynamic loading from Universal Dependencies GitHub repositories
Automatic train/dev/test split detection
CoNLL-U format parsing with conllu library
Dataset features: tokens, lemmas, UPOS, XPOS, features, heads, dependency relations
Metadata: language codes, families, licenses, genres

See MIGRATION.md for detailed migration instructions from v1.x to v2.0.

Loader version (e.g., 2.0.0): Reflects changes to the loading infrastructure and tooling
UD data version (e.g., 2.17): Reflects the Universal Dependencies release version

Both versions are tracked separately to provide clarity on what changed.