Datasets:

undertheseanlp
/

UDD-1

UDD-1 (Universal Dependency Dataset for Vietnamese) is a Vietnamese Universal Dependencies treebank with 40,000 sentences from 5 domains. The repo contains both the dataset files (CoNLL-U format) and the tooling pipeline for creating/validating them.

Domain Breakdown

Category	Source Dataset	Sentences	Sent ID Prefix
Wikipedia	`undertheseanlp/UVW-2026`	8,000	`uvw-`
News	`undertheseanlp/UVN-1`	8,000	`uvn-`
Legal	`undertheseanlp/UTS_VLC`	8,000	`vlc-`
Fiction	`undertheseanlp/UVB-v0.1`	8,000	`uvb-f-`
Non-fiction	`undertheseanlp/UVB-v0.1`	8,000	`uvb-n-`

Repository Structure

Root .conllu files: The published dataset splits (vi_udd-ud-{train,dev,test}.conllu)
data/: Parquet files for HuggingFace dataset hosting
src/: Pipeline scripts for data fetching, conversion, validation, and upload
tools/udtools/: Vendored copy of the Universal Dependencies tools package (validator + scorer)

Key Pipeline Scripts

Script	Purpose
`src/fetch_data.py`	Fetch 8,000 sentences from `undertheseanlp/UTS_VLC` (legal domain) → `sentences_vlc.txt`
`src/fetch_uvn_data.py`	Fetch 8,000 sentences from `undertheseanlp/UVN-1` (news domain) → `sentences_uvn.txt`
`src/fetch_uvw_data.py`	Fetch 8,000 sentences from `undertheseanlp/UVW-2026` (Wikipedia, quality_score >= 5) → `sentences_uvw.txt`
`src/fetch_uvb_data.py`	Fetch 8,000 fiction + 8,000 non-fiction from `undertheseanlp/UVB-v0.1` → `sentences_uvb.txt`
`src/build_dataset.py`	Combine all sentence files, assign sent_id prefixes, create stratified train/dev/test splits → `sentences_{train,dev,test}.txt`
`src/convert_to_ud.py`	Convert raw sentences to UD format using `underthesea` NLP toolkit (dependency parsing + POS tagging). Outputs JSONL and CoNLL-U
`src/statistics.py`	Compute dataset statistics from CoNLL-U files
`src/upload_to_hf.py`	Upload dataset splits to HuggingFace Hub with domain field (requires `HF_TOKEN` env var)
`src/run_conversion.sh`	Wrapper that runs conversion with GPU monitoring and timestamped results
`src/run_on_runpod.py`	Manage RunPod GPU instances for conversion (requires `RUNPOD_API_KEY`)
`src/fetch_ws_sentences.py`	Fetch 100K sentences (20K × 5 domains) for word segmentation dataset → `ws_sentences_*.txt`
`src/build_ws_dataset.py`	Convert sentences to BIO format via `word_tokenize` + `regex_tokenize` → `ws_{train,dev,test}.txt`
`src/fix_ws_errors.py`	Fix known WS errors (cross-boundary merges, always-split compounds) in BIO files → regenerate CoNLL-U
`src/check_ws_errors.py`	Rule-based WS error checker (7 rules: inconsistency, dictionary, long tokens, punctuation, numbers, anomalies) → `WS_CHECK_REPORT.md`

Pipeline Commands

1. Fetch sentences from all sources

python src/fetch_data.py        # Legal → sentences_vlc.txt
python src/fetch_uvn_data.py    # News → sentences_uvn.txt
python src/fetch_uvw_data.py    # Wikipedia → sentences_uvw.txt
python src/fetch_uvb_data.py    # Books → sentences_uvb.txt

2. Build combined dataset with splits

python src/build_dataset.py     # → sentences_{train,dev,test}.txt

3. Run UD conversion (GPU-optimized)

python src/convert_to_ud.py -i sentences_train.txt -o output/ -p train -b 64
python src/convert_to_ud.py -i sentences_dev.txt -o output/ -p dev -b 64
python src/convert_to_ud.py -i sentences_test.txt -o output/ -p test -b 64
# Or use the shell wrapper:
./src/run_conversion.sh <input_file> [batch_size]

4. Validate CoNLL-U files

cd tools/udtools
pip install -e .
python validate.py --lang vi vi_udd-ud-train.conllu

5. Run udtools tests

cd tools/udtools
python -m pytest tests/

6. Compute dataset statistics

python src/statistics.py

7. Upload to HuggingFace

export HF_TOKEN=<token>
python src/upload_to_hf.py

Word Segmentation Dataset Pipeline

Separate pipeline producing a 100K-sentence BIO-tagged dataset (VLSP 2013 compatible) for CRF word segmentation training in tree-1.

# Step 1: Fetch 20K sentences per domain (100K total)
uv run src/fetch_ws_sentences.py
# → ws_sentences_vlc.txt, ws_sentences_uvn.txt, ws_sentences_uvw.txt,
#   ws_sentences_uvb_f.txt, ws_sentences_uvb_n.txt

# Step 2: Convert to BIO format with stratified 80/10/10 split
uv run src/build_ws_dataset.py
# → udd-ws-v1.1-train.txt (~80K), udd-ws-v1.1-dev.txt (~10K), udd-ws-v1.1-test.txt (~10K)

# Step 3: Fix known segmentation errors (cross-boundary merges, always-split compounds)
uv run src/fix_ws_errors.py
# → fixes BIO files in-place, regenerates CoNLL-U, writes WS_FIX_REPORT.md

# Step 4: Run diagnostic checker (7 rules, no fixes applied)
uv run src/check_ws_errors.py                  # All rules (needs underthesea for dict rules)
uv run src/check_ws_errors.py --no-dict        # Skip dictionary rules (2, 3)
# → WS_CHECK_REPORT.md

Output BIO format (syllable\tB-W / syllable\tI-W, blank line between sentences) is compatible with tree-1's load_data_vlsp2013() which maps B-W→B, I-W→I.

Architecture Notes

Conversion Pipeline (`convert_to_ud.py`)

The core conversion flow: raw Vietnamese text -> underthesea.dependency_parse() + underthesea.pos_tag() -> Vietnamese POS mapped to Universal POS via UPOS_MAP -> syntax error post-processing via fix_syntax_errors() -> CoNLL-U output.

fix_syntax_errors() is a critical multi-pass function that corrects UD validation issues:

Redirects children of leaf-only relations (aux, case, punct, det, etc.)
Maps invalid deprels via DEPREL_MAP
Enforces UPOS/deprel consistency (e.g., det must be DET/PRON, advmod must be ADV)
Handles Vietnamese-specific auxiliary verbs (AUX_WORDS) and copula (la)
Fixes directional constraints (flat/conj/appos must be left-to-right)
Resolves multiple subjects/objects per predicate
Fixes non-projective punctuation attachment

udtools (Vendored UD Validator)

The tools/udtools/ directory is a vendored copy of the official UD tools. The validator class hierarchy is: Validator -> Level6 -> Level5 -> ... -> Level1. Each level adds progressively stricter UD compliance checks. The Validator class in validator.py is the main entry point.

Data Format

CoNLL-U: Standard 10-column UD format (ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, MISC)
JSONL: HuggingFace-compatible format with fields: sent_id, text, tokens, lemmas, upos, xpos, feats, head, deprel, deps, misc, domain
Sentence ID prefixes: vlc- = legal, uvn- = news, uvw- = wikipedia, uvb-f- = fiction, uvb-n- = non-fiction
Split ratios: Train (91.4%) / Dev (4.3%) / Test (4.3%), stratified by domain

Dependencies

underthesea: Vietnamese NLP toolkit (tokenization, POS tagging, dependency parsing)
torch: Required by underthesea models (GPU-accelerated)
datasets, huggingface_hub: For HuggingFace dataset operations
udtools dependencies: udapi>=0.5.0, regex>=2020.09.27 (see tools/udtools/pyproject.toml)

CLAUDE.md

Project Overview