commul/ud_genre

UD Genre Labels full-ud-v1.1.1-ud2.18

Derived sentence-level genre annotations for the universal-dependencies/universal_dependencies Universal Dependencies dataset. These labels are produced by the bootstrapping pipeline and are not authoritative gold annotations.

Dataset Summary

This dataset provides a sentence-level genre layer aligned to the universal-dependencies/universal_dependencies Parquet release. Each row contains one derived genre label for one UD sentence and can be joined back to the UD source data by (treebank, split, sent_id).

The export is a derived annotation layer, not a replacement for the UD treebanks and not a hand-validated gold genre dataset.

Label Coverage And Provenance

• Total UD sentences in artifact: 2,221,815
• Labeled sentences: 2,221,815 (100.0%)
• Unlabeled sentences: 0
• Genres exported: 18
• Metadata-derived labels: 1,334,339 (60.1% of labeled sentences)
• Clustering-derived labels: 887,476 (39.9% of labeled sentences)

Method	Meaning	Sentences	Share of labeled
cluster-derived	Assigned by cluster-to-reference genre similarity.	887,476	39.9%
single-genre-treebank	Directly inherited from a single-genre UD treebank.	736,887	33.2%
virtual-split	Directly inherited from sentence/document metadata in a mixed treebank.	597,452	26.9%

For cluster-derived rows, confidence is the top-1 cluster-label similarity score. Direct metadata-derived rows use confidence 1.0.

• Mean confidence: 0.9835
• Median confidence: 1.0000

Genre Distribution

Genre	Sentences	Share of labeled
news	688,568	31.0%
nonfiction	273,671	12.3%
fiction	252,860	11.4%
spoken	200,151	9.0%
wiki	110,856	5.0%
poetry	110,056	5.0%
academic	94,598	4.3%
grammar-examples	89,429	4.0%
bible	85,776	3.9%
blog	79,629	3.6%
legal	63,881	2.9%
social	47,339	2.1%
reviews	40,742	1.8%
web	34,128	1.5%
learner-essays	22,532	1.0%
government	12,728	0.6%
medical	9,372	0.4%
email	5,499	0.2%

Loading

from datasets import load_dataset

genres = load_dataset(
    "commul/ud_genre",
    revision="2.18",
    split="train",
)

The train split is the single exported split containing all sentence-level genre labels for this artifact.

For immutable provenance, load the artifact tag:

genres = load_dataset(
    "commul/ud_genre",
    revision="artifact/full-ud-v1.1.1/ud2.18",
    split="train",
)

Joining With Universal Dependencies

from datasets import load_dataset

genres = load_dataset(
    "commul/ud_genre",
    revision="2.18",
    split="train",
)

ud = load_dataset(
    "universal-dependencies/universal_dependencies",
    "en_ewt",
    revision="2.18",
    split="train",
)

genre_by_key = {
    (row["treebank"], row["split"], row["sent_id"]): row["genre"]
    for row in genres
    if row["treebank"] == "en_ewt" and row["split"] == "train"
}

first = ud[0]
genre = genre_by_key.get(("en_ewt", "train", first["sent_id"]))

Release Identity

• Train ID: full-ud-v1.1.1
• Artifact key: full-ud-v1.1.1-ud2.18
• Inventory status: default_hotfix
• HF branches: 2.18
• HF tag: artifact/full-ud-v1.1.1/ud2.18
• HF default branch: main
• HF repo: commul/ud_genre
• UD version: 2.18
• Scope: full
• Label schema: ud
• Artifact version: v1.1.1
• Source repo: https://github.com/bot-zen/ud-genre-bootstrap
• Source commit: 36188cb591a9d3122dd95b67258dc8e1fc7dcd4f
• Source branch: release/full-ud-v1
• Source tag: source/full-ud-v1.1.1
• Config SHA-256: ba9d3e8ec173f90659422a07e536ad74de9db2f848c3636b6d00321327a56eb8
• Canonical labels: academic, bible, blog, email, fiction, government, grammar-examples, learner-essays, legal, medical, news, nonfiction, poetry, reviews, social, spoken, web, wiki

Release Configuration

• Config: 2.18-community-release
• Run ID: full-ud-v1.1.1-ud2.18
• UD source: hf://universal-dependencies/universal_dependencies
• UD source revision: 2.18
• Embeddings: intfloat/multilingual-e5-large / mean
• Clustering: gmm
• Reference weighting: sentence_count

Output Columns

• treebank, split, sent_id: primary join key back to UD
• genre: derived sentence label
• confidence: top-1 similarity score for cluster-derived rows; 1.0 for direct metadata-derived rows
• method: single-genre-treebank, virtual-split, or cluster-derived
• ud_version, model, pooling, clustering_method, config_name, run_id: compact row-level provenance

Evaluation Framing

• paper_parity is used only for comparison with the original GMM+L paper protocol.
• End-user quality is tracked with sentence-level generalization metrics, which are stricter and more directly relevant for downstream annotation use.
• Known limitation: some paper-era treebank genre inventories are not fully recoverable from current sentence-level metadata subsets.

Evaluation Summary

• Baseline: UD v2.18 all_available generalization baseline
• Scope: Locked evaluation baseline measured on UD 2.18.
• Protocol: generalization
• Treebank set: all_available
• Config: configs/releases/full-ud-v1.1.0.yaml
• Source log: output/logs/2.18-all_available-generalization-e5_large-k10-anchor_combined-baseline.log

Metric	Value
Evaluated Sentences	629900
Multi-Genre Treebank Splits	79
Instance-Labeled Treebanks	44
Instance-Labeled Treebank Splits	79
Number of Folds	10
Overall Acc / Micro-F1	0.3324
Macro-F1	0.2699
Mean Fold Micro-F1	0.3373
Std Fold Micro-F1	0.1300
Mean Fold Macro-F1	0.2837
Std Fold Macro-F1	0.1246
Purity (PUR)	0.5190
Agreement (AGR)	0.5062
Overlap Error (Delta BC)	0.0760

• Missing anchor genres in this baseline: email

Release Summary

• Total sentences: 2,221,815
• Labeled sentences: 2,221,815
• Genres exported: 18
• Methods exported: cluster-derived, single-genre-treebank, virtual-split

Source Mapping Files

• configs/genre_mappings.json (sha256: a52bbd9d5be0ec267da69624930e5087f28359e0dea88d9f039534b0167cbe0e)
• configs/metadata_patterns.json (sha256: 2e0898fe402e72e6050a5efffd8f963696e6c693763325313c7b290540696d5c)
• configs/pud-patterns.json (sha256: 4b5ca652110604a9b567deea630115eda24a2b734359b4f48e71c0fee7b2034d)

Citation

Please cite the UD Workshop paper associated with this dataset: https://universaldependencies.org/udw26/papers/41_Paper.pdf

Contact

Point of Contact: appliedlinguisticsdevs@eurac.edu