KRadim/czech-punctuation-pos-syntax

Czech Punctuation, POS and Syntactic Dataset 🇨🇿

A High-Quality Dataset for Punctuation Restoration and Neuro-Symbolic LLM Grounding

This dataset is a structured, linguistically annotated corpus of the Czech language, specifically designed for Punctuation Restoration tasks, Part-of-Speech (POS) tagging, and token-level syntax embedding (such as nanoGPT custom metadata training).

Unlike pure raw text corpora, this dataset provides a deterministic 1:1 token-level mapping of morphological features (word class, grammatical case) and dependency syntax roles (subject, predicate, object) directly derived via Stanford's state-of-the-art Stanza pipeline (trained on the Prague Dependency Treebank standard).

---

📊 Dataset Structure

The dataset contains three official subsets split using a strict, reproducible 80% / 10% / 10% distribution with seed shuffling:

• Train: 80% of sentences
• Validation: 10% of sentences
• Test: 10% of sentences

Data Schema

Every split is delivered in the optimized Apache Parquet format utilizing the following strict schema layout:

Column Name	Type	Description
segment	string	The original raw Czech sentence structure (text segment).
punctuation_type	string	Classification target category based on the sentence ending/comma presence (čárka, tečka, otazník, vykřičník, žádná).
tokens_annotation	list (struct)	An array of structured token profiles (excluding raw punctuation marks).

Inside tokens_annotation Struct:

• slovo (string): The clean isolated token/word.
• slovni_druh (string): Czech morphological category (podstatne_jmeno, sloveso, zajmeno, prislovce, predlozka, jiny).
• vetny_clen (string): Syntactic dependency function (podmet, prisudek, predmet, jiny).
• pad (int64): Real Czech grammatical case number (1 to 7), or 0 for indeclinable word classes.

---

💻 Usage Example

You can instantly load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the full dataset splits
dataset = load_dataset("KRadim/czech-punctuation-pos-syntax")

# Access individual splits
train_data = dataset["train"]
validation_data = dataset["validation"]
test_data = dataset["test"]

# Inspect an example row
print(train_data[0])

JSON Record Preview

{
  "segment": "Co je na nich tak tajného?",
  "punctuation_type": "otazník",
  "tokens_annotation": [
    { "slovo": "Co", "slovni_druh": "zajmeno", "vetny_clen": "podmet", "pad": 1 },
    { "slovo": "je", "slovni_druh": "sloveso", "vetny_clen": "jiny", "pad": 0 },
    { "slovo": "na", "slovni_druh": "predlozka", "vetny_clen": "jiny", "pad": 6 },
    { "slovo": "nich", "slovni_druh": "zajmeno", "vetny_clen": "jiny", "pad": 6 },
    { "slovo": "tak", "slovni_druh": "prislovce", "vetny_clen": "jiny", "pad": 0 },
    { "slovo": "tajného", "slovni_druh": "jiny", "vetny_clen": "prisudek", "pad": 2 }
  ]
}

🛠️ Pipeline Construction Method

1. Ingestion & Balance: Aggregated raw multi-domain source datasets uniformly distributed across 5 structural punctuation categories.
2. Deep Neural Annotation: Processed locally using Stanford Stanza's Czech syntax parser. This ensures authentic dependency parsing and prevents the data hallucinations often introduced by Generative LLMs.
3. Symbolic Normalization: Native UPOS tags and feature vectors mapped cleanly to Czech vocabulary definitions (podmet, prisudek, pádové reprezentace).
4. Cloud-Native Serialization: Combined, shuffled (seed=42), sliced, and cast into an explicit PyArrow Table Layout with Snappy compression for instantaneous loading times.

Detailed procedure for creating a dataset:

1. Part 1: ⚙️ Create Czech text dataset for training
2. Part 2: 🔬 CZECH NEURO-SYMBOLIC DATASET TRANSFORMER (STANZA VERSION)
3. Part 3: 💾 FINAL PRODUCTION EXPORT & HUGGING FACE PUSH (80/10/10 SPLIT)

📜 Licensing & Acknowledgments

The text annotations and structural database components are compiled from open Czech resources and processed via the Stanford Stanza framework. This dataset is made available under the Open Data Commons Attribution License (ODC-BY v1.0). You are free to share, modify, and use the dataset for commercial or non-commercial purposes, provided that you attribute the source appropriately.