---
language:
- en
- zh
tags:
- chinese
- traditional-chinese
license: apache-2.0
task_categories:
- text-classification
pretty_name: Traditional Chinese vs Simplified Chinese Classification Dataset
size_categories:
- 100K<n<1M
---

# Traditional Chinese vs Simplified Chinese Classification Dataset

## Dataset Summary

This dataset contains **205,850** labeled and class-balanced Chinese text snippets in Traditional and Simplified Chinese, designed for training and evaluating models that classify Traditional Chinese variants (Taiwan Traditional vs. Mainland Traditional). It is carefully constructed to cover diverse sources, length distributions, and domains to encourage strong generalization.

### Key Characteristics

* **Total samples**: 205,850
* **Class balance**: perfectly balanced (Simplified: 102,925 / Traditional: 102,925)
* **File format**: Parquet (≈47 MB compressed; ~32.7% smaller than CSV)
* **Average text length**: 40–90 characters (includes long-text segments)
* **Multi-source**: sentence fragments, full sentences, paragraph segments
* **Quality control**: includes metadata such as Mainland-term score and lexical annotations

### Supported Tasks

* Traditional Chinese variant classification (Taiwan Traditional vs. Mainland Traditional)
* Text origin/variety identification
* Data cleaning and standardization pre-processing
* Cross-domain Chinese NLP research

## Dataset Structure

### Features

| Feature           | Type   | Description                                                          |
| ----------------- | ------ | -------------------------------------------------------------------- |
| `text`            | string | Main text content (for training and inference)                       |
| `text_length`     | int    | Character length of the text                                         |
| `language_type`   | string | Label (`簡體中文` or `繁體中文`)                                             |
| `total_score`     | int    | Mainland-term score (1–5; higher indicates stronger Mainland traits) |
| `mainland_terms`  | string | Detected Mainland-specific terms (may be empty)                      |
| `source_type`     | string | Source type (`sentence_fragment`, `full_sentence`, etc.)             |
| `fragment_length` | int    | Length of the fragment                                               |
| `fragment_index`  | int    | Index of the fragment in the original text                           |
| `fragment_start`  | float  | Start position of the fragment                                       |
| `fragment_end`    | float  | End position of the fragment                                         |
| `original_text`   | string | Original full text (if applicable)                                   |

### Splits

The default config provides the full dataset. Users are encouraged to create train/validation/test splits as needed:

```python
from datasets import load_dataset
from sklearn.model_selection import train_test_split

# Load dataset
dataset = load_dataset("renhehuang/traditional-chinese-classification", split="train")

# Recommended split: 80% train / 10% validation / 10% test
# Use stratify to maintain class balance
```

### Example

```json
{
  "text": "這組數據給我們老師提了個醒，以後對於咱自己的健康真得長點兒心，不能再忽視大意了。",
  "text_length": 40,
  "language_type": "簡體中文",
  "total_score": 3,
  "mainland_terms": "",
  "source_type": "sentence_fragment",
  "fragment_length": 40,
  "fragment_index": 30,
  "fragment_start": 2235.0,
  "fragment_end": 2275.0,
  "original_text": null
}
```

## Data Collection & Processing

### Sources

Aggregated from multiple public and internally curated corpora, including:

* 📰 News articles (both sides of the Strait)
* 💬 Social media (cleaned and de-identified)
* 📚 Book and literary excerpts
* 🛒 E-commerce reviews (product descriptions and user reviews)
* 🗣️ Forums and Q&A platforms

### Processing Pipeline

1. **Glyph-conversion validation**: OpenCC for consistent Simplified/Traditional conversion
2. **Lexical annotations**: automatic detection of Mainland/Taiwan region-specific terms (e.g., "軟件/軟體", "視頻/影片")
3. **Length-stratified sampling**: balanced coverage across text lengths
4. **Deduplication**: similarity-based removal of near-duplicates
5. **Quality filtering**: removal of samples with excessive foreign characters, symbols, or garbled text
6. **Balancing**: under/over-sampling to achieve class balance

### Labeling Method

* **Automatic**: rules based on origin, glyphs, and lexical features
* **Human review**: expert checks for borderline cases (~5% of samples)
* **Consistency checks**: cross-validation to ensure label accuracy

## Usage

### Basic Loading

```python
from datasets import load_dataset

# Load full dataset (Parquet format for faster IO)
dataset = load_dataset("renhehuang/traditional-chinese-classification")

# Inspect the first record
print(dataset['train'][0])
```

### Train/Validation Split

```python
from datasets import load_dataset

dataset = load_dataset("renhehuang/traditional-chinese-classification", split="train")

# 80/20 split while preserving class balance
split = dataset.train_test_split(test_size=0.2, stratify_by_column="language_type", seed=42)
train_dataset = split['train']
val_dataset = split['test']

print(f"Train: {len(train_dataset)}")
print(f"Validation: {len(val_dataset)}")
```

### Integration with Transformers

```python
from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("renhehuang/traditional-chinese-classification", split="train")
tokenizer = AutoTokenizer.from_pretrained("ckiplab/bert-base-chinese")

def preprocess(examples):
    # Map labels to integers
    labels = [0 if label == "簡體中文" else 1 for label in examples["language_type"]]

    tokenized = tokenizer(
        examples["text"],
        truncation=True,
        max_length=384,
        padding="max_length"
    )

    tokenized["labels"] = labels
    return tokenized

# Batch processing
tokenized_dataset = dataset.map(preprocess, batched=True, remove_columns=dataset.column_names)
```

### Long-Text Handling

```python
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("ckiplab/bert-base-chinese")

def chunk_text(text, max_length=384, stride=128):
    """Chunk long texts with sliding windows"""
    encoding = tokenizer(
        text,
        truncation=True,
        max_length=max_length,
        stride=stride,
        return_overflowing_tokens=True,
        return_attention_mask=True,
        padding="max_length"
    )
    return encoding

# Chunk an over-length sample
long_text_sample = dataset['train'].filter(lambda x: x['text_length'] > 384)
if len(long_text_sample) > 0:
    chunks = chunk_text(long_text_sample[0]['text'])
    print(f"#chunks: {len(chunks['input_ids'])}")
```

## Use Cases

### ✅ Recommended

* **Text classification training**: fine-tuning Transformers (BERT, RoBERTa, etc.)
* **Origin variety identification**: infer TW vs. Mainland usage
* **Data cleaning**: standardizing mixed-corpus resources
* **Cross-domain transfer learning**: as a pretraining resource
* **Benchmarking**: evaluation bed for Chinese NLP models

### ⚠️ Limitations & Notes

1. **Hong Kong Traditional excluded**: only Taiwan Traditional and Mainland Traditional
2. **Domain bias**: higher share of social media and news; fewer professional domains (medical, legal)
3. **Temporal coverage**: collected up to Oct 2025; latest internet slang not guaranteed
4. **Code-mixing**: small proportion of zh–en mixed texts may affect purely monolingual tasks
5. **Label ambiguity**: ~2–3% borderline samples (generic lexicon, literary styles)

### 🚫 Not Recommended

* Hong Kong Traditional classification (requires additional labeling)
* Classical Chinese/literary analysis (modern text focus)
* Ultra-short text (< 10 characters; consider filtering)
* Highly specialized domains (augment with domain data)

## Evaluation & Baselines

### Verified Model

Using this dataset, `renhehuang/bert-traditional-chinese-classifier` achieved:

* **Accuracy**: 87.71%
* **F1**: 0.8771
* **Training samples**: 156,824 (subset of this dataset)

### Recommended Metrics

* Accuracy (overall)
* F1 score (macro/weighted)
* Confusion matrix
* Stratified performance by text length

### Baseline Training Setup

```python
from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=4,
    weight_decay=0.01,
    save_strategy="epoch",
    load_best_model_at_end=True,
    metric_for_best_model="eval_accuracy",
)
```

## Ethics & Fairness

### Privacy

* ✅ All personally identifiable information (PII) removed or anonymized
* ✅ No sensitive political, religious, or racial content included
* ✅ Social media content de-identified

### Bias & Fairness

* ⚠️ Potential source-side bias (media stance)
* ⚠️ No balancing on gender/age demographics
* ⚠️ Skew toward certain topics (tech, entertainment)

### Usage Guidance

1. **Avoid discriminatory applications**: do not use to disadvantage specific regions
2. **Human-in-the-loop**: add human review for critical decisions
3. **Ongoing monitoring**: track fairness metrics after deployment
4. **Transparency**: disclose dataset source when training models

## License & Citation

### License

This dataset is released under **CC BY 4.0** (Creative Commons Attribution 4.0 International):

* ✅ Commercial use
* ✅ Modification and derivatives
* ✅ Distribution and re-licensing
* ⚠️ Attribution required

### Citation

If you use this dataset, please cite:

```bibtex
@misc{traditional-chinese-classification-dataset,
  title = {Traditional Chinese vs Simplified Chinese Classification Dataset},
  author = {Huang, Renhe},
  year = {2025},
  month = {10},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/renhehuang/traditional-chinese-classification}},
  note = {205,850 balanced samples for Traditional/Simplified Chinese classification}
}
```

### Related Resources

* 🤗 **Model**: [renhehuang/bert-traditional-chinese-classifier](https://huggingface.co/renhehuang/bert-traditional-chinese-classifier)
* 📊 **Project repo**: [GitHub](https://github.com/Edwarddev0723/bert-base-chinese-traditional-classifier)
* 📝 **Technical report**: see project README

## Maintenance & Updates

### Version History

* **v1.0** (2025-10-29): Initial release, 205,850 samples

### Known Issues

1. Most `original_text` fields are null (reserved for future expansion)
2. Some `mainland_terms` annotations may be incomplete
3. `fragment_start/end` can be NaN for full-sentence samples

### Roadmap

* [ ] Add Hong Kong Traditional label (planned v2.0)
* [ ] Expand professional domains (medical, legal, finance)
* [ ] Provide pre-split train/val/test sets
* [ ] Add multimodal data (images, ASR transcripts)

## Contact

* **Dataset issues**: open a Discussion on the [Hugging Face dataset page](https://huggingface.co/datasets/renhehuang/traditional-chinese-classification)
* **Model-related**: use [GitHub Issues](https://github.com/Edwarddev0723/bert-base-chinese-traditional-classifier/issues)
* **Collaboration**: contact via Hugging Face private messages

---

## Appendix: Quick Stats

```python
# Dataset statistics (summary)
{
  "total_samples": 205850,
  "labels": {
    "簡體中文": 102925,
    "繁體中文": 102925
  },
  "file_format": "parquet",
  "file_size_mb": 47.09,
  "compression": "snappy",
  "avg_text_length": 56.3,
  "median_text_length": 48,
  "min_text_length": 10,
  "max_text_length": 384,
  "source_types": {
    "sentence_fragment": "~85%",
    "full_sentence": "~10%",
    "paragraph": "~5%"
  }
}
```

## Acknowledgements

Thanks to all data providers, annotators, and the open-source community. Special thanks to:

* CKIP Lab (base BERT model)
* OpenCC (Simp/Trad conversion)
* Hugging Face (dataset hosting)
* All users who provided feedback and suggestions