README.md

---
dataset_info:
  features:
  - name: tokens
    list: string
  - name: segment_tags
    list:
      class_label:
        names:
          '0': B
          '1': I
  splits:
  - name: train
    num_bytes: 618033522
    num_examples: 544133
  - name: test
    num_bytes: 32877176
    num_examples: 28639
  download_size: 62316691
  dataset_size: 650910698
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
task_categories:
- token-classification
language:
- my
tags:
- myanmar
- burmese
- nlp
- sequence-labeling
- text-segmentation
- chunk-segmentation
pretty_name: Myanmar Text Segmentation Dataset
size_categories:
- 100K<n<1M
---

*Please visit to the [GitHub repository](https://github.com/chuuhtetnaing/myanmar-language-dataset-collection) for other Myanmar Langauge datasets.*

# Myanmar Text Segmentation Dataset

A token classification dataset for Myanmar (Burmese) chunk segmentation, formatted for sequence labeling tasks using the BIO tagging scheme.

📓 **Dataset Creation Notebook**: [myanmar-text-segmentation-dataset.ipynb](https://github.com/chuuhtetnaing/myanmar-language-dataset-collection/blob/main/Myanmar%20Text%20Segmentation/myanmar-text-segmentation-dataset.ipynb)

📓 **Fine-Tuning Notebook**: [myanmar-text-segmentation-fine-tuning.ipynb](https://github.com/chuuhtetnaing/myanmar-language-dataset-collection/blob/main/Myanmar%20Text%20Segmentation/myanmar-text-segmentation-fine-tuning.ipynb) (based on the [HuggingFace Token Classification Guide](https://huggingface.co/docs/transformers/en/tasks/token_classification))

🚀 **Try it out**: [Myanmar Text Segmentation Demo](https://huggingface.co/spaces/chuuhtetnaing/myanmar-text-segmentation-app)

## Dataset Description

This dataset is designed for **chunk segmentation** of Myanmar text. The input tokens are syllables (for Myanmar) or characters (for English), and the labels indicate chunk boundaries using B (Beginning) / I (Inside) tags.

For example, the unsegmented text `မြန်မာနိုင်ငံတွင်` is first broken into syllables `["မြန်", "မာ", "နိုင်", "ငံ", "တွင်"]`, then labeled as `[B, I, B, I, I]` to produce the segmented output `မြန်မာ နိုင်ငံတွင်`.

## Source Data

Derived from [chuuhtetnaing/myanmar-wikipedia-dataset](https://huggingface.co/datasets/chuuhtetnaing/myanmar-wikipedia-dataset).

### Processing Pipeline

1. **Paragraph extraction**: Each Wikipedia article is split by newlines, preserving full paragraphs as individual rows rather than sentence-by-sentence. This design allows models to handle multi-sentence inputs without requiring line-by-line splitting at inference time.

2. **Language filtering**: Each paragraph is classified using Facebook's [fastText language identification model](https://huggingface.co/facebook/fasttext-language-identification). Only paragraphs identified as Myanmar (`__label__mya_Mymr`) are retained.

3. **Tokenization**: Myanmar text is tokenized into syllables using regex-based rules that handle consonants, subscripts (္), and asat (်) markers. English text is tokenized into individual characters.

4. **Chunk boundary labeling**: Original spacing from the Wikipedia source text is converted to B/I sequence labels, where `B` marks the first token of each chunk and `I` marks continuation tokens.

5. **Deduplication**: Duplicate token sequences are removed from the final dataset.

## Dataset Statistics

| Split | Examples |
|-------|----------|
| Train | 544,133 |
| Test | 28,639 |

## Data Format

```python
{
    "tokens": ["မြန်", "မာ", "နိုင်", "ငံ", "တွင်"],  # List of tokens
    "segment_tags": [0, 1, 0, 1, 1]  # 0 = B (chunk start), 1 = I (chunk continuation)
}
```

## Features

- `tokens`: `Sequence[string]` - Input tokens (Myanmar syllables or English characters)
- `segment_tags`: `Sequence[ClassLabel]` - Chunk boundary labels (`B`=0, `I`=1)

## Usage

```python
from datasets import load_dataset
ds = load_dataset("chuuhtetnaing/myanmar-text-segmentation-dataset")

ds["train"].features["segment_tags"].feature.names
# ['B', 'I']
```

```python
def reconstruct(tokens, labels):
    result = []
    for token, label in zip(tokens, labels):
        if label == 0 and result:  # B tag (chunk boundary)
            result.append(" ")
        result.append(token)
    return "".join(result)

ds["test"][1018]['tokens']
# ['လျှပ်', 'စစ်', 'ငါး', 'ရှ', 'ဉ့်', '(Electrophorus)', 'သည်', 'မျိုး', 'ရင်း', 'Gymnotidae', 'ရှိ', 'ရေ', 'ချို', 'နေ', 'ငါး', 'ရှ', 'ဉ့်', 'မျိုး', 'စု', 'တ', 'ခု', 'ဖြစ်', 'သည်', '။', ...]

ds["test"][1018]['segment_tags']
# [0, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, ...]

reconstruct(ds["test"][1018]['tokens'], ds["test"][1018]['segment_tags'])
# 'လျှပ်စစ်ငါးရှဉ့် (Electrophorus) သည် မျိုးရင်း Gymnotidae ရှိ ရေချိုနေ ငါးရှဉ့် မျိုးစုတခုဖြစ်သည်။ ...'
```

## Intended Use

- Training chunk segmentation models for Myanmar NLP
- Token classification / sequence labeling experiments ([HuggingFace Token Classification Training Example](https://huggingface.co/docs/transformers/en/tasks/token_classification)).
- Myanmar language processing research