README.md

metadata

license: apache-2.0
tags:
  - bangla
  - text-augmentation
  - text-adversarial-attack
  - nlp
  - corrupted-text
  - robustness
language:
  - bn
pretty_name: Augmented Bangla Dataset for Robustness Testing
dataset_info:
  features:
    - name: '0'
      dtype: string
    - name: '1'
      dtype: string
  splits:
    - name: train
      num_bytes: 5930986
      num_examples: 7980
  download_size: 1501634
  dataset_size: 5930986
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for Augmented Bangla Dataset (TextAttack)

Dataset Summary

This dataset is an augmented (adversarial attack) version of the hishab/hishab-pr-bn-v1 dataset using TextAttack, designed to support the development and evaluation of robust Bangla NLP models. It has been enhanced using TextAttack, a textual adversarial transformation framework, to simulate noisy or corrupted text inputs.

For each original paragraph, four types of character-level transformations were applied on individual lines, and the results were combined into augmented paragraphs. This makes the dataset useful for tasks involving robustness testing, denoising models, and error correction in Bangla.

---

Supported Tasks and Use Cases

• Text Robustness Evaluation
• Noisy Text Correction
• Augmentation for Low-resource NLP
• Training and testing resilient Bangla language models

---

Languages

• Bangla (bn)

---

Data Generation Method

This dataset was created by augmenting each paragraph from the hishab/hishab-pr-bn-v1 dataset using four different character-level transformations from TextAttack:

🔧 Transformations Applied

1. Random Character Insertion
   A random bn character is inserted into a random word.

2. Random Character Deletion
   A character is randomly deleted from a random word.

3. Random Character Substitution
   A character in a random word is replaced with another bn character.

4. Random Character Swap
   Two adjacent characters in a random word are swapped.

Each line in the original paragraph was augmented individually, and four augmented paragraphs were generated from a single paragraph.

---

Dataset Structure

Each example in the dataset contains:

• augmented_paragraph_without_punctuation: The augmented paragraph without punctuation
• augmented_paragraph_with_punctuation: The augmented paragraph with punctuations

Example Augmentation

{
  "original": "ঠিক বলেছিস। চল, এগিয়ে যাই।",
  "augmented": "ঠিক উবলেছিস। চথল, এগিয়ে যাই।"
}