Welcome to PdD, the game-changing Character-Level Perturbation Generator! If you've ever wanted to push the boundaries of text data augmentation, you've come to the right place! PdD isn't just another tool—it’s your text manipulation playground, where every character can be twisted, swapped, and transformed to your heart’s content.
At its core, PdD is driven by three powerful principles: Probability distribution, Density, and Diversity. These metrics empower you to perturb text in ways that are dynamic, intelligent, and, most importantly, impactful! Whether you're enhancing neural network robustness, simulating real-world typos, or exploring creative text transformations, PdD gives you full control over your textual experiments.
PdD is based on our paper “Formalizing Robustness Against Character-Level Perturbations for Neural Network Language Models.” Our work ensures that the perturbations aren’t just random noise but structured, meaningful, and deeply insightful. You can cite our work using the following BibTeX entry:
@inproceedings{ma2023formalizing,
title={Formalizing Robustness Against Character-Level Perturbations for Neural Network Language Models},
author={Ma, Zhongkui and Feng, Xinguo and Wang, Zihan and Liu, Shuofeng and Ma, Mengyao and Guan, Hao and Meng, Mark Huasong},
booktitle={International Conference on Formal Engineering Methods},
pages={100--117},
year={2023},
organization={Springer}
}Imagine this simple sentence:
"I am the best student in the world."
Instead of processing whole words, PdD views this as a sequence of characters, unlocking powerful transformations:
['I', ' ', 'a', 'm', ' ', 't', 'h', 'e', ' ', 'b', 'e', 's', 't', ' ',
's', 't', 'u', 'd', 'e', 'n', 't', ' ', 'i', 'n', ' ',
't', 'h', 'e', ' ', 'w', 'o', 'r', 'l', 'd', '.']PdD gives you the power to customize perturbations with three key parameters:
- 🔄 Probability Distribution: Control how characters are perturbed. Choose between uniform or normal distributions (managed in
distribution.py). - 📊 Density: Define the fraction of characters to perturb. A density of
0.1means 10% of characters will be modified (adjustable insampler.py). - 🌀 Diversity: The fun part! Select from a range of perturbations—deletion, typos, invisible characters, and more! Explore the
diversityfolder for all the possibilities.
PdD isn’t just about adding random noise—it gives you full control over how text is perturbed. Here are some exciting options:
- ✂️ Deletion – Remove characters entirely (
''). - 🔳 Space Insertion – Turn letters into spaces (
' '). - 🔀 Case Switching – Flip uppercase and lowercase (
'a'↔'A'). - ⌨️ Keyboard Typos – Simulate real-world typos (e.g.,
'g'→'h'). - 🎭 Homoglyphs – Replace characters with visually similar ones (e.g.,
'a'→'а'). - 🫥 Invisible Characters – Insert invisible symbols that mess with models.
Make sure you're using Python 3.10 or later and install the required dependencies:
- torch
- tqdm
- scikit-learn
- transformers
- datasets
PdD is structured to be intuitive and easy to experiment with:
./diversity– The heart of PdD! Define custom perturbation rules here../example– Quick-start examples to see PdD in action../scripts– Scripts for running experiments based on our research../eval– Code for training and evaluating models on perturbed data../results– Stores original and perturbed outputs + evaluation results.
PdD is fully customizable. Want to tweak how certain characters get perturbed? Dive into the diversity folder and define your own perturbation rules. The power is in your hands! 💪
Need to perturb only certain characters at specific locations? PdD lets you define exactly where and how to apply perturbations, giving you unparalleled control over text transformations.
With PdD, text augmentation has never been this exciting. Whether you're testing the robustness of AI models, mimicking human typos, or diving into linguistic experiments, PdD is here to help you explore the untapped potential of character-level text transformations.
Ready to revolutionize the way you handle text? Let’s perturb some characters! 🔥