Parrot Paraphraser

Parrot is designed to accelerate training for Natural Language Understanding (NLU) models. Unlike standard paraphrasers, a good NLU augmentor must preserve intents and slots/entities.

In a typical NLU augmentation workflow:

1. An input utterance with annotations (intents and slots) is provided.
2. The augmentor generates $N$ paraphrases while attempting to preserve the original meaning and slot positions.
3. The output paraphrases are converted back into annotated data (e.g., Rasa format) by mapping the original slots to the new text.

Parrot is optimized for conversational interfaces, meaning it is trained on text samples with a maximum length of 32 tokens.

To use Parrot, initialize the Parrot class with a model_tag (e.g., prithivida/parrot_paraphraser_on_T5). Use the .augment() method to generate paraphrases for a given input phrase. Note that you should initialize the model only once in your application lifecycle.

from parrot import Parrot
import torch
import warnings
warnings.filterwarnings("ignore")

# Init models (make sure you init ONLY once if you integrate this to your code)
parrot = Parrot(model_tag="prithivida/parrot_paraphraser_on_T5")

phrases = ["Can you recommend some upscale restaurants in Newyork?"]

for phrase in phrases:
  print("Input_phrase: ", phrase)
  para_phrases = parrot.augment(input_phrase=phrase, use_gpu=False)
  for para_phrase in para_phrases:
   print(para_phrase)

If you are using an AMD GPU on Linux (tested on Ubuntu 22.04 with Radeon RX 6650 XT), you can install the project using the ROCm-specific requirements.

Follow these steps:

1. Clone the repository.
2. Install dependencies from requirements-rocm.txt.

To verify the installation, check if PyTorch can detect the GPU. If it returns False, you may need to set the HSA_OVERRIDE_GFX_VERSION environment variable to a compatible version (e.g., 10.3.0) to allow the ROCm driver to function correctly.

git clone https://github.com/PrithivirajDamodaran/Parrot_Paraphraser.git
cd Parrot_Paraphraser
pip install -r requirements-rocm.txt

# Verify installation
python3 -c 'import torch; print(torch.cuda.is_available())'

# If verification returns False, try:
export HSA_OVERRIDE_GFX_VERSION=10.3.0
python3 -c 'import torch; print(torch.cuda.is_available())'

Install the Parrot library directly from the GitHub repository using pip.

pip install git+https://github.com/PrithivirajDamodaran/Parrot_Paraphraser.git

When using Parrot, be aware of the following current limitations:

• Diversity Scores: Diversity scores are not normalized; different diversity rankers score paraphrases differently.
• Command Style Inputs: Some command-style input phrases may generate less adequate paraphrases.

If you use Parrot in your work, please cite it using the following BibTeX reference:

@misc{prithivida2021parrot,
  author       = {Prithiviraj Damodaran},
  title        = {Parrot: Paraphrase generation for NLU.},
  year         = 2021,
  version      = {v1.0}
}