This project explores the application of style transfer techniques in natural language processing. The goal is to transform the style of a given text while preserving its content. This can be useful in various applications such as text generation, sentiment transfer, and more.

This codebase uses uv, which can be installed on POSIX systems with,

curl -LsSf https://astral.sh/uv/install.sh | sh

(See uv for more details.)

Install dependencies with,

uv sync
uv pip install -e .

• Generate formal/informal text dataset

• Linear-probe evaluation of where a simple model is learning style (linear decoding of hidden layers to predict style class)

• Manipulate architecture to include a style latent
  • Freeze the entire model except for the style latent when fine-tuning, and train the style latent to predict the style class
• Demonstrate that manipulating the style latent can change the style of the text generated
• Evaluate the performance of the style transfer model with formality classification models
  • ...and human evaluation?