Unsupervised Text Style Transfer for Controllable Intensity

Quick Links

• Overview
• Requirements
• Data Preparation
• Experiments
• Results
• Citation

Overview

This repository contains the implementation of Unsupervised Text Style Transfer for Controllable Intensity, a framework for controlling the intensity of style transfer in text generation tasks. The project focuses on two main style transfer tasks:

• Readability Transfer: Adjusting text complexity and readability levels
• Sentiment Transfer: Modifying sentiment polarity with controllable intensity

The framework supports both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) approaches using the trlx library.

Figure 1: Overview of the proposed method for controllable intensity text style transfer

Model Architecture

Figure 2: Model architecture

Requirements

• Python >= 3.8
• PyTorch 2.0.1 with CUDA 11.8
• 2+ GPUs recommended for distributed training

Installation

# Create conda environment (recommended)
conda create -n readability_summ python=3.8
conda activate readability_summ

# Install PyTorch with CUDA 11.8
pip install torch==2.0.1+cu118 -f https://download.pytorch.org/whl/cu118/torch_stable.html

# Install dependencies
pip install -r requirements.txt

Main dependencies include:

• Transformers 4.46.3
• trlx 0.7.0
• DeepSpeed (for distributed training)
• Various NLP evaluation libraries (bert_score, textstat, etc.)

Environment Variables

The following environment variables can be used to configure the project:

Variable	Description	Default
SFT_MODEL_DIR	Path to the SFT checkpoint for RL training	checkpoints/sft_readability
OPENAI_API_KEY	API key for GPT-based data generation	(empty)
OPENAI_BASE_URL	Base URL for the OpenAI-compatible API	https://api.openai.com/v1/
TRAIN_FILE	Path to training data	data/train_summary_prompt_parallel.json
VAL_FILE	Path to validation data	data/val_summary_prompt_parallel.json

Data Preparation

Readability

The data should be prepared in JSON format with parallel text pairs. Expected format:

{
  "input_text_prompt": "source text with prompt",
  "output_text": "target text with desired readability"
}

Place your data files in the data/ directory:

• train_summary_prompt_parallel.json - Training data
• val_summary_prompt_parallel.json - Validation data

You can use the preprocessing script to generate readability-controlled data:

# Set your API key first
export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.openai.com/v1/"

python src/preprocess/generate_readability_by_gpt.py

Sentiment

Similar data preparation process for sentiment transfer tasks. Prepare parallel data with source and target sentiment levels.

Experiments

Training

1. Supervised Fine-Tuning (SFT)

Train the model using supervised fine-tuning:

bash scripts/train_sft_readability.sh

Key parameters (configurable via environment variables or script editing):

• Learning rate: 1e-4
• Batch size: 8 per device
• Max source length: 1024
• Training epochs: 20

2. Reinforcement Learning (RL)

Train with PPO reinforcement learning for better controllability:

# Set the SFT checkpoint path
export SFT_MODEL_DIR="checkpoints/sft_readability"

bash scripts/train_rl_readability.sh

Inference

Run inference on test data:

# SFT model inference
bash scripts/inference_sft_readability.sh

# RL model inference (runs all 4 readability levels in parallel)
bash scripts/inference_rl_readability.sh

Evaluation

The project includes various evaluation metrics:

• ROUGE scores for text quality
• BERTScore for semantic similarity
• Readability metrics: Flesch-Kincaid, Gunning Fog, Coleman-Liau
• Style similarity: TF-IDF based feature extraction and cosine similarity

Example Outputs

Figure 3: Examples of text style transfer with different intensity levels

Results

Figure 4: Experimental results - Part 1

Figure 5: Experimental results - Part 2

Project Structure

UTST-CI/
├── data/                # Training and validation data (not tracked)
├── figs/                # Figures and visualizations
├── scripts/             # Training and inference scripts
│   ├── train_sft_readability.sh
│   ├── train_rl_readability.sh
│   ├── inference_sft_readability.sh
│   └── inference_rl_readability.sh
├── src/
│   ├── utils/           # Shared utilities
│   │   ├── readability_utils.py   # Readability metrics and helpers
│   │   └── style_scorer.py        # TF-IDF style similarity scorer
│   ├── preprocess/      # Data preprocessing scripts
│   ├── train/           # Training scripts (SFT + RL)
│   └── inference/       # Inference scripts
├── requirements.txt
└── README.md

Citation

If you use this code in your research, please cite:

@article{gu2026unsupervised,
  title={Unsupervised Text Style Transfer for Controllable Intensity},
  author={Gu, Shuhuan and Tao, Wenbiao and Ma, Xinchen and He, Kangkang and Guo, Ye and Li, Xiang and Lan, Yunshi},
  journal={arXiv preprint arXiv:2601.01060},
  year={2026},
  url={https://arxiv.org/abs/2601.01060}
}

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.