anonymousrole / itda

Authorship Style Transfer with Inverse Transfer Data Augmentation

Authorship Style Transfer with Inverse Transfer Data Augmentation

This is the offficial implementation of the paper [Authorship Style Transfer with Inverse Transfer Data Augmentation].

Overview

Authorship style transfer aims to modify the style of neutral text to match the unique speaking or writing style of a particular individual. We propose an inverse transfer data augmentation ITDA method, leveraging GPT-3.5 to create (neutral text, stylized text) pairs. We use this augmented dataset to train a BART-base model adept at style transfer. Our experimental results, conducted across four datasets with distinct authorship styles, establish the effectiveness of ITDA over style transfer using GPT-3.5.

Evaluation Results

We evaluate ITDA on four benchmarks: Lin Daiyu, Shakespeare, Trump, Lyrics. We adopt four metrics: BLEU and BS (BERTScore) measure content preservation, SC measures style transfer strength, and GPT-4 measures overall performance. Since user-provided text often spans a range of topics, we also collect a new test set comprising neutral texts spanning diverse topics to do out-of-distribution evaluation.

Install the requirements

First, you need to create a virtual environment and activate it:

conda deactivate
conda create -n <env_name> python=3.8
conda activate <env_name>

Then, install the cuda version Pytorch：

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1

Finally, install the requirements:

conda install --file requirements.txt

Stylized Datasets

datasets/hlm, datasets/shakespeare, datasets/trump, datasets/lyrics

Training

(a) Cluster-based Demonstration Annotation

python kmeans.py

(b) Stylized Text Augmentation

python stylized_augmentation.py

(c) Inverse Transfer Data Augmentation

python dynamic_inverse_transfer.py

(d) Fine-tune a Compact Model

python ft_bart_en.py   #For English Datasets
python ft_bart_ch.py   #For Chinese Datasets

Inference

python bart_transfer.py

Classifier Training

python classifer_train_en.py  #For English Datasets
python classifer_train_ch.py  #For Chinese Datasets

Evaluation

python evaluation/eval_content.py (BLEU, BERTScore)
python evaluation/classifier_metrics_en.py (SC)   #For English Datasets
python evaluation/classifier_metrics_ch.py (SC)   #For Chinese Datasets
python evaluation/GPT4_judge.py (GPT-4 Score)

Fixed Prompting

Fixed Few-shot Prompting for Fowrad Transfer or Inverse Transfer with GPT-3.5

python fixed_transfer.py