Personality-Adapted Language Generation for Social Robots

Code for paper submission to RO-MAN 2023.

Repository structure

.
├── text_study_data                   # Includes all the data used, collected and analysis from the text-only study
├── video_study_data                  # Includes all the data used, collected and analysis from the video-based study
├── cross_study_data                  # Includes all the data used, and analysis done comparing the two studies
├── style_trasnfer_paraphrase         # Contains the paraphrase model for personality style transfer
│   ├── utils                         # Utilities
│   │   └── prepare_data.py           # Download and pre-process data
│   ├── style_transfer.py             # A wrapper to use the paraphrase model
│   ├── eval_personality_transfer.py  # A script to generate the transferred sentences
│   └── ...
├── emotions.py                       # A wrapper used to recognize emotions from text and express them on furhat
├── main.py                           # Connects to Furhat and executes the experiment
└── ...

Requirements

Python 3.7+. See requirements.txt for requirement on animating furhat with emotions.

Other specific requirements can be found in each subfolder. E.g. requirements for paraphrase model are in style_transfer_paraphrase/requirements.txt

Instructions for using STRAP personality transfer checkpoints

Start by cloning this repo.

git clone https://github.com/alessioGalatolo/PersonalityLanguageGeneration
cd PersonalityLanguageGeneration

Make sure pytorch (>1.9.0) is installed (see official documentation for the best way of doing this) in your working environment.

python3.7 -m venv env
source env/bin/activate
pip install torch torchvision torchaudio

Move to the style_transfer_paraphrase directory and install STRAP through pip

pip install --editable .

For use in inference the only other library needed is transformers, install with pip install transformers.

pip install transformers

Then go to the release page of this repo PersonalityLanguageGeneration/releases/ and download STRAP_essays.tar.gz. Extract it in the project folder.

wget https://github.com/alessioGalatolo/PersonalityLanguageGeneration/releases/download/v1.0.0/STRAP_essays.tar.gz
tar zxf STRAP_essays.tar.gz
rm STRAP_essays.tar.gz

Go to https://drive.google.com/drive/folders/1hB0lJt4MjuWbgdY7_I_2eISAoNmM6O9f and download the folder paraphraser_gpt2_large. Then extract it under style_transfer_paraphrase/paraphraser_gpt2_large (this has to be done manually).

Then use the interface in style_transfer directly inside python:

from style_transfer import PersonalityTransfer

to_extrovert = PersonalityTransfer("./essays/model_ext", "paraphraser_gpt2_large", top_p=0.6)
to_introvert = PersonalityTransfer("./essays/model_int", "paraphraser_gpt2_large", top_p=0.6)

# convert a short sentence
some_sentence = "Hello, this is an example sentence"
print("Example sentence is: ", some_sentence)
print("Extrovert version is: ", to_extrovert.transfer_style(some_sentence))
print("Introvert version is: ", to_introvert.transfer_style(some_sentence))

# for longer sentences we reccomend splitting them (e.g. using . ! and ?)
from re import findall
long_text = "This is a very long text. It contains multiple sentences and interesting facts! Did you know that all giant pandas in zoos around the world are on loan from China?"
print("Example sentence is: ", long_text)
print("Extrovert version is: ", to_extrovert.transfer_style("\n".join(findall(r'([^.?!]+[.?!])', long_text))))
print("Introvert version is: ", to_introvert.transfer_style("\n".join(findall(r'([^.?!]+[.?!])', long_text))))

NB: It is reccomended to run STRAP on a GPU, which is done by default. However, if you're GPU is not powerful enough you can still run it on the CPU by passing device='cpu' when initializing PersonalityTransfer:

from style_transfer import PersonalityTransfer

to_extrovert = PersonalityTransfer("./essays/model_ext", "paraphraser_gpt2_large", top_p=0.6, device='cpu')
to_introvert = PersonalityTransfer("./essays/model_int", "paraphraser_gpt2_large", top_p=0.6, device='cpu')

Acknowledgments

Part of this code was borrowed from martiansideofthemoon/style-transfer-paraphrase.