/polyjuice

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Polyjuice

This repository contains code for generating counterfactual sentences as described in the following paper:

Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models
Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, Daniel S. Weld Association for Computational Linguistics (ACL), 2021

Bibtex for citations:

@inproceedings{wu-etal-2021-polyjuice,
    title = "Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models",
    author = "Wu, Tongshuang  and
      Ribeiro, Marco Tulio  and
      Heer, Jeffrey  and
      Weld, Daniel",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.523",
    doi = "10.18653/v1/2021.acl-long.523",
    pages = "6707--6723",
}

Installation

From Pypi:

pip install polyjuice_nlp

From source:

git clone git@github.com:tongshuangwu/polyjuice.git
cd polyjuice
pip install -e .

Polyjuice depends on SpaCy and Huggingface Transformers. To use most functions, please also install the following:

# install pytorch, as here: https://pytorch.org/get-started/locally/#start-locally
pip install torch
# The SpaCy language package
python -m spacy download en_core_web_sm

Perturbation

from polyjuice import Polyjuice
# initiate a wrapper.
# model path is defaulted to our portable model:
# https://huggingface.co/uw-hai/polyjuice
# No need to change this unless you are using customized model
pj = Polyjuice(model_path="uw-hai/polyjuice", is_cuda=True)

# the base sentence
text = "It is great for kids."

# perturb the sentence with one line:
# When running it for the first time, the wrapper will automatically
# load related models, e.g. the generator and the perplexity filter.
perturbations = pj.perturb(text)

# return: ['It is bad for kids too.',
# "It 's great for kids.",
# 'It is great even for kids.']

More advanced APIs

Please see the documents in the main Python file for more explanations.

To perturb with more controls,

perturbations = pj.perturb(
    orig_sent=text,
    # can specify where to put the blank. Otherwise, it's automatically selected.
    # Can be a list or a single sentence.
    blanked_sent="It is [BLANK] for kids.",
    # can also specify the ctrl code (a list or a single code.)
    # The code should be from 'resemantic', 'restructure', 'negation', 'insert', 'lexical', 'shuffle', 'quantifier', 'delete'.
    ctrl_code="negation",
    # Customzie perplexity score. 
    perplex_thred=5,
    # number of perturbations to return
    num_perturbations=1,
    # the function also takes in additional arguments for huggingface generators.
    num_beams=3
)

# return: [
# 'It is not great for kids.', 
# 'It is great for kids but not for anyone.',
# 'It is great for kids but not for any adults.']

To detect ctrl code from a given sentence pair,

pj.detect_ctrl_code(
    "it's great for kids.", 
    "It is great for kids but not for any adults.")
# return: negation

To get randomly placed blanks,

random_blanks = py.get_random_blanked_sentences(
    sentence=text,
    # only allow selecting from a preset range of token indexes
    pre_selected_idxes=None,
    # only select from a subset of dep tags
    deps=None,
    # blank sub-spans or just single tokens
    is_token_only=False,
    # maximum number of returned index tuple
    max_blank_sent_count=3,
    # maximum number of blanks per returned sentence
    max_blank_block=1
)

Selection

For selecting diverse and surprising perturbations (for augmentation and explanation experiments in our paper), please see the notebook demo.