This repository contains code for compositional perturbations as described in the following paper:
Tailor: Generating and Perturbing Text with Semantic Controls
Alexis Ross*, Tongshuang Wu*, Hao Peng, Matthew E. Peters, Matt Gardner Association for Computational Linguistics (ACL), 2022
Bibtex for citations:
@inproceedings{ross-etal-2022-tailor,
title = "Tailor: Generating and Perturbing Text with Semantic Controls",
author = "Ross, Alexis and
Wu, Tongshuang and
Peng, Hao and
Peters, Matthew E and
Gardner, Matt",
booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics",
month = aug,
year = "2022",
address = "Online",
publisher = "Association for Computational Linguistics",
}From Pypi:
pip install tailor_nlpFrom source:
git clone https://github.com/allenai/tailor.git
cd tailor
pip install -e .See link for information on how to format Ontonotes 5.0 and train the Tailor generator.
See link for the data. (More information in Section 5 of the paper.)
See link for the perturbations used in our NLI data augmentation experiments. (More information in Section 6 of the paper.)
- See the tutorial notebook for a detailed walkthrough of the API.
- See the documents in the main Python file for more explanations.
- See Tutorial 02 to learn how to use the default perturbation function on NLI data.
- See Tutorial 03 to learn how to define a customized perturbation function for MATRES data.
# initiate a wrapper.
from tailor import Tailor
tl = Tailor()
text = "In the operation room, the doctor comforted the athlete."
# 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 = tl.perturb(text)
# return: [
# 'the athlete was comforted by the doctor .',
# 'In which case , the doctor comforted the athlete.',]To perturb with more controls,
perturbations = tl.perturb(
sentence=text,
selected_span = "In the operation room",
# can filter perturbations by their change type, as printed above.
allowed_perturbs=["change_content"],
# can reuse the detected strategies
candidate_inputs = perturb_strategies,
# filter out degeneration with gpt-2 perplexity score. If None, then this step is skiped.
perplex_thred=50,
# max number of perturbations to return.
num_perturbs=10
)
# return: ["In case of an injury , the doctor 's comforted the athlete.",
# "In case of a fatal accident , the doctor 's comforted the athlete.",
# "In case of a bruised hand , the doctor 's comforted the athlete."]To attach additional context,
tl.perturb_with_context(
"In the operation room, the doctor comforted the athlete.",
"In the operation room",
to_content="bridge",
verbalize=True
)
# return: ["Under the bridge , the doctor 's comforted the athlete.",
# "Under a bridge , the doctor 's comforted the athlete."]
tl.perturb_with_context(
"In the operation room, the doctor comforted the athlete.",
"In the operation room",
to_semantic_role="TEMPORAL",
verbalize=True
)
# return: ['When the doctor came into the operation room , the physician comforted the athlete.',
# "While the doctor was in the operation room , the physician 's comforted the athlete."]
tl.perturb_with_context(
"In the operation room, the doctor comforted the athlete.",
"comforted",
to_tense="future",
verbalize=True
)
# return: ['In the operation room , the doctor will comfort the athlete.',
# "In the operation room , the doctor 's will comfort the athlete."]