This repository contains the data collected for our EMNLP 2021 paper Contrasting Human- and Machine-Generated Word-Level Adversarial Examples for Text Classification.
The human- and machine-generated adversarial examples are in the file collected_adversarial_examples.csv. The file contains 1020 rows, representing the 170 sequences unperturbed and perturbed with each of the 5 attacks.
The columns are as follows:
- id: the sequence ID, which also identifies the attack used (or no attack)
- text: the corresponding text
- succ: whether the adversarial examples successfully flipped the classifier label
- label: the actual ground truth label of the sequence
- num_queries: the number of queries needed to generate the adversarial example
- sub_rate: the word substitution rate
The raw collected data from the crowdsourcing experiments corresponding to Task 4 of the first data collection stage (see Section 3.1 in the paper) can be found in task_4.json.
The collected ratings for each generated adversarial example can be found in ratings.json. For each rated text, the JSON provides the total amount of ratings for both naturalness and sentiment. For both criteria, the ratings are on a scale from 1 (very negative sentiment/very unnatural) to 5 (very positive sentiment/very natural).
If you find this repository useful, please consider citing our paper:
@inproceedings{mozes-etal-2021-contrasting,
title = "Contrasting Human- and Machine-Generated Word-Level Adversarial Examples for Text Classification",
author = "Mozes, Maximilian and
Bartolo, Max and
Stenetorp, Pontus and
Kleinberg, Bennett and
Griffin, Lewis",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.651",
doi = "10.18653/v1/2021.emnlp-main.651",
pages = "8258--8270",
}