Skip to content
No description, website, or topics provided.
Python Roff C Perl HTML Shell Other
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
code
.gitignore
LICENSE.txt
NOTICE.txt
README.md
top250.pdf
top250.png

README.md

Text Processing Like Humans Do: Visually Attacking and Shielding NLP Systems

This repository contains selected code and data for our NAACL 2019 long paper on Visually Attacking and Shielding NLP Systems.

Citation

@inproceedings{Eger:2019:NAACL,
            title = {Text Processing Like Humans Do: Visually Attacking and Shielding NLP Systems},
            year = {2019},
          author = {Steffen Eger and G{\"o}zde G{\"u}l {\c S}ahin and Andreas R{\"u}ckl{\'e} and Ji-Ung Lee and Claudia Schulz and Mohsen Mesgar and Krishnkant Swarnkar and Edwin Simpson and Iryna Gurevych},
       booktitle = {Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics},
           month = {Februar},
         journal = {NAACL 2019},
           pages = {1634--1647},
           url = {https://www.aclweb.org/anthology/N19-1165}
}

Abstract: Visual modifications to text are often used to obfuscate offensive comments in social media (e.g., ''!d10t'') or as a writing style (''1337'' in ''leet speak''), among other scenarios. We consider this as a new type of adversarial attack in NLP, a setting to which humans are very robust, as our experiments with both simple and more difficult visual perturbations demonstrate. We investigate the impact of visual adversarial attacks on current NLP systems on character-, word-, and sentence-level tasks, showing that both neural and non-neural models are, in contrast to humans, extremely sensitive to such attacks, suffering performance decreases of up to 82%. We then explore three shielding methods---visual character embeddings, adversarial training, and rule-based recovery---which substantially improve the robustness of the models. However, the shielding methods still fall behind performances achieved in non-attack scenarios, which demonstrates the difficulty of dealing with visual attacks.

Contact persons: Steffen Eger, eger@ukp.informatik.tu-darmstadt.de, Gözde Gül Sahin (sahin@ukp...), Andreas Rückle (rueckle@ukp...), Ji-Ung Lee (lee@ukp...), ...

https://www.ukp.tu-darmstadt.de/

https://www.tu-darmstadt.de/

Don't hesitate to send us an e-mail or report an issue, if something is broken (and it shouldn't be) or if you have further questions.

Project Description

Embeddings

We created diverse embeddings, including 576 dimensional visual Character Embeddings (called vce.normalized in the folder below). All embeddings and weights can be found on our public file server.

Below we include a tsne visualization of the visual character representations from vce.normalized (restricted to the first 250 chars).

Additional Resources

  • VIPER is our visual text perturber. See code/VIPER
  • Our G2P model is described in code/G2P
  • The code for our POS Tagging and Chunking experiments is available in code/POS_Chunk
  • The code to run the Toxic Comment Classification experiments can be found in code/TC

Semantic Similarity Sanity Check

We additionally performed a semantic similarity "sanity check" to check whether VELMo assigns higher cosine similarity to a pair of (clean,perturbed) sentences than SELMo. The pairs are of the form:

Greetings Hey douche bag wassup Grëȩtinɋs Ηey ḏoʋchḛ bag ✿ąssɥp

The code can be found in code/sanity_check

AllenNLP modifications

To run the visually informed models, we did some modifications on the AllenNLP code. The modified versions can be found in code/AllenNLP_Modifications and are necessary to obtain the respective embedding representations for toxic comment classification, chunking, and POS tagging.

You can’t perform that action at this time.