Adversarial NER Attacks

This is a demo website for the paper SeqAttack: On Adversarial Attacks for Named Entity Recognition. The video below demonstrates how the website works.

Check out SeqAttack on GitHub

Prediction

Predicts named entities in the input text using a bert-base-cased model fine-tuned on CoNLL2003. The recognized entities are person names (PER), locations (LOC), organizations (ORG) and miscellaneous (MISC)

Attack

Run an attack against a BERT-based model trained on CoNLL2003. Attacks are executed using our framework SeqAttack. The following attack strategies are supported:

DeepWordBug: character-level attack. Supports character insertions, deletions, replacements and swaps (ab → ba)
DeepWordBug (constrained): same as above, but the algorithm cannot attack named entities
BERT-Attack: word replacements using a bert-base-cased masked language model
SCPN: paraphrase generation

WARNING: BERT-Attack and SCPN may need a long compute time, thus attacks are limited to 30 seconds. After the timeout the attack will be considered failed.

Dataset Visualization

Visualize an attacked dataset. Sample datasets for various attack strategies can be downloaded at the links belows.
For each sample the top row contains the original text (with its ground truth prediction) and the bottom row its perturbed counterpart.
We recommend to reload the page before loading a new dataset

deepwordbug.json deepwordbug-constraint.json bert-attack.json scpn.json

Open Dataset