hiphop2311 / Making-Adversarial-perturbation-with-Text-classification

Making-Adversarial-perterbation-for-Text-classification

Dataset

• Consists of 2225 documents from the BBC news website corresponding to stories in five topical areas from 2004-2005.
• Class Labels: 5 (business, entertainment, politics, sport, tech)

Classification model using

• Multinomial Naive Baye

Preprocessing

• Convert all letter to lower case
• Remove punctuation
• Tokenize word
• Remove stopwords
• Remove stopwords
• Lemmatize

Make perturbation

• Random swap characters in each significant word collected from training Corpus (others are applied with fix probability )

• Change some characters to a similar one.

• Dropout some frequent words.

Run

python3 controller.py

Result

Input	Accuracy	F1
News report	0.964	0.963
News report with perturbation	0.424	0.396

Example program