Authors' implementation of "Adversarial Over-Sensitivity and Over-Stability Strategies for Dialogue Models" (CoNLL 2018) in TensorFlow (the code was built upon TF 1.4, but any version later than that should also work). Note that if you use any Python version <= 3.5, you will need to manually change all formatted string literals.
Includes code for adversarial testing and training of the VHRED model, adversarial testing datasets, as well as pretrained checkpoints for normal and adversarially-trained models.
Authors: Tong Niu, Mohit Bansal
(1) Download the preprocessed version of Ubuntu Dialogue Corpus, put the three raw .txt files under data/raw/, and put the vocab file Dataset.dict.pkl under data/).
(2) To generate the raw adversarial data for each strategy, run
python3 src/basic/generate_raw_adv.py [strategy]
where "strategy" is one of "normal_inputs", "random_swap", "stopword_dropout", "paraphrase", "grammar_errors", "add_not" and "antonym". We have also released the test set purturbed by each strategy under data/raw/. Note that to generate raw data for the Data-level Paraphrasing strategy, you need to download the small version of PPDB2.0 dataset and put it under data/; to generate raw data for Grammar Errors, you need to download the AESW dataset and put it under data/.
Note that raw adversarial data for the Generative-level Paraphrasing strategy need to be generated separately by downloading the implementation of the Pointer-Generator model by See et al. (2017), and training on the ParaNMT-5M dataset. After training, feed the contexts of the Ubuntu dataset to the trained model, and put the outputs under data/raw/.
(3) To index any raw data, run
python3 src/basic/preprocess.py [strategy]
where "strategy" is one of "normal_inputs", "random_swap", "stopword_dropout", "paraphrase", "generative_paraphrase", "grammar_errors", "add_not" and "antonym".
(4) To train the VHRED model by Serban et al. 2016 from scratch, run
python3 src/main.py --batch_size [batch_size]
(5) To test the VHRED model on the normal input and each adversarial strategy, run
python3 src/main.py --test_strategy [strategy] --start_epoch [start_epoch] --batch_size [batch_size]
where "strategy" is one of "normal_inputs", "random_swap", "stopword_dropout", "paraphrase", "generative_paraphrase", "grammar_errors", "add_not" and "antonym", and "start_epoch" is the epoch to restore from.
(6) To obtain the F1 scores for outputs generated based on the normal input and each adversarial strategy, use the evaluation code by Serban et. al. (2018) on the output generated by testing on each adversarial strategy.
(7) Expected Results:
The model should obtain the following results ("N" stands for "Normal" and "A" stands for "Adversarial"):
(1) To train on each adversarial strategy, run
python3 src/main.py --train_strategy [strategy] --start_epoch [start_epoch] --batch_size [batch_size]
where "strategy" is one of "normal_inputs", "random_swap", "stopword_dropout", "paraphrase", "generative_paraphrase", "grammar_errors", "all_should_not_change", "add_not" and "antonym", and "start_epoch" is the epoch to restore from.
(2) To test on the trained adversarial model, run
python3 src/main.py --train_strategy [strategy] --test_strategy [strategy] --start_epoch [start_epoch] --batch_size [batch_size]
(1) The adversarial data for each strategy is provided under data/raw/
(2) Alternatively, you can also download our pre-trained checkpoints (for the vhred model, as well as for each adversarially trained model) and put them under ckpt/ to reproduce the F1 results.
If you happen to use our work, please consider citing our paper.
@inproceedings{niu2018adversarial,
author = {Niu, Tong and Bansal, Mohit},
title = {Adversarial Over-Sensitivity and Over-Stability Strategies for Dialogue Models},
booktitle = {The SIGNLL Conference on Computational Natural Language Learning (CoNLL)},
year = {2018},
}