Robust Machine Comprehension Models via Adversarial Training

By Yicheng Wang, Mohit Bansal. 2018

125 citations recorded by the source corpus.

It is shown that many published models for the Stanford Question Answering Dataset (Rajpurkar et al., 2016) lack robustness, suffering an over 50% decrease in F1 score during adversarial evaluation based on the AddSent (Jia and Liang, 2017) algorithm. It has also been shown that retraining models on data generated by AddSent has limited effect on their robustness. We propose a novel alternative adversary-generation algorithm, AddSentDiverse, that significantly increases the variance within the adversarial training data by providing effective examples that punish the model for making certain superficial assumptions. Further, in order to improve robustness to AddSent’s semantic perturbations (e.g., antonyms), we jointly improve the model’s semantic-relationship learning capabilities in addition to our AddSentDiverse-based adversarial training data augmentation. With these additions, we show that we can make a state-of-the-art model significantly more robust, achieving a 36.5% increase in F1 score under many different types of adversarial evaluation while maintaining performance on the regular SQuAD task.

Subfields: Artificial Intelligence, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

ReCoRD: Bridging the Gap between Human and Machine Commonsense Reading Comprehension (2018, background): “Since it is hard to control the quality of crowdsourced questions, a large portion of questions in these datasets can be answered by word matching or paraphrasing (Jia and Liang, 2017; Rajpurkar et al., 2018; Wang and Bansal, 2018).”
Avoiding Reasoning Shortcuts: Adversarial Evaluation, Training, and Model Development for Multi-Hop QA (2019, methodology): “However, our study also differs from the previous works (Jia and Liang, 2017; Wang and Bansal, 2018) in two points.”
CLINE: Contrastive Learning with Semantic Negative Examples for Natural Language Understanding (2021, unknown): “…to adopt adversarial training on PLMs, which applies gradient-based perturbations to the word embeddings during training (Miyato et al., 2017; Zhu et al., 2020; Jiang et al., 2020) or adds high-quality adversarial textual examples to the training phase (Wang and Bansal, 2018; Michel et al., 2019).”
Adversarial Language Games for Advanced Natural Language Intelligence (2019, background): “(Zhang et al., 2019) aims at ﬁnding adversarial examples of NLP models to evaluate the model robustness (Jia and Liang, 2017; Michel et al., 2019) or to help to train a robust model (Wang and Bansal, 2018; Cheng et al., 2019b).”
There is Strength in Numbers: Avoiding the Hypothesis-Only Bias in Natural Language Inference via Ensemble Adversarial Training (2020, background): “Other semantics-preserving perturbations include typos (Hosseini et al., 2017), the addition of distracting sentences (Wang and Bansal, 2018; Jia and Liang, 2017), character-level perturbations (Ebrahimi et al., 2018), and paraphrasing (Iyyer et al., 2018).”
Adversarial Attacks on Deep Learning Models in Natural Language Processing: A Survey (2019, methodology): “[142] followed the previous works and also worked on SQuAD dataset.”
Beyond Leaderboards: A survey of methods for revealing weaknesses in Natural Language Inference data and models (2020, methodology): “…on these data acquire when evaluated on samples not drawn from the training distribution and (c) propose architectural (Sagawa et al., 2020) and training procedure (Wang and Bansal, 2018) improvements in order to achieve more robust generalisation beyond data drawn from the training distribution.”
Better Robustness by More Coverage: Adversarial Training with Mixup Augmentation for Robust Fine-tuning (2020, background): “Instead of adding continuous gradient perturbations, another type of AT directly adds discrete text adversarial examples into training (Jia and Liang, 2017; Wang and Bansal, 2018; Michel et al., 2019).”