sonqt / unanswerable-robustness

The Impacts of Unanswerable Questions on the Robustness of Machine Reading Comprehension Models (EACL 2023)

The Impacts of Unanswerable Questions on the Robustness of Machine Reading Comprehension Models

Introduction

This repository contains the source code for the architectures described in the following paper:

The Impacts of Unanswerable Questions on the Robustness of Machine Reading Comprehension Models
EACL 2023
Son Quoc Tran, Phong Nguyen-Thuan Do, Uyen Le, Matt Kretchmar
Computer Science Department, Denison University, Granville, Ohio
The UIT NLP Group, Vietnam National University, Ho Chi Minh City

1. Getting Started

Installing Java

Please refer to Java Download

Running StanfordCoreNLPServer

For further information, refer to CoreNLP

1. Download file stanford-correnlp-latest.zip
2. Unzip file

cd stanford-corenlp-4.4.0

Start stanfordCoreNLPServer

java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -annotators "tokenize,ssplit,pos,parse" -port 9000 -timeout 30000

2. Attack

Nearest word with Glove

cd src

Download Glove and use glove.6B.100d.txt

Find nearby words for words in dataset

python3 find_nearby_words.py

Attack

python3 attack_main.py

Examples of Adversarial Attack on Answerable and Unanswerable Questions:

Question Types	Question	Attacked Context	Answer
Answerable	What desert is to the south near Arizona?	To the east is the Colorado Desert and the Colorado River at the border with Arizona, and the Mojave Desert at the border with the state of Nevada. To the south is the Mexico–United States border. Sea is the name of the water body that is found to the west.	Colorado River
Unanswerable	What desert is to the south near Arizona?	To the east is the Colorado Desert and the Colorado River at the border with Arizona, and the Mojave Desert at the border with the state of Nevada. To the south is the Mexico–United States border. The desert ofedmonton desert is to the north near Burbank.

Examples of Negation Attack:

Question	In the effort of maintaining a level of abstraction, what choice is typically left independent?
Answer	encoding
Context	Even though some proofs of complexity theoretic theorems regularly assume some concrete choice of input encoding, one tries to keep the discussion abstract enough to be independent of the choice of encoding. [...] In the effort of maintaining a level of abstraction, base64 choice is typically left not independent.

Citation and Contact

If you found this repository helpful, please cite:

@inproceedings{tran-etal-2023-impacts,
    title = "The Impacts of Unanswerable Questions on the Robustness of Machine Reading Comprehension Models",
    author = "Tran, Son Quoc  and
      Do, Phong Nguyen-Thuan  and
      Le, Uyen  and
      Kretchmar, Matt",
    booktitle = "Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.eacl-main.113",
    pages = "1543--1557",
    abstract = "Pretrained language models have achieved super-human performances on many Machine Reading Comprehension (MRC) benchmarks. Nevertheless, their relative inability to defend against adversarial attacks has spurred skepticism about their natural language understanding. In this paper, we ask whether training with unanswerable questions in SQuAD 2.0 can help improve the robustness of MRC models against adversarial attacks. To explore that question, we fine-tune three state-of-the-art language models on either SQuAD 1.1 or SQuAD 2.0 and then evaluate their robustness under adversarial attacks. Our experiments reveal that current models fine-tuned on SQuAD 2.0 do not initially appear to be any more robust than ones fine-tuned on SQuAD 1.1, yet they reveal a measure of hidden robustness that can be leveraged to realize actual performance gains. Furthermore, we find that robustness of models fine-tuned on SQuAD 2.0 extends on additional out-of-domain datasets. Finally, we introduce a new adversarial attack to reveal of SQuAD 2.0 that current MRC models are learning.",
}

Please contact Son Quoc Tran at [email protected] if you have any questions.