CODAH: An Adversarially-Authored Question Answering Dataset for Common Sense

Michael Chen, Mike D’Arcy, Alisa Liu, Jared Fernandez, Doug DowneyNAACL 2019

TL;DR

This paper introduces CODAH, a new dataset for common sense question answering designed to challenge state-of-the-art AI models. By having humans create questions that specifically target model weaknesses, CODAH reveals a significant gap between current AI performance and human understanding, even on datasets previously considered challenging.

What Problem Does This Paper Address?

Building AI systems that can reason with common sense is a long-standing challenge. While large language models have achieved impressive results on existing benchmarks like SWAG, they often exploit dataset "shortcuts" rather than truly understanding common sense. This means that even when models perform well, they don't necessarily possess human-level common sense. The authors aim to address this by creating a more robust benchmark that forces models to demonstrate genuine common sense reasoning, moving beyond superficial pattern matching.

Methods

The core innovation is the adversarial question generation process. Instead of simply collecting questions, human annotators were given feedback on how a state-of-the-art question-answering model (BERT) performed on their submissions. They were incentivized to create questions that the model failed to answer correctly, both before and after fine-tuning. This "human-in-the-loop" adversarial approach ensures that the generated questions are genuinely difficult for current AI systems, as they are specifically designed to exploit model weaknesses.

Datasets

  • SWAG (Situations With Adversarial Generations): Used as a baseline and for pre-training models. SWAG consists of sentence completion questions based on video observations.
  • CODAH (Commonsense Dataset Adversarially-authored by Humans): The newly introduced dataset. It's a challenging extension of SWAG, containing 2,801 multiple-choice sentence completion questions.

Key Insights and Ideas

  • Adversarial Authoring is Key: The most significant insight is that having humans actively target model weaknesses is a powerful way to create challenging datasets. This moves beyond simply collecting data to actively "breaking" existing systems.
  • Current Models Exploit Data, Not Common Sense: The large performance gap between humans and models on CODAH, even after fine-tuning on SWAG, strongly suggests that models are not truly reasoning with common sense but rather exploiting dataset regularities.
  • Quantitative and Negation Reasoning are Hard: Specific categories of common sense reasoning, like quantitative comparisons and understanding negation, proved particularly difficult for the tested models.
  • Artifacts Don't Explain Everything: The authors found that even when models are trained on data without obvious artifacts, they still struggle with CODAH, further supporting the idea that the difficulty lies in genuine common sense.

Key Results

  • Human Accuracy: Humans achieve a remarkable 95.3% accuracy on CODAH.
  • Model Performance: The best performing models (OpenAI GPT-1 and BERT), even when fine-tuned on SWAG, achieve accuracies of only 65.3% and 64.5% respectively on CODAH. This represents a significant drop from their performance on SWAG.
  • Error Reduction: Human performance on CODAH shows a 7-fold reduction in error compared to the fine-tuned BERT model.
  • Dataset Size Impact: While CODAH is smaller than SWAG, experiments suggest that the difficulty stems from the nature of the questions rather than just size. When comparing subsets of equal size, CODAH still presents a significant challenge, especially for BERT.

Limitations and Future Work

  • Dataset Size: CODAH is relatively small (2,801 questions). The authors suggest that aggregating diverse, independently-gathered datasets in a similar format could lead to more challenging benchmarks.
  • Dynamic Adversarial Collection: For future work, they propose updating the adversarial collection system as new models emerge, ensuring contributors always target the strongest systems. This might involve adjusting reward structures to incentivize fooling the latest models.
  • Deeper Analysis of Categories: While they identified categories like Quantitative and Negation as difficult, further detailed analysis into why models struggle with these specific types of reasoning is needed.
  • Exploring Other Reasoning Types: The authors suggest that identifying and categorizing more distinctive types of common sense reasoning would be beneficial for directing future research.

Summary updated

BibTeX
@inproceedings{chen2019codah,
  author = {Michael Chen and Mike D’Arcy and Alisa Liu and Jared Fernandez and Doug Downey},
  title = {{CODAH: An Adversarially-Authored Question Answering Dataset for Common Sense}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure