[Paper Review] Avoiding Reasoning Shortcuts: Adversarial Evaluation, Training, and Model Development for Multi-Hop QA

Yichen Jiang, Mohit Bansal|arXiv (Cornell University)|Jun 17, 2019

Topic Modeling26 references4 citations

TL;DR

This paper identifies reasoning shortcuts in the HotpotQA dataset that allow models to answer multi-hop questions via single-hop word matching instead of true compositional reasoning. It proposes adversarial document generation to eliminate these shortcuts, introduces a control unit for dynamic question-guided multi-hop attention, and demonstrates that models with this architecture achieve significantly higher robustness and performance on adversarial evaluations compared to baselines.

ABSTRACT

Multi-hop question answering requires a model to connect multiple pieces of evidence scattered in a long context to answer the question. In this paper, we show that in the multi-hop HotpotQA (Yang et al., 2018) dataset, the examples often contain reasoning shortcuts through which models can directly locate the answer by word-matching the question with a sentence in the context. We demonstrate this issue by constructing adversarial documents that create contradicting answers to the shortcut but do not affect the validity of the original answer. The performance of strong baseline models drops significantly on our adversarial evaluation, indicating that they are indeed exploiting the shortcuts rather than performing multi-hop reasoning. After adversarial training, the baseline's performance improves but is still limited on the adversarial evaluation. Hence, we use a control unit that dynamically attends to the question at different reasoning hops to guide the model's multi-hop reasoning. We show that this 2-hop model trained on the regular data is more robust to the adversaries than the baseline model. After adversarial training, this 2-hop model not only achieves improvements over its counterpart trained on regular data, but also outperforms the adversarially-trained 1-hop baseline. We hope that these insights and initial improvements will motivate the development of new models that combine explicit compositional reasoning with adversarial training.

Motivation & Objective

• To expose and evaluate reasoning shortcuts in the HotpotQA dataset that allow models to bypass true multi-hop reasoning.
• To develop adversarial documents that eliminate these shortcuts by introducing conflicting answers while preserving semantic plausibility.
• To improve model robustness through adversarial training and introduce a control unit for dynamic, question-guided multi-hop attention.
• To demonstrate that explicit compositional reasoning mechanisms outperform standard models in adversarial settings.

Proposed method

• Adversarial document generation by perturbing answer spans and document titles to create conflicting answers that mislead shortcut-exploiting models.
• Construction of an adversarial dev set (adv-dev) and training set (adv-train) by replacing key entities in supporting documents with semantically plausible but incorrect alternatives.
• Design of a 2-hop model with a recurrent control unit that dynamically attends to relevant question words at each reasoning step to guide bi-attention mechanisms.
• Integration of supporting fact supervision during training to encourage evidence grounding and reduce reliance on surface-level matching.
• Adversarial training of both baseline and 2-hop models on the generated adversarial data to improve robustness.
• Evaluation of model performance on both standard and adversarial splits to measure generalization and reasoning robustness.

Experimental results

Research questions

• RQ1 To what extent do state-of-the-art models in HotpotQA rely on reasoning shortcuts rather than true multi-hop reasoning?
• RQ2 How effective are adversarial documents in exposing and measuring models' vulnerability to shortcut exploitation?
• RQ3 Can adversarial training improve model robustness on examples that eliminate reasoning shortcuts?
• RQ4 Does a control unit that dynamically guides attention over question words improve multi-hop reasoning performance?
• RQ5 Can a 2-hop model with explicit reasoning control outperform a 1-hop baseline after adversarial training?

Key findings

• The performance of a strong baseline model (Bi-attention + Self-attention) dropped significantly on the adversarial dev set, indicating heavy reliance on reasoning shortcuts.
• Supporting fact supervision improved the baseline's performance on the adversarial set, showing that evidence grounding enhances robustness.
• Adversarial training improved the baseline’s performance on the adversarial set, but it remained significantly below performance on the standard dev set.
• The proposed 2-hop model with a control unit outperformed the baseline on the adversarial evaluation when trained on regular data, demonstrating improved robustness.
• After adversarial training, the 2-hop model achieved better performance on the adversarial dev set than the adversarially-trained 1-hop baseline, indicating the effectiveness of explicit compositional reasoning.
• The 2-hop model with adversarial training showed further improvements, confirming that combining explicit reasoning with adversarial training enhances robustness.