Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yichen Jiang, Mohit Bansal

TL;DR

This work reveals that HotpotQA contains reasoning shortcuts that enable single-hop answers via word-matching, and demonstrates that standard QA models suffer when those shortcuts are removed with adversarial documents. It introduces AddDoc adversaries to induce multi-hop requirements and evaluates both regular and adversarial training, showing substantial robustness gaps in baseline models. To address this, the authors propose a 2-hop compositional model with a recurrent control unit and bridge-entity supervision, which improves robustness to adversaries and, with adversarial training, outperforms the adversarially-trained single-hop baseline. The results motivate developing models that integrate explicit compositional reasoning with adversarial training to foster true multi-hop inference and better dataset design.

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.

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

TL;DR

This work reveals that HotpotQA contains reasoning shortcuts that enable single-hop answers via word-matching, and demonstrates that standard QA models suffer when those shortcuts are removed with adversarial documents. It introduces AddDoc adversaries to induce multi-hop requirements and evaluates both regular and adversarial training, showing substantial robustness gaps in baseline models. To address this, the authors propose a 2-hop compositional model with a recurrent control unit and bridge-entity supervision, which improves robustness to adversaries and, with adversarial training, outperforms the adversarially-trained single-hop baseline. The results motivate developing models that integrate explicit compositional reasoning with adversarial training to foster true multi-hop inference and better dataset design.

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.

Paper Structure

This paper contains 30 sections, 5 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Adversarial Evaluation
  3. The HotpotQA Task
  4. Adversary Construction
  5. Models
  6. Encoding
  7. Single-Hop Baseline
  8. Compositional Attention over Question
  9. Bridge Entity Supervision
  10. Experimental Setup
  11. Adversarial Evaluation and Training
  12. Dataset and Metrics
  13. Training Details
  14. Results
  15. Regularly-Trained Models
  16. ...and 15 more sections

Figures (4)

  • Figure 1: HotpotQA example with a reasoning shortcut, and our adversarial document that eliminates this shortcut to necessitate multi-hop reasoning.
  • Figure 2: An illustration of our AddDoc procedure. In this example, the keyword "headquarter" appears in no distractor documents. Thus the reader can easily infer the answer by looking for this keyword in the context.
  • Figure 3: A 2-hop bi-attention model with a control unit. The Context2Query attention is modeled as in seo2016bidaf. The output distribution $cv$ of the control unit is used to bias the Query2Context attention.
  • Figure 4: A single-hop HotpotQA example that cannot be fixed with our adversary.