Doing Good or Doing Right? Exploring the Weakness of Commonsense Causal Reasoning Models

Mingyue Han, Yinglin WangACL 2021

TL;DR

This paper investigates why pre-trained language models (PLMs) excel at commonsense causal reasoning tasks like COPA, yet struggle with true causal understanding. The authors reveal that PLMs often rely on superficial semantic similarity rather than genuine causal inference, leading them to fail under specific adversarial conditions. They propose a simple regularization technique that significantly improves model robustness and causal reasoning ability.

What Problem Does This Paper Address?

Pre-trained language models (PLMs) like BERT and its successors have achieved remarkable performance on various natural language understanding tasks, including the Choice of Plausible Alternatives (COPA) task. COPA requires models to choose between two alternatives (a cause or an effect) given a premise. While PLMs achieve high accuracy on COPA, the authors question whether this success stems from genuine causal reasoning or from exploiting superficial cues present in the data. Specifically, they highlight the problem of "semantic similarity bias," where models might pick the more semantically related alternative rather than the one that logically follows causally. This bias can be exploited by adversarial attacks, revealing a vulnerability in current models' ability to perform true causal inference.

Methods

The authors employ two main probing experiments to uncover the weaknesses of PLMs in causal reasoning:

  • Perturbation with Distractors (Exp1): They introduce a third, distractive alternative to existing COPA samples. This distractor is semantically related to the premise but is not a valid cause or effect. The goal is to see if models' performance drops significantly when faced with a semantically plausible but causally incorrect option. They also introduce a "COPA-random" baseline to control for the increased difficulty of a ternary choice.
  • Masking Question Type (Exp2): They investigate whether models rely on the explicit "ask-for" cue (e.g., "cause" or "effect") in the prompt. By randomly shuffling the order of the premise and alternatives, they effectively mask the question type. A robust model should maintain performance even without this explicit cue.

To address the identified semantic similarity bias, the authors propose a novel regularization loss added to the standard cross-entropy loss during training. This regularization encourages the model to be "neutral" when presented with reversed cause-effect pairs, meaning it should not strongly favor one alternative over the other if the input order is flipped, thus discouraging reliance on superficial cues.

Datasets

  • COPA (Choice of Plausible Alternatives): The standard dataset used for training and evaluating commonsense causal reasoning.
  • BCOPA: A dataset introduced by Kavumba et al. (2019) with an unbiased token distribution between correct and incorrect alternatives, designed to mitigate bias from token distribution.
  • BCOPA-CE: A new dataset introduced in this paper. It's constructed from the original COPA-test set by generating new, plausible alternatives for the opposite question type (e.g., if the original asked for an effect, they generate a plausible cause). This dataset aims to evaluate causal reasoning ability under unbiased token distribution and is designed to be challenging.
  • COPA-test hard: A challenging subset of COPA with unbiased token distribution.

Key Insights and Ideas

  • Semantic Similarity Bias is Pervasive: PLMs, despite high COPA scores, are heavily influenced by semantic similarity. They can be easily fooled by distractors that are semantically related but not causally correct.
  • Token Distribution Debiasing Isn't Enough: Augmenting training data with methods like BCOPA (which balances token distribution) doesn't fully solve the problem and can even amplify semantic bias in some cases.
  • Question Type Matters (for PLMs): PLMs are sensitive to the explicit "ask-for" cue. Masking this cue significantly degrades their performance, indicating they don't fully grasp the underlying causal relationship independent of prompt structure.
  • A Simple Regularization Fix: Adding a straightforward regularization loss that penalizes strong predictions when the cause-effect order is reversed proves effective in mitigating semantic bias and improving robustness.
  • Causal Intuition vs. Co-reference: Human causal reasoning prioritizes the causal link, while PLMs might overemphasize co-reference or other superficial relationships.

Key Results

  • Perturbation Experiments (Exp1): Introducing a premise as a distractor significantly degrades PLM performance, with accuracy dropping substantially (e.g., BERT-large drops from 69.5% to 47.6%). The CS (causal knowledge) baseline is much more robust.
  • Masking Experiments (Exp2): PLMs' performance drops significantly when the question type is masked, indicating reliance on explicit cues. The authors suggest a robust model's accuracy should be around 50% + half of the original accuracy in this setting.
  • Proposed Regularization: The proposed model with regularization (e.g., db-l-reg) shows improved performance on the challenging BCOPA-CE set (e.g., 86.3% vs. 72.3% for db-l) and the COPA-test hard set (e.g., 89.7% vs. 88.6% for db-l). It also outperforms previous debiased models on COPA-hard.
  • Error Analysis: In a specific case, the regularized DeBERTa model correctly identifies the causal link, while the original DeBERTa is misled by superficial cues, highlighting the effectiveness of the proposed method.

Limitations and Future Work

  • Limitations: The paper acknowledges that while their regularization improves robustness, PLMs still don't achieve perfect causal understanding. The performance on the original COPA test set is slightly lower for the regularized models, suggesting a trade-off between generalization on the original task and robustness to adversarial attacks. The authors also note that previous debiased models on token distribution can amplify semantic bias, raising questions about the unintended consequences of debiasing strategies.
  • Future Work: The authors suggest further exploring the nature of semantic similarity bias and developing more sophisticated methods to achieve true causal reasoning. They also imply that the BCOPA-CE dataset can be a valuable tool for future research in this area.

Summary updated

BibTeX
@inproceedings{han2021doing,
  author = {Mingyue Han and Yinglin Wang},
  title = {{Doing Good or Doing Right? Exploring the Weakness of Commonsense Causal Reasoning Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure