[Paper Review] What If We Simply Swap the Two Text Fragments? A Straightforward yet Effective Way to Test the Robustness of Methods to Confounding Signals in Nature Language Inference Tasks

Haohan Wang, Da Sun+1 more|arXiv (Cornell University)|Sep 7, 2018

Topic Modeling33 references3 citations

TL;DR

This paper proposes a simple yet effective method—swapping premise and hypothesis in NLI tasks—to test model robustness against confounding signals in training data. By evaluating model accuracy on swapped test sets, the authors reveal that models relying on statistical patterns (e.g., word distributions) fail to generalize, while those using semantic understanding maintain consistent performance, especially for entailment and contradiction pairs.

ABSTRACT

Nature language inference (NLI) task is a predictive task of determining the inference relationship of a pair of natural language sentences. With the increasing popularity of NLI, many state-of-the-art predictive models have been proposed with impressive performances. However, several works have noticed the statistical irregularities in the collected NLI data set that may result in an over-estimated performance of these models and proposed remedies. In this paper, we further investigate the statistical irregularities, what we refer as confounding factors, of the NLI data sets. With the belief that some NLI labels should preserve under swapping operations, we propose a simple yet effective way (swapping the two text fragments) of evaluating the NLI predictive models that naturally mitigate the observed problems. Further, we continue to train the predictive models with our swapping manner and propose to use the deviation of the model's evaluation performances under different percentages of training text fragments to be swapped to describe the robustness of a predictive model. Our evaluation metrics leads to some interesting understandings of recent published NLI methods. Finally, we also apply the swapping operation on NLI models to see the effectiveness of this straightforward method in mitigating the confounding factor problems in training generic sentence embeddings for other NLP transfer tasks.

Motivation & Objective

• Investigate confounding signals in NLI datasets, particularly word distribution biases in hypotheses that enable models to predict labels without true semantic understanding.
• Propose a simple, interpretable sanity check: swapping premise and hypothesis in test sets to evaluate whether models rely on semantic reasoning or superficial statistical cues.
• Develop a training protocol using varying percentages of swapped sentence pairs to train models and measure performance deviation as a proxy for robustness to confounding factors.
• Evaluate whether sentence embeddings trained with swapping generalize better to downstream NLP transfer tasks by mitigating confounding signal biases.

Proposed method

• Apply premise-hypothesis swapping to test sets to create a stress test: if a model is truly robust, its accuracy should remain stable for neutral and contradiction pairs, and drop for entailment pairs under swapping.
• Use propositional logic to justify the expected behavior: entailment (P → H) is not symmetric, while contradiction (P → ¬H) and neutrality (P ⊥ H) are symmetric under swap.
• Train NLI models on multiple training sets with increasing percentages (0% to 100%) of swapped sentence pairs to assess model stability under data distribution shifts.
• Define 'deviation' as the fluctuation in model performance across different swap percentages; lower deviation indicates higher robustness to confounding signals.
• Fine-tune sentence embeddings using the swapping-augmented training data and evaluate them on downstream NLP transfer tasks (e.g., textual similarity, natural language inference).
• Apply the swapping-based training to multiple SOTA NLI models (e.g., DGA, KIM, ADV) and compare their robustness via performance deviation metrics.

Experimental results

Research questions

• RQ1 Do state-of-the-art NLI models truly understand semantic relationships, or do they exploit statistical biases in word distributions of hypotheses?
• RQ2 Can premise-hypothesis swapping serve as a reliable, lightweight sanity check to detect models relying on superficial cues rather than semantic inference?
• RQ3 Does training NLI models on data with varying degrees of swapped sentence pairs lead to more robust representations that generalize better to downstream tasks?
• RQ4 To what extent does the deviation in model performance across different swap percentages correlate with robustness to confounding signals in NLI data?
• RQ5 Can swapping during training improve the quality of sentence embeddings for transfer learning in other NLP tasks?

Key findings

• Models like DGA and KIM show low performance deviation across different swap percentages, indicating strong robustness to confounding signals and likely reliance on semantic understanding.
• The ADV model exhibits higher deviation, suggesting sensitivity to data distribution shifts and possible exploitation of statistical cues, though it remains promising with further analysis.
• Sentence embeddings trained with 25%–75% of training pairs swapped achieve higher performance on downstream NLP transfer tasks, demonstrating improved generalization.
• Models that fail the swapping stress test—showing significant accuracy drop on swapped entailment pairs—likely rely on lexical or distributional patterns rather than logical inference.
• The swapping-based evaluation reveals that several SOTA models perform well on original test sets but degrade substantially under swap, indicating overfitting to spurious correlations.
• The proposed deviation metric effectively quantifies robustness, offering a practical tool for model selection and benchmarking beyond standard accuracy.