Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference (1902.01007v4)

Abstract: A machine learning system can score well on a given test set by relying on heuristics that are effective for frequent example types but break down in more challenging cases. We study this issue within natural language inference (NLI), the task of determining whether one sentence entails another. We hypothesize that statistical NLI models may adopt three fallible syntactic heuristics: the lexical overlap heuristic, the subsequence heuristic, and the constituent heuristic. To determine whether models have adopted these heuristics, we introduce a controlled evaluation set called HANS (Heuristic Analysis for NLI Systems), which contains many examples where the heuristics fail. We find that models trained on MNLI, including BERT, a state-of-the-art model, perform very poorly on HANS, suggesting that they have indeed adopted these heuristics. We conclude that there is substantial room for improvement in NLI systems, and that the HANS dataset can motivate and measure progress in this area

Diagnosing Syntactic Heuristics in NLI Models: An Analysis with HANS

The paper "Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference" presents a comprehensive analysis of the failure modes of statistical NLI models. Authors R. Thomas McCoy, Ellie Pavlick, and Tal Linzen observe that high performance on standard datasets does not necessarily indicate true linguistic understanding but may rather reflect reliance on superficial syntactic heuristics.

Focus and Objectives

The paper specifically explores the hypothesis that NLI models often adopt three fallible syntactic heuristics:

1. Lexical Overlap Heuristic: Treats all hypotheses with words from the premise as entailed.
2. Subsequence Heuristic: Assumes that contiguous subsequences in the premise entail all hypotheses.
3. Constituent Heuristic: Assumes that all complete subtrees in the syntactic parse tree of the premise are entailed.

To empirically test these hypotheses, the authors introduce the HANS (Heuristic Analysis for NLI Systems) dataset. HANS is a controlled evaluation set designed to include examples where each heuristic fails, thereby enabling a granular assessment of NLI model behaviors.

Methods and Experimentation

The paper evaluates four popular NLI models: DA, ESIM, SPINN, and BERT. These models span a range of architectural approaches from bag-of-words (DA) to tree-based (SPINN) and transformer-based (BERT). Each model was trained on MNLI, a standard NLI dataset, and then evaluated on HANS.

Key Findings

The experimental results from HANS reveal substantial deficiencies in all models:

• Performance on MNLI: Despite achieving high accuracy on the MNLI test set (DA: 72%, ESIM: 77%, SPINN: 67%, BERT: 84%), the models perform poorly on HANS.
• Heuristic Failures: Models exhibit near-zero accuracy on HANS examples designed to invalidate the heuristics, specifically scoring below 10% in most cases.
• Architectural Susceptibilities: DA and ESIM, lacking sophisticated syntactic structure encoding, fail across all heuristic cases. SPINN performs relatively better on subsequence cases due to its tree-based nature but still falls short. BERT, with the most advanced architecture, shows slightly improved but still inadequate performance on HANS.

Analysis of Results

These findings indicate that high MNLI performance may not reflect genuine linguistic understanding but rather an exploitation of dataset-specific regularities. SPINN’s relative success on some subsequence and constituent cases hints at the potential advantages of incorporating richer syntactic structures, though its overall performance underscores the insufficiency of current training data in fostering true generalization.

Implications and Future Directions

The implications of this research extend to both practical and theoretical domains. Practically, the results underscore the importance of carefully curated evaluation sets like HANS for diagnosing model behaviors beyond traditional datasets. Theoretically, the findings motivate reconsideration of inductive biases and training data in NLI models to more robustly capture syntactic nuances and promote genuine generalization.

Moreover, the authors' successful enhancement of model performance via augmented training with HANS-like examples provides promising avenues for future work. It suggests that strategically designed training data can mitigate heuristic biases, fostering deeper linguistic competencies in models. This approach could lead to the development of more linguistically informed and resilient NLI systems, capable of handling diverse inference challenges.

Conclusion

In summary, this paper provides critical insights into the limitations of current NLI models through the lens of syntactic heuristics. The introduction of the HANS dataset as a diagnostic tool exposes the over-reliance on superficial heuristics by state-of-the-art models, advocating for enhanced training methodologies and evaluation strategies. This paper advances the understanding of NLI systems, guiding future research towards more intricate and linguistically robust architectures.