Reproduce “Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference” on OpenResearch

About this paper

R. Thomas McCoy, Ellie Pavlick, Tal Linzen

Researchers from Johns Hopkins and Brown Universities introduced HANS, a diagnostic dataset, to demonstrate that state-of-the-art Natural Language Inference models frequently rely on fallible syntactic heuristics. These models exhibited near-zero accuracy on HANS examples designed to contradict these heuristics, but their performance improved dramatically when trained on augmented data containing counter-examples.

Key insights

• Current Natural Language Inference (NLI) models, including state-of-the-art architectures like BERT, extensively rely on superficial syntactic heuristics (e.g., lexical overlap, subsequence, constituent) rather than robust linguistic understanding.
• Models can learn to overcome these fallible heuristics if provided with targeted counter-examples during training, demonstrating that the issue lies more with training data bias than inherent architectural limitations.
• Diagnostic datasets are essential for probing specific model weaknesses and understanding what models truly learn, moving beyond aggregated accuracy metrics on standard benchmarks.

Summary by alphaXiv.