Source-linked AI summary
Stress Test Evaluation for Natural Language Inference
Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, Graham Neubig
TL;DR
Standard NLI benchmarks show high model accuracy, but it remains unclear whether this reflects genuine semantic understanding. The paper introduces automatically constructed stress tests for targeted evaluation and finds that sentence-encoding models struggle with many phenomena despite high benchmark accuracy. These tests are intended to supplement standard evaluation and guide more informed model analysis and choices.
Problem
High accuracy on standard NLI datasets does not establish whether models make genuine inferential decisions rather than relying on sophisticated pattern matching.
Method
The paper automatically constructs fine-grained stress tests covering competence, distraction, and noise phenomena, using distinct construction techniques for each class.
Results
Four state-of-the-art sentence-encoding models struggle on many stress-test phenomena despite reporting high accuracy on NLI.
Takeaways & Limitations
Stress tests should supplement standard NLI evaluation to expose model strengths and weaknesses and support more informed model choices.
Takeaways & Limitations
The stress-test suite omits real-world knowledge, ambiguity, and unknown-error categories because they are difficult to construct automatically or do not correspond to particular phenomena.
Abstract
from arXiv · showhide
Natural language inference (NLI) is the task of determining if a natural language hypothesis can be inferred from a given premise in a justifiable manner. NLI was proposed as a benchmark task for natural language understanding. Existing models perform well at standard datasets for NLI, achieving impressive results across different genres of text. However, the extent to which these models understand the semantic content of sentences is unclear. In this work, we propose an evaluation methodology consisting of automatically constructed "stress tests" that allow us to examine whether systems have the ability to make real inferential decisions. Our evaluation of six sentence-encoder models on these stress tests reveals strengths and weaknesses of these models with respect to challenging linguistic phenomena, and suggests important directions for future work in this area.
1 Introduction
NLI evaluates whether a hypothesis is entailed by, contradicted by, or neutral to a premise, making it a benchmark for language understanding. The paper questions whether strong benchmark performance reflects genuine understanding or sophisticated pattern matching, and proposes fine-grained stress tests targeting linguistic phenomena.
- NLI requires classifying a hypothesis as entailment, contradiction, or neutral given a premise.
- NLI serves as a benchmark for natural language understanding because success requires reasoning about difficult linguistic phenomena.
- High accuracies from sentence-encoder models on SNLI and MultiNLI leave unclear whether models understand semantic content.
- The paper proposes automatically constructed stress tests to assess robustness to specific linguistic phenomena and competence in task-relevant reasoning.
- The methodology uses a challenging-model error typology to construct phenomenon-specific tests and benchmarks four state-of-the-art models on them.
2 Weaknesses of State-of-the-art NLI
The error analysis identifies lexical, semantic, numerical, grammatical, and data-related weaknesses in state-of-the-art NLI models. Word overlap is the largest listed category, while antonymy and numerical reasoning expose failures in core inference capabilities.
- 29% of analyzed errors involved word overlap, including wrong entailment predictions for unrelated pairs and neutral predictions when overlap was low.
- 13% of errors involved negation words causing contradiction predictions for neutral or entailed statements.
- 5% of errors involved failure to detect contradiction when premise-hypothesis pairs contained antonyms instead of explicit negation.
- 4% of errors involved inability to reason about numbers or quantifiers, while 3% involved length mismatch distractions.
- The analysis also identified grammaticality errors involving spelling mistakes or incorrect subject-verb agreement.
- Real-world knowledge accounted for 12% of errors, ambiguity for 6%, and unknown causes for 26%.
3 Stress Test Set Construction
The stress tests automatically probe NLI models on targeted linguistic phenomena using competence, distraction, and noise categories. Constructions preserve labels where required and include human checks for naturalness and correctness.
- Test selection: Stress tests target word overlap, negation, length mismatch, antonyms, spelling errors, and numerical reasoning.These phenomena were selected because they accounted for many errors in prior analysis.
- Scope and validation: The construction scope excludes real-world knowledge, ambiguity, and unknown categories, while grammaticality is represented only through spelling errors.Human evaluation found 91% agreement for word-overlap labels and 85% for length-mismatch labels with the gold labels.
- Test classes: Competence tests assess quantities and antonymy, distraction tests assess robustness to lexical similarity, negation words, and length mismatch, and noise tests assess spelling errors.The tests are organized by perceived difficulty and the type of challenge presented to models.
- Competence tests: Antonymy constructions substitute sampled WordNet antonyms for adjectives or nouns after word-sense disambiguation.The construction produces 1,561 matched and 1,734 mismatched premise-hypothesis pairs.
- Competence tests: Numerical-reasoning tests generate entailment, contradiction, and neutral hypotheses from premises containing numerical quantities and named entities.The source filtering yields 40,000 sentences before named-entity filtering and 2,500 premise sentences afterward; heuristic rules generate 7,596 pairs.
- Distraction tests: Distraction tests attach independently true tautologies to premises or hypotheses, preserving entailment, contradiction, or neutrality while introducing shallow cues.The tautologies are designed without topical word overlap with the original premise or hypothesis.
4 Experiments
Experiments show that sentence encoders lose accuracy across competence, distraction, and noise stress tests, exposing reliance on lexical cues and limited generalization from adversarial training.
- Overall performance: All six models lose performance across stress tests, with RC best on antonymy, CH best on numerical reasoning, and CH strongest on distraction tests.CH’s gated attention may help handle shallow word-level distractions.
- Competence tests: 36.4% and 32.8% are RC’s matched and mismatched antonymy accuracies, only slightly above random performance.Models overpredict entailment, accounting for 86.4% and 87.6% of errors on the matched and mismatched sets.
- Competence tests: 89.8% of easy antonym pairs and 57.2% of hard pairs appear as contradictions in training data, linking success partly to prior exposure.Successful cases are overwhelmingly complementary antonyms, while failures include more relational and gradable pairs.
- Competence tests: No model exceeds 33% accuracy on numerical reasoning, with 1,703 of 4,337 common errors misclassifying neutral pairs as entailment.The construction creates high word overlap in neutral pairs by flipping entailed premise-hypothesis pairs.
- Distraction tests: Accuracy decreases on word-overlap and length-mismatch tests, while false-neutral errors increase for all models.The results indicate that lexical similarity is a strong entailment signal and that models often default to neutral when it decreases.
- Noise tests: NB and IS degrade sharply on noise, whereas CH, RC, BiLSTM, and CBOW remain comparatively robust through subword modeling or mean pooling.Character-level CNNs can resist perturbations, while mean pooling may reduce the impact of single-word edits.
- Distraction training: Training on one distraction helps BiLSTM on that tautology but causes collapse on a different tautology.The result contrasts task-specific robustness with the human ability to ignore such distractions without training on each one.
5 Related Work and Discussion
The discussion situates stress testing within adversarial and error-analysis research, and identifies linguistically motivated modeling and transferable core competencies as future directions.
- Related work: Prior work advocates adversarial evaluation and relating model errors to well-defined linguistic phenomena to diagnose strengths and weaknesses.This motivates targeted stress tests beyond aggregate benchmark scores.
- Related work: Manual error analysis is positioned within earlier analyses of dependency parsing and information extraction, while the paper seeks a scalable automatic proxy.The broader literature includes both pre-neural NLI work and analyses of model errors in other NLP tasks.
- Discussion: The paper proposes explicit negation scope and semantic roles as examples of linguistically motivated additions to neural NLI models.These directions respond to the room for improvement revealed by the stress tests.
- Discussion: Identifying core competencies such as quantitative reasoning or antonymy may support performance across sentiment analysis, question answering, and relation extraction.The paper also proposes studying transfer from models with such competencies.
6 Conclusion
The paper presents stress tests as a complement to standard NLI evaluation, showing that high benchmark accuracy can coexist with weaknesses on targeted linguistic phenomena. It argues that these tests can deepen diagnosis and guide future model development.
- Conclusion: The paper introduces large-scale stress tests and automatic construction techniques for targeted NLI evaluation.The tests cover quantities, antonymy, shallow lexical cues, and random perturbations.
- Conclusion: Four state-of-the-art sentence encoders struggle on many stress-test phenomena despite high NLI accuracy.This establishes a gap between standard benchmark performance and targeted robustness.
- Conclusion: Stress tests should supplement rather than replace standard MultiNLI evaluation and can provide deeper insight into model strengths and weaknesses.The framework can be updated as new model forms create new evaluation needs.
- Conclusion: The released stress tests and resources are intended to promote models that move closer to true natural language understanding.The authors frame broader coverage and informed model choice as future benefits.
Appendix A. Error Analysis on Mismatched Set
The appendix describes how mismatched-set errors are categorized using the typology introduced earlier.
- Appendix A: Table 6 reports the proportions of mismatched-set misclassified examples assigned to each typology error category.The categories are defined in the paper’s earlier error-analysis typology.
Appendix B. Error Types on Antonymy
On the antonymy stress test, models frequently mistake contradiction for entailment because premise–hypothesis pairs share substantial lexical similarity.
- 86.4% and 87.6% of total errors on matched and mismatched sets, respectively, are false entailment errors on average.The antonymy test contains only one gold class: contradiction.
- All four models make many false entailment errors because they respond to high lexical similarity between the premise and hypothesis.
- Table 7 reports the percentages of C-E and C-N errors on the antonymy test.
Appendix C. Additional Experiments with Spelling Error Stress Tests
The spelling-error stress test applies character-level perturbations to hypotheses, and these perturbations do not significantly affect model performance.
- The stress test perturbs hypotheses by swapping adjacent characters or substituting a randomly selected alphabetical character.
- There is no significant performance effect from perturbing either a function word or a content word.
- Different perturbation types produce no considerable performance difference, consistent with models categorizing altered words as unknown words.