Source-linked AI summary
Annotation Artifacts in Natural Language Inference Data
Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman, Noah A. Smith
TL;DR
NLI datasets may encode label-predictive artifacts because crowd workers generate hypotheses under a structured annotation protocol. The paper tests this with premise-oblivious classifiers and analyzes linguistic patterns, finding that many examples are classifiable from hypotheses alone and that current NLI progress is overstated. It therefore releases harder evaluation sets while identifying limitations of artifact filtering.
Problem
Crowdsourced NLI datasets may contain hypothesis-only clues that permit label prediction without testing the intended premise–hypothesis inference relation.
Method
The paper trains a premise-oblivious fastText classifier and analyzes lexical and structural heuristics associated with inference classes.
Results
Hypothesis-only artifacts classify substantial portions of SNLI and MultiNLI, while high-performing NLI models perform dramatically worse on the remaining hard examples.
Takeaways & Limitations
NLI model performance can be inflated by annotation artifacts, motivating harder benchmarks such as the released Hard SNLI and MultiNLI test sets.
Takeaways & Limitations
Filtering easy examples may leave new artifacts and remove important inference phenomena, because artifacts reflect skewed sample distributions rather than incorrect individual examples.
Abstract
from arXiv · showhide
Large-scale datasets for natural language inference are created by presenting crowd workers with a sentence (premise), and asking them to generate three new sentences (hypotheses) that it entails, contradicts, or is logically neutral with respect to. We show that, in a significant portion of such data, this protocol leaves clues that make it possible to identify the label by looking only at the hypothesis, without observing the premise. Specifically, we show that a simple text categorization model can correctly classify the hypothesis alone in about 67% of SNLI (Bowman et. al, 2015) and 53% of MultiNLI (Williams et. al, 2017). Our analysis reveals that specific linguistic phenomena such as negation and vagueness are highly correlated with certain inference classes. Our findings suggest that the success of natural language inference models to date has been overestimated, and that the task remains a hard open problem.
1 Introduction
The paper argues that crowdsourced NLI hypotheses contain label-predictive artifacts, allowing premise-free classification and revealing weaknesses in current evaluations.
- NLI determines whether a hypothesis is entailed by a premise, and supports semantic tasks such as question answering and summarization.
- Crowd workers generate entailment, neutral, and contradiction hypotheses from a provided premise.
- The crowdsourcing process produces artifacts that can reveal the correct inference class without access to the premise.
- Entailed hypotheses tend to use gender-neutral references, purpose clauses signal neutrality, and negation correlates with contradiction.
- Models perform dramatically worse on examples where a hypothesis-only classifier fails, suggesting that NLI remains an open problem.
2 Annotation Artifacts are Common
The paper tests whether hypothesis wording alone exposes NLI labels, finding substantial premise-independent predictability and identifying annotation strategies behind these artifacts.
- Annotation artifacts are patterns produced when task framing influences crowd workers’ language-generation choices.
- Common strategies include removing gender or number information, adding purpose clauses, and introducing negation.
- A premise-oblivious fastText classifier predicts NLI labels from hypothesis words and bigrams alone.
- A significant portion of each test set is classified correctly without the premise, substantially above the most-frequent-class baseline.
- These artifacts allow strong dataset performance without modeling natural language inference itself.
3 Characteristics of Annotation Artifacts
Annotation artifacts arise from lexical and sentence-level heuristics in crowdsourced NLI hypotheses, creating class-predictive patterns without the premise. These patterns include class-linked word choices, guideline-induced phrasing, and systematic length differences.
- Overview: Crowd workers appear to use multiple heuristics, so even the most discriminative words are individually infrequent and artifacts remain diverse.The analysis focuses on lexical choice and sentence length.
- Lexical Choice: Entailed hypotheses favor generic or vague terms, approximate quantities, and gender-neutral references, with some cues tied to SNLI’s image-caption domain.Examples include animal, outdoors, some, and person.
- Lexical Choice: The SNLI annotation example may have primed workers to replace specific words with generic ones and add purpose clauses.The paper links dog-to-animal, scenery-to-outdoors, and purpose-clause patterns to the example.
- Lexical Choice: Neutral hypotheses correlate with modifiers, superlatives, and cause or purpose clauses, while contradiction correlates with explicit negation and words that oppose premise content.Because is highlighted among discourse markers for neutral hypotheses; nobody, no, never, and nothing indicate contradiction.
- Sentence Length: Neutral hypotheses tend to be longer than entailments: the neutral median is 9 tokens, while 60% of entailments have 7 tokens or fewer.Half of hypotheses with at least 12 tokens are neutral, whereas a similar portion at length 5 or under are entailments; length is less significant in MultiNLI.
- Sentence Length: 8.8% of entailed SNLI hypotheses are fully contained within their premises, compared with 0.2% of neutral and contradiction hypotheses.MultiNLI shows similar trends.
4 Re-evaluating NLI Models
The authors test whether high-performing NLI models rely on hypothesis artifacts by comparing performance on Easy and Hard examples. Their findings motivate caution about artifact-free evaluation and dataset construction.
- The authors partition each NLI test set into Easy examples correctly classified by the premise-oblivious model and Hard examples it misclassified.
- They evaluate DAM, ESIM, and DIIN on the full, Hard, and Easy test sets after retraining them on the original training data.
- Performance on the Hard subset is substantially lower than on the remaining instances, indicating that high-performing models depend heavily on annotation artifacts.
- Filtering Easy examples may not remove artifacts, because Hard examples can acquire new label-correlated patterns after filtering.The authors give the example of removing contradiction examples containing “no,” which changes its distribution across classes.
- Easy examples also contain important inference phenomena, so removing them could hinder learning; the authors instead suggest balancing artifacts across labels.
5 Discussion
The discussion places annotation artifacts within a broader pattern of dataset bias and argues that they inflate reported model performance. The authors release Hard evaluation sets and encourage more challenging benchmarks.
- Many datasets contain annotation artifacts: Annotation artifacts are presented as a recurring problem across NLP datasets, including entailment, question answering, visual question answering, and story cloze tasks.
- Supervised models leverage annotation artifacts: Supervised models can achieve strong performance by exploiting dataset artifacts rather than fully modeling the intended task.
- Annotation artifacts inflate model performance: Annotation artifacts inflate model performance because large portions of test sets can be solved using artifacts alone.
- Annotation artifacts inflate model performance: The authors release Hard SNLI and MultiNLI test sets for evaluating NLI models alongside the original benchmarks.
- Annotation artifacts inflate model performance: They encourage additional challenging benchmarks that expose the true performance levels of state-of-the-art NLI models.