learn.ifs.name

Annotation Artifacts in Natural Language Inference Data

Gururangan et al., 2018, Annotation Artifacts in Natural Language Inference Data (Proceedings of the 2018 Conference of the North American Chapter of ACL: Human Language Technologies, Volume 2 (Short Papers))

Subscriber deck — free during early access

18 cards

1 Introduction

  1. What is the main finding of “Annotation Artifacts in Natural Language Inference Data”?
    The crowdsourcing protocol of SNLI and MultiNLI leaves clues in the hypotheses, so the label can often be predicted from the hypothesis alone, without the premise; NLI models’ success had been overestimated and the task remains hard.
  2. How were SNLI and MultiNLI hypotheses created, and why does that matter here?
    Crowdworkers saw a premise and wrote three hypotheses: one definitely true (entailment), one that might be true (neutral), one definitely not true (contradiction). Writing to a label invites quick, repeatable heuristics that leave label-specific traces in the hypothesis text.

2 Annotation Artifacts are Common

  1. What are annotation artifacts?
    Patterns in a dataset produced by the framing of the annotation task and the strategies annotators adopt, here in how crowdworkers phrase hypotheses for each label, which correlate with labels independently of the actual inference.
  2. How accurate is a hypothesis-only classifier on SNLI and MultiNLI?
    A bag-of-words-and-bigrams fastText classifier that never sees the premise got 67.0% on SNLI (majority class 34.3%) and about 53% on MultiNLI (53.9% matched, 52.3% mismatched; majority about 35%). Other premise-oblivious classifiers gave similar results.

3.1 Lexical Choice

  1. How did Gururangan et al. find words indicative of each NLI class?
    With pointwise mutual information between word and class over training hypotheses, \(\text{PMI}(\text{word}, \text{class}) = \log \frac{p(\text{word}, \text{class})}{p(\text{word}, \cdot)\, p(\cdot, \text{class})}\), with aggressive add-100 smoothing to emphasize highly discriminative correlations.
  2. Which lexical artifacts mark entailed hypotheses?
    Generic words generalizing over specific premise words (animal for dog, instrument for guitar, outdoors/outside for beach), approximate quantities replacing exact numbers (some, at least, various), and gender-neutral references (person, human).
  3. Which lexical artifacts mark neutral hypotheses?
    Modifiers (tall, sad, popular) and superlatives (first, favorite, most) that add plausible but unsupported information, and cause or purpose clauses, raising the frequency of discourse markers like because.
  4. Which lexical artifacts mark contradictions?
    Negation words (nobody, no, never, nothing) and words that contradict almost any premise, such as sleeping (contradicts any activity), naked (any clothing description) and cat (many SNLI images show dogs).
  5. How might the SNLI annotation guidelines have caused some artifacts?
    The guideline example (“Two dogs are running through a field.”) had the entailment “There are animals outdoors.” (generic word, scenery word, number removed), the neutral “Some puppies are running to catch a stick.” (purpose clause), and a contradiction about pets on a couch, likely priming workers.
  6. Are the artifact words very frequent?
    No: even the most discriminative words appear in only a small share of hypotheses (e.g. nobody in 0.1% of contradictions, no in a few percent), so the artifacts are diverse, and workers used many different heuristics.

3.2 Sentence Length

  1. How does hypothesis length relate to the NLI label in SNLI?
    Neutral hypotheses tend to be long (median 9 tokens), entailments short (60% have 7 tokens or fewer). Half of hypotheses with at least 12 tokens are neutral, and about half of those with 5 or fewer are entailments, so length alone is a useful feature.
  2. What does word containment reveal about how entailed hypotheses were written?
    As bags of words, 8.8% of entailed SNLI hypotheses are fully contained in their premise, versus 0.2% of neutrals and contradictions, suggesting many entailments were written by deleting words from the premise.

4 Re-evaluating NLI Models

  1. How did Gururangan et al. split NLI test sets to test models?
    Into Easy examples, which the hypothesis-only classifier labelled correctly, and Hard examples, which it got wrong, then evaluated models trained on the original training sets on full, Hard and Easy subsets.
  2. How did strong NLI models perform on Easy versus Hard examples?
    Much worse on Hard: on SNLI, DIIN scored 93.4% on Easy but 72.7% on Hard (86.5% overall); on MultiNLI matched 87.6% vs 64.1%. Similar gaps for DAM and ESIM. The bulk of their success came from examples solvable by artifacts.
    SNLIfullHardEasy
    DAM84.769.492.4
    ESIM85.871.392.6
    DIIN86.572.793.4
  3. Why isn’t simply removing Easy examples a complete fix for annotation artifacts?
    Hard examples may carry new artifacts (removing contradictions with no leaves no mostly in other classes), and Easy examples contain genuine inference phenomena (animal really is a hypernym of dog). Artifacts don’t make examples wrong; the label distribution is skewed, so balancing artifacts across labels may be better.
  4. What did Gururangan et al. suggest for collecting future NLI data?
    Change prompts or worker training to encourage a wide range of strategies, or use baseline or adversarial systems during collection to flag examples that rely on over-represented heuristics.

5 Discussion

  1. Which other datasets were found to have similar artifacts?
    SICK (negation, overlap and hypernymy predict labels), CNN/Daily Mail (problems from automatic annotation), the ROC Stories cloze task (endings alone suffice), relation inference benchmarks biased toward the method used to build them, and other NLI sets (Poliak et al. 2018, hypothesis-only baselines).
  2. What practical lesson does this paper teach about evaluating on crowdsourced benchmarks?
    Check for shortcuts with partial-input baselines (e.g. hypothesis-only); a model scoring well may be exploiting annotation artifacts rather than solving the task, so report results on hard subsets and build more challenging benchmarks.

Source

  • Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman and Noah A. Smith. 2018. Annotation Artifacts in Natural Language Inference Data. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp. 107–112. License: CC BY 4.0. Flashcards adapted from this work. doi:10.18653/v1/N18-2017 https://aclanthology.org/N18-2017/