---
title: "Investigating Biases in Textual Entailment Datasets — Tan 2019"
description: "The authors train an RNN classifier that sees only the hypothesis. 64% accuracy. 13 references, cited by 2 papers here."
authors: ["Shawn Tan", "Yikang Shen", "Chin-wei Huang", "Aaron Courville"]
year: 2019
venue: "arXiv:1906.09635"
source: "https://arxiv.org/abs/1906.09635"
doi: "10.48550/arxiv.1906.09635"
arxiv: "1906.09635"
licence: "arXiv non-exclusive distribution licence (read only, no redistribution)"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/tan-2019-investigating-biases-textual-entailment-datasets/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tan-2019-investigating-biases-textual-entailment-datasets/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) / Investigating Biases in Textual Entailment Datasets

# Investigating Biases in Textual Entailment Datasets

**Authors:** Shawn Tan, [Yikang Shen](https://md.research.elamur.ai/free/en/authors/yikang-shen/), Chin-wei Huang and [Aaron Courville](https://md.research.elamur.ai/free/en/authors/aaron-courville/) · **Year:** 2019 · **Venue:** arXiv:1906.09635 · **Licence:** [arXiv non-exclusive distribution licence (read only, no redistribution)](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/1906.09635) · [PDF](https://arxiv.org/pdf/1906.09635) · [DOI 10.48550/arxiv.1906.09635](https://doi.org/10.48550/arxiv.1906.09635) · [arXiv 1906.09635](https://arxiv.org/abs/1906.09635) · [OpenAlex](https://openalex.org/W2952281677)

**Citations:** 5 citations (OpenAlex, 2026-10-14) · [2](https://md.research.elamur.ai/free/en/papers/tan-2019-investigating-biases-textual-entailment-datasets/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
The paper asks whether crowdsourced entailment corpora carry annotation artifacts similar to those found earlier in VQA, and whether NLI models exploit them. The authors train an RNN classifier that sees only the hypothesis. On SNLI it reaches 64% accuracy, far above the roughly 33% expected by chance. On MultiNLI (mismatched set) it reaches 51%. They also scramble premises at test time to see how much trained models still rely on the hypothesis alone. A bigram study ranks hypothesis bigrams by the entropy of their label distribution and exposes crowdworker writing habits. Negations such as "nobody is" signal contradiction, while added or removed details signal neutral or entailment. As a fix, they greedily prune 20% of SNLI training data using a Naive Bayes model that is retrained after each removal. The pruning lowers hypothesis-only accuracy and costs ESIM a few test points. The authors read this as evidence that benchmark scores overstate true inference ability. They argue for train and test splits with deliberately different biases.

## Key points
- **Hypothesis-only accuracy:** a classifier that never sees the premise scores 64% on SNLI and 51% on MultiNLI mismatched. The authors conclude that MultiNLI is the less biased corpus.
- **Shuffled premises:** when test premises are mismatched with their hypotheses, ESIM drops from 88% to 40.5%. An LSTM sentence-embedding model drops from 70% to 50%. Both stay above chance, and the embedding model leans more heavily on hypothesis cues.
- **Bigram scoring method:** bigrams occurring more than 200 times are smoothed with Laplace α = 1 and sorted by label entropy. In SNLI, "nobody is" favours contradiction about 222:1. In MultiNLI, the top cue "and never" favours contradiction about 8:1.
- **Fewer entailment cues:** strong entailment-specific cues are rare, since writers can produce entailed sentences simply by dropping details from the premise.
- **Pruning results** (hypothesis-only / ESIM train / ESIM test):

| Training set | h-only | ESIM train | ESIM test |
|---|---|---|---|
| ORIGINAL | 64% | 93% | 88% |
| RANDOM (20% removed) | 59% | 94% | 87% |
| GREEDY (20% removed) | 56% | 81% | 85% |

## Interesting ideas
- Feed a model only part of the input it logically needs; any score above chance reveals a shortcut in the data rather than reasoning.
- At test time, pair each input with a wrong partner and measure how much accuracy survives; the leftover above chance quantifies dependence on one side alone.
- Rank surface features by how sharply they predict a label, with a frequency floor and smoothing, to get a readable list of annotation artifacts.
- Remove biased training items one at a time with a cheap, exactly updatable model (Naive Bayes counts are just decremented), rescoring after each removal so new dominant cues are caught.
- Compare a targeted data reduction against random removal of the same size, so effects of bias removal are separated from effects of having less data.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (learn): Add partial-input runs as standard negative controls in evaluation, e.g., answering school-test questions without the passage, to flag questions solvable by surface cues.
- **I2** (answer): Shuffle which fable a question is attached to and check that the readers' accuracy collapses, proving answers come from the built world of the text.
- **I3** (learn): Audit silver data and LLM-teacher outputs for lexical patterns that predict labels too well before they leak into training.
- **I4** (learn): Apply the same greedy, count-based rebalancing to silver treebank or frame data so no single cue is over-represented.
- **I5** (learn): Whenever repair rules or filters drop training items, train a size-matched random-drop baseline as the control.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Université de Montréal](https://md.research.elamur.ai/free/en/universities/universite-de-montreal/) — University, Montreal, Canada · 47 works · 73,845 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works

🧭 Topics (1): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 510 works

## Cite

Cite as: Shawn Tan, Yikang Shen, Chin-wei Huang and Aaron Courville. 2019. Investigating Biases in Textual Entailment Datasets. arXiv:1906.09635.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{tan2019investigating,
  title = {{Investigating Biases in Textual Entailment Datasets}},
  author = {Shawn Tan, Yikang Shen, Chin-wei Huang and Aaron Courville},
  year = {2019},
  doi = {10.48550/arxiv.1906.09635},
  eprint = {1906.09635},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1906.09635}
}
```

- [← previous Guess Who's Coming (and Who's Going): Bringing Perspective to the Rational Speech Acts Framework](https://md.research.elamur.ai/free/en/papers/anderson-2019-guess-who-s-coming-who/)
- [next → Apertium-fin-eng–Rule-based Shallow Machine Translation for WMT 2019 Shared Task](https://md.research.elamur.ai/free/en/papers/pirinen-2019-apertium-fin-eng-rule-based/)
- [↑ back to papers of 2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/)

## References

13 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2017. Don’t just assume; look and answer: Overcoming priors for visual question answering. arXiv preprint arXiv:1712.00377. [↗](https://arxiv.org/abs/1712.00377)
2. Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433. [↗](https://doi.org/10.1007/s11263-016-0966-6)
3. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.1508.05326)
4. Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced lstm for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 1657–1668. [find](https://scholar.google.com/scholar?q=Qian+Chen%2C+Xiaodan+Zhu%2C+Zhen%2DHua+Ling%2C+Si+Wei%2C+Hui+Jiang%2C+and+Diana+Inkpen%2E+2017%2E+Enhanced+lstm+for+natural+language+inference%2E+In+Proceedings+of+the+55th+Annual+Meeting+of+the+Association+for+Computa)
5. Yichen Gong, Heng Luo, and Jian Zhang. 2017. Natural language inference over interaction space. arXiv preprint arXiv:1709.04348. [↗](https://arxiv.org/abs/1709.04348)
6. Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. [Making the v in vqa matter: Elevating the role of image understanding in visual question answering](https://md.research.elamur.ai/free/en/papers/goyal-2016-making-v-vqa-matter-elevating/). In CVPR, volume 1, page 9. [↗](https://doi.org/10.1007/s11263-018-1116-0)
7. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324. [↗](https://arxiv.org/abs/1803.02324)
8. Bill MacCartney and Christopher D Manning. 2009. An extended model of natural logic. In Proceedings of the eighth international conference on computational semantics, pages 140–156. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Bill+MacCartney+and+Christopher+D+Manning%2E+2009%2E+An+extended+model+of+natural+logic%2E+In+Proceedings+of+the+eighth+international+conference+on+computational+semantics%2C+pages+140%E2%80%93156%2E+Association+for+Co)
9. Marco Marelli, Stefano Menini, Marco Baroni, Luisa Bentivogli, Raffaella Bernardi, Roberto Zamparelli, et al. 2014. [A sick cure for the evaluation of compositional distributional semantic models](https://md.research.elamur.ai/free/en/papers/marelli-2014-sick-cure-evaluation-compositional-distributional/). In LREC, pages 216–223.
10. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. arXiv preprint arXiv:1805.01042. [↗](https://arxiv.org/abs/1805.01042)
11. Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426. [↗](https://arxiv.org/abs/1704.05426)
12. Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67–78. [↗](https://doi.org/10.1162/tacl_a_00166)
13. Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2016. Yin and yang: Balancing and answering binary visual questions. In Computer Vision and Pattern Recognition (CVPR), 2016 IEEE Conference on, pages 5014–5022. IEEE. [find](https://scholar.google.com/scholar?q=Peng+Zhang%2C+Yash+Goyal%2C+Douglas+Summers%2DStay%2C+Dhruv+Batra%2C+and+Devi+Parikh%2E+2016%2E+Yin+and+yang%3A+Balancing+and+answering+binary+visual+questions%2E+In+Computer+Vision+and+Pattern+Recognition+%28CVPR%29%2C+2016)

*Updated 24 October 2026*