---
title: "Which Spurious Correlations Impact Reasoning in NLI Models? A"
description: "The authors present an interactive, human-in-the-loop dashboard for finding these cues… 30% accuracy. 17 references, cited by 3 papers here."
authors: ["Robin Chan", "Afra Amini", "Mennatallah El-Assady"]
year: 2023
venue: "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)"
source: "https://aclanthology.org/2023.acl-demo.44/"
doi: "10.18653/v1/2023.acl-demo.44"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/chan-2023-which-spurious-correlations-impact-reasoning/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/chan-2023-which-spurious-correlations-impact-reasoning/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/15/) / Which Spurious Correlations Impact Reasoning in NLI Models? A Visual Interactive Diagnosis through Data-Constrained Counterfactuals

# Which Spurious Correlations Impact Reasoning in NLI Models? A Visual Interactive Diagnosis through Data-Constrained Counterfactuals

**Authors:** Robin Chan, [Afra Amini](https://md.research.elamur.ai/free/en/authors/afra-amini/) and Mennatallah El-Assady · **Year:** 2023 · **Venue:** [Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2023](https://md.research.elamur.ai/free/en/venues/acl/2023/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2023.acl-demo.44/) · [PDF](https://aclanthology.org/2023.acl-demo.44.pdf) · [DOI 10.18653/v1/2023.acl-demo.44](https://doi.org/10.18653/v1/2023.acl-demo.44) · [OpenAlex](https://openalex.org/W4385565551)

**Citations:** 1 citations (OpenAlex, 2026-10-14) · 3 in our library, of the 149,439 papers whose bibliographies we read

## Summary
NLI models trained on crowdsourced corpora such as SNLI often latch onto surface cues that merely co-occur with a label. The authors present an interactive, human-in-the-loop dashboard for finding these cues and writing examples that break them. Their starting point is "data-constrained counterfactuals" (DCC), which are training items that already contradict a pattern. A DCC must meet two conditions. Its nearest neighbours, measured by cosine similarity of the model's [CLS] embeddings, carry a different label. It also sits in the hard-to-learn or ambiguous zone of a Data Map.

The dashboard guides the user through three phases. The user first understands the DCC, then diagnoses it against similar items with both matching and differing labels, and finally refines new examples. In the refining phase, GPT-3 (davinci-003) supplies few-shot suggestions, and the trained model estimates where each new example would land on the Data Map.

The diagnosis surfaced three families of spurious correlation: Semantic Relevance, Logical Fallacies and Bias. It also exposed a set of annotation artifacts. A small batch of generated examples pushed a robust WaNLI-trained model down to roughly 30% accuracy. This suggests the workflow can produce useful adversarial test suites and augmentation data.

## Key points
- A DCC is defined by two conditions:
  - Similar training items exist that carry another label.
  - The model struggles with it, falling in the hard-to-learn or ambiguous Data Map region.
- To avoid picking mislabeled items, only SNLI samples with multiple annotations and at least 75% agreement are kept. Most such samples have four annotations.
- The dashboard tasks follow the explAIner framework:
  - U1–U2 show the DCC and its Data Map position.
  - D1–D2 show nearest neighbours with differing and matching labels, plus their labels on the map.
  - R1–R2 cover GPT-3 suggestions and model feedback.
- The GPT-3 prompt works as follows:
  - Same-label neighbours come first, sorted by rising similarity to the DCC, and the DCC is placed last.
  - The premise and hypothesis are linked by a label word: Implication, Possibility, or Contradiction.
- `roberta-large-wanli` scored around 30% on the generated examples. The tool is available at dcc.lingvis.io.

## Interesting ideas
- Rather than inventing counterfactuals from scratch, first look for items already in the corpus that defeat a pattern the model leans on, and use them as seeds. This targets real weaknesses instead of random edits.
- Combine a similarity condition with a training-dynamics condition (low or unstable confidence across checkpoints) to select informative examples. Requiring high annotator agreement on top of this filters out items that are hard simply because they are mislabeled.
- Show the user a hard example side by side with its closest neighbours under the same label and under a different label. This contrast makes the shared surface cue visible to a human.
- Give instant model feedback on a newly written example by estimating how hard it is before it is accepted. This steers annotators toward genuinely challenging items.
- Use a small set of curated counterfactuals as an adversarial test suite, since even a robust model can collapse on it.

## For Mova

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

- **I1** (learn): Mine silver data for sentences whose nearest neighbours carry a different analysis yet the parser handles them poorly, and treat them as seeds for targeted repair rules.
- **I2** (learn): Track per-sentence confidence across perceptron or gparse epochs, and flag hard-but-agreed items (where witness parsers concur) separately from likely label noise.
- **I3** (answer): For failed school-test or fable QA items, retrieve similar solved and unsolved items to expose which cue the reader used instead of the actual frame reasoning.
- **I4** (learn): When an LLM teacher proposes new training sentences, have the current model score its expected difficulty and keep only those it is likely to get wrong.
- **I5** (frames): Build entailment-style probes where extra unstated details or misleading lexical cues should not change who-did-what-to-whom, and use them as negative controls for the frame layer.

*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

- [ETH Zurich](https://md.research.elamur.ai/free/en/universities/eth-zurich/) — University, Zurich, Switzerland · 84 works · 9,181 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (1): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,685 works

## Cite

Cite as: Robin Chan, Afra Amini and Mennatallah El-Assady. 2023. Which Spurious Correlations Impact Reasoning in NLI Models? A Visual Interactive Diagnosis through Data-Constrained Counterfactuals. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations).

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

```bibtex
@inproceedings{chan2023which,
  title = {{Which Spurious Correlations Impact Reasoning in NLI Models? A Visual Interactive Diagnosis through Data-Constrained Counterfactuals}},
  author = {Robin Chan, Afra Amini and Mennatallah El-Assady},
  booktitle = {Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2023},
  doi = {10.18653/v1/2023.acl-demo.44},
  url = {https://doi.org/10.18653/v1/2023.acl-demo.44}
}
```

- [← previous When to generate hedges in peer-tutoring interactions](https://md.research.elamur.ai/free/en/papers/abulimiti-2023-when-generate-hedges-peer-tutoring/)
- [next → "You are Cursed by the God YHW": an early Hebrew inscription from Mt. Ebal](https://md.research.elamur.ai/free/en/papers/stripling-2023-cursed-god-yhw-early-hebrew/)
- [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/15/)

## References

17 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. [A large annotated corpus for learning natural language inference](https://md.research.elamur.ai/free/en/papers/bowman-2015-large-annotated-corpus-learning-natural/). In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/d15-1075)
2. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. [Language models are few-shot learners](https://md.research.elamur.ai/free/en/papers/brown-2020-language-models-few-shot-learners/). In Advances in Neural Information Processing Systems, volume 33, pages 1877–1901. Curran Associates, Inc.
3. Cleo Condoravdi, Dick Crouch, Valeria de Paiva, Reinhard Stolle, and Daniel G. Bobrow. 2003. Entailment, intensionality and text understanding. In Proceedings of the HLT-NAACL 2003 Workshop on Text Meaning, pages 38–45. [find](https://scholar.google.com/scholar?q=Cleo+Condoravdi%2C+Dick+Crouch%2C+Valeria+de+Paiva%2C+Reinhard+Stolle%2C+and+Daniel+G%2E+Bobrow%2E+2003%2E+Entailment%2C+intensionality+and+text+understanding%2E+In+Proceedings+of+the+HLT%2DNAACL+2003+Workshop+on+Text+Me)
4. Smith. 2021. Competency problems: On finding and removing artifacts in language data. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1801–1813, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. [↗](https://doi.org/10.48448/8kgp-zg94)
5. William Huang, Haokun Liu, and Samuel R. Bowman. 2020. [Counterfactually-augmented SNLI training data does not yield better generalization than unaugmented data](https://md.research.elamur.ai/free/en/papers/huang-2020-counterfactually-augmented-snli-training-data/). In Proceedings of the First Workshop on Insights from Negative Results in NLP, pages 82–87, Online. Association for Computational Linguistics.
6. Robin Jia and Percy Liang. 2017. [Adversarial examples for evaluating reading comprehension systems](https://md.research.elamur.ai/free/en/papers/jia-2017-adversarial-examples-evaluating-reading-comprehension/). In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021–2031, Copenhagen, Denmark. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/d17-1215)
7. Daniel Khashabi, Tushar Khot, and Ashish Sabharwal. 2020. More bang for your buck: Natural perturbation for robust question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 163–170, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Daniel+Khashabi%2C+Tushar+Khot%2C+and+Ashish+Sabharwal%2E+2020%2E+More+bang+for+your+buck%3A+Natural+perturbation+for+robust+question+answering%2E+In+Proceedings+of+the+2020+Conference+on+Empirical+Methods+in+Nat)
8. Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, Zhiyi Ma, Tristan Thrush, Sebastian Riedel, Zeerak Waseem, Pontus Stenetorp, Robin Jia, Mohit Bansal, Christopher Potts, and Adina Williams. 2021. Dynabench: Rethinking benchmarking in NLP. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4110–4124, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Douwe+Kiela%2C+Max+Bartolo%2C+Yixin+Nie%2C+Divyansh+Kaushik%2C+Atticus+Geiger%2C+Zhengxuan+Wu%2C+Bertie+Vidgen%2C+Grusha+Prasad%2C+Amanpreet+Singh%2C+Pratik+Ringshia%2C+Zhiyi+Ma%2C+Tristan+Thrush%2C+Sebastian+Riedel%2C+Zeerak+)
9. Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. 2022. [WANLI: Worker and AI collaboration for natural language inference dataset creation](https://md.research.elamur.ai/free/en/papers/liu-2022-wanli-worker-ai-collaboration-natural/). In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6826–6847, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
10. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. [RoBERTa: A robustly optimized BERT pretraining approach](https://md.research.elamur.ai/free/en/papers/liu-2019-roberta-robustly-optimized-bert-pretraining/).
11. Ido Dagan, Oren Glickman, and Bernardo Magnini. 2006. The PASCAL recognising textual entailment challenge. In Machine Learning Challenges. Evaluating Predictive Uncertainty, Visual Object Classification, and Recognising Tectual Entailment, pages 177–190, Berlin, Heidelberg. Springer Berlin Heidelberg. [find](https://scholar.google.com/scholar?q=Ido+Dagan%2C+Oren+Glickman%2C+and+Bernardo+Magnini%2E+2006%2E+The+PASCAL+recognising+textual+entailment+challenge%2E+In+Machine+Learning+Challenges%2E+Evaluating+Predictive+Uncertainty%2C+Visual+Object+Classificati)
12. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. [Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference](https://md.research.elamur.ai/free/en/papers/mccoy-2019-right-wrong-reasons-diagnosing-syntactic/). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448, Florence, Italy. Association for Computational Linguistics. Matt Gardner, William Merrill, Jesse Dodge, Matthew Peters, Alexis Ross, Sameer Singh, and Noah A.
13. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. [Adversarial NLI: A new benchmark for natural language understanding](https://md.research.elamur.ai/free/en/papers/nie-2020-adversarial-nli-new-benchmark-natural/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885–4901, Online. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/2020.acl-main.441)
14. Thilo Spinner, Udo Schlegel, Hanna Schäfer, and Mennatallah El-Assady. 2020. explAIner: A visual analytics framework for interactive and explainable machine learning. IEEE Transactions on Visualization and Computer Graphics, 26(1):1064–1074. [↗](https://doi.org/10.1109/tvcg.2019.2934629)
15. Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275–9293, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Swabha+Swayamdipta%2C+Roy+Schwartz%2C+Nicholas+Lourie%2C+Yizhong+Wang%2C+Hannaneh+Hajishirzi%2C+Noah+A%2E+Smith%2C+and+Yejin+Choi%2E+2020%2E+Dataset+cartography%3A+Mapping+and+diagnosing+datasets+with+training+dynamics%2E+)
16. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. [Transformers: State-of-the-Art natural language processing](https://md.research.elamur.ai/free/en/papers/wolf-2020-transformers-state-art-natural-language/). In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/2020.emnlp-demos.6)
17. Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel Weld. 2021. [Polyjuice: Generating counterfactuals for explaining, evaluating, and improving models](https://md.research.elamur.ai/free/en/papers/wu-2021-polyjuice-generating-counterfactuals-explaining-evaluating/). In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6707–6723, Online. Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.2101.00288)

*Updated 1 November 2026*