Elamur Research⌕
☰
home / papers / 2022 / Capture Human Disagreement Distributions by Calibrated Networks for Natural Language Inference

Capture Human Disagreement Distributions by Calibrated Networks for Natural Language Inference

Paper · 2022 · Yuxia Wang, Minghan Wang, Yimeng Chen et al.

Summary

NLI items often have no single right answer: when many people label the same premise–hypothesis pair, their votes split. Earlier work (AmbiNLI) tried to model this split by training directly on crowd-sourced label distributions, which are costly to collect. The authors propose a different explanation. Any model whose confidence matches its real correctness rate, a calibrated model, should already reproduce the human vote distribution. Their view is that soft crowd labels help mainly because they act as a form of label smoothing. They train NLI classifiers on SNLI and MultiNLI and recalibrate them with label smoothing or with post-hoc temperature scaling. They evaluate on all of ChaosNLI (100 votes per item) using JSD, ECE, NLL and accuracy. Calibrated models reach divergence scores close to those obtained with gold ambiguity labels, and accuracy does not drop. Tests on MedNLI show the same pattern across domains, although calibration quality depends on the data. In practice, cheap calibration can stand in for expensive multi-annotation.

Key points

  • Evaluation uses all of Chaos-SNLI (1,514 items) and Chaos-MNLI (1,599 items). Training uses SNLI/MultiNLI, along with AmbiNLI subsets, which carry only about 5 annotations per item.
  • Temperature scaling (T=1.2 on Chaos-SNLI, T=1.5 on Chaos-MNLI) lowered ECE, JSD and NLL compared with the untuned smnli baseline.
  • In a case study where T was picked on the test set, raising T on Chaos-MNLI from 1.5 to 4.0 cut JSD from 0.29 to 0.23. Lower ECE went together with lower JSD.
  • On the MedNLI test set:
    • The smnli model had accuracy 0.6125 and ECE 0.2104.
    • Training on AmbiSM with label smoothing (α=0.8) gave 0.6188 accuracy and 0.0687 ECE.
    • In-domain fine-tuning reached 0.7771 accuracy.
  • Soft labels do not always help. For models that are already calibrated or under-confident, and for clear-cut data, extra soft-label training can make under-confidence worse.

Interesting ideas

  • I1. Treat a model's confidence as a stand-in for how much humans would disagree, provided the model is calibrated; the question "how to model ambiguity" then becomes the simpler question "how to calibrate".
  • I2. Temperature scaling adjusts one scalar on held-out data and never changes the argmax, so accuracy stays exactly the same while confidences become honest.
  • I3. Label smoothing gives a cheap version of the regularisation that soft crowd labels provide, because it stops logit gaps from growing without limit.
  • I4. Calibration is a property of the data, not of the model: ECE and JSD both get worse under genre or domain shift.
  • I5. Evaluate against a distribution of human answers (JSD against many votes) rather than a single gold label.

For Mova

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

  • I1 (tree): Calibrated arc and label probabilities in gparse could flag attachments that annotators would genuinely dispute, so these get kept as alternatives instead of being forced into one tree.
  • I2 (tokens): Fit a temperature on the tagger's and lemmatizer's dev outputs so that downstream steps receive trustworthy POS and morphology probabilities without any retraining.
  • I3 (learn): When training on silver data from witness parsers, smooth the targets, or better, use the witnesses' vote split as the soft target, so the student is not over-confident on contested items.
  • I4 (learn): Report calibration error separately for each text genre (fables, school tests, maths problems) as part of evaluation, and recalibrate per genre instead of assuming one global setting.
  • I5 (answer): For school-test or fable QA items that are open to interpretation, compare the reader's answer distribution with several human answers, so that reasonable hedging is not counted as an error.

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

🎓The University of MelbourneMelbourne · Australia60 works · 21,199 citations🏢Huawei TechnologiesShenzhen · China29 works · 9,991 citations

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

🧭 Topics 1 Reasoning · 510 works

Cite

Cite as Yuxia Wang, Minghan Wang, Yimeng Chen, Shimin Tao, Jiaxin Guo, Chang Su, Min Zhang and Hao Yang. 2022. Capture Human Disagreement Distributions by Calibrated Networks for Natural Language Inference. Findings of the Association for Computational Linguistics: ACL 2022.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.findings-acl.120

@inproceedings{wang2022capture,
  title = {{Capture Human Disagreement Distributions by Calibrated Networks for Natural Language Inference}},
  author = {Yuxia Wang, Minghan Wang, Yimeng Chen, Shimin Tao, Jiaxin Guo, Chang Su, Min Zhang and Hao Yang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2022},
  year = {2022},
  doi = {10.18653/v1/2022.findings-acl.120},
  url = {https://doi.org/10.18653/v1/2022.findings-acl.120}
}

References

15 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. our card
  2. Dominik Maria Endres and Johannes E. Schindelin. 2003. A new metric for probability distributions. IEEE Trans. Inf. Theory, 49(7):1858–1860. ↗
  3. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330. PMLR. ↗
  4. Weizhi Li, Gautam Dasarathy, and Visar Berisha. 2020. Regularization via structural label smoothing. In The 23rd International Conference on Artificial Intelligence and Statistics, AISTATS 2020, 26-28 August 2020, Online [Palermo, Sicily, Italy], volume 108 of Proceedings of Machine Learning Research, pages 1453–1463. PMLR. ↗
  5. Michal Lukasik, Srinadh Bhojanapalli, Aditya Krishna Menon, and Sanjiv Kumar. 2020. Does label smoothing mitigate label noise? In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 6448–6458. PMLR. ↗
  6. Johannes Mario Meissner, Napat Thumwanit, Saku Sugawara, and Akiko Aizawa. 2021. Embracing ambiguity: Shifting the training target of nli models. arXiv preprint arXiv:2106.03020. ↗ our card
  7. Rafael Müller, Simon Kornblith, and Geoffrey Hinton. 2019. When does label smoothing help? In 33rd Conference on Neural Information Processing Systems (NeurIPS 2019). find
  8. Yixin Nie, Xiang Zhou, and Mohit Bansal. 2020. What can we learn from collective human opinions on natural language inference data? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9131–9143, Online. Association for Computational Linguistics. our card
  9. Giorgio Patrini, Frank Nielsen, Richard Nock, and Marcello Carioni. 2016. Loss factorization, weakly supervised learning and label noise robustness. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 708–717. JMLR.org. ↗
  10. Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent disagreements in human textual inferences. Trans. Assoc. Comput. Linguistics, 7:677–694. ↗ our card
  11. Dataset SNLI MNLI Alexey Romanov and Chaitanya Shivade. 2018. Lessons from natural language inference in the clinical domain. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 1586–1596. Association for Computational Linguistics. ↗ our card
  12. Lingxi Xie, Jingdong Wang, Zhen Wei, Meng Wang, and Qi Tian. 2016. Disturblabel: Regularizing CNN on the loss layer. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4753– 4762. IEEE Computer Society. find
  13. Shujian Zhang, Chengyue Gong, and Eunsol Choi. 2021. Capturing label distribution: A case study in NLI. CoRR, abs/2102.06859. ↗
  14. Xiang Zhou, Yixin Nie, and Mohit Bansal. 2021. Distributed NLI: learning to predict human opinion distributions for language reasoning. CoRR, abs/2104.08676. Used By \#Samples \#Labels Train UNLI 55,517 1r Dev. UNLI ChaosNLI AmbiS 3,040 1,514 9,842 1r 100 5 Test UNLI AmbiS 3,040 9,824 1r 5 Dev. M. ChaosNLI AmbiM 1,599 9,815 100 5 Dev. Mism. AmbiM 9,832 *Table 9: Datasets of ambiguous labels. “1r” denotes a regression label in the [0,1], refer to AmbiNLI original paper.* Gabriel Pereyra, George Tucker, Jan Chorowski, Łukasz Kaiser, and Geoffrey Hinton. 2017. Regularizing neural networks by penalizing confident output distributions. In The International Conference on Learning Representations (ICLR 2017). ↗ our card
  15. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112–1122. Association for Computational Linguistics. Split *Figure 4: Reliability diagram for AmbiMNLI validation set.* find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2022-capture-human-disagreement-distributions-calibrated/