Elamur Research⌕
☰
home / papers / 2019 / Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets

Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets

Paper · 2019 · Nelson F. Liu, Roy Schwartz and Noah A. Smith

Summary

Challenge (adversarial) datasets often cause sharp drops in model accuracy. A drop alone does not show whether the model architecture is limited or whether the original training data simply lacked certain patterns. Liu, Schwartz and Smith address this with "inoculation by fine-tuning." A model trained on the standard dataset continues training on a small sample of challenge examples. Training halts after five consecutive epochs without any gain in accuracy on the original development set. The model is then re-tested on both test sets, and one of three outcomes follows:

  • Outcome 1: the gap closes, which points to a data gap.
  • Outcome 2: nothing changes, which points to a model limitation.
  • Outcome 3: original performance drops, which suggests the two data distributions conflict.

The authors tested five NLI stress tests from Naik et al. using ESIM and decomposable attention trained on MultiNLI. They also tested Adversarial SQuAD using BiDAF and QANet, with fine-tuning sets of 50 to 1K examples. All three outcomes appeared. This shows that challenge sets differ widely in what they actually diagnose. The method is cheap, task-independent and useful for interpreting robustness benchmarks.

Key points

  • Word overlap and negation (Outcome 1): about 100 challenge examples, under 0.03% of the MultiNLI training data, let ESIM recover over 90% of the gap. Original accuracy stays intact.
  • Spelling errors and length mismatch (Outcome 2): neither model improves much. A character-aware ESIM variant behaves the same way, which hints at a weakness beyond word representations.
  • Numerical reasoning (Outcome 3): ESIM closes the whole gap with 100 examples and DA with 750, but original accuracy falls. The data turns out to be shallow:
    • a 3-rule baseline exceeds 80%;
    • 35% of hypotheses contain "more/less than," and 95% of those are labeled neutral.
  • Adversarial SQuAD (Outcome 3): 400 examples let BiDAF close more than 60% of the gap, at roughly a 7% cost on original SQuAD; QANet shows the same pattern. The likely cause is that the distractor sentence always sits at the paragraph's end.
  • Limitation: the method is uninformative for label-skewed challenge sets, such as the antonym test, where every example is labeled contradiction. Code is publicly released.

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

🎓University of WashingtonSeattle · United States263 works · 180,581 citations🔬Allen Institute for Artificial IntelligenceSeattle · United States219 works · 89,712 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

Cite

Cite as Nelson F. Liu, Roy Schwartz and Noah A. Smith. 2019. Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets. 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).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n19-1225 · arXiv 1904.02668

@inproceedings{liu2019inoculation,
  title = {{Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets}},
  author = {Nelson F. Liu, Roy Schwartz and Noah A. Smith},
  booktitle = {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)},
  year = {2019},
  doi = {10.18653/v1/n19-1225},
  eprint = {1904.02668},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/n19-1225}
}

References

18 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In Proc. of ICLR. find
  2. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proc. of EMNLP. find
  3. Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced LSTM for natural language inference. In Proc. of ACL. find
  4. Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proc. of EMNLP. find
  5. Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2017. Bidirectional attention flow for machine comprehension. In Proc. of ICLR. our card
  6. Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proc. of NAACL. find
  7. Adams Wei Yu, David Dohan, Minh-Thang Luong, Rui Zhao, Kai Chen, Mohammad Norouzi, and Quoc V. Le. 2018. Qanet: Combining local convolution with global self-attention for reading comprehension. In Proc. of ICLR. our card
  8. Zhengli Zhao, Dheeru Dua, and Sameer Singh. 2018. Generating natural adversarial examples. In Proc. of ICLR. find
  9. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2018. Hotflip: White-box adversarial examples for text classification. In Proc. of ACL. our card
  10. Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. 2018. AllenNLP: A deep semantic natural language processing platform. In Proc. of NLP-OSS. our card
  11. Max Glockner, Vered Shwartz, and Yoav Goldberg. 2018. Breaking NLI systems with sentences that require simple lexical inferences. In Proc. of ACL. our card
  12. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R. Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language inference data. In Proc. of NAACL. find
  13. Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proc. of EMNLP. our card
  14. Pramod Kaushik Mudrakarta, Ankur Taly, Mukund Sundararajan, and Kedar Dhamdhere. 2018. Did the model understand the question? In Proc. of ACL. our card
  15. Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Penstein Rosé, and Graham Neubig. 2018. Stress test evaluation for natural language inference. In Proc. of COLING. find
  16. Ankur P. Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In Proc. of EMNLP. our card
  17. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proc. of EMNLP. our card
  18. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Semantically equivalent adversarial rules for debugging NLP models. In Proc. of ACL. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2019-inoculation-fine-tuning-method-analyzing/