Elamur Research⌕
☰
home / papers / 2018 / Higher-Order Coreference Resolution with Coarse-to-Fine Inference

Higher-Order Coreference Resolution with Coarse-to-Fine Inference

Summary

First-order coreference systems score only pairs of mentions. Each link is decided on its own, so a cluster can be made of links that each look plausible but don't fit together. In the paper's example, an ambiguous "you" connects a singular "I" with a plural "all of you". The authors extend the end-to-end span-ranking model of Lee et al. (2017) with a differentiable approximation of higher-order inference. The predicted antecedent distribution acts as attention and produces an expected antecedent vector. A learned gate merges this vector into each span's representation, and the spans are then rescored. Repeating this step is expensive, so they add a cheap bilinear "coarse" factor. It ranks every candidate antecedent and passes only the best few to the costly "fine" scorer, and the whole model is trained with a single objective. On the English CoNLL-2012 (OntoNotes) data the system reaches 73.0 average F1, a new state of the art at the time, while scoring far fewer antecedents per span. The work shows that a model can enforce consistent clusters and still be efficient.

Key points

  • Baseline: Lee et al. (2017) with ELMo embeddings and tuned hyperparameters, scoring 72.3 average F1 (MUC, B³, CEAFφ4) on the test set. Adding coarse-to-fine pruning gives 72.6, and adding second-order inference gives 73.0.
  • The coarse-to-fine model sends only 50 antecedents per span to the fine scorer, compared with the 250 nearest used by the distance heuristic. On the dev set, cutting K from 250 to 50 costs under 0.2 F1 with coarse-to-fine pruning but almost 5 F1 with the heuristic.
  • Using the bilinear score on its own instead of the full pairwise scorer loses more than 3 F1. It works as a pre-filter, not as the only scorer.
  • Second-order inference adds 0.8 F1 on the dev set, and a third iteration adds just 0.1. The gain comes mainly from higher precision, while coarse-to-fine pruning mainly raises recall.
  • The code is public at github.com/kentonl/e2e-coref.

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

Fields: Linguistics & languages

Topics: Coreference resolution

Cited by

240 papers of our library cite it (of the 140673 whose bibliographies we read); 82 of them have a page here, 158 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n18-2108 · arXiv 1804.05392

@inproceedings{lee2018higher,
  title = {{Higher-Order Coreference Resolution with Coarse-to-Fine Inference}},
  author = {Lee, Kenton and He, Luheng and Zettlemoyer, Luke},
  booktitle = {Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)},
  year = {2018},
  doi = {10.18653/v1/n18-2108},
  eprint = {1804.05392},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/n18-2108}
}

References

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

  1. Eric Bengtson and Dan Roth. 2008. Understanding the value of features for coreference resolution. In EMNLP. find our card
  2. Anders Björkelund and Jonas Kuhn. 2014. Learning structured perceptrons for coreference resolution with latent antecedents and non-local features. In ACL. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lee-2018-higher-order-coreference-resolution-coarse/