Elamur Research⌕
☰
home / papers / 2018 / Deep Neural Networks for Coreference Resolution for Polish

Deep Neural Networks for Coreference Resolution for Polish

Summary

This work presents the first coreference resolver for Polish built on deep neural networks. Coreference resolution means grouping text fragments that point to the same discourse entity. The authors train a feed-forward classifier that scores pairs of mentions and returns the probability that they corefer. Mentions are then joined into chains by either a mention-based or an entity-based clustering step, with a tunable linking threshold.

The baseline input combines 50-dimensional word embeddings, mention-type flags and pair-level cues such as distance, string identity and acronym matching. The authors then test several variants:

  • 300-dimensional embeddings
  • a richer feature set taken from earlier Polish systems, including plWordNet- and Wikipedia-based knowledge features
  • a Siamese architecture
  • more negative training examples
  • global mention clustering

All systems are evaluated on the Polish Coreference Corpus with gold mentions, using MUC, B3, CEAFE and their CoNLL average. Larger embeddings and the Siamese design gave no gains, whereas the extra features clearly helped. The best result came from combining the neural model with a multi-pass sieve of rule-based resolvers, which sets a new state of the art for Polish. The resulting tool is released as Corneferencer.

Key points

  • Data: Polish Coreference Corpus (PCC), built on texts from the National Corpus of Polish and the Rzeczpospolita Corpus. It contains about 1,800 documents in 14 genres, with 540K tokens, 180K mentions and 128K clusters. The authors used a genre-balanced 90/10 train/test split.
  • Baseline network:
    • 1,147 input features feeding three ReLU hidden layers (500/300/100 units) and a sigmoid output
    • trained with Adam for 2 epochs, mini-batch 128, batch normalization and dropout 0.2
    • implemented in Keras on TensorFlow
  • Training used 426K mention pairs, balanced between positive and negative. The baseline classified about 72.27% of 40K test pairs correctly.
  • The best configuration (neural model plus sieve-based resolvers) reaches 81.23 CoNLL F1, which is 0.53 points above the previous Polish state of the art.
  • The models and the Corneferencer tool are publicly available at zil.ipipan.waw.pl/Corneferencer.

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

5 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 3 not yet.

Cite

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

@inproceedings{niton2018deep,
  title = {{Deep Neural Networks for Coreference Resolution for Polish}},
  author = {Nitoń, Bartłomiej and Morawiecki, Paweł and Ogrodniczuk, Maciej},
  booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)},
  year = {2018},
  doi = {10.63317/3zspjtf5bc5p},
  url = {https://doi.org/10.63317/3zspjtf5bc5p}
}

References

19 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Murray, D. G., Steiner, B., Tucker, P., Vasudevan, V., Warden, P., Wicke, M., Yu, Y., and Zheng, X. (2016). Tensor-Flow: A system for large… find
  2. Springer Berlin Heidelberg, Berlin, Heidelberg. Bagga, A. and Baldwin, B. (1998). Algorithms for Scoring Coreference Chains. In Proceedings of The 1 st International Conference on Language Resources and Evaluation Workshop on Linguistics Coreference, pages 563– 566. find our card
  3. Bromley, J., Guyon, I., Lecun, Y., Säckinger, E., and Shah, R. (1994). Signature Verification using a "Siamese" Time Delay Neural Network. In In NIPS Proc. find
  4. Chollet, F. et al. (2015). Keras. https://github.com/fchollet/keras. ↗
  5. Clark, K. and Manning, C. D. (2016). Deep reinforcement learning for mention-ranking coreference models. arXiv preprint arXiv:1609.08667. find our card
  6. Haghighi, A. and Klein, D. (2007). Unsupervised Coreference Resolution in a Nonparametric Bayesian Model. In John A. Carroll, et al., editors, Proceedings of the 45 th Annual Meeting of the Association of Computational Linguistics, pages 848–855. The Association for Computational Linguistics. find our card
  7. Ioffe, S. and Szegedy, C. (2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In Francis R. Bach et al., editors, ICML, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. find
  8. JMLR.org. Kingma, D. P. and Ba, J. (2014). Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6980. Kopeć, M. and Ogrodniczuk, M. (2012). Creating a Coreference Resolution System for Polish. In Nicoletta Calzolari, et al., editors, Proceedings of the 8 th International Conference on Languag… find
  9. European Language Resources Association. Luo, X. (2005). On Coreference Resolution Performance Metrics. In Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural Language Processing, HLT ’05, pages 25–32, Vancouver, Canada. Association for Computational Linguisti… find our card
  10. Nair, V. and Hinton, G. E. (2010). Rectified Linear Units Improve Restricted Boltzmann Machines. In Johannes Fürnkranz et al., editors, Proceedings of the 27th International Conference on Machine Learning (ICML-10), pages 807–814. Omnipress. Nitoń, B. and Ogrodniczuk, M. (2017). Multi-pass sieve co… find
  11. Walter De Gruyter. Ogrodniczuk, M. (2013). Translation- and Projection-Based Unsupervised Coreference Resolution for Polish. In Mieczysław A. Kłopotek, et al., editors, Proceedings of the 20 th International Conference Intelligent Information Systems, volume 7912 of Lecture Notes in Computer Science… find
  12. Springer-Verlag, Berlin, Heidelberg. Pradhan, S., Ramshaw, L., Marcus, M., Palmer, M., Weischedel, R., and Xue, N. (2011). CoNLL-2011 Shared Task: Modeling Unrestricted Coreference in OntoNotes. In Proceedings of the 15 th Conference on Computational Natural Language Learning: Shared Task, CONLL Sha… find our card
  13. Wydawnictwo Naukowe PWN, Warsaw. Recasens, M. and Hovy, E. H. (2011). BLANC: Implementing the Rand Index for Coreference Evaluation. Natural Language Engineering, 17:485–510. find
  14. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research, 15:1929–1958. find
  15. Vilain, M., Burger, J., Aberdeen, J., Connolly, D., and Hirschman, L. (1995). A Model-Theoretic Coreference Scoring Scheme. In Proceedings of the 6 th Message Understanding Conference (MUC-6), pages 45–52. find our card
  16. Wawer, A., (2015). Sentiment Dictionary Refinement Using Word Embeddings, pages 186–193. find
  17. Springer International Publishing, Cham. Wiseman, S., Rush, A. M., and Shieber, S. M. (2016). Learning Global Features for Coreference Resolution. arXiv preprint arXiv:1604.03035. Woliński, M. (2014). Morfeusz reloaded. In Nicoletta Calzolari, et al., editors, Proceedings of the 9 th International… find our card
  18. Language Resource References Maciej Ogrodniczuk. (2015). Polish Coreference Corpus. Institute of Computer Science, Polish Academy of Sciences, http://zil.ipipan.waw.pl/ ↗
  19. PolishCoreferenceCorpus. Presspublica. (2002). Rzeczpospolita Corpus. Dawid Weiss, http://www.cs.put.poznan.pl/ dweiss/rzeczpospolita. ↗

↑ back to top

This page in Markdown →