Elamur Research⌕
☰
home / papers / 2016 / Joint Event Extraction via Recurrent Neural Networks

Joint Event Extraction via Recurrent Neural Networks

Summary

The paper tackles ACE-style event extraction: finding trigger words with their event subtypes and labeling the roles that given entity mentions play in those events. Earlier strong systems either used learned CNN features in a pipeline, where trigger errors propagate to argument classification, or ran joint structured prediction over hand-engineered discrete features, which generalize poorly to unseen words. The authors combine the strengths of both. Their model, JRNN, encodes each sentence with a bidirectional GRU. Each token's input joins a word embedding, an entity-type embedding using BIO labels, and a binary vector of dependency relations. Triggers and argument roles are then predicted jointly. Memory vectors and matrices record earlier decisions so that later predictions can depend on them. On the ACE 2005 blind test set, JRNN gets the best trigger and argument F1 among the compared systems. Its largest gains come on sentences that contain several events. The paper reports this as the first neural joint model for event extraction.

Key points

  • Task setup follows ACE 2005, with 8 event types, 33 subtypes and 36 argument roles in total. As in prior work, gold entity mentions, times and values are supplied as argument candidates.
  • JRNN reaches 69.3 F1 on triggers and 55.4 F1 on arguments.
    • On argument roles it beats DMCNN (Chen et al., 2015) by 1.9 points.
    • It beats Li et al. (2013)'s joint structured perceptron by 1.8 points on triggers and 2.7 points on arguments.
  • On sentences with multiple events, JRNN outperforms DMCNN by 13.9 F1 on triggers (64.8 vs 50.9) and 6.5 F1 on arguments (55.2 vs 48.7).
  • On single-event sentences, JRNN still leads on triggers (75.6) but trails DMCNN on arguments (50.0 vs 54.6). The authors attribute this to DMCNN's position embeddings and to the trigger–argument memory matrix having no effect when a sentence contains only one event.
  • A concatenation-based variant of CBOW (C-CBOW), which gives each context word its own weight, worked best for pretraining the word embeddings.

Interesting ideas

  • I1. Memory of earlier predictions in a joint tagger: storing past trigger and role decisions in memory vectors or matrices lets a neural model capture dependencies between outputs. Structured perceptrons used to get this from global features. The trick could carry over to any joint labeling task, such as mention detection combined with relation extraction.
  • I2. Hybrid inputs: sparse linguistic signals can sit alongside dense embeddings. Here that means a binary dependency-relation vector and entity-type embeddings. This is a cheap way to inject parser output for low-resource languages, where embeddings alone may not generalize well.
  • I3. No position features in joint decoding: when the whole sentence is labeled at once, there is no single anchor word, so the authors leave out relative-position embeddings. This is a useful check when moving from pipelined models to joint ones.
  • I4. Evaluation split by number of events: reporting scores separately for single-event and multi-event sentences reveals where a joint model actually helps. The same slicing could be applied to any structured extraction benchmark.
  • I5. Order-aware CBOW: concatenating context-word vectors instead of averaging them gives each position its own weight. This is a simple embedding tweak worth testing when training task-specific vectors.

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: Natural language processing

Cited by

234 papers of our library cite it (of the 140884 whose bibliographies we read); 15 of them have a page here, 219 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n16-1034

@inproceedings{nguyen2016joint,
  title = {{Joint Event Extraction via Recurrent Neural Networks}},
  author = {Nguyen, Thien Huu and Cho, Kyunghyun and Grishman, Ralph},
  booktitle = {Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2016},
  doi = {10.18653/v1/n16-1034},
  url = {https://doi.org/10.18653/v1/n16-1034}
}

References

31 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events. Marco Baroni, Georgiana Dinu, and Germán Kruszewski. 2014. Don’t count, predict! a systematic comparison of context-counting vs. context-predicting semantic vectors. In… find our card
  2. Yoshua Bengio, Patrice Simard, and Paolo Frasconi. 1994. Learning long-term dependencies with gradient descent is difficult. In Journal of Machine Learning Research 3. Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A neural probabilistic language model. In Journal of Ma… find
  3. Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multi-pooling convolutional neural networks. In ACL-IJCNLP. find our card
  4. Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In EMNLP. find
  5. Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. In arXiv preprint arXiv:1412.3555. find
  6. Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: deep neural networks with multitask learning. In ICML. find
  7. Felix Gers. 2001. Long short-term memory in recurrent neural networks. In PhD Thesis. find
  8. Ralph Grishman, David Westbrook, and Adam Meyers. 2005. Nyus english ace 2005 system description. In ACE 2005 Evaluation Workshop. find
  9. Prashant Gupta and Heng Ji. 2009. Predicting unknown time arguments based on cross-event propagation. In ACL-IJCNLP. find
  10. Sepp Hochreiter and Jurgen Schmidhuber. 1997. Long short-term memory. In Neural Computation. find
  11. Yu Hong, Jianfeng Zhang, Bin Ma, Jianmin Yao, Guodong Zhou, and Qiaoming Zhu. 2011. Using cross-entity inference to improve event extraction. In ACL. find our card
  12. Ruihong Huang and Ellen Riloff. 2012. Modeling textual cohesion for event extraction. In AAAI. find
  13. Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. In ACL. Rafal Józefowicz, Wojciech Zaremba, and Ilya Sutskever. 2015. An empirical exploration of recurrent network architectures. In ICML. find our card
  14. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In EMNLP. find our card
  15. Qi Li, Heng Ji, and Liang Huang. 2013. Joint event extraction via structured prediction with global features. In ACL. find our card
  16. Qi Li, Heng Ji, Yu Hong, and Sujian Li. 2014b. Constructing information networks using one single model. In EMNLP. find
  17. Shasha Liao and Ralph Grishman. 2010. Using document level cross-event inference to improve event extraction. In ACL. find our card
  18. Shasha Liao and Ralph Grishman. 2011. Acquiring topic features to improve event extraction: in pre-selected and balanced collections. In RANLP. find
  19. David McClosky, Mihai Surdeanu, and Christopher Manning. 2011. Event extraction as dependency parsing. In BioNLP Shared Task Workshop. find
  20. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. In ICLR. find
  21. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013b. Distributed representations of words and phrases and their compositionality. In NIPS. find
  22. Thien Huu Nguyen and Ralph Grishman. 2015a. Relation extraction: Perspective from convolutional neural networks. In The NAACL Workshop on Vector Space Modeling for NLP (VSM). find our card
  23. Thien Huu Nguyen and Ralph Grishman. 2015b. Event detection and domain adaptation with convolutional neural networks. In ACL-IJCNLP. find our card
  24. Siddharth Patwardhan and Ellen Riloff. 2009. A unified model of phrasal and sentential evidence for information extraction. In EMNLP. find
  25. Hoifung Poon and Lucy Vanderwende. 2010. Joint inference for knowledge extraction from biomedical literature. In NAACL-HLT. find
  26. Sebastian Riedel and Andrew McCallum. 2011a. Fast and robust joint models for biomedical event extraction. In EMNLP. find
  27. Sebastian Riedel and Andrew McCallum. 2011b. Robust biomedical event extraction with dual decomposition and minimal domain adaptation. In BioNLP Shared Task 2011 Workshop. find
  28. Sebastian Riedel, Hong-Woo Chun, Toshihisa Takagi, and Jun’ichi Tsujii. 2009. A markov logic approach to bio-molecular event extraction. In BioNLP 2009 Workshop. find
  29. Joseph Turian, Lev-Arie Ratinov, and Yoshua Bengio. 2010. Word representations: A simple and general method for semi-supervised learning. In ACL. find our card
  30. Deepak Venugopal, Chen Chen, Vibhav Gogate, and Vincent Ng. 2014. Relieving the computational bottleneck: Joint inference for event extraction with high-dimensional features. In EMNLP. find
  31. Matthew D. Zeiler. 2012. Adadelta: An adaptive learning rate method. In CoRR, abs/1212.5701. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nguyen-2016-joint-event-extraction-recurrent-neural/