Adversarial Training for Weakly Supervised Event Detection
Paper · 2019 · Xiaozhi Wang, Xu Han, Zhiyuan Liu et al.
Summary
Event detection (ED) means finding trigger words and assigning event types, and it usually depends on costly manual annotation. Existing weakly supervised alternatives label data automatically with hand-crafted rules and knowledge bases. The resulting data is noisy, covers few topics, and is skewed toward some of them. Wang et al. address this in two steps. First, a trigger-based latent instance discovery strategy treats every sentence containing a known trigger word as a possible event instance, which yields a broad candidate set without manual rule design. Second, a GAN-style scheme separates reliable from unreliable data. A discriminator learns to recognize correctly labeled, informative instances, and a generator picks the most confusing unreliable instances to mislead it. Once the two reach balance, the discriminator is more robust to noise and the generator surfaces useful new examples. With DMCNN and DMBERT encoders, the method beats prior systems in both distantly supervised and semi-supervised settings. This shows that ED training data can be expanded cheaply while keeping quality high.
Key points
- Encoders are CNN and BERT with dynamic multi-pooling around the candidate trigger, named DMCNN and DMBERT. Code and data are available at github.com/thunlp/Adv-ED.
- In the semi-supervised setting, ACE-2005 triggers (599 documents, 8 types, 33 subtypes) seed candidate harvesting from the New York Times corpus. The filtered data then extends the ACE-2005 training set.
- On ACE-2005, DMBERT+Boot reaches 75.1 F1 (DMBERT: 74.6), the best among compared systems including GCN-ED (73.1). DMCNN+Boot gains 1.7 points over DMCNN (70.8 vs 69.1).
- A manual check had four annotators judge 150 sampled instances. Precision was 91.7% after the first iteration and 87.5% after the second, versus 88.9% for Chen et al. (2017) and 91.0% for Zeng et al. (2018).
- In distantly supervised experiments, the adversarial method outperformed all baselines on every reported metric, including +MIL and +NA variants.
📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 9,085 worksArtificial Intelligence · 8,888 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksComputer Vision and Pattern Recognition · 1,282 works
🧭 Topics 1Research subjects and languages the work is about.Information extraction · 109 works
Cite
Cite as Xiaozhi Wang, Xu Han, Zhiyuan Liu, Maosong Sun and Peng Li. 2019. Adversarial Training for Weakly Supervised Event Detection. 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).
@inproceedings{wang2019adversarial,
title = {{Adversarial Training for Weakly Supervised Event Detection}},
author = {Xiaozhi Wang, Xu Han, Zhiyuan Liu, Maosong Sun and Peng Li},
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/n19-1105},
url = {https://doi.org/10.18653/v1/n19-1105}
}References
- Event-Type: Justice Subtype: Sue In ACE-2005 find
- Dell sued for ”bait and switch” and false promises. Discovered 1. The lawyers for the four former state officials who have been sued told the jurors . . . 2. But litigation held up the project until . . . . find
- David Ahn. 2006. The stages of event extraction. In Proceedings of ACL Workshop on Annotating and Reasoning about Time and Events, pages 1–8. *Table 5: The examples with highlighting triggers.*
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2019-adversarial-training-weakly-supervised-event/