Modeling Skip-Grams for Event Detection with Convolutional Neural Networks

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 886–891, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Modeling Skip-Grams for Event Detection with Convolutional Neural Networks Thien Huu Nguyen Computer Science Department New York University New York, NY 10003 USA thien@cs.nyu.edu Ralph Grishman Computer Science Department New York University New York, NY 10003 USA grishman@cs.nyu.edu Abstract Convolutional neural networks (CNN) have achieved the top performance for event de- tection due to their capacity to induce the underlying structures of the k-grams in the sentences. However, the current CNN-based event detectors only model the consecutive k-grams and ignore the non-consecutive k- grams that might involve important structures for event detection. In this work, we propose to improve the current CNN models for ED by introducing the non-consecutive convolu- tion. Our systematic evaluation on both the general setting and the domain adaptation set- ting demonstrates the effectiveness of the non- consecutive CNN model, leading to the signif- icant performance improvement over the cur- rent state-of-the-art systems. 1 Introduction The goal of event detection (ED) is to locate event triggers of some specified types in text. Triggers are generally single verbs or nominalizations that evoke the events of interest. This is an important and challenging task of information extraction in natu- ral language processing (NLP), as the same event might appear in various expressions, and an expres- sion might express different events depending on contexts. The current state-of-the-art systems for ED have involved the application of convolutional neural net- works (CNN) (Nguyen and Grishman, 2015b; Chen et al., 2015) that automatically learn effective fea- ture representations for ED from sentences. This has overcome the two fundamental limitations of the tra- ditional feature-based methods for ED: (i) the com- plicated feature engineering for rich feature sets and (ii) the error propagation from the NLP toolkits and resources (i.e, parsers, part of speech taggers etc) that generate such features. The prior CNN models for ED are characterized by the temporal convolution operators that linearly map the vectors for the k-grams in the sentences into the feature space. Such k-gram vectors are ob- tained by concatenating the vectors of the k con- secutive words in the sentences (Nguyen and Grish- man, 2015b; Chen et al., 2015). In other words, the previous CNN models for ED only focus on modeling the consecutive k-grams. Unfortunately, such consecutive mechanism is unable to capture the long-range and non-consecutive dependencies that are necessary to the prediction of trigger words. For instance, consider the following sentence with the trigger word “leave” from the ACE 2005 corpus: The mystery is that she took the job in the first place or didn’t leave earlier. The correct event type for the trigger word “leave” in this case is “End-Org”. However, the previous CNN models might not be able to detect “leave” as an event trigger or incorrectly predict its type as “Movement”. This is caused by their reliance on the consecutive local k-grams such as “leave ear- lier”. …

Content truncated to 500 words