Event Detection: Gate Diversity and Syntactic Importance Scores for Graph Convolution Neural Networks

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5405–5411, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5405 Event Detection: Gate Diversity and Syntactic Importance Scores for Graph Convolution Neural Networks Viet Dac Lai1, Tuan Ngo Nguyen1 and Thien Huu Nguyen1,2 1 Department of Computer and Information Science, University of Oregon, Eugene, OR 97403, USA 2 VinAI Research, Vietnam {vietl,tnguyen,thien}@cs.uoregon.edu Abstract Recent studies on event detection (ED) have shown that the syntactic dependency graph can be employed in graph convolution neural net- works (GCN) to achieve state-of-the-art per- formance. However, the computation of the hidden vectors in such graph-based models is agnostic to the trigger candidate words, po- tentially leaving irrelevant information for the trigger candidate for event prediction. In addi- tion, the current models for ED fail to exploit the overall contextual importance scores of the words, which can be obtained via the depen- dency tree, to boost the performance. In this study, we propose a novel gating mechanism to filter noisy information in the hidden vec- tors of the GCN models for ED based on the information from the trigger candidate. We also introduce novel mechanisms to achieve the contextual diversity for the gates and the importance score consistency for the graphs and models in ED. The experiments show that the proposed model achieves state-of-the-art performance on two ED datasets. 1 Introduction Event Detection (ED) is an important task in Infor- mation Extraction of Natural Language Processing. The main goal of this task is to identify event in- stances presented in text. Each event mention is associated with a word or a phrase, called an event trigger, which clearly expresses the event (Walker et al., 2006). The event detection task, precisely speaking, seeks to identify the event triggers and classify them into some types of interest. For in- stance, consider the following sentences: (1) They’ll be fired on at the crossing. (2) She is on her way to get fired. An ideal ED system should be able to recog- nize the two words “fired” in the sentences as the triggers of the event types “Attack” (for the first sentence) and “End-Position” (for the second sen- tence). The dominant approaches for ED involve deep neural networks to learn effective features for the input sentences, including separate models (Chen et al., 2015) and joint inference models with event argument prediction (Nguyen and Nguyen, 2019). Among those deep neural networks, graph convo- lutional neural networks (GCN) (Kipf and Welling, 2017) have achieved state-of-the-art performance due to the ability to exploit the syntactic depen- dency graph to learn effective representations for the words (Nguyen and Grishman, 2018; Liu et al., 2018; Yan et al., 2019). However, two critical is- sues should be addressed to further improve the performance of such models. First, given a sentence and a trigger candidate word, the hidden vectors induced by the current GCN models are not yet customized for the trig- ger candidate. As such, the trigger-agnostic rep- resentations in the GCN …

Content truncated to 500 words