![](https://csdnimg.cn/release/download_crawler_static/86283650/bg1.jpg)
Graph Convolutional Networks with
Argument-Aware Pooling for Event Detection
Thien Huu Nguyen
Department of Computer and Information Science
University of Oregon
Eugene, Oregon 97403, USA
thien@cs.uoregon.edu
Ralph Grishman
Computer Science Department
New York University
New York, NY 10003 USA
grishman@cs.nyu.edu
Abstract
The current neural network models for event detection have
only considered the sequential representation of sentences.
Syntactic representations have not been explored in this area
although they provide an effective mechanism to directly link
words to their informative context for event detection in the
sentences. In this work, we investigate a convolutional neural
network based on dependency trees to perform event detec-
tion. We propose a novel pooling method that relies on en-
tity mentions to aggregate the convolution vectors. The exten-
sive experiments demonstrate the benefits of the dependency-
based convolutional neural networks and the entity mention-
based pooling method for event detection. We achieve the
state-of-the-art performance on widely used datasets with
both perfect and predicted entity mentions.
Introduction
Event Detection (ED) is an important information extraction
task of natural language processing that seeks to recognize
instances of specified types of events (event mentions) in
text. Each event mention is often presented within a single
sentence in which an event trigger is selected to associate
with that event mention. Event triggers are generally sin-
gle verbs or nominalizations that serve as the main words
to evoke the corresponding events. The event detection task,
more precisely stated, aims to detect event triggers and clas-
sify them into specific types of interest. For instance, con-
sider the following sentence with two words “fired”:
“The police officer who fired into a car full of teenagers was
fired Tuesday”
In this example, an ED system should be able to realize that
the first occurrence of “fired” is an event trigger of type At-
tack while the second “fired” takes End-Position as its event
type. ED is a challenging task, as an expression might evoke
different events depending on contexts (illustrated in our
previous example with the word “fired”), and the same event
might be presented in various expressions (e.g, the trigger
words “killed”, “shot”or“beat for the event type Attack).
The current state-of-the-art approach for ED employs
deep learning models in which convolutional neural net-
works (CNN) are the typical architectures (Nguyen and
Grishman 2015; Chen et al. 2015; Nguyen and Grishman
Copyright
c
2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
2016b). In the basic implementation, CNNs apply the tem-
poral convolution operation over the consecutive k-grams
1
in the sentences, attempting to generate the latent structures
that are informative for ED (Nguyen and Grishman 2015;
Chen et al. 2015). The disadvantage of such consecutive
convolution is the inability to capture the non-consecutive
k-grams that can span words far apart in the sentences.
Those non-consecutive k-grams are necessary to recog-
nize event triggers in some situations. For example, in the
example above, the non-consecutive 3-grams “officer was
fired” should be considered to correctly identify the event
type End-Position for the second word “fired”. The non-
consecutive CNN model (NCNN) in (Nguyen and Grish-
man 2016b) seeks to overcome this problem by operating the
temporal convolution over all the non-consecutive k-grams
in the sentences, leading to the state-of-the-art CNN model
for ED.
Unfortunately, due to the consideration of all possible
non-consecutive k-grams, the non-consecutive CNN archi-
tecture might model unnecessary and noisy information, po-
tentially impairing the prediction performance for ED. In
particular, NCNN utilizes the max-pooling operation to ag-
gregate the convolution scores over all the non-consecutive
k-grams. As such k grams might include irrelevant or mis-
leading sequences of words, the max-pooling might incor-
rectly focus on those k-grams and make a wrong final pre-
diction for ED. One example is the non-consecutive 3-gram
“car was fired” in the example above. In contrast to the cor-
rect 3-gram “officer was fired”, “car was fired” suggests the
event type Attack for the second word “fired”, causing the
confusion or failure of NCNN to predict the true event type
of “End-Position” in this situation.
One way to circumvent this issue for NCNN is to notice
that “police officer” is the subject of the second word “fired”
while “a car” does not have much direct connection with
the second “fired” in this example. Guided by this intuition,
in this paper, we propose to perform the convolution opera-
tion over the syntactic dependency graphs of the sentences to
perform event detection. Syntactic dependency graphs rep-
resents sentences as directed trees with head-modifier de-
pendency arcs between related words (Mcdonald and Pereira
2006; Koo, Carreras, and Collins 2008). Each word in such
1
k is often chosen to be some fixed value.
The Thirty-Second AAAI Conference
on Artificial Intelligence (AAAI-18)
5900