192 Y. Chen et al.
a representation of the entire sentence is needed, which can be acquired by apply-
ing a max-pooling layer over the RNNs [15]. However, in event extraction, one
sentence may contain two or more events, and these events may share the argu-
ment with different roles, as shown in Fig. 1. We apply a dynamic multi-pooling
layer in our LSTM-based framework, which can capture the valuable semantics
of a whole sentence automatically and reserve information more comprehensively
to extract events [4].
Besides the problem stated above, both of traditional approaches [1,13,18]
and deep neural network based approaches [4,20] did not model the interaction
between the candidate arguments and predict them separately. However, these
interactions are important to predict arguments. For example, in Fig. 1,ifwe
know “his wife” and “unborn son” are paralleled in the sentence, we are easy to
predict they play same role in the corresponding event. Thus we propose a tensor
layer to explore the interaction between candidate arguments automatically.
In this paper, we present a novel framework dubbed Bidirectional Dynamic
Multi-Pooling Long Short-Term Memory Tensor Neural Networks (BDLSTM-
TNNs) for event extraction, which can automatically induce valuable clues for
event extraction without complicated NLP preprocessing and predict candidate
arguments simultaneously. We propose a Bidirectional Long Short-Term Mem-
ory Network with Dynamic Multi-Pooling (BDLSTM) to extract event triggers
and arguments separately, which can capture meaningful semantics of words
with taking the context words into consideration and capture more valuable
information for event extraction within a sentence automatically. And we devise
a tensor layer, which aims to explore interaction between candidate arguments
and predict them jointly. We conduct experiments on a widely used ACE2005
event extraction dataset, and the experimental results show that our approach
outperforms other state-of-the-art methods.
2 Event Extraction Task
In this paper, we focus on the event extraction task defined in Automatic Content
Extraction
1
(ACE) evaluation, where an event is defined as a specific occurrence
involving participants. First, we introduce some ACE terminology to understand
this task more easily:
– Event mention: a phrase or sentence within which an event is described,
including a trigger and arguments.
– Event trigger: the word that most clearly expresses the occurrence of an
event.
– Event argument: an entity mention, temporal expression or value (e.g. Job-
Title) that is involved in an event (viz., participants).
– Argument role: the relationship between an argument to the event in which
it participates.
1
http://projects.ldc.upenn.edu/ace/.