In the previous phase, co-occurrenc

In the previous phase, co-occurrence pairs of event expressions are acquired from the web corpus. Next, we extract the events’ predicate-argument structures (i.e., “[Subj] [Pre] [Obj]” instances) relied on dependency syntactic structures of event expressions.
This idea is inspired by the event extraction work in (Zhao, 2008), which recognized eventarguments mainly based on the dependency-path feature employed in a maximum entropy classifier. Compared with the well-known phrase-structure, the dependency structure of a sentence is more likely to reflect the semantic relations between contiguous or noncontiguous words. And we further find that, the dependency structure can map to our event-argument structure in limited corpus, in spite of its weakness in generic semantic role labeling tasks (Xue, 2008). So, dependencies are used as the syntactic theory of choice.
Here, we use the dependency representations proposed in (NLP Toolkit 2011) to describe our method. As shown in Figure 2, a dependency tree composes of some contiguous words in a sentence. Every word in sentence can be viewed as a node of tree. Two nodes holding a dependency relation constitute a dependency pair, in which one node is the head (e.g. “criticize”) while the other is a dependent (e.g. “student”). A dependency relation is represented by a directed arc pointing from the head to the dependent with a functional category label (e.g. “SBV”). So, a dependent can be viewed as


a child node of its head. The core word has a “HED” dependency to the virtual root, and we define it the level-1 node.