guments corresponding to locations and sites con- sidered in Task 2.
Event Extraction System
Figure 1: Event Extraction.
sults shows that performance against gold stan- dard annotations is not always correlated with event extraction performance. We further find that the dependency types and overall structures employed by the different dependency representa- tions have specific advantages and disadvantages for the event extraction task.
Bio-molecular Event Extraction
For evaluation, we apply the system of Miwa et al. (2010b). The system was originally developed for finding core events (Task 1) using the native out- put of the Enju and GDep parsers. The system consists of three supervised classification-based modules: a trigger detector, an event edge detec- tor, and a complex event detector. The trigger detector classifies each word into the appropriate event types, the event edge detector classifies each edge between an event and a candidate participant into an argument type, and the complex event de- tector classifies event candidates constructed by all edge combinations, deciding between event and non-event. The system uses one-vs-all sup- port vector machines (SVMs) for classification.
In this study, we adopt the event extraction task defined in the BioNLP 2009 Shared Task (Kim et al., 2009) as a model information extraction task. Figure 1 shows an example illustrating the task of event extraction from a sentence. The shared task provided common and consistent task defi- nitions, data sets for training and evaluation, and evaluation criteria. The shared task defined five simple events (Gene expression, Transcription, Protein catabolism, Phosphorylation, and Local- ization) that take one core argument, a multi- participant binding event (Binding), and three reg- ulation events (Regulation, Positive regulation, and Negative regulation) used to capture both bi- ological regulation and general causation. The participants of simple and Binding events were specified to be of the general Protein type, while regulation-type events could also take other events as arguments, creating complex event structures. We consider two subtasks, Task 1 and Task 2, out of the three defined in the shared task. Task 1 focuses on core event extraction, and Task 2 involves augmenting extracted events with sec- ondary arguments (Kim et al., 2009). Events are represented with a textual trigger, type, and ar- guments, where the trigger is a span of text that states the event in text. In Task 1 the event argu- ments that need to be extracted are restricted to the core Theme and Cause roles, with secondary ar-
The system operates on one sentence at a time, building features for classification based on the syntactic analyses for the sentence provided by the two parsers as well as the sequence of the words in the sentence, including the target candi- date. The features include the constituents/words around entities (triggers and proteins), the depen- dencies, and the shortest paths among the enti- ties. The feature generation is format-independent regarding the shared properties of different for- mats, but makes use also of format-specific infor- mation when available for extracting features, in- cluding the dependency tags, word-related infor- mation (e.g. a lexical entry in Enju format), and the constituents and their head information.
We apply here a variant of the base system in- corporating a number of modifications. The ap- plied system performs feature selection removing two classes of features that were found not to be beneficial for extraction performance, and applies a refinement of the trigger expressions of events. The system is further extended to find also sec- ondary arguments (Task 2). For a detailed descrip- tion of these improvements, we refer to Miwa et al. (2010a).
Parsers and Representations
Six publicly available parsers and three depen- dency formats are considered in this paper. The