Building Patterns for Biomedical Event Extraction

Generally, Event Extraction is to identify any instance of a particular class of events in a natural language text, to extract the relevant arguments of the event, and to represent the extracted information into a structured form.1Let us define Event on the binary relation between two entities for special event verbs which are predefined by biologists. Here, Entity means biomedical entities such as proteins, genes, cells, tissues, etc. According to the definition of event, our event extraction system considers only such sentences which contain at least one event verb and two entities. The training consists of two procedures (Figure 1). First, the preprocessor involves chunking, named entity tagging, dependency relation tagging and sentence normalization with special items for building patterns. Special items are entities, event verbs, non-event verbs, prepositions, relatives, conjunctions and symbols. Second, all possible candidate events are extracted from the training corpus and the corresponding patterns are also generated. At this time, we utilize the following assumptions: one event can be described by one or more patterns in the whole documents and one pattern also can be generated by one or more events. Therefore, the event and the pattern information has reciprocal relation. We use the event score (Equation 1) to measure the reliability of extracted events and the pattern score (Equation 2) to measure the reliability of extracted patterns. The scores are iteratively updated in a co-updating method. Updating the event score causes reranking of candidate events and the iteration is continued until the ranking of events is no longer changed. The result of the training is a set of generated patterns and their scores. The events in training corpus are also extracted as the by-product of the training.