Biomedical Event Extraction With Dual Decomposition

Bio-molecular events concern the detailed behavior of bio-molecules. When extracting bio-molecular event, for each event, its text trigger, class, and arguments are extracted. In the biomedical domain, events can be arguments of other events, resulting in a nested structure, such as regulation events.This thesis describes a system for extracting complex bio-molecular events from biomedical literature with dual decomposition, which jointly extracts bio-molecular events. This thesis is characterized by a wide array of features based on dependency parse graphs and the use of word embedding to leverage semantic and syntactic information. The system mainly consists of four components:trigger recognition, argument detection, jointly inference with dual decomposition and semantic post-processing. By separating trigger recognition from argument detection, this system can use methods familiar from named entity recognition to tag words as event triggers tached with confidence score; argument detection is to predict each trigger-trigger or trigger-entity pair whether it corresponds to an actual event argument tached with confidence score. Both steps can thus be approached as classification tasks with online aglorithm. Then we jointly infer with dual decomposition. At last, a rule-based post-processing step is used to refine the output in accordance with the restrictions of event arguments.The proposed system leverages semantic and syntactic information in trigger detection and argument detection, rusulting in better performance in both trigger detection and argument detection, which improves the overall performance. On the development set of BioNLP’09, the system achieves an F-score of59.77%on the primary task, which is0.96%higher than the best system. On the development set of BioNLP’11, the F-score is57.08%,0.35%higher than the best published result. On the test set of BioNLP’13, F-score is53.19%, outperforming2.22%than the best published result.