Hierarchical Modular Event Detection Based on Dependency Graph

Event Detection (ED) is a very important sub-task in the field of information extraction, which studies how to correctly identify the trigger words that trigger event generation from unstructured text containing event information. We can regard ED as a token-based multi-classification task and sequence labeling task. However, in the previous methods, the ED task is performed for fine-grained types of events, ignoring the more abstract information of coarse-grained event types, which leads to missing conceptual semantic information about the class hierarchy of events. We propose a new ED method (Hierarchical Modular Event Detection Based on Dependency Graphs, HMED) in this paper. First, we implement dynamic modeling of multi-order dependency label information between words, which is used to generate the fine-grained representations of event types. Then we design an upper-level conceptual module based on the characteristics of the ACE corpus to compute the coarse-grained representations of event types and fuse fine-grained and coarse-grained event conceptual semantic information through global attention. On the widely used ACE2005 corpus, our hierarchical module can significantly improve the performance when compared with the most current state-of-the-art results.