An adaptive model for sequential labeling systems

There are many levels in the task of information extraction. Level 1 deals with named entities such as PERSON, ORG, DATE, etc. Level 2 concerns the role played by the named entities wrt a specific event. For instance, in a management change event, a PERSON might be either the new coming person to the company or the leaving one. Building learning models for event extraction without considering the different levels is completely misleading. In this paper, the reasons for considering these levels are explained, and an adaptive model for event extraction is proposed. It could be applied on any sequence labeleling system, e.g., CRF-based classifier, RNN, LSTM, etc. The experimental results show that the adaptive model outperforms the direct model in terms of efficiency and gives comparable results compared to GLA2E, an expert’s pattern based event extractor.