DEGREE: A Data-Efficient Generation-Based Event Extraction Model

Anonymous

Abstract: Due to the high cost of human annotations, learning a data-efficient event extraction model that can be trained with only a few labeled examples has become a crucial challenge. In this paper, we focus on low-resource end-to-end event extraction. We propose DEGREE, a model that formulates event extraction as a conditional generation problem. Given a passage and a manually designed prompt, DEGREE learns to summarize the event happening in the passage into a natural sentence that follows a predefined pattern. The final event structure predictions are then extracted from the generated sentence with a deterministic algorithm. DEGREE has the following advantages to learn well with less training data. First, with our design of prompts, DEGREE obtains semantic guidance by leveraging label semantics and thus better captures the argument roles. In addition, the proposed model is capable of using additional weakly-supervised information, such as the description of events. Finally, learning triggers and argument roles in an end-to-end manner encourages the model to better utilize the shared knowledge and dependencies between them. Our experimental results and ablation studies demonstrate the strong performance of DEGREE for low-resource event extraction.

Paper Type: long