DEGREE: A Data-Efficient Generative Event Extraction Model

  • 2021-09-14 17:27:15
  • I-Hung Hsu, Kuan-Hao Huang, Elizabeth Boschee, Scott Miller, Prem Natarajan, Kai-Wei Chang, Nanyun Peng
  • 0


Event extraction (EE) aims to identify structured events, including eventtriggers and their corresponding arguments, from unstructured text. Most of theexisting works rely on a large number of labeled instances to train models,while the labeled data could be expensive to be obtained. In this work, wepresent a data-efficient event extraction method by formulating eventextraction as a natural language generation problem. The formulation allows usto inject knowledge of label semantics, event structure, and outputdependencies into the model. Given a passage and an event type, our modellearns to summarize this passage into a templated sentence in a predefinedstructure. The template is event-type-specific, manually created, and containsevent trigger and argument information. Lastly, a rule-based algorithm is usedto derive the trigger and argument predictions from the generated sentence. Ourmethod inherently enjoys the following benefits: (1) The pretraining of thegenerative language models help incorporate the semantics of the labels forgenerative EE. (2) The autoregressive generation process and our end-to-enddesign for extracting triggers and arguments force the model to capture thedependencies among the output triggers and their arguments. (3) The predefinedtemplates form concrete yet flexible rules to hint the models about the validpatterns for each event type, reducing the models' burden to learn structuresfrom the data. Empirical results show that our model achieves superiorperformance over strong baselines on EE tasks in the low data regime andachieves competitive results to the current state-of-the-art when more databecomes available.


Quick Read (beta)

loading the full paper ...