Trigger or not Trigger: Dynamic Thresholding for Few Shot Event Detection

Recent studies in few-shot event trigger detection from text address the task as a word sequence annotation task using prototypical networks. In this context, the classification of a word is based on the similarity of its representation to the prototypes built for each event type and for the “non-event” class (also named null class). However, the “non-event” prototype aggregates by definition a set of semantically heterogeneous words, which hurts the discrimination between trigger and non-trigger words. We address this issue by handling the detection of non-trigger words as an out-of-domain (OOD) detection problem and propose a method for dynamically setting a similarity threshold to perform this detection. Our approach increases f-score by about 10 points on average compared to the state-of-the-art methods on three datasets.