Towards Fine(r)-grained Identification of Event Coreference Resolution Types

Abstract

In this paper we present initial efforts to study complex event-event relations or event coreference in the Dutch language. We are primarily interested in the event-subevent relations between event pairs, in which one event is part of another (larger) encompassing event. We detail how event coreference is defined and annotated in the Dutch ENCORE corpus, after which the corpus is used as training data. Two experiments are conducted in order to gauge the possibility of integrating event-event relationships in ongoing research on Dutch event coreference resolution. The first experiment consists in classifying the nature of the coreferential relations between two gold-standard events. This task is used as a stepping stone for the second experiment, in which we attempt to predict whether pairs of textual events corefer and, if so, what the nature of this coreferentialrelation is. Our baseline experiments consist of fine-tuning various transformer language models, after which model ensembles are created to gauge the combined performance. Initially, the best results were achieved with these ensembles. However, in a second step, we also applied self-ensembling and self-distillation techniques to improve the fine-tuning process of the existing monolingual language models. Here we demonstrated that adding a warmup parameter in the self-ensembling process and a temperature in the self-distillation algorithm can have a noticeable effect on model performance, leading to on par or better performance than the ensembles.