Exploiting a More Global Context for Event Detection Through Bootstrapping

Over the last few years, neural models for event extraction have obtained interesting results. However, their application is generally limited to sentences, which can be an insufficient scope for disambiguating some occurrences of events. In this article, we propose to integrate into a convolutional neural network the representation of contexts beyond the sentence level. This representation is built following a bootstrapping approach by exploiting an intra-sentential convolutional model. Within the evaluation framework of TAC 2017, we show that our global model significantly outperforms the intra-sentential model while the two models are competitive with the results obtained by TAC 2017 participants.