The Importance of Character-Level Information in an Event Detection Model

This paper tackles the task of event detection that aims at identifying and categorizing event mentions in texts. One of the difficulties of this task is the problem of event mentions corresponding to misspelled, custom, or out-of-vocabulary words. To analyze the impact of character-level features, we propose to integrate character embeddings, that can capture morphological and shape information about words, to a convolutional model for event detection. More precisely, we evaluate two strategies for performing such integration and show that a late fusion approach outperforms both an early fusion approach and models integrating character or subword information such as ELMo or BERT.