Event factuality identification via hybrid neural networks

Abstract

Event factuality identification aims at determining the factual nature of events, and plays an important role in NLP. This paper proposes a two-step framework for identifying the factuality of events in raw texts. Firstly, it extracts various basic factors related with factuality of events. Then, it utilizes a hybrid neural network model which combines Bidirectional Long Short-Term Memory (BiLSTM) networks and Convolutional Neural Networks (CNN) for event factuality identification, and considers lexical and syntactic features. The experimental results on FactBank show that the proposed neural network model significantly outperforms several state-of-the-art baselines, particularly on speculative and negative events.