Abstract

ABSTRACT There is a certain discrepancy between real-world events and their representations in text (linguistic events or event mentions). The event mentions often refer to future or hypothetical events that have not actually occurred or whose occurrence is uncertain. In this paper we address the problem of predicting event factuality in Croatian texts using supervised machine learning. For each event mention, we aim to predict its polarity (whether the denoted event has actually happened) and its certainty (the level of confidence that the denoted event has happened). We use only lexically-based features, in order to investigate how well this problem may be addressed for a resource-poor language such as Croatian. Our preliminary results suggest that while predicting event polarity using only lexicallybased features is feasible, predicting event certainty mandates the use of more sophisticated features.