A Methodology for the Automatic Annotation of Factuality in Spanish

Resumen

In the last decade, factuality has undeniably been an area of growing interest in Natural Language Processing. This paper describes a rule-based tool to automatically identify the factual status of events in Spanish text, understood with respect to the degree of commitment with which a narrator presents situations. Factuality is represented compositionally, considering the following semantic categories: commitment, polarity, event structure, and time. In contrast with neural machine learning approaches, this tool is entirely based on manually created lexico-syntactic rules that systematize semantic and syntactic patterns of factuality. Thus, it is able to provide explanations for automatic decisions, which are very valuable to guarantee accountability of the system. We evaluate the performance of the system by comparison with a manually annotated Gold Standard, obtaining results that are comparable, if not better, to machine learning approaches for a related task, the FACT 2019 challenge at the IBERLEF evaluation forum.