Factuality Classification Using Multi-facets Based on Elementary Discourse Units for News Articles

摘要

Factuality classification is used for classifying information based on degrees of certainty. It has been actively used in different applications including information extraction, textual entailment, finding semantic uncertainty and certainty, or fact extraction. In this paper, we propose an approach to improve factuality classification by analyzing information in Elementary Discourse Units (EDUs) and their relations. We use news articles as our case study since it contains information that has various degrees of certainty or factuality values (i.e., information about certain events or uncertain information from factual and opinionated information). In this work, we use five sets of facets for factuality classification, which are (1) Epistemic Modality set, (2) Subjectivity Type set, (3) Rhetorical Structure Theory (RST) set, (4) Semantic Implicative and Factive Patterns set and (5) Weasel Words set. Unlike previous works on factuality classification, we use multiple facets of EDU to examine certainty and unambiguity level of information. We performed experiments based on news articles in FactBank corpus. We evaluated our method by comparing with several state-of-the-art factuality classification techniques and the results clearly show that our method can improve accuracy in terms of precision, recall and Fl-measure as 94.1%, 93.9% and 93.9%, respectively.