Detecting Factual and Non-Factual Content in News Articles

Abstract

News articles are a major source of facts about the current state and events of our surrounding world. However, not all news articles are equally rich in presenting the facts. In this paper, we consider the problem of detecting factual and non-factual parts in news articles. We present a comprehensive survey on the existing literature on fact classification on news articles as well as a related and more widely studied problem of subjectivity vs objectivity classification of statements. Combining these techniques and some new features we design a framework for classifying facts and non-facts in news articles. We present extensive experiments on this task using several features and combinations of those on two datasets, one of which was used for subjectivity classification in previous works. We show that standard textual dataset dependent features such as n-grams produce good results on both datasets, but more general features such as part of speech tags and entity types produce inconsistent results. We analyze the results based on the nature of the datasets to present insights on the usefulness of the features and their applicability in the classification task we are considering.