Fact checking: detection of check worthy statements through support vector machine and feed forward neural network

Abstract

Detection of check-worthy statements is a subtask in the fact-checking process, automation of which would decrease the time and burden required to fact-check a statement. This paper proposes an approach focused on the classification of statements into check-worthy and not check-worthy. For the current paper, a dataset is constructed by consulting different fact-checking organizations. It contains debates and speeches in the domain of politics. Thus, even the ability of check worthy approach is evaluated on this domain. It starts with extracting sentence-level and context features from the sentences, and classifying them based on these features. The feature set and context were chosen after several experiments, based on how well they differentiate check-worthy statements. The findings indicated that the context in the approach gives considerable contribution in the classification, while also using more general features to capture information from the sentences. The results were analyzed by examining all features used, assessing their contribution in classification, and how well the approach performs in speeches and debates separately to detect the check worthy statements to reduce the time and burden of fact checking process.