A machine-learning based framework for detection of fake political speech

Abstract

Daily news is one of the primary needs of modern society to keep in touch with the world. Unfortunately, social media platforms have notably become a politicians’ tool for spreading propaganda campaigns and disparage opponents, which leads to side effects such as amplifying social discord. In order to thwart fake news, independent journalists have maintained a fact-checking organization and shared their checking results on political speeches on their website, which has raised public awareness for upholding democratic values. Meanwhile, researchers have proposed various types of machine-learning and deep-learning based approaches as well as linguistic based approaches by using various types of information for the detection of fake news. Some of them have shown promising results on the detection of fake news. However, they focused on the detection of hoaxes, hateful speech, attractive headlines, political astroturfs, and satirical news or posts. In this paper, we propose an automated framework for the detection of fake political speech. It uses different classification methods for extracting features from political speech statement and its metadata including speech subject, location, speaker’s profile, speaker’s credibility, and speech context information. The features are then used to train a machine learning model with automatic feature selection and parameter tuning. On the Liar dataset, our trained Support Vector Machine (SVM) model has achieved 74% detection accuracy. The evaluation results show that our framework is effective in the detection of fake political speech.