A Hybrid Semantic–Sentiment Framework for Automatic Fake News Detection Using Doc2Vec and Machine Learning

Abstract

The spread of misinformation through digital plat-forms such as social media sites and news portals has posed a problem of maintaining information credibility and building trust. Manual approaches alone cannot help cope with the huge volumes of information uploaded on these platforms every day. An automatic approach for fake news detection which utilizes NLP, sentiment analysis, semantic embedding methods and several machine learning algorithms has been developed in this paper. The news headlines collected from FakeNewsNet dataset have been pre-processed via tokenization, stop-word elimination, lemmatization, and n-grams extraction. Doc2Vec approach has been employed to extract semantic vectors whereas sentiment analysis has been done with the help of VADER tool. These semantic vectors have been provided as input to various machine learning algorithms such as Logistic Regression, Linear SVM, Random Forest, Gradient Boosting, XGBoost, LightGBM, Naïve Bayes and K-Nearest Neighbor. Experimental results suggest that ensemble learning models outperform other forms of machine learning techniques. Out of all the tested algorithms, ExtraTrees performed with the highest classification accuracy (77.54%) whereas XGBoost produced the highest macro F1-Score (0.5059).