FAKE NEWS DETECTION SYSTEM BASED ON PROBABILISTIC SENTIMENT SCORE AND SENTENCE EMBEDDING

Abstract

Analyzing of news truthfulness is a challenging problem in today’s era because there is a massive information on the social networking sites (SNS) which turns out to be very difficult to manually analyze. Moreover, the impact of fake or negative news is tremendously huge to the internet users. In this complex field, scientists use sophisticated computer algorithms and neural network structures to examine and distinguish between the truthfulness of textual content that is distributed via various media channels. As a result, academic research related to filtering and banning fake news has been highly demanding since very recent years. Although there are some significant results and improvements made using different feature extraction methods and classification algorithms it still has some gaps to meet the important necessities to detect fake news because each method has some biases, variances and generalization errors. This research contributes to this area by using probabilistic sentiment score and sentence embedding, marks a significant advance forward in the accuracy of detecting fake news. It differs significantly from traditional approaches such as TF-IDF or bag- of-words (BOW) representation, which frequently ignore complex semantic and contextual nuances. The system first implements probabilistic sentiment model to get probabilistic sentiment score using TF-IDF, mutual information and logistic regression methods. Secondly, the system applies sentence embedding method to extract semantic and contextual feature vectors. The system finally uses Naïve Bayes and Support Vector Machine classifiers based on concatenated features (Probabilistic sentiment score and sentence embedding feature vector) for classification process. The system performs the experiments upon ISOT dataset and other fake news dataset from Kaggle. The effectiveness of the proposed method is remarkable 99% accuracy rate, which outperforms other models. Moreover, the results prove that the proposed concatenated feature is superior not only Naïve Bayes but also Support Vector Machine classifiers.