Multi-layer perceptron based fake news classification using knowledge base triples

Abstract

Recent attempts to detect fake news have relied on the implementation of machine or deep learning models that have been trained on text. These models, on the other hand, are insufficient for classifying knowledge base facts or triples as fake or true. However, it is critical to assess the credibility of facts before they are included to the knowledge base. Hence, this paper suggests using a Multi-layer Perceptron to categorize a given triple as fake or true. Furthermore, extant works embed the features using either frequency or prediction based word embedding models, and thus both document and word level features are not captured. To address this issue, a data modeling approach is proposed that vectorizes the triples using two cutting-edge word embedding models, Wrod2Vec and GloVe, as well as TF-IDF and Counter Vectorizer. Empirical results show that the Multi-layer Perceptron with GloVe and count vectorizer outperforms the baseline model in terms of accuracy. Moreover, named entity tags associated with the entities, such as PERSON, add an extra feature for training the models. As a result, an algorithm that jointly extracts the triples along with named entity tags is also proposed. Experiments demonstrated that models trained on triples with named entity tags produce high accuracy. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.