Tensor Embeddings for Content-Based Misinformation Detection with Limited Supervision

Abstract

Web-based technologies like social media have become primary news outlets for many people in recent years. Considering the fact that these digital outlets are extremely vulnerable to misinformation and fake news which may impact a user’s opinion toward social, political, and economic issues, the necessity of robust and efficient approaches for misinformation detection task comes to light more than ever. The majority of misinformation detection approaches previously proposed leverage manually extracted features and supervised classifiers which require a large number of labeled data which is often infeasible to collect in practice. To meet this challenge, in this work we propose a novel strategy mixing tensor-based modeling of article content and semi-supervised learning on article embeddings for the misinformation detection task which requires very few labels to achieve state-of-the-art results. We propose and experiment with three different article content modeling variations which target article body text or title, and enable meaningful representations of word co-occurrences which are discriminative in the downstream news categorization task. We tested our approach on real world data and the evaluation results show that we achieve 75% accuracy using only 30% of the labeled data of a public dataset while the previously proposed and published SVM-based classifier results in 67% accuracy. Moreover, our approach achieves 71% accuracy on a large dataset using only 2% of the labels. Additionally, our approach is able to classify articles into different fake news categories (clickbait, bias, rumor, hate, and junk science) by only using the titles of the articles, with roughly 70% accuracy and 30% of the labeled data.

The authors contributed equally to this paper.