Research Area:  Machine Learning

Abstract:

Automatic detection of fake news is an important issue given the disproportionate effect of fake news on democratic processes, individuals and institutions. Research on automated fact-checking has proposed different approaches based on traditional machine learning methods, using hand-crafted lexical features. Nevertheless, these approaches focus on analyzing the text claim without considering the facts that are not explicitly given but can be derived from it. For example, external evidence that is retrieved from the Web as a knowledge source of the claim can provide complementary context of the claim and gives convincing reasons from it to support or oppose. Recent approaches study this deficit by incorporating supportive evidence (article) corresponding to the claim. However, these methods are either requiring substantial feature modeling, not considering several supporting evidences, or even not analyzing the language of the supporting evidence deeply. To this end, we propose an end-to-end framework, named Automatic Fake News Classification Through Self-Attention (ACT), which exploits different supportive articles to a claim which mimics manual fact-checking processes. The model presents an approach that computes the claim credibility by aggregating over the prediction generated by every claim-retrieved article pair. The article input is represented by using self-attention on the top of a bidirectional LSTM neural network. By using the self-attention, the model concentrates on nuanced linguistic features and does not require any feature engineering, lexicons or any other manual intervention. Moreover, different aspects of the supporting article are extracted into multiple vector representations. Hence, different meaningful article representations can be extracted into a two-dimensional matrix to represent the article. In the end, a majority vote over the several external articles of a given claim is applied to assess the claim-s credibility. We conduct experiments on three different real-world datasets, compare them to the state-of-the-art approaches and analyze our results, which shows performance improvements.

Keywords:  

Author(s) Name:  Nujud Aloshban

Publisher name:  ACM

DOI:  https://doi.org/10.1145/3394231.3397901

Volume Information:  Pages 115–124

Paper Link:   https://dl.acm.org/doi/abs/10.1145/3394231.3397901