A forthcoming paper reports a new state-of-the-art (SOTA) benchmark performance for fake news detection. The work, undertaking with support from the COURAGE project, describes an ensemble-based approach that combines a range of natural language processing techniques with the overall aim of identifying whether a news story is to be believed or not. One of the novelties presented is the utilisation of different automated text summarization methods for the task of fake news detection. The key idea is to be able to feed the input into transformer-based classifiers whose main weakness remains the fact that they can only process texts of a limited overall length. By first applying text summarization the idea is to distill the original text into a shorter version that still encapsulates all the signals to classify it as fake or not and which is short enough to be processed by BERT.

Philipp Hartl, a Master student at the University of Regensburg, is the lead author of this paper entitled "Applying Automatic Text Summarization for Fake News Detection” which has been accepted to appear at the 13th Language Resources Evaluation Conference (LREC) later this month. The conference will be held in Marseille (https://lrec2022.lrec-conf.org/). A pre-print can be found on arXiv (https://arxiv.org/abs/2204.01841).