Tobb Etu at Checkthat! 2020: Prioritizing English and Arabic Claims Based on Check-Worthiness

Abstract

Misinformation has many negative consequences on our daily life. While spread of misinformation is very fast, investigating veracity of claims is slow. Therefore, we urgently need systems helping human fact checkers in the combat against misinformation. In this paper, we present our participation in check-worthiness tasks (i.e., Task 1 and Task 5) of CLEF-2020 Check That! Lab. For English Task 1, we use logistic regression with fined-tuned BERT predictions, POS tags, controversial topics and a hand-crafted word list as features. For English Task 5, we again use logistic regression with fined-tuned BERT predictions and word embeddings as features. For Arabic Task 1, we use a hybrid approach of fined-tuned BERT model with the model used for English Task 5. For the Arabic task, we use AraBert as our Bert model. In the official evaluation of primary submissions, our primary models a) ranked 3rd in Arabic Task 1 based on P@30 and shared the 1st rank with another group based on P@5, b) ranked 5th in English Task 1 based on average precision and shared the 1st rank with five other groups based on reciprocal rank, P@1, P@3 and P@5 metrics, and c) ranked 3rd in Task 5 based on average precision. Copyright © 2020 for this paper by its authors.