Multilingual Models for Checkworthy Social Media Posts Detection

Anonymous

Abstract: This work presents an extensive study of transformer-based NLP models, dedicated to detection of social media posts containing verifiable factual claims and harmful claims. The document summarizes activities carried out during the pipeline execution, which led to the design of the NLP models for post detection. These activities included dataset collection, dataset pre-processing, architecture selection, setting up the experiments, model training (fine-tuning), model testing and implementation. Comprehensive analysis of various models was conducted. Special attention was focused on multilingual models, which are capable of processing English social media posts and simultaneously posts of low-resource languages, like: Polish, Czech, Slovak and Bulgarian. The obtained results were validated with state-of-the-art models and the comparison proved the robustness of the created models. The work's novelty consists in multi-label, multilingual classification models, which can efficiently perform simultaneous detection of harmful posts and of social media posts containing verifiable factual claims.

Paper Type: long

Research Area: NLP Applications