Data mining; Fake news; Natural language processing; Quantitative linguistics; Text classification

In this contribution, we present an introductory study on language credibility assessment via semantic attributes.To check whether a given piece of news is legitimate or fake, one can verify its claims, or analyse the language.We focus on the latter, employing the tools of machine learning. We construct a synthetic set of 30 semanticattributes designed to categorise texts according to their credibility. Using the annotated dataset Verifee ofCzech news, these attributes are evaluated for each text in the dataset by a large language model, utilising itscapability to understand the context. The ability of the attributes to discriminate credible texts from the others ismeasured by statistical means on a training dataset, and their weights are set accordingly. The final evaluation isperformed on a validation dataset. In addition to a higher accuracy achieved compared to a fine tuned LLM-basedclassification, the interpretability of our classification approach is guaranteed since the set of attributes comesfrom the natural language.

In this contribution, we present an introductory study on language credibility assessment via semantic attributes.To check whether a given piece of news is legitimate or fake, one can verify its claims, or analyse the language.We focus on the latter, employing the tools of machine learning. We construct a synthetic set of 30 semanticattributes designed to categorise texts according to their credibility. Using the annotated dataset Verifee ofCzech news, these attributes are evaluated for each text in the dataset by a large language model, utilising itscapability to understand the context. The ability of the attributes to discriminate credible texts from the others ismeasured by statistical means on a training dataset, and their weights are set accordingly. The final evaluation isperformed on a validation dataset. In addition to a higher accuracy achieved compared to a fine tuned LLM-basedclassification, the interpretability of our classification approach is guaranteed since the set of attributes comesfrom the natural language.