Identifying Media Bias beyond Words: Using Automatic Identification of Persuasive Techniques for Media Bias Detection

Resumen

Detecting media bias is a challenging task due to the complexity and ambiguity of language. Current approaches are limited in their ability to generalise across regions and styles of journalism. This paper proposes a new approach that focusses on detecting rhetorical linguistic techniques rather than just analysing words or contextual representations. We compare three different systems based on different techniques for identifying media bias, including a lexical-based system, a language transformers-based system, and a cascade transformers system that relies on persuasive techniques detection. We have evaluated these systems using a Ukraine crisis news dataset and splitting it by according to the country to generate training and test sets, i.e. different sets for each country. The results of the cascade system outperforms by at least a 6% the other approaches in identifying media bias when evaluating with different countries setup. Our results suggest that models capable of detecting rhetorical and persuasive linguistic techniques are necessary to generalise media bias effectively