Detecting offensive language by integrating multiple linguistic phenomena

Abstract

Social media have grown into the prirnary means of communicating between people, allowing users to have conversations and share their opinions. The rise in digital social connections has led to the dissemination of harmful communicatton. The Natural Language Processing arises for the development of computational systems to interpret human language. Giving computers this skill offers a plethora of benefits, including the potential to moderate harmful conduct on social media. This thesis relies on advanced methods based on transfer learning to tackle the offensive language detection problem. We have generated appropriate resources to enable us to train Machine Learning systems, particularly for Spanish , for which we discovered a significant lack of resources. Moreover, we have identified different linguistic phenomena that could occur in the expression of offensive language and proposed a novel methodology that relies on integrating these phenomena into a Multi-Task Learning system to detect more accurate this problem.