Transfer Learning Across Arabic Dialects for Offensive Language Detection

Abstract

The Arabic language is spoken by a wide range of countries across Asia, however, it is a low-resource language that has a minimal number of linguistic resources. Moreover, the large spread of Arabic speakers spans several countries and cultures, which creates a complex variation in its dialectal form. This variation makes it very challenging to analyze online Arabic content, particularly for offensive language detection. We propose a transfer learning approach for dialectal Arabic offensives language detection based on the BERT model. The results demonstrate the effectiveness of the proposed system in improving the performance of some Arabic dialects, such as the Tunisian and the Egyptian.

Topics & Concepts