Ontology-Boosted Deep Learning for Multi-Label Classification of Arabic Abusive Messages on Social Networks
Salma Abid Azzi, Chiraz Ben Othmane Zribi · Procedia Computer Science · 2024
With the interconnection of today’s world and the exponential growth of users’ generated data on social network, an unprecedented propagation of abusive content has emerged, particularly within the Arabic-speaking community. Deep learning models have shown promise in tackling this issue, yet they demand a substantial amount of data while having a ”black-box” nature and a limited interoperability. To address these deficits, we propose to create a representative ontology of abusive messages in Arabic as a way to replicate the domain knowledge of human beings. As part of a broader study, we have already presented a deep learning multi-label model for detecting Arabic abusive messages. However, it is worth noting that this model necessitates a substantial volume of data to demonstrate its efficacy. Our objective here is to enhance our initial contribution by creating a structured and abstract knowledge representation that aims to enrich and specialize our multi-label model while remedying to the aforementioned problems. Exploiting the ontology allows the extraction of additional features that will be fed to the initial model. These features encompass not only effective semantic knowledge but also textual and linguistic forms. Experimental results showed that our model attained a 91% micro-averaged F1-score, marking a 6-point increase compared to the initial deep learning model. Furthermore, it achieved a precision of 91% and a recall of 89%.