A Transformer-based Deep Learning Model to Enhance Hope Speech Detection

As the volume of information and communication increases on the internet there has been great effort in the reduction of negatively focusing or abusive materials. In as much as negative communication is avoided and controlled, there is need to also look for positive content and promotion of such contents as well. The objective of this research is to study “hope speech” that manifests itself where multilingual and imbalance datasets are available. To that end, we present a machine learning-enabled system that employs the transformer-based model DeBERTa-V3-small to categorise social media texts as hope speech and non-hope speech classes after conducting a rigorous preprocessing and random oversampling to manage imbalance data. According to the evaluation results, our proposal outperforms two well-known benchmarks including BERT and Random Forest. It points to the effectiveness of transformer-based approach DeBERTa-V3 in strengthening the positive discourse and brings insightful prospects for the future studies of identifying and promoting hope speech online.