A Twitter-based Arabic Mental Health Disorder (MHD) Dataset

Sentiment classification is a dominant task in the sentiment analysis field. This task requires a huge, annotated corpus to feed into training models. Manual annotation is the optimal technique for this task, but it is a time-consuming and extensive process that is also prone to human bias. In this paper, we introduce automatic annotation for a Twitter-based Arabic Mental Health Disorder (MHD) dataset by employing transfer learning. We have utilized the existing manual annotation datasets with three cutting-edge Arabic language models. To validate the MHD dataset, we performed a manual annotation on it and calculated the inter-annotator agreement metric between the manual and proposed approaches using Cohen's Kappa statistic. According to the findings, the MHD dataset has a Cohen's Kappa of k = 0.85, which indicates a strong agreement between both annotation approaches. In addition to that, we conducted different baseline models for which we present the results.