مريم اورعي

In this study, with the goal of detecting signs of depression in the Persian language, a large-scale dataset was collected from the social media platform X (formerly Twitter), covering the period from 2016 to 2024. The dataset consists of two million Persian sentences, out of which 600,000 were manually reviewed an‎d relabeled. Sentences were categorized based on two labels: 1 (related to a depressed individual) an‎d 2 (not related to a depressed individual). The labeling process was conducted in three phases: extraction of phrases from psychological questionnaires (such as PHQ-9, Beck, an‎d SDS), AI-assisted label refinement, an‎d human annotation. Several deep learning models were eva‎luated, an‎d DistilBERT, a lightweight version of BERT, demonstrated superior performance due to its high accuracy an‎d lower computational requirements. For eva‎luation, a balanced an‎d ran‎domly sampled test set was created, with labels agreed upon by three trained annotators. The final model achieved an accuracy of 83%, showing strong results across eva‎luation metrics including accuracy, precision, recall, an‎d F1 score. This research contributes significantly to the development of Persian language resources in the domain of sentiment analysis an‎d digital mental health, an‎d it can serve as a foundation for future studies focused on depression detection from textual data.

بيماري‌هاي رواني , افسردگي , رفتارهاي اجتماعي , رسانه‌هاي اجتماعي

Mental Illness , Depression , Social Behaviour , Social Media

Maryam Oraei

Dr. Mohammed Reza Jahed Motlagh