Detection of Depression Using Depression Indicators

Abstract

This article addresses Task ”2a” of the MentalRiskES 2023 competition, which focuses on detecting depression in Telegram texts. The approach involves creating eight specialized corpora based on the PHQ-9 questionnaire to fine-tune pre-trained BERT models for identifying signs of depression in messages. Each corpus is designed to represent the different indicators of the PHQ-9 questionnaire, using three classes: positive (texts that reflect an indicator), negative (texts opposing the indicator), and neutral (texts unrelated to the indicator). These classes allow for the creation of clear and representative examples of texts associated with each indicator. The pre-trained BERT models, specialized for each indicator, evaluate the texts of each Telegram user to generate a vector that will feed into a multilayer perceptron (MLP) neural network for final classification. The results achieved a macro-F1 score of 0.77 with proposed model Ind-bert-base-spanish, surpassing the best result in the competition by 5 %. This performance highlights the effectiveness of combining advanced natural language processing techniques, such as BERT models and MLP networks, to address mental health challenges in digital communication.