Evaluation and Comparison of Text Classifiers to Develop a Depression Detection Service

Abstract

Depression is a mental disorder which can become a serious health problem. This research line is focused on creating a depression detection service from text analysis. Sentiment Analysis and Natural Language Processing methods will be used to develop this service. The service will classify text in positive or negative depending on the emotions inferred from user’s input texts. In this study, five classifiers have been evaluated to determine which one fits better for this purpose. Naïve Bayes, Decision Tree, Naïve Bayes based on Bernoulli model and Maximum Entropy are the classifiers analyzed. A specific corpus which includes sentences that fit research needs was designed for testing purposes, whereas Sentiment140’s corpus was used with training purposes. An evaluation methodology formed by three tests was designed. Results show a promising starting point, but further analysis will be needed. Future related works will focus on expanding classification to get user’s mood and not only a binary classification. It will allow to include this service as input for developing personalized intervention and education mHealth systems.