Application Of Text Mining Techniques For Pattern Recognition In Distress Analysis Interview Corpus DAIC-WOZ

Abstract

The WHO defines depression as a frequent mental disorder, in which the individual experiences guilt, loss of self-esteem, sleep problems, poor concentration, in consolation, permanent melancholy, lack of interest, changes in appetite and fatigue (WHO, 2022).

In this paper, we generate predictive models by making use of data mining and machine learning techniques.

The data used for this research corresponds to DAIC-WOZ (University of Southern California, 2019), a real world data set provided by the University of Southern California, which has clinical interviews in different formats: audio, video and questionnaire responses. We use different vectorization techniques (TF/IDF and BERT Vectorizer) and   apply different supervised learning techniques and Deep learning, namely: BERT, Decision Tree, Logistic Regression for global features, and also the combined techniques, BERT/Logistic Regression and Decision Tree/Logistic Regression.  We use accuracy metric to assess the quality of the models obtained.

We identify that the BERT approach has a good performance over Logistic Regression model. Deep learning opens the doors to work with new deep learning and PLN techniques to analyze structured and unstructured information.