Rotulação de sintoma de depressão utilizando aprendizado ativo e processamento de linguagem natural

Abstract

The approach to caring for psychological disorders, especially depression and anxiety, is seen as one of the current major concerns in mental health. Social networks, such as Twitter, not only allow individuals to maintain contact and promote mutual support, but are also often the subject of research aiming to identify individuals with potential depressive profiles or to classify depressive posts. In this context, it is observed that most of the current Natural Language Processing (NLP) systems operate based on models whose success is closely linked to the quality and amount of specific training data available. However, acquiring substantial amounts of annotated data is generally a costly process, especially considering the arduous and complex nature of labeling for NLP activities. In this scenario, some approaches have been proposed in an attempt to mitigate the cost of generating quality training sets. Active Learning seeks to achieve high accuracy with a reduced number of annotated data, allowing a learning algorithm to suggest which observations the expert should label to be used in the training process. In this project, an initial investigation of active learning strategies using a model commitee in the automatic classification of a symptom of depression (sadness or depressive mood) in Twitter posts is carried out, using the SetembroBR corpus, achieving an F1 value up to 10 percentage points higher with the Consensus Entropy query compared to random sampling.