Major Depressive Disorder (MDD) is one of the most common and comorbid mental disorders that impacts a person’s day-to-day activity. In addition, MDD affects one’s linguistic footprint, which is reflected by subtle changes in speech production. This allows us to use natural language processing (NLP) techniques to build a neural classifier to detect depression from speech transcripts. Typically, current NLP systems discriminate only between the depressed and non-depressed states. This approach, however, disregards the complexity of the clinical picture of depression, as different people with MDD can suffer from different sets of depression symptoms. Therefore, predicting individual symptoms can provide more fine-grained information about a person’s condition. In this work, we look at the depression classification problem through the prism of the symptom network analysis approach, which shifts attention from a categorical analysis of depression towards a personalized analysis of symptom profiles. For that purpose, we trained a multi-target hierarchical regression model to predict individual depression symptoms from patient–psychiatrist interview transcripts from the DAIC-WOZ corpus. Our model achieved results on par with state-of-the-art models on both binary diagnostic classification and depression severity prediction while at the same time providing a more fine-grained overview of individual symptoms for each person. The model achieved a mean absolute error (MAE) from 0.438 to 0.830 on eight depression symptoms and showed state-of-the-art results in binary depression estimation (73.9 macro-F1) and total depression score prediction (3.78 MAE). Moreover, the model produced a symptom correlation graph that is structurally identical to the real one. The proposed symptom-based approach provides more in-depth information about the depressive condition by focusing on the individual symptoms rather than a general binary diagnosis.

In this study we address the problem of automated word stress detection in Russian using character level models and no partspeech-taggers. We use a simple bidirectional RNN with LSTM nodes and achieve the accuracy of 90% or higher. We experiment with two training datasets and show that using the data from an annotated corpus is much more efficient than using a dictionary, since it allows us to take into account word frequencies and the morphological context of the word.