Ανίχνευση Κατάθλιψης από Ψυχοθεραπευτικές Συνεδρίες με Μηχανική Μάθηση και Νευρωνικά Δίκτυα

Abstract

Depression is a common and serious medical illness that affects the way affected people feel, think and act. It can lead to a variety of physical, social and emotional problems and can decrease people’s ability to work and function. Fortunately, there is a plethora of available medications able to treat depression. However, it is crucial that clinicians diagnose early the signs of the disorder and prescribe the suitable treatment. The diagnosis procedure is not an easy task. In many cases, the symptoms may not be indicative and thus complicate the process. In this work, we explore the task of Depression Detection from transcribed therapy sessions and propose models and methods that address the task in hand. Firstly, we consider the transcribed dialogues derived from the sessions between a therapist and a client. In order to leverage the hierarchical structure of documents, we propose a Hierarchical Attention Network and perform document classification. Our task is a binary classification task, so the model decides upon the depression status of clients. However, therapy sessions provide valuable insights of the cognitive and behavioral functioning of clients, which can not be easily captured through processing of the raw document. Indeed, our analysis shows that depressed people use affective language to a greated extent than not-depressed. Therefore, we leverage behavioral and psycholinguistic cues of the client and therapist language to enhance the performance of our models. In particular, we integrate prior word-level psycholinguistic knowledge extracted from affective lexica, into the network architectures. The integration is performed into the self-attention mechanism of the system, which can force higher values for attention weights corresponding to salient affective words. In addition, we also incorporate the summary attributed to each session into the proposed architectures. Our approach improves the performance of the proposed architecture in the General Psychotherapy Corpus and the DAIC-WoZ 2017 depression datasets, achieving state-of-the-art 71.6 and 70.3 using the test set, F1-scores respectively. Next, we experiment with the problem of dialogue modeling in the context of detecting depression. We present related work for the task of emotion recognition in conversations and propose a model that introduces the speaker-role of utterances into the encoding process. The resulting performance is comparable to the baseline network. Finally, we propose future directions for incorporating the inter-speaker dependencies. Overall, our work addresses the task of depression detection from therapy sessions and proposes methods that improve the results of our networks, especially in the case we have small amount of data. This fact results in high performing models and improved robustness across two corpora. This work is summarized into the [81] research paper, which is submitted to the Interspeech 2020 conference.