Current students


SCOTTI VINCENZOCycle: XXXV

Section: Computer Science and Engineering
Advisor: SBATTELLA LICIA
Tutor: BARESI LUCIANO

Major Research topic:
Empathetic, Generative, Data-Driven Conversational Agents for Mental Healthcare

Abstract:
The research work included in this thesis aims at exploring the possibilities of empathy in human-computer interaction in mental healthcare scenarios through a conversational agent for psychotherapy sessions. The peculiarity of this research stands in the choice of leveraging empathy at multiple interaction levels in order to ease the development of a good therapeutic relationship between the user and the psychotherapy agent to improve the wellbeing of the patient. Moreover, we aim at embody the agent through a voiced interaction channel, so that better clinical results can be reached via complex multimodal interactions.

The agent will rely on the cognitive constructivist-relational model as psychotherapy approach, a novelty with respect to all the other automatic systems developed for this purpose, which usually leverage Cognitive-Behavioral-Therapy (CBT) or specific cognitive-behavioral techniques The aim of this alternative approach is to reach self-knowledge understanding the meanings each person gives to events rather than ‘suggesting and reprogramming behaviours or thoughts’ as in cognitive-behavioral approaches. The final object of this psychotherapeutic approach is to bring the subject to a better self-awareness and transformation of hers/his own cognitive, relational, imaginative processes. . The supervisor of the thesis (Prof licia Sbattella) is responsible for the design of the adopted psychotherapeutic model. The thesis will adopt the model and implement the agent using anonymized recordings of psychotherapeutic sessions based on the model.

This research aims at going beyond simple “emotional resonance” in the empathetic response, leveraging a more complex, multilevel empathic mechanism as driver for the generation of the reply. Through this multi-level architecture, which will help to provide an effective strategy for social interaction, we expect to achieve a long-term deep-engagement with the user and better clinical results. Emotion will still play an important role in the interaction, but we are completing the empathic stack with two layers: one sub-layer (the physical one, mainly working on prosodic and temporal aspects – to deal with agreement or contrast, synchronization or desynchronization) and one over layer (the semantic one, related to the contents, the meaning, the imaginative dimensions of the interaction).
The study will concentrate on the usage of voice as communication channel. The choice is done in order to have a better detection of the empathetic dimensions of the input, as well as to arouse the perception of empathy in the user.

The agent will be developed leveraging an architecture combining traditional model-driven Artificial Intelligence (AI) solutions for Natural Language Processing (NLP) and more recent data-driven solutions based on Deep Neural Networks (DNN). To be more detailed, we will use data-driven approaches in the analysis of the inputs (both voice and text) as well as in the output generation (again for both modalities). The latter aspect, to our knowledge, has never been taken into account. The core component, referred to as controller, will leverage the input information at the various empathy stack levels in order to select the best response, always according to the selected dimensions of empathy. This componet, designed with a model-driven approach, will guide the generation of the output response.

With this hybrid agent, together with the empathic response stack, we aim at getting rid of the “rigidity” characterizing current systems, which will surely result in a better therapy experience; such experience will also be improved by the cognitive constructivist-relational model guiding the clinical interaction.