Description
Date depot: 1 janvier 1900
Titre: Context-sensitive generation of multimodal behaviours
Directrice de thèse:
Catherine PELACHAUD (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
The PhD will take part of the European project ARIA - VALUSPA (Affective Retrieval Interface Assistants – using Virtual Agents with Linguistic Understanding, Social skills and Personalised Aspects). The Affective Retrieval of Information Assistants, ARIA, are Embodied Conversational Agents with Linguistic Understanding, Social skills, and Personalised Aspects. They should be able to communicate using the same verbal and non-verbal modalities used in human-human interaction, have interpersonal skills akin to those of humans, and adapt to the user in terms of learning their preferences, personality, and manner of interaction. The agents should also be capable of dealing with unexpected situations, such as the user suddenly changing topic or task, the social group changing when a second user arrives, the user interrupting the ECA, or even the user suddenly changing its attitude.
The aim of the PhD is to develop a computational model of the behaviours of the ECA that would convey information at different levels:
• Interpersonal stance,
• overall communicative behaviours,
• emergence of synchrony with user's behaviour,
• multimodal response to unexpected situations
The work will make use of an existing ECA platform, Greta (Ochs et al, 2013). In particular we will extend our previous model of multimodal behaviours (Chollet et al., 2014) where we apply sequence mining on data from a corpus to extract frequent sequences for different types of attitude and communicative expressions and to use them as data to generate non-verbal behaviours for the ECAs.
To create an adaptive ECA, we will use a reinforcement algorithm to update the efficiency of a non-verbal behaviour used to communicate a given intention and/or emotional state. The reinforcement signal will be the achievement of the communicative intention and/or emotional state. The non-verbal behaviour of the ECA will be selected based on its efficiency computed dynamically during the interaction.
Doctorant.e: Melzi Soumia