Narrative Embodied Conversational Agent – NARECA

One of the greatest challenges in AI is to enable humans to communicate with the machine, without having to use linguistic structures linked with a particular application. Since most of the existing systems do not take into account human interaction specificities, the user behaviour is limited by the machine constraints. The main issue in to adapt the machine to the human behavior, particularly when the user is a child, with cognitive and communicative still in developpement. Designing such an interface requires to improve existing dialogue models and to integrate it into an Embodied Conversational Agent (ECA). However, it is currently not possible to design a generic ECA taking into account the scientific issues to overcome (speech, prosody, emotion, nonverbal behavior, size of vocabulary, linguistic phenomena, etc.).
The goal of NARECA is to improve the modeling of dialogic interactions and to exploit these models in an ECA. The challenge is to extract dialogue models directly from real from real dialogues (methodology and tools). Our approach is firstly to automatically annotate the dialogues using NLP and learning methods and secondly to mine these annotations in order identify recurrent behavioral patterns. The output of the project will be a publically available software platform for automatic dialogue analysis and an applicative demonstration integrating the collected dialogic patterns: a narrative affective ECA.
NARECA follows and is built on the preliminary results obtained during the project ACAMODIA (2011-2013), funded by the CNRS as a PEPS INS2I-INSHS.