The concept of the "Council of Coaches" is inherently new. A very strong focus from day 1 of the project has been on developing tangible prototypes to demonstrate this concept. Overall, the concept of Council of Coaches was successfully demonstrated by means of the functional demonstrator. Users showed good levels of engagement with the demonstrator.
Furthermore, the project has realized an integrated, fully open-source software platform that allows content creators to develop applications supporting multiple AI-driven embodied conversational agents that constitutes a major step beyond the state of the art, both in terms of innovative aspects, as well as its fully open-source nature.
In summary, Council of Coaches has pushed the state of the art in the following ways:
- The novel idea of a group of virtual coaches for coaching on health and wellbeing was successfully demonstrated to provide a promising method of engaging end-users with their health.
- The WOOL Platform was developed and released as open-source platform for authoring- and integrating scripted dialogues in web- and mobile applications. Where existing tools are often proprietary and never focus on the authoring process from a non-technical domain-expert’s point of view, WOOL is taking a place in the market for dialogue systems for serious applications (www.woolplatform.eu).
- A methodology and vision on how to practically implement RRI principles in the daily workings of a RIA project has been established.
- The coach-as-a-sensor concept has been demonstrated to measure data that is harder to measure through sensors, integrated through dialogues with the end user to capture e.g. emotional and cognitive data.
- A holistic behavior framework was developed that takes data from different sources (mobile phones, off-the-shelf sensors, and the coaches themselves), and uses it to predict long-term behaviours.
- An integrated dialogue system that advances the state of the art by combining expertise, models, and insights from the fields of argumentation and social conversation. The novelty of this aspect of the system lies in the way that various modules in the system work together to move from models of coaching topics and content, to structural models of dialog to determine the content of successive moves in the conversation, via models of social conversational intents that are needed to deliver a single piece of content, to actual delivery of the content in the form of agent behaviour.
- A computational model to control the multimodal behaviours of virtual agents was developed based on studies to understand the relationship between task, social cohesion, and multimodal behaviours. A new corpus on group interactions has been developed in which video of a multi-party coaching sessions was analysed and annotated with dialogue and gesture moves.
- A computational model of communicative gestures was also developed. While previous computational models work either with only beat gestures (which marks speech rhythm) or with only illustrative gestures (including iconic, metaphoric and deictic), we have developed a model which works for both types of gestures. Moreover, our model also considers eyebrow movements which are linked to speech prosody. Our model was trained on data which has been annotated with gesture types. The data was also segmented into gesture phases (preparation, stroke, retraction, hold). It allows our model to learn where to place a gesture stroke such that it aligns with the speech.
