About the GraphDial Project:
The two central objectives of GraphDial are:
- The development of new neural approaches to dialogue management (dialogue state tracking and action selection) using probabilistic graphs as core representation of the dialogue state. Relying on graphs to capture the dialogue state makes it possible to capture rich conversational contexts including varying numbers of entities and relations.
- The use of weak supervision to learn the parameters of these graph-based models from indirect data, which are much easier to collect than annotated dialogue data. The project will integrate weak supervision signals extracted from heuristic rules, grounding responses, and structural constraints.
The project will develop generic models of dialogue management that are applicable to a broad range of domains. The experimental evaluation will be done in the context of human-robot interaction tasks and a secondary objective is to improve dialogue modelling techniques in this particular domain.
The Norwegian Computing Center (NR) will be responsible for the overall management of GraphDial. The project leader is Pierre Lison.
The project is planned for a duration of four years and is divided in four work packages. The timeline is illustrated below:
- WP-0: Project management and dissemination:
- Task 0.1: Project management. Coordination of research activities, PhD supervision, organisation of project-related events, reporting, quality assurance and administrative duties.
- Task 0.2: Dissemination. Research publications, participation to international conferences, public outreach activities, demonstration of prototypes to stakeholders, etc.
- WP-1: Baseline system for HRI:
- Task 1.1: Dialogue architecture. Integration of core dialogue architecture into Pepper and its onboard modules (person detection, speech recognition and synthesis, navigation, etc.).
- Task 1.2: Experimental design. Design of interaction scenarios and evaluation metrics to be used for collecting data and assessing the performance of dialogue management models. Implementation of rule-based baseline system for these interaction scenarios.
- WP-2: Dialogue state tracking:
- Task 2.1: Neural state tracking on graphs. Tracking models to update the dialogue state (represented as a probabilistic graph) upon new observations through graph transformations.
- Task 2.2: Weak supervision for dialogue state tracking. State tracking models trained with multiple supervision signals (heuristic rules, constraints, grounding/clarification responses).
- Task 2.3: NLU integration. Dialogue state tracking on structured inputs (semantic parses).
- WP-3: Action selection:
- Task 3.1: Graph-based action selection model. Development of action selection models to determine the most appropriate action to perform using graph neural networks.
- Task 3.2: Weak supervision for action selection. Training of action selection model based on a combination of weak supervision signals, including heuristic rules and user feedback.
The software developed for the project will be released under an open-source license and published on public repositories.
The full 10-pages project description is available online.