Work progress 2005

Work performed

The partners have successfully completed a range of key tasks in the first two years (2004 and 2005) of the project, both in advancing the theory and developing implementations. Many of these tasks have involved close collaboration between the sites, and a variety of new techniques and ideas have emerged as a result.

The active project tasks in 2005 were in the areas of extending the TALK grammar library to include multimodality (D1.2); generating Statistical Language Models (SLMs) from grammars (D1.3); integrating multiple modalities with SLMs; proof-of-concept systems in the in-home domain; integrating ontology-based dialogue management with the ISU approach (D2.1); developing reconfigurable systems; programming by voice; extending Information State modelling (D3.1) for multimodal turn planning and developing modality specific resources; developing a baseline dialogue system exhibiting Reinforcement Learning (D4.2); developing a prototype Bayes Net dialogue manager (D4.3); software infrastructure and system integration for the in-car and in-home showcases (D5.2); a formative evaluation of the showcase in-car dialogue system (D6.3); and annotation guidelines for data collections (D6.2).

User testing the in-case showcase

Significant progress has been made in each of these active tasks, measurable by the fact that each one has produced either a deliverable or a status report. In the second year, especially strong progress has been made in system development and integration (D2.1, D4.2, D5.2), system evaluation (D6.3), and theoretical advances have been made in the areas of multimodal grammars (D1.2), SLM generation (D1.3), extended Information States (D3.1), and novel machine learning techniques (D4.3).

 

Results achieved

The project has made good progress on the research issues which were planned for the second year, and has built solidly on a successful first year. In the areas of system development, integration, and evaluation, multimodal grammars and SLM generation, and in combining learning with ISU dialogue management, progress has been particularly strong.

For example, 2 new multimodal dialogue systems have been developed (one is the first ISU system to employ learned dialogue strategies, see D4.2), and several existing ones have been enhanced. We have shown in early results that users of the prototype showcase in-car system show good satisfaction scores and obtain high task completion rates (see D6.3). New methods for SLM generation from grammars have found an excellent compromise between the ease of grammar development and the robustness of statistical language models in the first stages of system development (see D1.3).