COREDIAL · Spoken dialog management that combines corpus-based statistical learning and reinforcement learning with a constraint-based core
FP7 — People (Marie Curie Actions)
- Duration
- —
- EU contribution
- €45,000
- Participants
- 1
- Scheme
- MC-ERG
Lines connect the coordinator with its partners.
Results in brief
Spoken dialog management that combines corpus-based statistical learning and reinforcement learning with a constraint-based core
Spoken dialog systems (SDS) have made great progress over the last 10 years. Their use has become wide-spread in many areas, especially call centre automation, but also for interacting with car- based dialog systems or robots, for example. Despite this progress, significant challenges remain: human-machine communication is in practice quite different from human-human communication. The COREDIAL project addresses dialog management, a core task for spoken dialog systems that deals with making an action decision, and extends the approach to include language understanding and generation in a complete spoken dialog system. The objective is to provide a constructive proof that corpus-based statistical methods can be combined with reinforcement learning for dialog management. In the reporting period, the principal researcher Dr Sebastian Varges worked on crucial components of the planned work programme: 1) Final evaluation of experimental data of a SDS. 2) Textual natural language generation component. This part of the work program is already completed but will be revised if other system components require changes. 3) Content-to-speech generation to drive speech synthesis: Text generation is being extended to generate text with prosodic mark-up to drive a text-to-speech system. This means that the voice can empharise parts of the speech more like a human being. 4) Dialog system architecture: We extended a proven dialog system architecture to handle the demands of statistics-based dialog management.
Data: CORDIS, © European Union
Project objective
Spoken dialog systems (SDS) have made great progress over the last 10 years. Their use has become wide-spread in many areas, especially call center automation, but also for interacting with car- based dialog systems or robots, for example. Despite this progress, significant challenges remain: human-machine communication is in practice quite different from human-human communication.We will address dialog management, a core task for spoken dialog systems that deals with making an action decision, and extend the approach to include language understanding and generation in a complete spoken dialog system. The objective of CoreDial project is to provide a constructive proof that corpus-based statistical methods can be combined with reinforcement learning for dialog management. Both methods will be used as (re)ranking methods for lexically realized dialog moves that are generated by a constraint-based, overgenerating core system. This involves several components (and thus sub-goals/objectives):a) Constraint-based core dialog manager and generator using syntactic mechanisms to produce lexically realized dialog moves,b) Corpus-based statistical ranker employing a novel dialog language model that uses methods from statistical machine translation,c) Reinforcement Learning based (re)-ranker to optimize the overall dialog and in particular handle noise, e.g. deciding about clarification questions,d) Dialog system architecture and statistical model combination to integrate the components.This system and its evaluation will be the main result of the project.The Reintegration Grant will allow the applicant to integrate himself into the host organization, maintain links to the current Marie Curie host, and prepare himself for his future professional development.
Original text from CORDIS.
Participants
- UNIVERSITAET POTSDAM · PotsdamCoordinatorGermany
Links
Data: CORDIS, © European Union