Applications of Memory-Based learning to spoken dialogue system development

Lamont, Owen (2012) Applications of Memory-Based learning to spoken dialogue system development. PhD thesis, Murdoch University.

Abstract

This thesis introduces a new Spoken Dialogue System (SDS) development tool called the Generalising Memory-Based Learner (GMBL). The principle components and techniques, namely Hierarchical Hidden Markov Models (HHMM) and Memory-Based Learning (MBL), are described. We will argue that GMBL is an elegant, fast and intuitive way of constructing SDSs capable of answering questions and responding to commands that control connected equipment. Expediting the building of SDSs is the primary focus.

It is argued that GMBL is elegant by virtue of the regular, modular nature of the HHMMs used extensively throughout its architecture. These HHMM modules are straightforward to visualise and are easily comprehensible to developers, thus making development intuitive and simple.

The ease of SDS construction was evaluated by having a developer construct a train timetable application using GMBL, from a specification in a previous experiment. GMBL was clearly faster than a hand-coded application but was comparable though slightly slower than the reference system. With further development, this hierarchical memory-based paradigm shows promise for practical dialogue system development.

The usability of a GMBL-developed SDS was empirically evaluated by having participants use two interfaces to a simulation of an automated house. The GMBL-developed speech interface was compared with a conventional handheld remote control style interface. This experiment found that user satisfaction of the GMBL-developed speech interface compared well with the remote control interface and in general users successfully completed more tasks more quickly using the GMBL-developed interface.

The technical contributions of this thesis include demonstrations of novel applications for HHMMs in Spoken Dialogue Systems, novel ways of constructing HHMM topologies, and new modifications to HHMM that increase their power of representation when applied to adjacency pairs (question-answering) and instruction-action pairs.

In conclusion GMBL does demonstrate the promise of hierarchical memory-based learning architectures for dialogue modelling, although there were difficulties with minimising developer effort, not all of which were resolved satisfactorily. This work further highlights the under-reporting of development effort in dialogue modelling literature and argues that the developer, rather than the user, should be the crucial stakeholder in natural language system development.

Item Type:	Thesis (PhD)
Murdoch Affiliation(s):	School of Information Technology
Notes:	Note to the author: If you would like to make your thesis openly available on Murdoch University Library's Research Repository, please contact: repository@murdoch.edu.au. Thank you.
Supervisor(s):	Mann, Graham
URI:	http://researchrepository.murdoch.edu.au/id/eprint/41654

Item Control Page