Deliverables
Automated dialogue systems have evolved from heavily structured, inflexible, system initiative services to allow a looser structure with more mixed user and system initiative and greater flexibility, for example fast track for experienced users and context sensitive help for novice users. However, even the most advanced of today's dialogue systems are still rigid in their structure and users with wide ranging preferences and capabilities are generally required to interact with the same inflexible telephone based service, despite a growing corporate interest in particle marketing and personalisation of automated services offered via other channels such as the Internet.
The demonstrators developed for this project were concerned with the development of personalised, adaptive dialogue systems, which can sensitise themselves to each user based on historical customer information, and also on present behaviour. Many aspects of the dialogue systems were broadly configured according to user demographics such as age, gender and income. The prompt messaging, for example, featured voices considered to be appropriate for the type of user. The language of the messages was also adjusted in terms of directness, openness, simplicity and formality. The dialogues incorporated types of help, degrees of system or user initiative and structured vs. fast-track services according to each user's technographic profiles. Different degrees of explicitness in confirmation were used according to a confidence profile for each user. In addition to such data, information based on each user's experiences with the service considerably enriched each user's profile and the system's potential to exploit it in terms of the services offered and the relative prominence given to popular choices and also in terms of the underlying speech recognition technology, via the use of dynamic grammars and accent modelling to optimise the system in broad terms for each user¿s style of speech and language.
As well as being configurable according to personalised criteria, the dialogue systems were also adaptive during the course of a call altering, for example, the degree of system/user initiative, the style of confirmation or error recovery and the prompt wordings according to the range of responses valid and invalid given by users at various stages. Any such run-time adaptations feed back into each user's permanent profile with the system. For system adaptation, key research issues were the classification of possible responses into various adaptation triggers and the range and degree of adaptation associated with these triggers.
Adaptation to the conversation partner is, of course, a feature of natural human-human dialogues, and so achieving it will contribute to perceived naturalness and usability in spoken HLT services. A related aspect of human dialogue behaviour is the flexibility of turn-taking, whereby a conversation does not consist simply of a sequence of utterances by the partners in turn, with their ends signalled by silence (as in conventional spoken dialogues with automated services), but exhibits the phenomena of rapid response (anticipating the end of the conversation partner's turn and hence responding as soon as it ends, without waiting for a period of silence), interruption (where it is judged more helpful, for purpose of the task in hand, not to wait for the end of the partner's turn), and back-channel feedback (in which the listener provides signals of understanding, agreement etc during an utterance, in the form of short utterances or vocalisations ('yes', 'uhuh', and the like in English), but the speaker's turn continues uninterrupted). Implementing the full range of human capabilities in this respect in an automated system is a challenging task, but some aspects were explored, in particular the implementation of a capability for back-channel feedback by the system during the user's speech, and the capability to distinguish between interruption (causing the service to stop speaking) and back-channel feedback by the user during a service utterance.
The project has created a Design Guide for spoken dialogues, which is based on a wide range of usability trials data for the various languages and applications created for the project. The Guide details the design options in the creation of a successful spoken HLT service. It has been made available to all interested user enterprises.
Over 20 important elements in the design of spoken HLT services have been identified and described in the Design Guide, a number of which are accompanied by example speech files drawn from the demonstrators implemented during the project.
The design guide covers the following topics:
Discourse markers
System personality
General principles of message wording
Open and closed prompting styles
Personalised services - adapting the system to the user
The number of pieces of information to capture at one time
Confirmation of user input
Recorded and synthesised speech
Attitudes to services with and without tone prompts
Individual differences and service usage
Priming metaphors as aids to use
Interface metaphors as aids to navigation
The effectiveness of aid strategies
Politeness - don't blame the user
The voice used in recorded messages
Tones and chimes in automated telephone services
Keypad choices for DTMF input
Barge-in
Barge-in and priming
Requirements for barge-in
Mixing modalities
Landmarks
menus.
Searching for OpenAIRE data...
There was an error trying to search data from OpenAIRE
No results available