Tasks, Games and Domain Knowledge in Dialogue Management Martin Beveridge, John Fox and David Milward Cancer Research UK, London WC2A 3PX
{martin.beveridge, john.fox}@cancer.org.uk,
[email protected]
1 Problem
2 Solution
The inexorably rising cost of healthcare provision is a major problem worldwide. A cost-effective way of improving quality of care while containing costs is to manage clinical decision-making and workflow by means of computerized care pathways. These pathways are increasingly represented using machine-readable formalisms, and this creates important opportunities for developing flexible natural language interfaces for capturing clinical data (e.g. taking clinical histories) and giving advice (e.g. recommending clinical actions). Natural language interaction could also offer patients more control of their own care, by allowing them to obtain advice whenever and as often as they need it. Voice interaction seems a particularly suitable approach as speech is a natural way for people to communicate and does not require the patient to use any technology other than a telephone. Spoken dialogue systems have great potential for delivery of healthcare services. However medical applications that provide advice to patients or to practitioners pose particular problems relative to standard information seeking dialogues such as flight booking or banking. In most current commercial dialogue systems, the dialogue designer specifies the exact interactions which can take place. Manual-coding allows precise control of what can occur within a dialogue, but is an expensive process, especially for complex dialogues. In the medical domain, the knowledge structures are particularly complex, and the system requires complex reasoning and decision-making to respond to the user. It therefore seems necessary to integrate technologies such as medical guidelines and advice systems directly with the dialogue system so that dialogues can be generated automatically to reflect user behaviour and changes in clinical context.
Cancer Research UK (CR-UK) has developed a dialogue system (Beveridge and Milward, 2003a; Beveridge and Milward, 2003b; Milward and Beveridge, 2003) as part of the EU HOMEY project (Home Monitoring through an Intelligent Dialogue System, IST-2001-32434). The first application of the system is intended for use by medical General Practitioners to determine whether a patient with suspected breast cancer should be referred to a cancer specialist (Bury et al., 2001). The dialogue system therefore needs to be closely integrated with medical domain knowledge, in this case in the form of an ontology for the breast cancer domain provided by Language & Computing n.v. (Ceusters et al., 2002), and knowledge of clinical tasks and processes, in this case the PROforma process specification language (Fox et al., 2003). In order to allow integration with these domain representations, our dialogue model is divided into high- and low-level representations. The low-level representation defines a finite-state network of communication acts represented by a VoiceXML specification. The high-level representation captures information regarding the intentional and informational structures underlying the dialogue, along with its current attentional state (Grosz and Sidner, 1986; Mann and Thompson, 1988). The information in the high-level representation is in turn derived from the underlying domain specification, with the intentional structure deriving from decisions, plans and other tasks in the medical care pathway and the informational structure deriving from the medical ontology. In order to make use of these representations in a practical system we have employed a multi-level architecture, similar to 3layer hybrid agent architectures (Gat, 1998) as shown in Figure 1.
Reactive Skills
Client Browser VoiceXML, XHTML, etc
Sequencing
Move Engine HLDS Game Engine ATS
Deliberation
References
Domain Manager
Plan Execution Engine
tiative, and varied groupings of clinical data by the game engine (an animated demonstration of the system is available.) Formal evaluations of the technology are to be carried out in the domain of cancer referral decisions (described earlier) and in the collection of family history data for genetic risk assessment and management. We are also investigating the potential for the technology to scale up to a larger medical domain encompassing the diagnosis, treatment and long-term follow-up of patients with breast cancer.
Ontology Engine
Figure 1. Dialogue System Architecture The deliberative layer is provided by a domain manager that creates an abstract task specification (ATS) based on the outputs of the plan execution and ontology engines. The sequencing layer includes a Game Engine which determines the conversational games (Lewin, 2000) to be played in order to complete the tasks in the ATS. The Game Engine uses ontological knowledge to reorder games. For example, a game concerning a refinement of a concept will, if possible, be ordered to appear immediately after the concept has been introduced to ensure maximal dialogue coherence. The resulting High-Level Dialogue Specification (HLDS) is used by a Move Engine to generate the sequence of low-level communicative acts (moves) that can be made by either participant at the current point in the dialogue. The reactive layer interprets the low-level specification (VoiceXML) in order to complete the specified dialogue segment, and handles lowlevel reactive behavior required to support communication, e.g. repeating prompts, changing the speech volume etc. If an event occurs which the reactive layer cannot handle (e.g. the user taking the initiative) then it is passed back to the sequencing and deliberative layers to be processed. The complete spoken dialogue system appears to be flexible and robust, resolving lexical and terminological ambiguities in real time using ontological knowledge, accommodating mixed ini-
Beveridge, M., and Milward, D. 2003a. Definition of the High-Level Task Specification Language. Deliverable D11, EU HOMEY Project, IST-200132434, http://www.acl.icnet.uk/lab/homey. Beveridge, M., and Milward, D. 2003b. Combining Task Descriptions and Ontological Knowledge for Adaptive Dialogue. To appear in Proc. Text, Speech and Dialogue (TSD), Ceske Budejovice, Czech Republic. Bury, J., Humber, M., and Fox, J. 2001. Integrating Decision Support with Electronic Referrals. In R. Rogers, R. Haux, and V. Patel (eds). Medinfo, IOS Press, Amsterdam. Ceusters, W., Beveridge, M., Milward D., and Falavigna, D. 2002. Specification for Semantic Dictionary Integration, Deliverable D9, HOMEY Project, IST2001-32434, http://www.acl.icnet.uk/lab/homey. Fox, J., Beveridge, M., and Glasspool, D. 2003. Understanding Intelligent Agents: Analysis and Synthesis. AI Communications (in press). Gat, E. 1998. On Three-Layer Architectures. In D. Kortenkamp, R. Bonasso and R. Murphy (eds) AI and Mobile Robots, AAAI Press, Menlo Park, CA. Grosz, B., and Sidner, C. 1986. Attention, Intention and the Structure of Discourse. Computational Linguistics, 12(3):175-204. Lewin, I. 2000. A Formal Model of Conversational Game Theory, In Proc. Gotalog-00, 4th Workshop on the Semantics and Pragmatics of Dialogue, Gothenburg, Sweden. Mann, W., and Thompson, S. 1988. Rhetorical Structure Theory: Towards a Functional Theory of Text Organization. Text, 8(3):243-281. Milward, D., and Beveridge, M. 2003. OntologyBased Dialogue Systems. To appear in Proc. IJCAI 3rd Workshop on Knowledge and Reasoning in Practical Dialogue Systems, Acapulco, Mexico.
