UPML: A framework for knowledge system reuse
Dieter Fensel V. Richard Benjamins Enrico Motta Bob Wielinga
Institute AIFB Dep. SWI Knowledge Media Institute Dep. SWI
Univ. of Karlsruhe Univ. of Amsterdam The Open University Univ. of Amsterdam
76128 Karlsruhe 1018 WB Amsterdam M
Germany The Netherlands
Abstract
Problem-solving methods provide reusable
architectures and components for implementing the
reasoning part of knowledge-based systems. The
Unified Problem-solving Method Development
Language, UPML, has been developed to describe
and implement such architectures and components
and to facilitate their semiautomatic reuse and
adaptation. In a nutshell, UPML is a framework for
developing knowledge-intensive reasoning systems
based on libraries of generic problem-solving
components. The paper describes the components,
architectural constraints, development guidelines,
and tools provided by UPML. Our focus is hereby
on the meta ontology that has been developed to
formalize the architectural structure and elements
of UPML.
1 Introduction
Problem-solving methods (PSMs) for knowledge-based
systems (KBSs) (cf. [Schreiber et al., 1994]; [Benjamins &
Fensel, 1998]) decompose the reasoning task of a KBS in a
number of subtasks and inference actions that are connected
by knowledge roles. Several problem solving method
libraries are now available [Breuker & van de Velde, 1994],
[Motta & Zdrahal, 1998]. The IBROW project [Benjamins et
al., 1998] has been set up with the aim of enabling
semiautomatic reuse of PSMs. This reuse is provided by
integrating PSM libraries in an internet-based environment.
A software broker selects and combines PSMs from different
libraries and provides a knowledge engineer with semi-
automated support for configuring a reasoning system.
Hence, a description language for these reasoning
components (i.e., PSMs) must provide human-
understandable, high-level descriptions, which should also
be grounded on a formal representation, to allow automated
support by the broker. To this purpose we have developed the
Unified Problem-Solving Method Development Language,
UPML [Fensel et al., 1999b]. UPML is a software
architecture for knowledge-based systems providing
components, adapters and a configuration (called
architectural constraints) of how the components should be
connected using the adapters. Finally design guidelinesK7 6AA, Milton Keynes 1018 WB Amsterdam
United Kingdom The Netherlands
specify how to develop a system constructed from the
components and connectors that satisfies the architectural
constraints.
In knowledge engineering terms UPML provides a
meta-ontology for describing knowledge-based systems. The
different elements of a specification correspond to concepts
of this ontology and the architectural constraints are axioms
in this ontology.
In this paper we outline the main features of the
approach we have taken to define a framework for
knowledge sharing and reuse. In particular we illustrate the
basic meta-ontology of UPML, its underlying architecture,
support tools and development guidelines. Because of space
constraints we can only provide a limited number of
technical details. Hence, the paper is better seen as an
overview report on the main issues we are facing and the
solutions we are developing.
The paper is organized as follows. In Section 2, we will
briefly sketch the overall structure of UPML. Then we will
discuss the (meta-)ontology that can be used to formalize
UPML. Section 4 introduces the architectural constraints of
UPML and Section 5 shows various ways in which tools for
developing, selecting, and combining PSMs can make use of
the (meta-)ontology. Section 6 briefly mentions the
development guidelines of UPML. Conclusions, related
work and outlook are discussed in Section 7.
2 The Overall Structure of UPML
[Fensel et al., 1999a] introduce the four components types of
a UPML specification:
• Tasks define the problems that should be solved by the
KBS.
• PSMs define the reasoning process of a KBS in
domain-independent terms.
• Domain models describe the domain knowledge of the
KBS.
• Ontologies provide the terminology used in tasks,
PSMs and domain definitions.
Each of these elements is described independently to enable
the reuse of task descriptions in different domains, the reuse
of PSMs across different tasks and domain, and the reuse of
domain knowledge for different tasks and PSMs. Therefore,

search
-
search
with preference
a
c
i
gdiagnosis
complete and
parsimonious diagnoses
anesthesiology anesthesiology
hill
Initi
hill-
Upda
for h
ontology domain model task task refiner bridtaskFig 1. The overall structure
adapters are required to adjust the (reusable) parts to each
other and to the specific application problem. UPML
provides two types of adapters: bridges and refiners.
• Bridges explicitly model the relationships between
two distinct parts of an architecture, e.g. between
domain and task or task and PSM.
• Refiners can be used to express the stepwise
specialization of a class of elements of a specification,
e.g. a task is refined or a PSM is refined.
Very generic PSMs and tasks can be refined to more specific
ones by applying a sequence of refiners (cf. [Fensel, 1997]).
Again, separating generic and specific parts of a reasoning
process maximizes reusability.
Together, the six UPML building blocks define a software
architecture. The overall structure of a UPML specification
is presented in Figure 1 (a more detailed discussion of the
example can be found in [Fensel et al., 1999b]). A task
called “complete and parsimonious diagnoses” is defined by
importing an ontology called “diagnosis”. The PSM applied
to solve the task is “hill climbing”. A bridge is required to
connect the generic terminology of hill climbing with the
diagnostic task: states and states transitions of the method
have to be rephrased in terms of the task ontology.
Hill climbing is only one possible refinement of a generic
search method that decomposes an entire search task intosearch
Derive Successor

Initialize

Select Node

Stop

Update Nodes
climbing
lize for
limbing

te Nodes
ll-climbing
Nodes
PSM PSM
e
import
decompose PSM refiner of a UPML specification.
five more elementary subtasks: Initialize, Derive Successor
Nodes, Select Node, Stop and Update Nodes. Hill climbing
can be derived from this generic search method by (i)
refining one of its subtasks (i.e., update node forgets all
earlier nodes and only processes the currently derived
successors further) and (ii) introducing a preference
ordering.
PSM-mediated task decomposition and PSM specialization
through a refiner are analogous to the part-of and is-a
constructs of knowledge representation formalisms.
Subtasking corresponds to the part-of construct because it
decomposes a task into subtasks. The refinement of
problem-solving methods, as introduced in [Fensel, 1997],
corresponds to the is-a relationship of knowledge
representation formalisms - e.g., Hill-climbing is a
specialization of a general search method by refining some
of its attributes (i.e., subtasks).
3 The Meta Ontology of UPML
We used PROTÉGÉ-II [Puerta et al., 1992] to develop a
meta ontology of UPML. PROTÉGÉ-II is a knowledge
acquisition tool-generator. After defining an ontology it
semiautomatically generates a graphical interface for
collecting the knowledge that is described by the ontology.
The ontology can be described in terms of classes and