Ontologies for Flexible Integration: multi-collection inquiry in large digital libraries
Invited talk,
Conference on Information, Knowledge, and Knowledge Management (IuK2000). March 2000, Darmstadt, Germany.
Powerpoint slides available by request

The fundamental technical challenge for e-business today is flexible integration: to get computers to work together despite conceptual differences and changes over time. For example, digital libraries need interfaces that query multiple collections whose metadata is not entirely compatible. A basic strategy is for computers to interpret requests, rather than tell each other exactly what to do. For example, digital-library users will phrase queries in familiar language, which remote collections interpret in terms of their own terminology. Requests that represent relatively more of their meaning more closely constrain interpretation, potentially yielding more accurate responses. Ontologies are domain models that define terminology. The tags in XML messages, for example, may also be defined as ontology concepts -- making the semantics of the messages much more explicit. Thus, ontologies enable relatively flexible integration.

My goal in this talk is to describe ontology-based multi-collection inquiry as concretely as possible. First, I present a formal ontology for digital libraries. This ontology models bibliographic relations among works. Second, I describe how I automatically generated a catalog of ontological metadata by transforming ordinary catalog data from the University of Michigan library. Third, I introduce several kinds of computational reasoning that can map across ontologies with increasing degrees of differentiation. Finally, I provide examples of multi-collection inquiries. These show how ontological metadata can help users find information with better precision and recall than is available today.

Comparing Concepts in Differentiated Ontologies
Peter Weinstein and
William P. Birmingham
Proceedings of the Twelfth Workshop on Knowledge Acquisition, Modeling and Management (KAW'99). October 1999, Banff, Alberta, Canada.
Available in pdf, and the submission draft is available in html.

Concepts in differentiated ontologies inherit definitional structure from concepts in shared ontologies. Shared, inherited structure provides a common ground that supports measures of "description compatibility." These algorithms are the primary contribution of this paper. The description-compatibility measures compare concepts to predict semantic compatibility, the probability that an instance of a recommendation will satisfy a request. The description-compatibility measures cross a spectrum regarding their knowledge of the semantics of roles in concept definitions. Some of the measures identify and analyze correspondences among elements of the definitions, and are thus a form of analogical reasoning. We use simulations to evaluate the description-compatibility measures in detail. Description compatibility can be used to rank alternative query translations, and to guide search for capabilities across communities that subscribe to differentiated ontologies.

Agent Communication with Differentiated Ontologies: eight new measures of description compatibility
Peter Weinstein and
William P. Birmingham
Technical Report CSE-TR-383-99, Department of Electrical Engineering and Computer Science, University of Michigan, 1999.

We propose an approach to achieve appropriate exchange of services and data in distributed systems subject to semantic heterogeneity. We assume differentiated ontologies: that terms have formal definitions as concepts related to other concepts, that local concepts inherit from concepts that are shared, and that most or all primitives are shared. We then develop measures of description compatibility using the structure of the source and target definitions. We evaluate these measures by generating description-logic ontologies in artificial worlds. In our simulations, the "meaning" of a concept is its denotation in a finite universe of instances. The accuracy of the description compatibility measures can thus be judged by their success in predicting the overlap of concept denotations. Description compatibility can be used to guide agent search for services across communities that subscribe to differentiated ontologies.

Creating Ontological Metadata for Digital Library Content and Services
Peter Weinstein and William P. Birmingham
International Journal on Digital Libraries, special issue on artificial intelligence for digital libraries, 2(1): 19-36, Springer-Verlag, 1998.
Available in compressed postscript.

We use formal ontologies to represent knowledge about digital-library content and services. Formal ontologies define concepts with logic in a frame-inheritance structure. The expressiveness and precision of these structures supports computational reasoning that can be used in important ways. This paper focuses on the creation of ontological metadata.

We create ontological content metadata by generating it from MARC (MAchine Readable Cataloging) data. MARC contains much information that is hard to exploit computationally. In particular, relationships between works are implicit in shared values and natural-language notes. The conversion process involves specifying an ontological model, mapping MARC to the ontology, and reasoning about the data to create explicit links between works.

Service metadata will be supplied by providers who wish to participate fully in a digital library that is implemented as a decentralized multi-agent system. Agents advertise by describing their services in terms of ontologically defined concepts. We reason about these descriptions to organize them into subsumption taxonomies. Agents can then find the best available services to meet their needs by describing their needs, without requiring a priori knowledge of other agents. This infrastructure has demonstrated its usefulness in a multi-agent system organized as a computational economy.

Agent-Based Digital Libraries: Decentralization and Coordination
Peter Weinstein, William P. Birmingham, Edmund Durfee
IEEE Communications Magazine 37(1): 110-115, January 1999.
Available in postscript and RTF.

This paper describes agent-based systems and explains why digital libraries should be built with this type of architecture. The primary advantage of agent-based architecture is decentralization, which enables scaling, flexibility, and extensibility. The corresponding requirement is the need to coordinate agent activity. We describe the approach taken by the University of Michigan Digital Library project.

Ontology-Based Metadata: Transforming the MARC Legacy.
Peter Weinstein
Proceedings of the Third International ACM Digital Library conference, Pittsburgh, PA, USA, June 1998.
Available in postscript and RTF.

We propose a new catalog, based on a formal ontological model of bibliographic relations. A hierarchy of five central concepts describes the creation of work. Each kind of relation between works occurs at a particular level in the hierarchy. Related works share data at some level of the hierarchy, yielding a tree structure that reduces redundant representation of shared attributes.

To show that ontology-based metadata is practical, we generated a knowledge base of metadata from a sample of MARC records. We implemented the ontology in description logic (Loom), mapped MARC attributes and values to the ontology, and loaded the data into Loom with all values treated as separate instances. We then unified matching instances, and deduced relations between works. Thus, the process converts relationships implicit in MARC into explicit relations amenable to manipulation by computers.

Our web interface permits browsing by navigating relations between works. Ontology-based metadata can also support user inquiry and digital-library operation in other important ways.

Service Classification in a Proto-Organic Society of Agents.
Peter Weinstein and William P. Birmingham.
Proceedings of the IJCAI-97 Workshop on Artificial Intelligence in Digital Libraries, August 1997, Nagoya, Japan.
Available in postscript and RTF.

Service classification helps agents select services in societies of agents that are large, dynamic, and evolving. Services are capabilities that agents make available to other agents. Existing systems do not use complex service descriptions, or do not handle change well. Our Service Classifier Agent (SCA) maintains a dynamic ontology, using description logic to organize services defined by agents at runtime. Agents requiring a service select the best available, utilizing knowledge derived with subsumption and other types of inference. Thus, new agents can immediately participate in the society, without requiring modification of existing agents. The SCA has demonstrated its usefulness in a society of agents organized as a computational economy.

Seed Ontologies: growing digital libraries as distributed, intelligent systems.
Peter Weinstein and Gene Alloway.
Proceedings of the Second International ACM Digital Library conference, Philadelphia, PA, USA, July 1997.
Available in postscript and RTF.

Ontologies are more than a particularly elaborate approach to the description and classification of information. They can be used to support the operation and growth of a new kind of digital library, implemented as a distributed, intelligent system. We describe the design and use of ontologies in the University of Michigan Digital Library. These ontologies will model all aspects of the digital library, including content, services, and licenses. We have extended and refined the IFLA hierarchy for the realization of work, and are starting to use ontologies to support reasoning about content search. We have also used the ontologies to classify the capabilities of computational elements of the system (agents), in a dynamic way that sustains functionality as new agents are added to the system.

Matching Requests for Agent Services with Differentiated Vocabulary.
Peter Weinstein and William P. Birmingham.
Student abstract, Proceedings of the Fourteenth National Conference on Artificial Intelligence (AAAI-97).
Available in postscript and RTF.

Runtime Classification of Agent Services.
Peter Weinstein and William P. Birmingham.
Proceedings of the AAAI-97 Spring Symposium on Ontological Engineering at Stanford, Palo Alto, CA, USA, March 1997.
Available in postscript and RTF.

The Service Classifier Agent maintains a dynamic ontology of agent capabilities. To advertise their services, agents define concepts at runtime. These concepts are automatically classified with description logic. Agents requesting services can select the best available to meet their needs, using queries that exploit rich knowledge about services and their relations to other services.

Runtime classification of agent services encourages the development of agents to provide new services. New agents may be utilized immediately upon joining a society, without requiring modification or even notification of existing agents.

Customized Metadata for Internet Information.
Brian Dunkel, Nandit Soparkar, and Peter Weinstein.
Proceedings of the First International Conference on Conventional and Knowledge-Based Intelligent Electronic Systems (KES '97), Adelaide, Australia, May 1997.
I am first author of an earlier version.

Social Structure and the Development of Self-Esteem in Young Children
Aaron M. Pallas, Doris R. Entwisle, Karl L. Alexander, Peter Weinstein
Social Psychology Quarterly, 1990, Dec Vol 53(4) 302-315.

Investigated the nature of self-esteem in urban Black and White children over their 1st 4 yrs of schooling, using a confirmatory factor analytic model. Data were available for 687 Ss in 1983, 592 Ss in 1984, and 553 Ss in 1986. Ss responded to a questionnaire that included items measuring children's self-esteem. The rationale and conceptual basis for the model used led to a 5-dimensional structure (character, personal responsibility, academic, appearance, athletic). Estimating the model for age, gender, race, and socioeconomic status (SES) subsamples suggests that this structure applied well to all Ss and that the dimensions became differentiated more clearly with time. Ss differed in their average levels on the 5 dimensions: Boys were higher than girls on the athletic and appearance dimensions, and girls were higher than boys on the other dimensions. Blacks surpassed Whites in the athletic and appearance domains.

Other Papers

Agent Communication in Semantically Heterogenous Societies (1997).

Matching Expressions with Differentiated Vocabulary: Technology for Knowledge Sharing in Partially Focused Communities (1996).

An Expert System for Substantive Support of Project Based Education (1995).

Computer Interpretation of Natural Language Explanations in the Model-It Learning Tool (1994).

Learning: A Model of Personal Growth (1985).

A Model of Change: with applications to software engineering and social organization (1983).

Return to homepage.