Bruce W. Porter

Concept Learning and Heuristic Classification in Weak-Theory Domains

Abstract This paper describes a successful approach to concept learning for heuristic classification. Almost all current programs for this task create or use explicit, abstract generalizations. These programs are largely ineffective for domains with weak or intractable theories. An exemplar-based approach is suitable for domains with inadequate theories but raises two additional problems: determining similarity and indexing exemplars. Our approach extends the exemplar-based approach with solutions to these problems. An implementation of our approach, called Protos, has been applied to the domain of clinical audiology. After reasonable training, Protos achieved a competence level equaling that of human experts and far surpassing that of other machine learning programs. Additionally, an “ablation study” has identified the aspects of Protos that are primarily responsible for its success.

A library of generic concepts for composing knowledge bases

Building a knowledge base for a given domain traditionally involves a subject matter expert and a knowledge engineer. One of the goals of our research is to eliminate the knowledge engineer. There are at least two ways to achieve this goal: train domain experts to write axioms (i.e., turn them into knowledge engineers) or create tools that allow users to build knowledge bases without having to write axioms. Our strategy is to create tools that allow users to build knowledge bases through instantiation and assembly of generic knowledge components from a small library.In many ways, creating such a library is like designing an ontology: What are the most general kinds of events and entities? How are these things related hierarchically? What is their meaning and how is it represented? The pressures of making the library usable by domain experts, however, leads to departures from the traditional ontology design goals of coverage, consensus and elegance. In this paper we describe our component library, a hierarchy of reusable, composable, domain-independent knowledge units. The library emphasizes coverage (what is an appropriate set of components for our task), access (how can a domain expert find appropriate components) and semantics (what knowledge and what kind of representation permit useful composition). We have begun building a library on these principles, influenced heavily by linguistic resources. In early evaluations we have put the library into the hands of domain experts (in Biology) having no experience with knowledge bases or knowledge acquisition.

Knowledge entry as the graphical assembly of components

Despite some successes, the lack of tools to allow subject matter experts to directly enter, query, and debug formal domain knowledge in a knowledge-base still remains a major obstacle to their deployment. Our goal is to create such tools, so that a trained knowledge engineer is no longer required to mediate the interaction. This paper presents our work on the knowledge entry part of this overall knowledge capture task, which is based on several claims: that users can construct representations by connecting pre-fabricated, representational components, rather than writing low-level axioms; that these components can be presented to users as graphs; and the user can then perform composition through graph manipulation operations. To operationalize this, we have developed a novel technique of graphical dialog using examples of the component concepts, followed by an automated process for generalizing the users graphically-entered assertions into axioms. We present these claims, our approach, the system (called SHAKEN) that we are developing, and an evaluation of our progress based on having users encode knowledge using the system.

Automated modeling of complex systems to answer prediction questions

Abstract A question about the behavior of a complex, physical system can be answered by simulating the system—the challenge is building a model of the system that is appropriate for answering the question. If the model omits relevant aspects of the system, the predicted behavior may be wrong. If, on the other hand, the model includes many aspects that are irrelevant to the question, it may be difficult to simulate and explain. The leading approach to automated modeling, “compositional modeling”, constructs a simplest adequate model for a question from building blocks (“model fragments”) that are designed by knowledge engineers. This paper presents a new compositional modeling algorithm that constructs models from simpler building blocks—the individual influences among system variables—and addresses important modeling issues that previous programs left to the knowledge engineer. In the most rigorous test of a modeling algorithm to date, we implemented our algorithm, applied it to a large knowledge base for plant physiology, and asked a domain expert to evaluate the models it produced.

Supporting Start-to-Finish Development of Knowledge Bases

Developing knowledge bases using knowledge-acquisition tools is difficult because each stage of development requires performing a distinct knowledge-acquisition task. This paper describes these different tasks and surveys current tools that perform them. It also addresses two issues confronting tools for start-to-finish development of knowledge bases. The first issue is how to support multiple stages of development. This paper describes Protos, a knowledge-acquisition tool that adjusts the training it expects and assistance it provides as its knowledge grows. The second issue is how to integrate new information into a large knowledge base. This issue is addressed in the description of a second tool, KI, that evaluates new information to determine its consequences for existing knowledge.

Experimental Goal Regression: A Method for Learning Problem-Solving Heuristics

This research examines the process of learning problem solving with minimal requirements for a priori knowledge and teacher involvement. Experience indicates that knowledge about the problem solving task can be used to improve problem solving performance. This research addresses the issues of what knowledge is useful, how it is applied during problem solving, and how it can be acquired. For each operator used in the problem solving domain, knowledge is incrementally learned concerning why it is useful, when it is applicable, and what transformation it performs. The method of experimental goal regression is introduced for improving the learning rate by approximating the results of analytic learning. The ideas are formalized in an algorithm for learning and problem solving and demonstrated with examples from the domains of simultaneous linear equations and symbolic integration.

Representing roles and purpose

Ontology designers often distinguish Entities (things that are) from Events (things that happen). It is not obvious how this division admits Roles (things that are, but only in the context of things that happen). For example, Person might be considered an Entity, while Employee is a Role. A Person remains a Person independent of the Events in which he participates. Someone is an Employee only by virtue of participating in an Employment Event. The problem of how to represent Roles is not new, but there is little consensus on a solution. In this paper, we present an ontology that finds a place for Roles as well as a representation that allows Roles to be related to Entities and Events to express the teleological notion of purpose.

Protos: an examplar-based learning apprentice

Building Protos, a learning apprentice system for heuristic classification, has forced us to scrutinize the usefulness of inductive learning and deductive problem solving. While these inference methods have been widely studied in machine learning, their seductive elegance in artificial domains (e.g. mathematics) does not carry-over to natural domains (e.g. medicine). This paper briefly describes our rationale in the Protos system for relegating inductive learning and deductive problem solving to minor roles in support of retaining, indexing, and matching exemplars. The problems that arise from “lazy generalization” are described along with their solutions in Protos. Finally, an example of Pro tos in the domain of clinical audiology is discussed.

Capturing and answering questions posed to a knowledge-based system

As part of the ongoing project, Project Halo, our goal is to build a system capable of answering questions posed by novice users to a formal knowledge base. In our current context, the knowledge base covers selected topics in physics, chemistry, and biology, and our question set consists of AP (advanced high-school) level examination questions. The task is challenging because the questions are linguistically complex and are often incomplete (assume unstated knowledge), and because the users do not have prior knowledge of the systems contents. Our solution involves two parts: a controlled language interface, in which users reformulate the original natural language questions in a simplified version of English, and a novel problem solver that can elaborate initially inadequate logical interpretations of a question by selecting relevant pieces of knowledge in the knowledge base. An evaluation of the work in 2006 showed that this approach is feasible and that complex, multisentence questions can be posed and answered, thus illustrating novel ways of dealing with the knowledge capture impedance between users and a formal knowledge base, while also revealing challenges that still remain.

Enabling experts to build knowledge bases from science textbooks

The long-term goal of Project Halo is to build an application called Digital Aristotle that can answer questions on a wide variety of science topics and provide user- and domain-appropriate explanations. As a near-term goal, we are focusing on enabling subject matter experts (SMEs) to construct declarative knowledge bases (KBs) from 50 pages of a science textbook in the domains of Physics, Chemistry, and Biology in a way that the system can answer questions similar to those in an Advanced Placement (AP) exam in the respective discipline. The textbook knowledge is a mixture of textual information, mathematical equations, tables, diagrams, and domain-specific representations such as chemical reactions. In this paper, we explore the following question: Can we build a knowledge capture system to enable SMEs to construct KBs from the knowledge found in science textbooks and use the resulting KB for deductive question answering? We answer this question in the context of a system called AURA that supports knowledge capture from science textbooks.