Ferrets and topic maps: knowledge engineering for an analytical engine

The “Ferret” analytical engine, developed originally by the Y-12 National Security Complex of the U.S. Department of Energy to seek classified data and associations in documents and present its findings in the light of formal rules, requires a structured information base that represents not just individual facts but a set of implications and a collection of rules. The fundamental knowledge base is evolving towards forms that enhance flexibility and portability. The developers early realized that the knowledge base can be captured in XML by a series of trees that represent taxonomies, analytical structures, and specific indicative facts, but over this a topic map is needed to express links across the trees. Above this, the classification rules could form another topic map that points into the lower layers. In its latest form, however, the knowledge base has come to be entirely represented in a topic map. The “Ferret” engine combines sophisticated searching with rule-driven analysis and reporting. In its original application, the Ferret engine performs the equivalent of 5,000 simultaneous searches while reading documents at several thousand words per second. The analysis traces implications of concepts discovered in searching and applies the rules for interpreting implications and the actions to be taken when a significant piece of information is found. Because the topic maps that represent this knowledgecan be switched easily, Ferret can be reprogrammed to many tasks, including selection and categorization, scanning of e-mail and newsfeeds, diagnostics, and query expansion, in addition to the original classification application. Information Classification and the Origins of the Ferret System When the Y-12 National Security Complex (Y-12), a manufacturing facility of the U.S. Department of Energy (DOE) in Oak Ridge, Tennessee, started developing tools to support its management of classified documents, it was faced with the task of capturing the knowledge of how to identify classified information. Once captured, such knowledge would have to be stored in a maintainable fashion that was also accessible to Ferret, the automated analytical tool that we had developed. The Ferret project team initially developed a knowledge base as part of the program development. Since this hand-built base was difficult for anyone other than the original developer to maintain, the team soon settled on a knowledge base in XML that depends on developer to maintain, the team soon settled on a knowledge base in XML that depends on some familiar techniques, like tables and hierarchical trees, and adds to them an adaptation of the new techniques of topic maps (ISO/IEC 13250:2000). The knowledge base is now in transition to a topic map representation based on the XTM (XML Topic Map, www.topicmaps.org) specification. Since the original classification project, the applications for both the Ferret engine and the knowledge-engineering techniques have expanded. Although Y-12 is no longer involved in the original function for which it was created as part of the Manhattan Project during World War II—the final enrichment of weapons-grade uranium—it has retained a major role in the making and maintaining of components for the U.S. thermonuclear stockpile. Accordingly, much of the information handled at the plant is classified and must be protected. Decisions about what is actually classified are made by DOE on a national basis and adapted to specific local situations by facilities like Y-12. Day-to-day classification decisions are made on the basis of this approved guidance by authorized derivative classifiers (ADCs), who form the front line of defense for classified information. The first application of the Ferret engine, developed as a tool to support the ADCs in their work, reads electronic documents and highlights potentially classified passages, displaying along with each portion of text the proposed classification and the rules from the guidance that support the classification. Although the work of the ADCs is grounded in the formal classification rules for identifying classified information, the practical application of those rules depends on much more detailed knowledge than is contained in the published guidance. Recognition of significant information depends on knowledge of the manufacturing process, the design of the products, and the properties of the materials of which the products are made. It also depends on an awareness of what decisions have been made in the past and what information is available to the general public at the unclassified level. Finally, the ADC must be able to draw inferences from the combined collection of information. In addition to the details of product designs and manufacturing, the classification process must recognize numerous pieces of indirect information. Many parts and materials have been given codenames so that they can be discussed without revealing classified data. To elaborate on one of these codenames might constitute a breach of security. There are many specific facts, such as the inventories of certain materials and the rates at which they are used in manufacturing, that may be classified. Some facts are not themselves classified, but in combination they can add up to classified data. For example, mentioning a particular product in conjunction with certain buildings might reveal something of the product’s components if those buildings are known to process only certain materials. Mentioning a geometric attribute might imply the overall shape or configuration of a part. General properties of materials, such as metals and plastics, constitute a large part of the knowledge. Individually, most of the facts about materials—things that might be learned from any chemistry or physics text—are not classified. But in the particular context of Y-12’s products, these unclassified facts may suggest sensitive information. Part of the role of the ADCs, and thus of the Ferret system that supports them, is to recognize when such combinations have occurred in our context. How Ferret Works: The Classified Automobile Classification analysis is generally done by comparing the information in question to formal guidance that has been developed by the appropriate authorities. Guidance is usually written in guidance that has been developed by the appropriate authorities. Guidance is usually written in terms of general concepts, such as the high-level design of our products and the materials used in them, that we need to protect. While some broad guidance is written in narrative form, most of the specific guidance is presented in tabular form. Each rule in a table states a condition to be evaluated and associates with it a resulting classification to be applied if the document under evaluation meets the condition in question. Frequently these rules form a series of conditions reflecting increasing detail to be sought in candidate documents and thus increasing levels of sensitivity and need for protection. If we were in the automotive industry, we might have classification rules that look something like the following table: