ABSTRACT

This paper briefly describes sentient technology, a patented process for representing knowledge in a computer, and two recent applications that were developed using this technology. The methods used to capture and manage the body of knowledge for each application are discussed. The effort illustrates the successes and difficulties encountered in constructing natural classes for domain-specific applications, using a combination of manual and automated techniques. Lessons learned and possible future refinements to the technology are also discussed.

The application domains are at opposite ends of the knowledge spectrum and range from well structured formalisms to highly unstructured aggregations of ideas. The well structured domain consists of schematic drawings for petrochemical processing plants. The highly unstructured domain is the knowledge corpus on agility, a recent management concept that deals with virtual enterprise and global competition operating under conditions of unplanned change (Goldman, Nagel and Preiss, 1995).

Within the well structured petrochemical plant domain, the totality of knowledge represented in 7,000 schematic drawings was captured. This allowed the user to completely dispose of the drawings and recreate each from knowledge as a virtual document. In addition, a reduction of approximately 1000:1 in the amount of disk storage required to represent the physical configuration of the plant was achieved.

Within the unstructured domain of agility, a map of approximately 1,500 concepts and 50,000 relationships was created. The map resides between a search engine and a corpus of about 150 free text documents. Users with little or no domain knowledge can browse through the concepts and gain an in-depth understanding of their meaning. The concepts are represented through a combination of virtual views, text and digital multimedia artifacts.

AN OVERVIEW OF SENTIENT TECHNOLOGY

Sentient technology is being developed by Cogito, Inc. of Richland, WA, and is currently implemented in a Prolog-based tool called Cogito_TM (Cogito, 1996). The technology has five main aspects that provide a means for:

1) representing knowledge

2) automatically deriving views of knowledge

3) enabling user interaction (including editing) of knowledge through views

4) searching and navigating the knowledge through natural language queries

5) applying learning algorithms to support automated knowledge acquisition.

Cogito's architecture represents knowledge as a network of concepts (nodes) and relationships (links). The representation of concepts follows object-oriented notions of classes and instances; links are represented by a type of concept known as attributes. Under this structure a concept does not change its behavior until it comes into contact with another concept via a relationship. System concepts are special nodes used to define behaviors such as inheritance, learning patterns, identifiers, etc., and provide anchors for a massively interconnected web of nodes and relationships (see Figure 1). The result is a rooted acyclical graph of arbitrary complexity that forms a basis for the representation of natural classes within a network architecture.

A unique aspect of the embodiment is that language is treated separately from knowledge. The software stores each concept as a record the contents of which are simply a unique identifier and pointers to other records. This results in a knowledge representation scheme that is language neutral. For any concept(s), a set of linguistic expressions can be generated by using pre-defined vocabularies, grammars and subsetting constraints. Vocabularies define the symbologies, grammars define the rules for proper arrangement of the symbologies, and subsetting constraints define the portion of the knowledge to be viewed.

The current implementation of the software only supports well structured languages such as schematic drawings. Examples include hierarchies, matrices, process and instrumentation diagrams, flow diagrams and termination diagrams. Although these languages were originally developed for the petrochemical domain, they proved extremely useful in the agility domain as well.

The ability to automatically derive views of knowledge has two major benefits:

1) multiple users can view the same knowledge through an infinite variety of linguistic expressions

2) any view can serve as a point of entry for editing and/or expanding the knowledge base.

This means that updates need only be entered one place, one time. This is a major departure from the way computers are traditionally used, in which views and documents are considered real and the knowledge represented by the documents is virtual. Here the knowledge is real and the views and documents are virtual.

The richness of the relationships determines the extent to which the knowledge map can be navigated. Natural language queries identify concepts that result in a match or near match of the query. The user can explore relationships of nearest neighbors to determine the extent to which the concept is of interest or importance.

Learning patterns observe user behaviors while a knowledge map is being constructed, and provide a means of directing the user when similar structures are being entered. The learning patterns are domain-independent, thereby allowing users to quickly build concept maps across different domains. Examples of applications of the learning patterns will be provided in the next section.

A WELL STRUCTURED DOMAIN: CHEMICAL PROCESSING PLANTS

A recent project involved a petrochemical plant with a need to digitize over 7,000 schematic drawings. When electronically encoded as CAD files, the drawings required 20 gigabytes of hard disk storage. After capturing the totality of the knowledge contained within those 7,000 drawings, along with the symbols and grammars for expressing that knowledge, the entire document corpus could be faithfully recreated using only 20 megabytes of knowledge. As a result, all 7,000 documents could be completely eliminated, since the system could accurately recreate each drawing as well as any number of heretofore undefined views as virtual drawings. Because the virtual drawings retain their connection to the knowledge map, updates to the knowledge can be made through any view. If a connection to a flow control valve changes, and that change is entered on the termination diagram, it will automatically appear on the instrument loop and process instrumentation diagrams, as well as any other diagrams, plans and schedules showing that particular valve.

The construction of the petrochemical plant knowledge base was initially labor intensive. It involved about a half dozen data entry clerks from a temporary employment agency, with the occasional consultation of a chemical engineer to answer questions and reconcile conflicting information. It also required a knowledge engineer conversant in Prolog to define the grammars for the drawings. It was determined that not all 7,000 drawings had to be entered. For instance, all of the knowledge represented in a process flow diagram for a particular set of components is also contained in the process and instrumentation diagram (P&ID). Likewise, all of the knowledge in a cable list is contained in the termination diagram, and the knowledge in the termination diagram is contained in the instrument loop diagram. The implications of this are significant. Rather than storing multiple artifacts representing the same knowledge, the knowledge need only be stored once, along with any languages for expressing that knowledge.

Once the entire ANSI standard schematic symbol set had been entered, it became possible to electronically scan a large portion of the drawing collection. Learning algorithms were applied to system concepts such as composition (e.g., the system observed what types of components make up a subsystem via parent/child relationships in a compositional hierarchy). Connection learning patterns were set up to observe which class notions are typically associated with other class notions through an attribute. For example, the system observed that shielded pairs are usually connected to pressure indicator controllers. As each successive instance of a pressure indicator controller is entered, the system automatically prompts the user to make a positive, negative and neutral connection to the nearest terminal block. This saves time and eliminates the need to explain to each data entry clerk the procedures for making electrical connections. Because the learning algorithms are domain independent, similar prompts are generated when making any type of connection, whether it be hydraulic, functional, logical or organizational.

The cost benefits of this approach are significant. To manually build a P&ID for a receiving tank can take an experienced CAD engineer one or more hours, even when using a high-end graphics workstation. Once the knowledge has been captured, the same drawing, along with all of the related drawings, plans and schedules, can be created in a matter of a few minutes. When an operator is finished using a drawing, it vanishes until it is ready to be used again. As long as the knowledge base is properly maintained, each virtual drawing created will represent the reality of the plant configuration at that moment. Typically, a change to a flow control valve requires that 30 different CAD drawings and schedules be updated. By storing the valve as knowledge, changes need only be entered once, thereby resulting in significant savings in time spent for configuration management, reducing the chance for error and eliminating the need for manually checking consistency among the various documents.

A HIGHLY UNSTRUCTURED DOMAIN: AGILITY

In this section we describe a DARPA/NSF sponsored project recently completed for the Agility Forum (formerly the Agile Manufacturing Enterprise Forum). The purpose was to build and test a prototype knowledge map using sentient technology. The project had two main goals:

1) to capture and manage the rapidly evolving body of knowledge on agility

2) to make that knowledge easily available to the Forum membership.

Prior to the start of the project, the Forum had already begun the implementation of a text retrieval based document management system. The document collection consisted of news and journal articles, multimedia clips, electronic post-it notes, case studies, reports, conference proceedings and a book. In this system the user enters a search string in natural language and retrieves a list of documents, each weighted according to the strength of the "hit." The user must visually scan each selected document to verify the content and extract the knowledge. At any point the user can either:

1) say "yes, that's what I want," and order the document

2) highlight a portion of the text in the retrieved document and say to the search engine "show me more of this"

3) say "no, that's not what I want," and enter another keyword or phrase.

The third aspect turned out to be the most problematic. For a relatively new body of knowledge such as agility, it is not always apparent to the user what to ask for. We addressed this problem by creating a map of concepts that the user could easily browse. Upon finding a concept of interest, the user clicks the mouse button (we call this the "more" command) and the system reveals progressively deeper explanations, along with links to related concepts. The user can enter the map from any number of starting points: agility concept, Forum work group activity, company, industry, person, time frame, location, etc. For example, if the user enters the letters "UK," the system returns <United Kingdom>. Next to United Kingdom the user sees <UK Fine Chemicals, Ltd.>, and next to that, <type 4 virtual enterprise>. The user clicks on virtual enterprise and sees a definition, some more examples of companies that are operating as virtual enterprises, tools such as collaborative software that are successfully applied by these companies, and so on. In essence, the user is able to obtain a great deal of knowledge without ever opening a document. The documents still remain accessible from any point in the browsing process for users desiring more in-depth knowledge about a particular concept.

Our methodology for constructing and maintaining the map was as follows:

1) build an initial concept map through workshops and interviews with the Agility Forum "brain trust"

2) use the map to automatically search the document collection for agile content and link the documents to concepts in the map

3) use the results of the document search to expand and enrich the map.

We built an initial map of approximately 600 concepts. Many of these were created in groupware mode, in which participants were able to observe the creation and expansion of the knowledge map in real time. As participants entered concepts for their particular subdomains, they could observe other subdomains being created and, where appropriate, create links to those subdomains.

The use of different terminology was not a problem. The Agility Forum is a diverse organization---one group expresses a particular set of concepts in management terms, another expresses the exact same concepts in engineering terms. Because of the language neutrality of the system we were able to accommodate these differences in a manner transparent to the groups. We simply created two separate languages for the same set of concepts. In total, we created 14 different virtual views, some of which (tables, flow diagrams, connection matrices) were borrowed from the chemical processing application discussed earlier.

The word "agility" is a term whose definition has evolved over the past several years. Our approach enabled us to capture and represent all previous definitions. The most recent definition was kept in the main portion of the map, while the former definitions were maintained in an archive, but still linked to all other relevant concepts.

Another aspect of the map was that it allowed the Forum staff to identify "holes" in the document collection, something that is not easy to do with a text retrieval engine. For example, in one exercise a user entered a search for virtual collaboration. A "galaxy" view revealed the concepts closely related to virtual collaboration, along with links to the document collection. One component of virtual collaboration was <customer-supplier relations>, which indicated a large cluster of related documents. Next to this was a similar concept called <supplier-supplier relations> with no related documents.

During the knowledge elicitation process, it was determined that supplier-supplier relations are an important aspect of virtual collaboration, however, there were no documents in the collection to which a user could refer for more information. This alerted the Forum to direct some of its activities to collecting data, case studies etc. focusing on supplier-supplier relations. A pure data mining approach would not have discovered this hole, since the concept itself did not reside in the document corpus.

Data mining approaches were useful, however, in discovering concepts in the document collection that were not represented in the knowledge map. By using the data mining features of Cogito's discovery module, we were able to expand the initial map from 600 to about 1,500 concepts.

We learned that the technology had other useful features. Users could take the Forum's body of knowledge on agility and build customized maps for their own organization. Users could also use the map to create customized documents assembled from views generated as a result of their knowledge browsing. Virtual briefing charts were also created. As terminology changed and new information was incorporated, tables, charts and matrices would be automatically updated to reflect those changes. Like in the petrochemical plant example, changes need only be entered one place one time; there is no need to search through megabytes of briefing charts, figures and tables to accommodate an update.

FUTURE DIRECTIONS

In terms of automated knowledge extraction, we believe we had a reasonably good capability for manually identifying a critical mass of concepts and relationships, and then using automated techniques to expand the network into a natural class for the application domain. While the automated techniques were extremely useful for discovering new classes, they fell short in discovering relationships (such as causality, sequence, similarity, orthogonality, differentiation, etc.) among the classes. For our next phase, we plan to incorporate a set of semantic primitives (see Wierzbicka, 1996) into the discovery process. This will allow us to better identify and categorize class notions along with their many complex inter-relationships.

A more serious drawback was the user interface, which had an extremely limited set of fonts and styles. The system was also not usable over the web, which was a significant disadvantage to the Agility Forum, which uses its web page as its primary knowledge dissemination medium. While many prototyping efforts focus on the interface and on maximizing user-system interaction, we chose to concentrate our efforts on the core engine and the knowledge representation issues, at the expense of the interface design. In our user evaluation sessions we assembled a collection of user interface design specifications that will make the system more appealing to users at large. Most of the user interface problems arose in the unstructured domain; users in the well structured domain, primarily engineers, were quite satisfied with the interface. However, both domains were limited in the ability to interface with other applications. Future versions will require a more open architecture with seamless interfaces to existing applications.

SUMMARY

Based upon our experience to date, we believe sentient technology has two significant implications, depending upon whether the application domain is well structured or highly unstructured. For the well structured domain, sentient technology is a major step toward the elimination of documentation. This represents a radical shift in the focus of computing away from the storage of digital artifacts toward the storage of knowledge and language, treated as separate and distinct notions.

For the unstructured domain, sentient technology can be used to create a middle layer between a user and a corpus of documents. The middle layer is a concept map that allows a user to navigate the concept space and learn about the domain through progressive deepening, rather than manually trying to assimilate knowledge from documents retrieved by a search engine.

Although the domains are quite different, the approach used to build a natural class for each was strikingly similar. In each case, a critical mass of knowledge had to be accumulated using a combination of manual data entry and knowledge engineering. For the well structured domain much of the knowledge acquisition was performed manually by data entry clerks with little or no domain experience, augmented by occasional consultations with domain experts and supported by domain-independent learning modules. For the unstructured domain, manual entry by the user organization's staff members and "brain trust" using groupware in a workshop setting proved to be a tenable method.

In both cases, after a critical mass of knowledge had been acquired (approximately 600 concepts) the knowledge acquisition process gradually became more automatic. This occurred at a point in time when:

1) a reasonably large set of high level concepts was defined (similar to the "binning" process referred to by Prueitt (1996) in constructing an inverted index of themes)

2) the vocabularies and grammars became solidified.

Through both the manual and automated acquisition processes, a knowledge engineer and/or domain expert had to be available on a consultative basis to resolve discrepancies and to perform periodic validation audits. The need for such consultations became much less frequent in the later stages of development.

Much interesting work is being done in the construction of natural classes without a priori knowledge about the context of the document corpus (Wise, et al., 1995). This is probably useful for large collections of free text documents. The strategy described in this paper appears to be an approach more suited for non textual collections, especially schematic drawings, and for smaller text document corpora, where there may not be enough occurrences of certain themes to form a meaningful aggregation. As we continue on the road to the automated acquisition and management of knowledge, all of these methods, and others yet to be developed, will likely be necessary.

ACKNOWLEDGMENTS

The work described in this paper was supported in part by DARPA/NSF project DDM#932095I, along with in-kind contributions from Cogito, Inc., and Telart Technologies.

REFERENCES

Cogito, Inc., The Sentient Primer, Richland WA, 1996.

Goldman, Steven L., Roger N. Nagel and Kenneth Preiss, Agile Competitors and Virtual Organizations, Van Nostrand Reinhold, 1995.

Prueitt, Paul S., "Ontology Based Document Understanding," Notate '96 Conference, Washington, DC, May 21, 1996.

Wierzbicka, Anna, Semantics: Primes and Universals, Oxford University Press, New York, 1996.

Wise, James A., et al., "Visualizing the Non-Visual: Spatial Analysis and Interaction With Information From Text Documents," IEEE, 1995.