Paradigm Shift in Artificial Intelligence

John F. Sowa

During the 1980s, the dominant approach to knowledge acquisition required two kinds of highly trained, highly paid professionals. At the top of Figure 1, a knowledge engineer is interviewing a subject matter expert in order to capture her knowledge and encode it in the arcane formats of an AI system. Meanwhile, computational linguists, who were designing natural-language tools, tried to make them translate NL documents into similar encodings without requiring any human intervention. At the bottom of Figure 1, a physician who is examining a patient scribbles some notes on a sheet of paper, which some clerk will later transcribe for the computer. Then the NL tools will attempt to convert those notes to the formats specified by the knowledge engineer.

Subject matter expert talking to a knowledge engineer

Figure 1: Twentieth century approaches to knowledge acquisition and NL processing

There are two things wrong with Figure 1: the top row requires far too much human effort, and the bottom row is expected to process unrestricted natural language without any human assistance. To reduce the cost of two high-priced experts, some developers merged the two roles at the top row into one: either the subject matter expert learned knowledge engineering, or the knowledge engineer learned enough about the subject matter to extract knowledge from documents. Yet people with expertise in both fields became even more expensive to find, hire, and train. Figure 2 shows a better alternative: simplify the tools and the training required by the people who use them. Instead of designing complex NL tools that process documents without human intervention, AI researchers developed simpler knowledge extraction (KE) tools that can extract knowledge from documents with assistance from just one human editor. Furthermore, the editor communicates with the KE tools in a controlled natural language, which people can read without special training.

A knowledge editor using KE tools

Figure 2: Replacing two experts with one editor

The editor in Figure 2 represents various people who at different times might play different roles with respect to the subject matter, the computer system, and the people and activities involved with them. Each of the three people in Figure 1 has a different kind of expertise. Any of them might use KE tools to edit their knowledge or to write a note, a report, or a book that someone else might edit with the aid of KE tools. Following are the three kinds of knowledge:

From an editor's point of view, a KE system looks like an intelligent word processor combined with sophisticated tools for searching, classifying, summarizing, and paraphrasing. After the output has been revised by a human editor, who might be the original author of the documents, the result can be stored in a knowledge base or be written as an annotation to the documents.

Doug Skuce (1995, 1998, 2000) has designed an evolving series of knowledge extraction systems, which he called CODE, IKARUS, and DocKMan (Document-based Knowledge Management). All the input to the knowledge base, whether generated by the KE tools or entered directly by an editor, is represented in a version of controlled natural language called ClearTalk. The KE tools have the following advantages over the older systems represented by Figure 1:

As an example, the students in Skuce's operating systems course used the KE tools to map information from on-line Linux manuals to a knowledge base for a Linux help facility. The people who wrote the manuals were experts, but the students who edited the knowledge base were novice users of both Linux and the KE tools. As another example, Skuce built a simple knowledge base about animals for his 9-year-old daughter's school project. She and her class could browse the knowledge base on the web, and they had no difficulty in understanding every fact presented in ClearTalk.

The oldest logic patterns expressed in controlled natural language are the four types of statements used in Aristotle's system of syllogisms. Each syllogistic rule combines a major premise and a minor premise to draw a conclusion. Following are the four logic patterns:

  1. Universal affirmative. Every employee is a person.

  2. Particular affirmative. Some employees are customers.

  3. Universal negative. No employee is a competitor.

  4. Particular negative. Some customers are not employees.
These patterns and the syllogisms based on them are used in ClearTalk and many other controlled language systems. For inheritance in expert systems and object-oriented systems, the major premise is a universal affirmative statement with the verb is, and the minor premise is either a universal affirmative or a particular affirmative statement with is, has, or other verbs. For constraints in databases and knowledge bases, the major premise is a universal negative statement that expresses an illegal subject-predicate combination. Another important class of logic patterns are the if-then rules used in expert systems. Following are two such rules for a library database, written in Attempto Controlled English (Fuchs et al. 1998; Schwitter 1998):
If a copy of a book is checked out to a borrower
   and a staff member returns the copy
then the copy is available.
If a staff member adds a copy of a book to the library
   and no catalog entry of the book exists
then the staff member creates a catalog entry
        that contains the author name of the book
           and the title of the book
           and the subject area of the book
   and the staff member enters the id of the copy
   and the copy is available.
The Attempto system translates rules of this form to an executable program in the Prolog language.

Over the past thirty years, many natural-language query systems have been developed that are much easier to use than SQL. Unfortunately, one major stumbling block has prevented them from becoming commercially successful: the amount of effort required to define the vocabulary terms and map them to the appropriate fields of the database is a large fraction of the effort required to design the database itself. However, if appropriate KE tools are used to design the database, the vocabulary needed for the query system can be generated as a by-product of the design process. As an example, the RÉCIT system (Rassinoux 1994; Rassinoux et al. 1998) uses KE tools to extract knowledge from medical documents written in English, French, or German and translates the results to a language-independent representation in conceptual graphs. The knowledge extraction process defines the appropriate vocabulary, specifies the database design, and adds new information to the database. The vocabulary generated by the KE process is sufficient for end users to ask questions and get answers in any of the three languages.

Translating an informal diagram to a formal notation of any kind is as difficult as translating informal English specifications to executable programs. But it is much easier to translate a formal representation in any version of logic to controlled natural languages, to various kinds of graphics, and to executable specifications. Walling Cyre and his students have developed KE tools for mapping both the text and the diagrams from patent applications and similar documents to conceptual graphs (Cyre et al. 1994, 1997, 1999). Then they implemented a scripting language for translating the CGs to circuit diagrams, block diagrams, and other graphic depictions. Their tools can also translate CGs to VHDL, a hardware design language used to specify very high speed integrated circuits (VHSIC).

Design and specification languages have multiple metalevels. As an example, the Unified Modeling Language has four levels: the metametalanguage defines the syntax and semantics of the UML notations; the metalanguage defines the general-purpose UML types; a systems analyst defines application types as instances of the UML types; finally, the working data of an application program consists of instances of the application types. To provide a unified view of all these levels, Olivier Gerbé and his colleagues at the DMR Consulting Group implemented design tools that use conceptual graphs as the representation language at every level (Gerbé et al. 1995, 1996, 1997, 1998, 2000). For his PhD dissertation, Gerbé developed an ontology for using CGs as the metametalanguage for defining CGs themselves. He also applied it to other notations, including UML and the Common KADS system for designing expert systems. Using that theory, Gerbé and his colleagues developed the Method Repository System as an authoring environment for editing, storing, and displaying the methods used by the DMR consultants. Internally, the knowledge base is stored in conceptual graphs, but externally, the graphs can be translated to web pages in either English or French. About 200 business processes have been modeled in a total of 80,000 CGs. Since DMR is a Canadian company, the language-independent nature of CGs is important because it allows the specifications to be stored in the neutral CG form. Then any manager, systems analyst, or programmer can read them in his or her native language.

No single system discussed in this paper incorporates all the features desired in a KE system, but the critical research has been done, and the remaining work requires more development effort than pure research. Figure 3 shows the flow of information from documents to logic and then to documents or to various computational representations. The dotted arrow from documents to controlled languages requires human assistance. The solid arrows represent fully automated translations that have been implemented in one or more systems.

A knowledge assistant using KE tools

Figure 3: Flow of information from documents to computer representations

For all these tools, the unifying representation language is logic, which may be implemented in different subsets and notations for different tools. Aristotelian syllogisms together with if-then rules provide sufficient expressive power to specify a Turing machine, and they support efficient computational mechanisms for executing the specifications. For database queries and constraints, statements in full first-order logic can be translated to SQL. All these subsets, however, use the same vocabulary of natural-language terms, which map to the same ontology of concepts and relations. From the user's point of view, the system communicates in a subset of natural language, and the differences between tools appear to be task-related differences rather than differences in language.


For more detail and references, see the paper Ontology, Metadata, and Semiotics.