Researchers in computational linguistics have long acknowledged that understanding language requires understanding the world. It is people’s extensive commonsense knowledge of the world that enables them to perform with ease so many of the linguistic tasks that are difficult for computers: resolving ambiguity, finding referents for pronouns, determining a speaker’s true intent--in general, determining which of many possibilities makes the most sense. Most natural language systems have worked only in small, toy domains, because of the daunting task of providing the massive knowledge bases needed for larger domains. The Cyc project to encode large bodies of knowledge for use in AI systems [1] now offers the opportunity to build a large-scale natural language system and see if methods developed for small systems do indeed scale up, to see whether a large knowledge base does make all the difference.

The Knowledge-Based Natural Language (KNBL) project is the natural language companion to the Cyc project. It has two main components: Lucy converts an English sentence to a representation in CycL, the language of the Cyc knowledge base, and Koko produces English from CycL expressions. These systems combine a number of state-of-the-art techniques in language understanding and add a few new techniques. For example, the idea of a chart, a two-dimensional structure used in parsing, is generalized to a three-dimensional structure in which the third dimension represents alternatives used in semantic processing. The linguistic knowledge in Lucy and Koko is carefully separated from the world knowledge in the Cyc knowledge base.

The interesting question, of course, is what Lucy gains from having Cyc available as a resource. While the authors necessarily gloss over details in this overview paper, it seems that Lucy is able to get from Cyc the information necessary to solve such puzzles as lexical ambiguity, nominal compounds, and metonymy. For example, it can determine that “corn oil” is made from corn, because Cyc knows that corn is the sort of stuff that can be processed into oil. On the other hand, “baby oil” is oil for the use of babies. This kind of reasoning is not new, but KNBL may be the first system in which the knowledge base that supports it is not built specifically for this purpose.

In addition to Lucy and Koko, the authors describe a tool, Luke, that helps a system builder add words and their senses to the system dictionary. Luke is designed so that its user need not be a natural language expert; Luke uses the meanings of words to guess many of their syntactic properties and presents these guesses for confirmation. Because it can be difficult to navigate in a large lexical structure, Luke may be instructed in English.

Much work remains to be done before it can be said whether the use of a massive knowledge base will do what we hope it will for natural language systems. The KNBL project is reason for optimism.