Knowledge Acquisition Incorporating Interactive NL Understanding

Introduction A central issue in building knowledgebased systems is making them available to and understandable by naive users. This demonstration shows how a user and a knowledge-based system can collaborate to both create new knowledge and to extract existing knowledge from a knowledge base via natural language. KRAKEN1 is an interface designed for users who have expert knowledge of some field, but no special training in knowledge representation, logic, or the like. Our basic assumption is that the interfacing medium between the user and the knowledge base should be as close as possible to natural language (though see (Clark et al. 2001) for another possible approach). The demonstration will show how we parse from English into CycL (the logical representation used by the Cyc ontology and inference engine) using a variety of different parsing methods and post-processing steps, and will also demonstrate how this versatile representation language can be rendered back into English that is suitable for lightly trained users. These NLP tools are integrated with a large number of knowledge-engineering tools, since we (both Cycorp and the NLP community at large) have not succeeded in building a natural language system that can do everything necessary to construct knowledge via simple back-and-forth NL dialogue with a user. Cyc’s natural language understanding abilities consist of a lexicon with syntactic and semantic information, a hybrid top-down/bottom-up parsing system, a reformulation module, and a generation system.