Acquiring the Know-How of an Expert

For START to be expert at retrieving Web based assets it will need to have as much of the knowledge of an expert in the domain as possible. We hope to make START useful not only at retrieving documents, but also at interpreting information told to it about developing situations. As new facts are told to it, START should initiate problem solving efforts to understand the impact of the information in light of START's knowledge of Indicators and Warnings, causal chains, and the like. Effective problem solving will depend on possesing the problem solving know-how of an expert in the domain. As above, we believe that this information will be both volunteered and acquired interactively in reponse to breakdowns:

• Volunteered expertise: Many users will be willing to explain to HAWK how to go about solving typical problems. Such problem solving procedures are usually thought of as networks of subgoals which collectively are expected to solve the main goal. Ideally, these are best described using a mixture of diagrams and natural language (see section on the Intelligent Room below) but we believe that language alone can be a useful medium; however, there is the danger that the volunteered procedure may be overly constrained, because language tends to linearize the description.

• Expertise solicited in response to breakdowns: When a user submits annotations, HAWK will encourage her to also submit a set of questions which should be answerable. If these questions cannot be answered, or if the answer doesn't include the annotated object, then a breakdown has occurred. HAWK can present to the user the partial chain of reasoning which failed to reach the goal; this will help the user understand what information is missing. At this stage, new information or new rules may be provided. Rule-like information is provided as English If-Then sentences; HAWK will examine the vocabulary used in the rule and show the user the lexical categories taxonomically above each significant word. The user will then tell HAWK how much to generalize each vocabulary element in forming the rule.