Information gathering systems do not typically find all the information desired by the user on their first try, and so a cycle of query refinement occurs. Information retrieval systems -- which classify documents as either relevant or irrelevant to a user’s query -- allow the user to refine a query either directly by changing the wording of the query, or indirectly by examining search results and accepting or rejecting documents, integrating these selections into the query as positive and negative search terms. Studies of how library patrons interacted with (human) librarians suggest that queries evolving over time is a natural component of the information gathering process (e.g., (Bates in press), (Twidale & Nichols 1998)). Query refinement or evolution is not well supported, however, in many of the current information extraction (IE) systems -- which locate specific query-relevant information in documents and present this information to the user. It is our belief that the common "naive" user would benefit from a more flexible cycle of interaction and query refinement during information extraction, a cycle in which a shared representation for system knowledge would enable the user and system to negotiate their roles and vary their levels of initiative within a given task (Vanderheyden & Cohen 1998).