Int. J. Man-Machine Studies (1987) 27, 127-144 Knowledge elicitation using discourse analysis N. J. BELKIN AND H. M. BROOKS The School of Communication, Information, and Library Studies, Rutgers University, New Brunswick, N.J. 08903 U.S.A. P.J. DANIELS Admiralty Research Establishment, Ministry of Defence, Queens Road, Teddington, Middx. TWII OL9, U.K. This paper is concerned with the use of discourse analysis and observation to elicit expert knowledge. In particular, we describe the use of these techniques to acquire knowledge about expert problem solving in an information provision environment. Our method of analysis has been to make audio-recordings of real-life information interactions between users (the clients) and human intermediaries (the experts) in document retrieval situations. These tapes have then been transcribed and analysed utterance-by-utterance in the following ways: assigning utterances to one of the prespecified functional categories; identifying the specific purposes of each ut- terance; determining the knowledge required to perform each utterance; grouping utterances into functional and focus-based sequences. The long-term goal of the project is to develop an intelligent document retrieval system based on a distributed expert, blackboard architecture. 1. Introduction We are concerned with the problem of designing intelligent automated interfaces to mediate between people who feel they require information, and the (usually) computer-based knowledge resources which might contain information which could be of use to them. These three elements: user; intermediary; and knowledge resource, and the relations among them, constitute the general information system which is our focus of attention. The intermediary and knowledge resource elements together constitute the information provision mechanism (IPM). Such systems arise in, for instance, social security benefits offices, student advisory services and bibliographic retrieval systems. At present, almost all such information systems require a human intermediary. We assume that any automated interface will need to perform at least some of the functions that human intermediaries perform (as well as being capable of recognising situations when human intermediaries are necessary); such an interface would perforce be intelligent. Human intermediaries are required in such systems for a number of reasons. Generally speaking, the users in such systems are unfamiliar with the contents, structure and access mechanisms of the data base, nor should they be required to be, for they are at best intermittent participants in any one such system. The intermediary's role in this respect is to use his/her knowledge to choose the appropriate data base(s), to structure the search appropriately to the data base, and to formulate and apply the relevant query. These tasks might seem somewhat 127 0020-7373/87/080127+ 18503.00/0 9 1987 Academic Press Limited