Phrasing a text in terms the user can understand* John A. Bateman and Cecile L. Paris USC/Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 Abstract When humans use language, they show an es- sential, inbuilt responsiveness to their hear- ers/readers. When language is generated by machine, it is similarly necessary to ensure that that language is appropriate for its intended audience. Much of previous research on text generation and user modelling has focused on building a user model and selecting appropriate information from the knowledge base to present to the user. It is important, however, that the phrasing of a text be also tailored to the hearer - otherwise it may be just as ineffec- tive as texts which wrongly direct attention or which rely on knowledge that the hearer does not have. This research proposes a new mech- anism which allows the text planning process to specifically tailor syntactic phrasing to the hearer type. This is done in the context of an expert system explanation facility that needs to produce explanations of the expert system's behavior for a variety of different users - users who differ in goals, expectations, and expertise concerning both the expert system and its do- main. 1 Tailoring - the importance of making language appropriate for its audience Humans show an essential, inbuilt responsiveness to their hearers/readers in their use of language. It is simi- larly necessary to ensure that the language generated by machine is appropriate for its intended audience. Much text generation research in the past has focused on the selection of text content and organization in or- der to accomplish speakers' goals (e.g., [McKeown, 1985, Ilovy, 1988]). The presentation of that content is gener- ally also made responsive to hearers' states of focus of attention (e.g., [Grosz and Sidner, 1986]). More recently, it has been recognized that it is necessary in addition to make generated text sensitive to the hearers' goals and knowledge about domains, that is to take a user model into consideration (e.g., [McKeown et a/., 1985, Appelt, 1985, Jameson, 1987, Ilovy, 1988, Paris, 1988, Carberry, 1988]). These are all important factors if the text generated is to be both informative and understand- able to the user. The language used by specific groups of people, how- ever, often possesses syntactic patterns and lexical fea- tures that are distinctive to those groups; question an- swering systems can only be effective, therefore, if they appropriately customize their phrasing as well as their content and textual organization according to each dis- tinctive group of users - otherwise generated texts may be just as ineffective as texts which wrongly direct at- tention or which rely on knowledge that the hearer does not have. In contrast to most previous research, which has focused on the selection and organization of infor- mation from the knowledge base for presentation to the user, this research addresses the issue of expressing the selected information in language specifically tailored to the hearer. 1 This is done in the context of an expert system explanation facility. 2 What is involved in 'tailoring' In tailoring a response, whether it be to a user's goals for asking a question or to that user's level of exper- tise in a domain, a generation system first has to choose which information from the knowledge base at hand is most appropriate and organize its overall text struc- ture. The result of this phase is an organized collec- tion of the particular propositions to be expressed in English. Most generation systems take these proposi- tions as the inputs for the realization components of grammatical and lexical selection (e.g., [McKeown, 1985, Moore and Swartout, 1989]). However, the output of this phase is typically not detailed enough to control the many possibilities for expression that current grammar components provide. There is a large gap between the *The research described in this paper was supported in part by the Defense Advanced Research Projects Agency (DARPA) under a NASA Ames cooperative agreement num- ber NCC 2-520, by AFOSR contract F49620-87-C-0005, and by DARPA contract MDA903-87-C-641. 1 This issue was partially addressed in the HAM-ANS sys- tem [Morik, 1985], where, based on a user model, the system would decide whether to produce an anaphora. The work presented here is different as we are more concerned about systematic linguistic differences that exist in the language used by various groups of users. Bateman and Paris 1511