Expert model in Compounds, an ITS for second language learn- ing Pascale S´ ebillot, Fr´ ed´ eric Danna Irisa, Campus de Beaulieu, F-35042 Rennes C´ edex, France Paul Boucher Cerlico, Universit´ e de Nantes, F-44000 Nantes, France Keywords: knowledge representation in ITS, expert model, English compounds 1 Introduction In this paper, we present a part of Compounds, an Intelligent Tutoring System (ITS) which addresses the problem of learning English as a foreign language and deals specifically with the formation and interpretation of compounds. Its aim is to bring French students to understand, produce and translate English compounds as fluently as a native Englishman. The reasons that led us to develop this ITS are of two types: on the one hand, the compounding process is an essential part of English and an extremely productive process in this language. Therefore, a French student who wants to learn English needs to know how to produce and understand existing and new compounds. On the other hand, French students have a lot of difficulties to deal with English compounding process; these problems are due to a large extend to the fact that English is a word composition language while French is a stem composition language [Blo33, Ben74]. Moreover, recent linguistic litterature devotes relatively little space to compounding processes in English, and, among the works in this domain, very few concern language pedagogy. An ITS is usually composed of four components [NV88]: the expert model which contains the knowledge of the domain, the student model which contains information on the student and on his knowledge, the teaching module which manages the teaching plans and the interface which permits the communication between the student and the ITS. Here, we focus on the first developed part of Compounds: the expert (English) knowledge and its representation. The linguistic competence of the fluent speaker is idealised as the ability to generate the appropriate combination of words corresponding to a given semantic relationship, and conversely, as the ability to analyse a given surface configuration in terms of its underlying semantic relationship. Among the different possible patterns for English compounds, we have decided to focus firstly on the most productive and easily interpreted ones : N-N (dog-fish), the primary compounds V- N (cut-throat) and N-V (spoon-feed), and the synthetic compounds N-Ving (data-processing), Ving-N (swimming-pool), N-Ver (window-washer) and Ver-N (killer-shark) with right heads. We present the linguistic theories that we have studied in order to determine the rules that permit to find a definition for a given compound and, conversely, to find a compound corresponding to a definition. We have extended these theories to treat all the designated compounds. We explain how these strong linguistic bases enable us to build and represent the expert knowledge in Compounds and we briefly describe the implementation and its results. 2 The linguistic framework We have studied the works of different linguists to determine the expert knowledge for our ITS. Downing [Dow77] is the only one who really treats the problem of N-N compounds. She