In this paper we discuss a proposed user knowledge modeling architecture for the ICICLE system, a language tutoring application for deaf learners of written English. The model will represent the language pro ciency of the user and is designed to be referen

sessions. New users will require the system to initialize the model according to the input they provide in the rst sample of writing they enter3 . In the list of information sources above, we mention both explicit and implicit information provided to the model by user input. In the direct sense, a user's writing is a uniquely rich source of language pro ciency information. In comparison to the techniques other systems use to determine user knowledge such as polling, where one question is only likely to reveal one point of data (either the user understands or does not understand the concept in question), even a short multi-sentential piece is going to o er many points of data per utterance. Every grammatical construct successfully or unsuccessfully used, from determiner choice to word order, provides information about the user. These points can be correlated to provide a map of those constructs consistently used, those which are experiencing variation, and those which are absent therefore, even during the initial tutorial session, we are provided with a fairly rich source of explicitly-derived data about this individual, compared to what we could obtain from questioning the student. Relying on a subjective categorization of language ability from a teacher would also be less accurate, as it is di cult to classify discrete levels of achievement in this domain, judgments are likely to vary between instructors, and categories would translate roughly at best to tags on the myriad individual language features. Once the user placement and initial notation has occurred, implicit information can also be obtained if a given feature is highlighted as within the ZPD for a student, this implies indirectly that features indicated as adjacent by lateral links to the other hierarchies are also in the ZPD, and that features above or below the ZPD are unknown or well-known respectively. In the absence of direct evidence to contradict these conclusions, the user model allows for this inferencing to produce reasonably certain conclusions. This is how we plan to exploit the\stereotypic reasoning" suggested by (Wahlster3 Unfortunately, the rst session of error identi cation has to proceed without the assistance of a user model, but it will be aided by other data such as the expected co-occurrence of certain errors mentioned earlier.

and Kobsa, 1986), since the feature organization in SLALOM is based on a stereotypic acquisition order. Note that we do not recommend explicitly marking the inferred knowledge in the model following the lead of other explanation systems, implicit information in the user model can be derived at any time through inferencing, and thus should not be marked in the model so that it may be distinguished from explicit and con rmed information. Stereotypic information may not hold true for every individual, and we wish to distinguish between that information which we know from actual user performance and that

which we infer from our pro le of a typical learner. When the system makes reference to this model during tutorial response generation, it must take note of whether it is drawing from an explicit or implicit source the implicit information is less reliable, and our planner will again follow the conventions of similar systems and mark such inferences directly in its text plan for the purposes of recovery should they turn out false. It is hoped that the use of implicit user model information will be constrained to the early sessions with a given user only, since as argued above the ICICLE user model should be rich with explicit information and should be well- lled with direct information from user input in very few sessions. However, with revision of the representation over time, some tags may become less certain and the ability to infer additional information may be useful.5.2

In ICICLE, the responsibility for updating the model of a user lies with the error identi cation module, since that facet of the system processes all of the major parts of user performance. Each new analysis provides new (and potentially different) information that should be directed to the model. Because the user's knowledge is expected to change over time, so must the model. There is also the possibility that the user model is incorrect even a rich model such as the one proposed for ICICLE may contain faulty data, so the system must be capable of revising earlier notations. A model that can be overwritten over time gives rise to the question of whether new data should always champion over the old. The out-

Updating the Model