コネクショニストアプローチによるインタフェースのユーザモデル

抄録

To design intelligent and flexible computer interfaces the present paper develops models of a user which aim at guessing intention behind actions of the user. There are two types of user models : rule-based models and connectionist models. The author chooses the latter here, because they have various advantages over traditional rule-based models. Basic mechanism of the connectionist models being multiple constraint satisfaction, they go well with the requirement for guessing the intention of the user based on fragmentary information from various sources. Another advantage of the connectionist models is their flexibillty in learning and generalization. Developed in the paper are two kinds of connectionist user models. One is a model for typing error corrections of UNIX commands. The other is a group of models for extracting features from a sequence of UNIX commands. Various kinds of typing errors can be categorized into two : temporal and spatial errors. Representing these two kinds of typing errors on a network is the key to modelling the error correction problem. To represent the temporal ambiguities in typing a coarse coded layer and cross links between this layer and an input layer are employed. To represent the spatial ambiguities in typing special coding scheme is used for representing characters in a command : Hamming distance between a pair of characters be the same as the spatial distance between the corresponding pair of keys on a keyboard. Many pairs of erroneous and corrected commands are given to an input and output layers of the network as training patterns. It turns out that error correction capability of this model is not so good for short commands due to insufficient clue to restoration. To improve its performance this model assimilates other sources of information. The first one is the contextual information : history of commands and statistical features of the sequence. This, however, does not work satisfactorily due to large variation of statistical features even within the same user. The second source of information is the typing error characteristics of the user : frequency of each kind of typing errors. This information contributes to the restoration of erroneous typed commands. For extracting features from a sequence of commands, a model for obtaining command occurrence probabilities and models for command transition probabilities are developed using the so-called buffer model. These models are expected to suggest users candidates of useful macro commands. Much remains for further study, but the performance of the proposed connectionist models demonstrates their potential effectiveness in the user interfaces.