Research Summary

Alfred Kobsa

My research over the past 20 years has focused on the development of different kinds of software that facilitates people's use of computers. It spans the whole research cycle from problem analysis (e.g., user needs analysis) to software design, implementation, empirical evaluation, and industrial takeup. I have made significant scientific contributions in the areas of User Modeling, Artificial Intelligence (knowledge representation, expert systems, natural-language systems), Human-Computer Interaction (information visualization, universal design) and expert finders, and to a minor extent in the philosophy of science. These contributions will be detailed below.

1. User modeling and user-adapted interaction

One important way to make computer applications easier to use by people is to enable them to adapt to each individual user. While at the beginning of my research this claim was an unproven hypothesis, it has meanwhile been empirically confirmed for Educational Hypermedia Systems and Customer Relationship Management Systems on the World-Wide Web (see E6 and J14 for surveys), and to some extent for other application domains. In order to be able to cater to individual users, user-adaptive systems must construct user models that contain assumptions about, e.g., users' background knowledge, interests, preferences and misconceptions.

I was one of the first researchers world-wide in this area, and my research contributions both concern foundational aspects (generic user modeling systems, privacy and security) and prototypical applications in several areas (natural-language systems, educational hypermedia systems and universal design). These contributions will be described below. I also helped develop this field by launching the first international workshop and later conference series (P1, P3), by co-authoring and co-editing the first comprehensive surveys (J5, E1, E2), organizing a German national workshop series (P2), founding a scientific society, and by launching the leading scientific journal in this area (E4). The aims of the field (but not necessarily the research results) were recently picked up by industry, with about 50 small companies being launched by the beginning of 2000.

1.1 Generic user modeling systems

The development of user-adaptive application systems is very cumbersome if the user modeling component must be developed from scratch each time. After gaining extensive experiences from integrating such components into Natural-Language Dialog Systems (A2, C5, C6) and explanation components of Expert Systems (J7, C10), I decided to start condensing frequently-used representation and reasoning mechanisms into a generic (i.e., application-independent) user modeling system. In this vein, I defined requirements for their representational and inferential expressiveness (B8, W6) and developed a workbench for a terminological logic (B9, W4, W5, O8) that could be used as a representational basis. When first published, the resulting system BGP-MS (W3, J9, J10) spurred parallel research, most of which is collected in (E5).

Since 1998, commercial systems with related aims have been developed specifically for the personalization of e-commerce applications. As I analyzed in J13 and J15 however, neither these systems nor the academic developments are theoretically and practically satisfactory. In (J17) I describe a completely new paradigm (which is e.g. based on directory systems rather than the previously employed knowledge representation or database systems) that caters to both scientific and application-oriented demands. An implementation of this design is being commercially distributed.

1.2 Privacy and security

User-adaptive applications can adapt better to users the more data they possess about them, and therefore tend to collect as much data as possible and to "lay them in stock" for possible future usage. Their data collection is also deliberately unintrusive, so that users may frequently not be aware of it. Both is in conflict with privacy concerns of computer users that became manifest in numerous recent consumer polls, and with privacy laws that are in effect in many countries. I analyzed this problem more than ten years ago when user-adaptive applications were far from real-world deployment and privacy laws were still few and unspecific regarding online data collection (C9, J8). Since both have dramatically changed in the meantime, I resumed this research recently. To reconcile the benefits of personalized interaction with user concerns and privacy laws, I proposed and partly implemented (a) a security architecture for pseudonymous access to user-adaptive systems, which is generally not affected by privacy laws (C16, J18), and (b) a tailorable architecture for identified access that caters to privacy wishes of each individual user as well as the privacy laws of the country where the user resides (C25, J19). I recently presented both lines of research in invited talks at the international Adaptive Hypermedia and Adaptive World Wide Web Conference and the international User Modeling Conference.

1.3 Applications in Natural-Language Systems

My first application of user modeling components was in two systems that interact with users in a natural language (German and English), namely a "general" Natural-Language Dialog System (A2, C5, C6) and an explanation component of Expert Systems (J7, C10). The latter was a completely new application area. I developed theories on how to construct user models from different types of user input (J2, C4, W9, W10) and to generate appropriate system responses (J2, J6, C5) These application areas served as valuable testbeds for user modeling since their requirements on the depth of modeling and the representational expressiveness are very high. These insights directly impacted the requirements analysis for generic user modeling systems described in Section 1.1.
I recently returned to the problem of user-oriented explanation and presented a model that for the first time distinguishes between knowledge and capabilities that a user needs to carry out a plan and considers both when generating and presenting a plan to the user (C21, C23, C26).

1.4 Applications in Educational Hypermedia Systems

Educational hypermedia is much more constrained in its user modeling requirements than "unlimited" natural-language systems and thus more amenable to practical deployment. I developed the KN-AHS system that adapts the presentation of tutorial text and of images to the presumed knowledge level of each individual student. My paper on this system (C13, reprinted in J11, B13; also B12) was very influential since it divided the tutorial application into a "server" that performed the user modeling tasks, and a client that performed the hypermedia adaptation. This pre-WWW design anticipated the architecture of current web-based educational systems. When tested in a chiropractics domain, however, no significant differences could be found between students that used the adaptive tutorial and students that used a non-adaptive version of the system, which was in stark contrast to results that others had obtained at that time. Analyses that I performed several years later (W11) indicate that the relation between user knowledge and beneficial adaptation are more complex than what research had taken for granted before.

1.5 Applications in Universal Design

People with visual, auditory and motor impairments (including many elderly users) often have difficulties accessing contemporary user interfaces. The current "post-hoc" remedy is to design access systems (e.g. "screen readers" for vision-impaired people) that intercept interfaces on the operating system level, to obtain system output and translate it for the user (e.g., into speech), or to realize non-standard user input. While the leading operating systems fortunately became increasingly open in the past few years for this kind of access, this approach is nevertheless not satisfactory since the development of access systems always lags behind the development of operating systems by two years or more, and since access is only provided to a small part of the functionality of the OS. C15, C17,C18 and J11 describe my work in the paradigm of Universal Design that demands that interfaces should be designed on a very high level of abstraction that is then "translated" into a concrete interface based on a requirement specification of the user's needs (at installation time, or possibly at runtime). My research on the adaptation of information to users with different types of disabilities shows that runtime adaptation can be well realized with the user modeling and adaptation methods described in Section 1. This result was also presented in an invited talk at a recent WebNet conference. In a related effort, C20 describes work on using computers as an empowering technology that can help people partially overcome their handicaps.

2. Expert finders

Expert finding research is a sub-discipline of Knowledge Management that develops systems that assist people who seek the advice of experts. Automated expert finders do not access skills databases, but monitor electronic resources (in an intranet or on the Internet) for evidences of people's expertise, and build expert profiles which they exploit when giving recommendations. I recently analyzed common characteristics of such systems and discussed the advantages and disadvantages of different generic architectures to realize them (J16, B14, W13, W14). Many considerations paralleled those in the area of generic user modeling systems that I described in Section 1.1. This analysis made me convinced that the modeling of user expertise for personalization purposes and the modeling of expertise for expert recommendation have very much in common. A single architecture can be employed that serves both purposes, and the currently used modeling methods seem to complement each other. I recently submitted a grant proposal to further pursue this idea.

3. Information visualization

I studied information visualization from several different angles. In B9, J10, and J11, I used elaborate graphical editors for knowledge representation systems. They provided a far superior access to these systems than the textual interfaces that others researchers in terminological logics used in these days. C7 (later refined in C8) constitutes the first piece of work on the integrated analysis of referential natural language expressions and accompanying pointing gestures for the identification of objects in graphical representations. Such work became again popular a few years ago in the context of multimodal access. Recently I started empirical work in information visualization (C24), led by the belief that we need more and particularly more significant experimentally verified principles that will guide us in the design of visualization systems. Finally, I also contributed to the design of a visualization system for multidimensional data and its commercial deployment.

In a relatively early "former life" of mine, I also studied the relationship between research in Artificial Intelligence and Cognitive Psychology (B2) and the epistimological status of AI programs. In the latter area, I investigated whether AI programs are theories as some researchers believed. I showed that in a deductive-nomological model of explanation, AI programs are not theories in an interesting sense, but may be a step towards interesting theories (C1, C2; J3, reprinted in B7). This work attracted interest in the Philosophy of Science community.