Exploiting Conceptual Knowledge for Querying Information Systems

Exploiting Conceptual Knowledge for Querying Information Systems Joachim Selke and Wolf-Tilo Balke Institute for Information Systems, Technische Universität Braunschweig, Germany Abstract. Whereas today‟s information systems are well-equipped for efficient query handling, their strict mathematical foundations hamper their use for everyday tasks. In daily life, people expect information to be offered in a personalized and focused way. But currently, personalization in digital systems still only takes explicit knowledge into account and does not yet process conceptual information often naturally implied by users. We discuss how to bridge the gap between users and today‟s systems, building on results from cognitive psychology.

1. Introduction Given today‟s information overload querying digital information in a personalized manner has become a more and more demanding problem. To a large degree this information is stored in structured repositories and databases allowing for efficient retrieval. Commonly used query languages require users to explicitly specify hard constraints to describe their information needs. In most cases however, this is extremely difficult for the user, since information needs are often vague and not all relevant parameters are known a priori, i.e., before interacting with the information system. In fact, much knowledge is embodied in the user or perceived as being common knowledge (Balke and Mainzer 2005). Moreover, when evaluating such queries over a repository or database, very often empty results are returned in case of overspecified queries, or the user is flooded by irrelevant information when posing underspecified queries. Recently, these problems have stimulated research in the area of personalized query processing. But still only few approaches actually focus on the cognitive processes underlying human information search. A first step to human-centered information search are preference-based retrieval models, which build on simple pairwise item comparisons, in the form of „items with property A are better than those showing property B.‟ This kind of statement is very intuitive for users and easy to articulate. Characteristic properties of items in a collection usually can be described within a fixed attribute scheme. For example, used cars on sale can be described in terms of brand, model, age, color, or top speed. Users now are generally able to specify their preferences on attribute level like „I like VWs better than Ferraris‟ or „I prefer red cars to yellow ones.‟ Often such characteristic attributes are simple and intuitive representations of concepts that would be hard to describe otherwise. Considering the sample statements above, the user might have an understanding that VWs are typical „normal‟ cars, whereas Ferraris are typical examples of sports cars. On the other hand, the above color preference does not carry any information about other properties a user might want for his/her car (Leslie, 2007; Leslie, 2008). Obviously, integrating notions of typicality and user expectations into query processing would heavily improve the result quality in information searches as well as the user‟s search experience. Currently, no information system is able to process such implicit knowledge. In this paper we discuss the ingredients needed for designing information systems shaped for personalized and intuitive querying. 2

2. Overview of Current Preference-Based Query Processing Current preference-based information systems rely on classical preference models, which expose strong mathematical properties. These properties are intended to capture rational behavior being expressed as internal coherence and logical consistency within a system of beliefs and preferences (e.g., requiring preference relations to be transitive and constant over time, contexts, and occasions); choices are consequential, and options are evaluated using prior assessments of beliefs and values (Mellers, Schwartz, & Cooke, 1998). There are two major approaches in today‟s systems: utility functions (Fishburn, 1968; Fishburn, 1970; Keeney & Raiffa, 1976) and relational preference structures (Öztürk, Tsoukiàs, & Vincke, 2005; Kießling, 2002; Chomicki, 2003). A utility function assigns a numerical score to each data item. The score of an item depends only on the item‟s properties and represents its overall desirability. Relational preference structures link pairs of items through the notions of „is preferred to‟ and „is equally preferable as,‟ thus leading to qualitative preference orderings. Note that every utility function implies some simple- structured preference ordering. The problem of actually building personalized preference- based information systems has been investigated in various settings such as querying databases, getting automated product recommendations, or receiving support in decision making. During the 1990s, the top-k retrieval parad