Arbeitspapiere der GMD 699

MERIT (“Multimedia Extensions of Retrieval Interaction Tools”) provides a user centered interface to a database covering research programs, projects, and consortia in the field of information technology. The system supports the user by offering a selection of sample retrieval strategies – called cases – which are modifiable to meet situative requirements. Then the system is able to offer a situation-dependent query form. The process of selecting search terms is assisted by a semantic component proposing additional search terms that are derived from an initially given one by associative reasoning. The system generates graphical presentations of retrieval results which support – in accordance to the current case – a survey or detail oriented reception of the retrieved data, thus providing visual cues for the user’s relevance assessment. Additionally, the system employs interactive maps to display geographical data, and can provide scanned-in documents. 4 Towards a User-Centered Interface for Information Retrieval Thiel et al. 1 The Concept of User Support in MERIT The development of object-oriented databases with semantic schemata requires new approaches to the design of user interfaces for these systems. Due to the intensive interrelation between the information objects, the stored structures often show such a high degree of complexity that traditional query languages cannot provide an adequate way of handling the structures. As an alternative, we propose a user interface which supports the exploration of a relevant part of the database. MERIT (“Multimedia Extensions of Retrieval Interaction Tools”) offers access to a subset of the CORDIS databases covering research programs, projects and consortia in the field of information technology.1 The system features graphical presentations of retrieval results, employs interactive maps to display geographical data, and can provide scanned-in documents, e.g. photos. A guiding principle of MERIT is that the system should be adapted to the needs of the human user, and not the converse (Norman and Draper 1986). This means, that the database user, with his strong and even more with his weak points concerning competence and knowledge, has to be the central parameter in the interface design process. An ideal system would be one which presented the user a familiar environment, one in which he is –––––––––– 1. The CORDIS databases are offered online by ECHO, which is the official host organization of the CEC – Commission of the European Community. For our experiments, we use a downloaded subset of the original databases covering about 900 projects. These are funded by several different research programs. About 6000 organizations are involved as partners in project consortia. required to learn little about the formal details of the system in order to work effectively with it. For instance, human beings possess a complex system of interactive behavior which is acquired during the processes of language acquisition and which is continually being extended by new experiences. A system which responds to the user in even a small way as a human interlocutor can already employ procedures with which even naive users are implicitly familiar. Another successful design approach relies on the user’s ability to handle objects located in space: Direct manipulation interfaces are based on the illusion that the (virtual) objects which are presented on the screen are located in a virtual space and can be accessed by deictic gestures. While pointing to an object complex operations can be invoked by function keys, pop-up menus etc. Thus graphical interaction can be modeled akin to deicticvisual interaction between humans where gestures etc. support and replace verbal utterances. This user-centered approach enhances the design of flexible dialogues in which the system acts as an information broker who tries to assist the user in the following tasks: Finding a successful retrieval strategy: Taking into account that information seeking interactions are not determined by a predefined, standardized task structure, but by individual strategies of users to which the system should adapt, we use a “case-based” approach to user guidance. The general idea of case-based reasoning is to “solve new problems by adapting solutions that were used to solve old problems.” (Riesbeck and Schank 1989, p. 25). In accordance with this approach we Thiel et al. Towards a User-Centered Interface for Information Retrieval 5 describe the thematic level of informationseeking dialogues by a set of alternative dialogue plans that get modified during the dialogue depending on specific dialogue knowledge and the preceding discourse. Clarification of information needs: A conclusion of the user-centered approach is to require that interfaces to complex information systems adapt themselves as much as possible to the user’s view of the domain. Thus, in a databasequery system, the user’s view should be accommodated. This is counter to standard database applications, which typically require the user to learn the system’s way of organizing information before this information can be effectively accessed. MERIT assists the user in accessing a database without knowing the particular terminology or conventions used during the compilation of the database. Relevance assessment of retrieved items: In complex information systems, which are used for document retrieval or knowledge base access, the problem of relevance assessment arises: the user has to decide whether a retrieval result fits his information need. In this situation the system’s response should display the items found in the database in a way that supports this decision process. The presentation of information to a human user should take account of the fact that human perceptual processes operate in a highly parallel fashion. Presentation of information in a form which is serially consumed by the user (such as lengthy texts) are combined with visualizations that provide survey information at a glance. Because graphics and pictures are perceived in a holistic way, they complement the textual information. The user-centered approach of MERIT employs presentation forms which exhibit the cues that are required for a convenient relevance assessment. As a result, the system responds to the user’s needs addressing strategic, terminological and cognitive aspects. In the remainder of this paper we will outline the features of MERIT that contribute to this quality. We start, just as in a MERIT dialogue session, with the user’s determination of the information need that is to be satisfied by the system. We outline the system’s components, which provide strategical and terminological assistance to the user. Next, the cognitive adequate presentation forms for retrieved data are discussed with respect to their effects on the user’s relevance assessment. The dialogue guidance used in MERIT applies a case-based strategic dialogue planner. The dialogue plans allow the system to pursue a certain communicative goal. They are, however, flexible enough to cope with common dialogue problems. The next section of the paper concentrates on more internal aspects of MERIT, shortly sketching the architectural features while observing the data flow resulting from the dialogue. The paper concludes with an account of the achieved results and the future work. 2 User Support for Database Access At the beginning of a MERIT session the user has the opportunity to determine the global strategy of the dialogue by choosing a specific case. After having selected a case, he is given 6 Towards a User-Centered Interface for Information Retrieval Thiel et al. a query form in order to formulate the query in an easy way. For some special query attributes the system offers a terminology support which is provided by a component, called ‘Knowledge Explorer’. 2.1 Determination of the Global Strategy Formulating a complex query is often an exploratory and incremental process. In particular, in the beginning of a retrieval dialogue the user needs help or suggestions by the system. As well, some open problems effect even the whole retrieval process, e.g. ‘What are the query attributes to restrict the data to a relevant subset?’ or ‘How to restrict the attributes using which query operators or search terms?’ or ‘What is the best presentation form to present the requested data?’ or ‘Which attributes should be presented? Present the complete information at once in a single presentation, or step by step?’ and so on. Based on the assumption that formulating an information need and presenting the data is an incremental process (cf. Belkin and Marchetti 1990), a clear goal description is missing. Often there are conflicting constraints which have to be resolved. Therefore, starting with a general top-down problem solving strategy seems to be an inadequate approach. We propose a bottom-up procedure: starting with a selected set of examples, so-called ‘cases’ (cf. Riesbeck and Schank 1989, Tißen 1991), combined with dynamic and interactive modifications. Cases are derived from previous retrieval sessions, which were successful for a specific task in the application domain. A case is represented by a sequence of dialogue steps in a dialogue plan. On the strategic level of dialogue modeling, which is covered by the case-based approach, we distinguish between query and presentation steps. Each dialogue step, either for querying or for presenting retrieved data, is defined by a list of features, e.g. specifying the input and output of a dialogue step. The most important feature – called ‘perspective’ – describes subsets of the database relevant in a specific context, for the query component as well as for the presentation component. A hierarchical representation of perspectives together with a set of functions enables the system to adapt a query or presentation step directly to a new situation. For instance, the perspective concerning project information, focusing on project partners, is a sub-perspective of organizational aspects of project information. In this perspective only a small subset of database attributes are relevant for the query form. Because dialogue plans and perspectives are represented explicitly in knowledge bases, they can easily be modified during the ongoing dialogue. Modifications are either automatically generated by the system or interactively in a subdialogue with the user. This explicit dialogue model is the basis for a flexible dialogue control in an exploratory retrieval process. Querying the database to satisfy a new information need is based on experience from old solutions. In addition to an explicit dialogue model we employ a subsystem, consisting of a ‘retriever’ to find a good case, a ‘modifier’ for adaptation of old solutions to new problems, and a ‘storer’ to increase the case library by new, successful cases. This approach of a ‘case-based dialogue manager’ (the related implementation Thiel et al. Towards a User-Centered Interface for Information Retrieval 7 is called ‘CADI’) is described in detail in (Tißen 1991). 2.2 Query Formulation After the case selection, the user is presented a query form (cf. McAlpine and Ingwersen 1989) listing attributes of the selected perspective. In the next step the user has the opportunity for posing a query and by this way reduce the amount of relevant instances of the current concept class. This is achieved by stating attribute restrictions. Each line of the query form sheet represents a restriction consisting of a comparison operator selected from a menu that provides an attribute-specific choice, and a constant to which attribute values of instances in the knowledge base are to be compared. If all restrictions are fulfilled for a given instance, then this instance will belong to the set of responses. Figure 1 shows the query form for the thematic perspective of projects, i.e. it does not include organizational attributes like project partners, contact persons, location, dates, financial data, etc. By the filled-in query all projects will be retrieved which contain the terms “author” or “document” or Figure 1: A sample query form 8 Towards a User-Centered Interface for Information Retrieval Thiel et al. “text” in their general project information and at the same time the terms “natural language” or “multi-media” or “hypermedia” in their project objectives. The main features of this form-based query interface are: The user is freed from learning the syntax of a formal query language. He does not need to be familiar with the structure of the database, because the perspective provides him with a list of the relevant attributes. Although the query interface generates quite complex SQL-expressions internally, the underlying relational database system is hidden from the user. Quite on the contrary, the user is encouraged to take an object-oriented view of the data. This is further intensified by the presentation forms discussed in section 3. Query forms may contain multiple free text fields, for example a description of projects. The user may require that a text field of a retrieved instance must or must not contain a given search term. If search terms are given in different lines of a multiple line field they are implicitly connected by the boolean or. 2.3 Support of Free Text Search The last feature of the list above, the querying of free text, may cause a problem which is dealt with in the rest of this section. Originally, the user wants to do a content based search, but instead he has to deal with terms which just are the surface representation of the content. But every content information can be expressed by various concepts, all of which the user has to include in his search to avoid missing instances. Speaking technically, in order to improve the recall of his query the user generally has to add several synonyms or similar concepts to each search concept. Instead of leaving the task of finding these additional search concepts to the user, we employ a module called Knowledge Explorer (KX) which has conceptual knowledge of the database domain and suggest such supplementary concepts. This conceptual knowledge is stored as a fuzzy association network (cf. Kracker 1992). The meaning of a concept is solely determined by its relationships to other concepts. There are four types of such relationships: A positive association connects concepts which are semantically similar or often used in the same context, a negative association is used to express some kind of opposition, a generalization links one concept to another which is more general in a semantic or partitive sense, and the specialization is the inverse of the generalization relationship. The feature which distinguishes our concept network from a conventional semantic net is the fact that each relationship has a value out of [0, 1] assigned which is interpreted as the strength of the relationship. A positive association with a strength close to 1, for example, identifies a relation between two very similar concepts. Because the relationships between concepts are not crisp, but the relationship strength may be distinct differently, we call it a fuzzy network. We assume the positive association, the generalization and the specialization to be transitive relationships. Thiel et al. Towards a User-Centered Interface for Information Retrieval 9 We can use the inherent transitivity of these relationships to compute relationships between two concepts even if they are not directly connected. The operator to combine transitive relationships which is most commonly used are the max-min-operator or max-product-operator. Max means that only the path with the maximal strength of all paths between two concepts is regarded when determining the derived relationship. The minimum and the arithmetic product are used to aggregate the strength values of the relationships which have to be combined. There are many more (in fact, an unlimited number of) such T-norm functions [0, 1] × [0, 1] → [0, 1] suitable for combining transitive relationships. They differ in the strength of the resulting relationship. Opposed to (Bezdek, Biwas and Hunag 1986) where a single T-norm function is used for all types of relationships, we will employ several T-norm functions, for example the following three: T1 (a, b) = max (0, a + b – 1) T2 (a, b) = ab T3 (a, b) = min (a, b). It is easy to show that T1(a, b) ≤ T2(a, b) ≤ T3(a, b) holds. T1 is the most conservative function returning the smallest values. T3, on the contrary, is the makes the most optimistic conclusions and infers the strongest relationships. T2 assumes the arguments to be independent and treats them like independent probabilities. Depending on the types of the relationships to be combined the most suitable T-norm can be used (see table in figure 2). P: positive association N: negative association G: generalization S: specialization association categorization