The current image retrieval engine (MirrorSEEk) consists of two major parts addressing content and presentation personalization respectively: An information retrieval (IR) system implemented on a relational DBMS and an user interface management system (UIMS). Content management is part of the Mirror project, presentation personalization is investigated in Dynamo. Figure 1 shows the overall picture. The system works as follows: The UIMS gets requests from browsers, analyses the input and creates a query in a simple XML language for the Mirror back-end. The Mirror system processes the query, executes it, transforms the results back to XML and sends the XML results back to the UIMS. The UIMS chooses among a set of stylesheets and transforms the XML results into a Web presentation language format that is then send to the browser.
|
| Figure 1: Overview of the architecture of the current image search engine |
The next section will give some background on the Mirror project. The rest of this short document will discuss the UIMS component and conclude with some remarks about interaction support in SMIL and interaction personalization.
The Mirror DBMS is a research database system that combines information (inexact matching) and data (exact matching) retrieval. The Mirror DBMS uses an extensible structural object-oriented logical data model and query algebra, the MOA object algebra that is mapped on a binary relational physical database (Monet). This separation of the logical object-oriented data model and the physical data model brings the notion of physical data independence to the world of object-oriented databases and provides an excellent basis for algebraic query optimization. The Mirror DBMS extends the MOA object algebra by defining new structures and operations on these structures for information retrieval that can be used in combination with the existing basic structures to create powerful queries that can manage both the logical and layout structure of multimedia documents. This way the Mirror DBMS separates the multimedia retrieval functionality from the actual multimedia retrieval application.
Since digital libraries typically involve several players with conflicting wishes and needs (content providers, users, access providers) and many simultaneous users, a single database would soon become a bottleneck in a large digital library setting and difficult to maintain. An open, distributed environment avoids many of these problems. New components can easily be added or removed without taking the entire system down and could also balance workload much more efficiently. The Mirror DBMS is therefore part of an open distributed architecture consisting of several components such as daemons which can perform all sorts of complex tasks, end-user devices, media-content servers (typically webservers) and one or more meta-data databases or search engines (the actual Mirror DBMS), all connected via a common software bus (CORBA). The right hand side of figure 1 gives an overview of the Mirror part: The MirrorWrapper, MOA and Monet blocks constitute a Mirror database. The Mirror server keeps track of one or more Mirror databases and makes sure that the request from the UIMS is send to the right database(s). The indexer and crawler maintain Mirror databases and are implemented as daemons.
At the moment the Mirror DBMS supports only text and image retrieval. Possible extensions for the future are feature extraction algorithms for video and sound (retrieval), support for feature grammars (indexing), semi-structured querying (MOA extensions) and use of the semantic Web for improving crawling, indexing and retrieval.
The UIMS runs as a Java Servlet on a Apache Web server and contains separate query formulation, connection management and presentation generation modules. The query formulation unit transforms the request of the user (HTTP request by the browser) into a simple XML query language (specific for this image retrieval application). The presentation generation module applies an XSL stylesheet on the XML results from the Mirror server to transform the XML into a Web presentation language. At the moment the query formulation and presentation generation modules are application dependent and the resulting presentation is customized rather than personalized: The user (or the system on behalf of the user) can select from three different stylesheets (HTML, HTML+TIME or SMIL) in the user interface.
Interactive applications like the image retrieval application mentioned above and hypermedia portals in general need browsers and language profiles that have interaction support. Currently this support is lacking in all SMIL browsers, rendering them useless for these kinds of applications. Microsoft's HTML+TIME solution is an alternative although less elegant than SMIL and not designed for multimedia Web presentations in particular.
Query formulation effects presentation generation but the reverse is also true if the interaction style itself can also be personalized which may be needed for different application domains or interaction modes: The user input and query formulation units then need to know if the interaction elements in the newly generated presentation are different from the ones in the previous presentation otherwise the explicit feedback might be wrongly interpreted. This increases the complexity of presentation personalization further.