Classification and automatic indexing in a persistent object environment

Principal investigators:
Robert Godin and Brigitte Kerhervé
Université du Québec à Montréal

Co-investigator:
James M. Turner, professeur adjoint
École de bibliothéconomie et des sciences de l'information
Université de Montréal

• In this paper, "digital libraries" refers to distributed multimedia information systems;
• Typically such systems must manage very large amounts of data;
• File sizes for digital images, audio, and video are large, and the files have no alphabet to make text searching possible;
• Until techniques for edge-detection, object-recognition, matching strategies are perfected, all retrieval is dependent on the metadata;
• Metadata = an old computer science term for what information professionals have usually called a representation or document surrogate;
• For purposes of storage and retrieval, words are always associated with still and moving pictures and with sound materials;
• The index is used in direct querying as a retrieval mechanism, but both browsing and classification are important in providing the user with adequate search strategies and tools.

• Work on using automatic conceptual clustering of documents for retrieval purposes is an approach to this problem (Godin et al. 1993);
• The conceptual hierarchy generated is used as a navigational space where browsing and direct querying can be integrated;
• Browsing in data spaces has long been recognized as an attractive strategy, particularly for casual users and exploration of new domains (Marchionini and Shneiderman 1988);
• Browsing is also widely recognized as an important strategy in retrieving pictures (e.g. O'Connor 1986, Jussim 1977, Schuller 1993);
• Classification of the metadata for images is the focus of the project.

• What is needed is a general classification for images, but none exists (however, Iconclass, for art pictures, is fairly widely used for art);
• Attempts to adapt book classifications (e.g. UDC at the BBC) have ended in failure;
• Because of important technical obstacles, empirical study of storage and retrieval issues for moving images is still difficult, so the research focusses on stills;
• Essentially, what is needed is effective ways for relating words to nontextual information objects;
• One approach is to use automatic indexing techniques on the text, to generate a browsable semantic net;
• Perhaps more effective is to link textual representations of indexing concepts to pictorial representations in a classification showing whole-part relationships;
• This permits searchers to use words and pictures together to retrieve visual information objects;
• A visual dictionary may work for this purpose.

• To test the effectiveness of the dictionary as a front end to an information system;
• To test the effectiveness of the whole-part text component as a controlled vocabulary for the system;
• To determine whether the classificatory structure used in the dictionary can be applied to picture databases with either general or specialised content.

Main menu | EBSI | Comments

Other presentation