Search results

The purpose of this paper is to describe a new approach to a well-known problem for digital libraries, how to search across multiple unrelated libraries with a single query.

The approach involves creating new Dewey Decimal Classification terms and numbers from existing Dublin Core records. In total, 263,550 records were harvested from three digital libraries. Weighted key terms were extracted from the title, description and subject fields of each record. Ranked DDC classes were automatically generated from these key terms by considering DDC hierarchies via a series of filtering and aggregation stages. A mean reciprocal ranking evaluation compared a sample of 49 generated classes against DDC classes created by a trained librarian for the same records.

The best results combined weighted key terms from the title, description and subject fields. Performance declines with increased specificity of DDC level. The results compare favorably with similar studies.

The metadata harvest required manual intervention and the evaluation was resource intensive. Future research will look at evaluation methodologies that take account of issues of consistency and ecological validity.

The method does not require training data and is easily scalable. The pipeline can be customized for individual use cases, for example, recall or precision enhancing.

The approach can provide centralized access to information from multiple domains currently provided by individual digital libraries.

The approach addresses metadata normalization in the context of web resources. The automatic classification approach accounts for matches within hierarchies, aggregating lower level matches to broader parents and thus approximates the practices of a human cataloger.