An analysis and comparison of keyword recommendation methods for scientific data

To classify and search various kinds of scientific data, it is useful to annotate those data with keywords from a controlled vocabulary. Data providers, such as researchers, annotate their own data with keywords from the provided vocabulary. However, for the selection of suitable keywords, extensive knowledge of both the research domain and the controlled vocabulary is required. Therefore, the annotation of scientific data with keywords from a controlled vocabulary is a time-consuming task for data providers. In this paper, we discuss methods for recommending relevant keywords from a controlled vocabulary for the annotation of scientific data through their metadata. Many previous studies have proposed approaches based on keywords in similar existing metadata; we call this the indirect method. However, when the quality of the existing metadata set is insufficient, the indirect method tends to be ineffective. Because the controlled vocabularies for scientific data usually provide definition sentences for each keyword, it is also possible to recommend keywords based on the target metadata and the keyword definitions; we call this the direct method. The direct method does not utilize the existing metadata set and therefore is independent of its quality. Also, for the evaluation of keyword recommendation methods, we propose evaluation metrics based on a hierarchical vocabulary structure, which is a distinctive feature of most controlled vocabularies. Using our proposed evaluation metrics, we can evaluate keyword recommendation methods with an emphasis on keywords that are more difficult for data providers to select. In experiments using real earth science datasets, we compare the direct and indirect methods to verify their effectiveness, and observe how the indirect method depends on the quality of the existing metadata set. The results show the importance of metadata quality in recommending keywords.