Automatically Controlled-Vocabulary Indexing for Text Retrieval

The IR society has made efforts in free-term indexing for a long time. By contrast, few efforts are made in controlled-vocabulary indexing. A new model for controlled-vocabulary indexing is proposed in this paper. This proposed model, TF×OSDF×CSIDF, distinguishes subject-specific words from common words and domain-specific words in documents. 60,400 MEDLINE records are used as training data and testing data and 100 MeSH subject headings are used as the testing controlled vocabularies. The preliminary experiments show good results. The precision and the recall concurrently exceed 90% using abstracts as training materials. The precision reaches 90% and the recall still keeps at 70% using title only. The problem of indexer's consistency could be alleviated using the proposed model to automatically generate index terms.