Classification of MEDLINE Abstracts

This paper provides the preliminary result in our experiments to automatically assign MeSH terms to MEDLINE abstracts. Every year about 100,000 documents are added to MEDLINE, index terms are assigned by hand to each document from a controlled vocabulary called MeSH. This is necessarily time consuming and may lead to inconsistent indexing due to the large size of MeSH. Our purpose is to explore the feasibility of automating this indexing. To achieve the purpose, we apply two documents classification methods, based on SVMV [1] and AdaBoost [4], which show good results in classification of news corpora and analyze their results. We assumed a class consists of the abstracts which have the same MeSH term. Although MeSH terms have a hierarchical structure, each class is regarded to be independent. We used MeSH terms previously assigned by specialists as answer and compared the answer with the assigned MeSH term by application of SMVM and AdaBoost.