National Diet Library Subject Headings (NDLSH) with sub-headings do not have in principle their corresponding classification numbers, unlike those without sub-headings. This study tries to identify the proper combination among the NDLSH with NDC number pairs extracted from bibliographic re-cords by utilizing machine learning methods.
　First, adequacy of an individual pairing of a subject heading with a NDC number in a bibliographic record was judged manually for over 10,000 pairs. They were then used as training and evaluation data in machine learning experiments. About 80 percent of the pairs were eventually judged as proper com-binations, which had either a) the same classification numbers as, or b) the classification numbers being forward-matched with, the corresponding numbers for the main headings included in subject headings.
　Then, machine learning experiments were conducted with the manually judged pairs of subject headings and NDC numbers, whose results were evaluated in a cross validation manner. Two ways of establishing training data sets, five inclusive attribute sets of individual pairs, and seven major machine learning methods, were adopted in the experiments. The results showed that the machine learning approach to the issue had a certain effectiveness, but was not highly effective.

View full abstract