We propose an automatic method of extracting bibliographies for academic articles scanned with OCR markup. The method uses conditional random fields (CRF) for labeling serially OCR-ed text lines on an article's title page as appropriate names for bibliographic elements. Although we achieved excellent extraction accuracies for some Japanese academic journals, we needed a substantial amount of training data that had to be obtained through costly manual extraction of bibliographies from printed documents. Therefore, this paper reports an empirical evaluation of active sampling applied to the CRF-based extraction of bibliographies to reduce the amount of training data. We applied active sampling techniques to three academic journals published in Japan. The experiments revealed that the sampling strategy using the proposed criteria for selecting samples could reduce the amount of training data to less than half or even a third of those for two academic journals. This paper also reports the effect of pseudo-training data that were added to training.