Augmenting Historical Manuscripts with Automatic Hyperlinks

Hyperlinks are so useful for searching and browsing modern digital collections that researchers have longer wondered if it is possible to retroactively add hyperlinks to digitized historical documents. There has already been significant research into this endeavor for historical text; however, in this work we consider the problem of adding hyperlinks among graphic elements. While such a system would not have the ubiquitous utility of text-based hyperlinks, as we will show, there are several domains where it can significantly augment textual information. While OCR of historical text is known to be a difficult problem, the actual words themselves are inherently discrete. Thus, two words are either identical or not. This means that off-the-shelf machine learning algorithms, including semi-supervised learning, can be easily used. However, as we shall demonstrate, semi-supervised learning does not work well with images, because we cannot expect binary matching decisions. Rather we must deal with degrees of matching. In this work we make the novel observation that this ¿degree of matching¿ biased algorithms make overly confident predictions about simple shapes. We show that a simple technique for correcting this bias, and demonstrate through extensive experiments that our method significantly improves accuracy on diverse historical image collections.