On Lexical Resources for Digitization of Historical Documents

Many European libraries are currently engaged in mass digitization projects that aim to make historical documents and corpora online available in the Internet. In this context, appropriate lexical resources play a double role. They are needed to improve OCR recognition of historical documents, which currently does not lead to satisfactory results. Second, even assuming a perfect OCR recognition, since historical language differs considerably from modem language, the matching process between queries submitted to search engines and variants of the search terms found in historical documents needs special support. While the usefulness of special dictionaries for both problems seems undisputed, concrete knowledge and experience are still missing. There are no hints about what optimal lexical resources for historical documents should look like. The real benefit reached by optimized lexical resources is unclear. Both questions are rather complex since answers depend on the point in history when documents were born. We present a series of experiments which illuminate these points. For our evaluations we collected a large corpus covering German historical documents from before 1500 to 1950 and constructed various types of dictionaries. We present the coverage reached with each dictionary for ten subperiods of time. Additional experiments illuminate the improvements for OCR accuracy and Information Retrieval that can be reached, again looking at distinct dictionaries and periods of time. For both OCR and IR, our lexical resources lead to substantial improvements.