DATING HISTORICAL DOCUMENTS USING DEEP LEARNING

The date of a historical document, if it was published, can be recognized through the date of its preface or copyright page. However, dating an unpublished manuscript is much more difficult. In the case of old Hangeul documents, various linguistic features can be used to estimate the approximate date of the document, but as the number of features increases, the task tends to go beyond the purview of an individual human researcher, and become more appropriate to AI. This paper shows how artificial neural networks can be trained to estimate the date of documents using material whose date is known. For this purpose, various kinds of neural networks are examined: Bag-of-words model, CNN, RNN and Transformer. In addition, these models can be further sub-divided: unigram or bigram, character(syllable)-based or grapheme(phoneme)-based. After trained on documents with known dates, these models are applied to new (unseen) data, and the results are evaluated.