Rule-based normalisation of historical text - A diachronic study

Language technology tools can be very useful for making information concealed in historical documents more easily accessible to historians, linguists and other researchers in humanities. For many languages, there is however a lack of linguistically annotated historical data that could be used for training NLP tools adapted to historical text. One way of avoiding the data sparseness problem in this context is to normalise the input text to a more modern spelling, before applying NLP tools trained on contemporary corpora. In this paper, we explore the impact of a set of hand-crafted normalisation rules on Swedish texts ranging from 1527 to 1812. Normalisation accuracy as well as tagging and parsing performance are evaluated. We show that, even though the rules were generated on the basis of one 17th century text sample, the rules are applicable to all texts, regardless of time period and text genre. This clearly indicates that spelling correction is a useful strategy for applying contemporary NLP tools to historical text.