POS-tagging of Historical Dutch

POS-tagging of Historical Dutch Dieuwke Hupkes, Rens Bod Institute for Logic, Language and Computation University of Amsterdam Science Park 107 D.hupkes@uva.nl, rens.bod@uva.nl Abstract We present a study of the adequacy of current methods that are used for POS-tagging historical Dutch texts, as well as an exploration of the influence of employing different techniques to improve upon the current practice. The main focus of this paper is on (unsupervised) methods that are easily adaptable for different domains without requiring extensive manual input. It was found that modernising the spelling of corpora prior to tagging them with a tagger trained on contemporary Dutch results in a large increase in accuracy, but that spelling normalisation alone is not sufficient to obtain state-of-the-art results. The best results were achieved by training a POS-tagger on a corpus automatically annotated by projecting (automatically assigned) POS-tags via word alignments from a contemporary corpus. This result is promising, as it was reached without including any domain knowledge or context dependencies. We argue that the insights of this study combined with semi-supervised learning techniques for domain adaptation can be used to develop a general-purpose diachronic tagger for Dutch. Keywords: POS-tagging, historical Dutch, parallel corpora 1. Introduction To extract information from a (historical) text, it is often helpful to know the grammatical categories (or part-of- speech tags) of the words in this text. High-performance automatic part-of-speech taggers (POS-taggers) can be trained when large amounts of annotated training data are available, but automatically POS-tagging low-resource lan- guages for which such data do not exist has proved to be a challenging task. When aiming to automatically tag his- torical data, one is confronted with an additional difficulty: standardisation of orthography is relatively recent and thus historical corpora often contain a large variation in spelling, which effectively increases the amount of annotated train- ing data necessary to learn a good model. One approach to address this orthographical variation is to use a respelling tool to normalise/modernise the spelling of a text, prior to tagging it with a tagger trained on modern data of the same language. Rayson et al. (2007) found that for Middle English normalising the spelling of a text in- creases the accuracy of a rule based modern English tagger from just under 0.82 to 0.85. For manual modernisation, they report an accuracy of 0.89, indicating that to obtain state-of-the-art tagging results for Middle English also lex- ical and/or syntactical variation should be considered. Another approach to POS-tagging historical text is to trans- fer annotation via parallel corpora. Positive results for this technique have been reported for obtaining annota- tions for closely related languages (e.g., Bentivogli et al. (2004; Van Huyssteen and Pilon (2009; Yarowsky et al. (2001)). Moon and Baldridge (2007) report good results for this method for tagging historical English. However, the applicability of this approach is quite limited, as it requires the availability of a parallel corpus with a similar language for which a good POS-tagger is available. In this study we focus on POS-tagging 17th-century Dutch texts. …

Content truncated to 500 words