Authors
D. Hupkes
L.W.M. Bod
Date (dd-mm-yyyy)
2015-06
Title
Using Parallel Data to Improve Part-of-speech Tagging of 17th Century Dutch
Publication Year
2015-06
Document type
Abstract
Abstract
If one wants to extract information from a (historical) text, it is often useful to know the grammatical categories (or part-of-speech tags) of all the words in this text. Tools to automatically assign high accuracy POS-tags are freely available, but require large amounts of annotated training data. Automatically POS-tagging languages for which little annotated data is available has proved to be a challenging task, and when developing a POS-tagger for historical Dutch one is confronted with an additional difficulty: a large variation in spelling. Currently, digital humanities researchers working with historical Dutch texts often resort to taggers trained on contemporary Dutch (e.g., Van den Bosch et al., 2007), but the accuracy of the resulting tags is generally low.

In this paper, we explore methods for generating higher accuracy tags for historical
corpora. In particular, we investigate how information extracted from a diachronic
parallel corpus consisting of Dutch Bible texts from 1637 and 1977 can be used to
improve POS-tagging of 17th century Dutch texts from several domains. We explore
the possibility of using this corpus to rewrite/translate words prior to tagging, as well
as the opon of creating an annotated 17th century corpus by projecting tags from the
contemporary version of the corpus to its historical counterpart and using this corpus
to train a new tagger.

We show that even without applying methods to account for context dependencies,
both methods result in great improvements over a baseline of tagging the texts with
a tagger trained on contemporary Dutch. Furthermore, they significantly outperform
using simple rewrite rules to normalise/modernise spelling before tagging. The improvement subsists across domains, but the within domain results are significantly better than the results for other domains, suggesting that incorporating knowledge about the domain of the text can lead to further improvement.

The results of this study can be of direct use to digital humanities researchers
working with historical Dutch texts of the last 3 centuries, as the tags assigned by our
retrained tagger are of much higher quality than the current standard. Furthermore,
it shows that using parallel data to exploit the similarities between contemporary and
historical texts is a very promising path to developing diachronic taggers. Similar techniques could also be applied to other languages that have diachronic parallel corpora available, as well as to improve results on lemmatisation of historical Dutch texts. In future work, we will focus on domain adaptation techniques for a better consistency across domains.
Permalink
https://hdl.handle.net/11245/1.505708