Size is not Everything Genre Balance in Bootstrapping a Swedish PoS Tagger

Part-of-speech tagging is a basic component of natural language processing, and as such, needs to be as accurate as possible, or any subsequent processing will suffer. For Swedish, most tagger models are trained on the Stockholm-Umeå Corpus (SUC Ejerhed et al., 2006). As SUC is a balanced corpus, SUC models are better representatives for general language than models trained on news texts only, which is a common scenario for other languages. On the other hand, the corpus is a bit too small for tagger training, considering the size of the tagset needed to express the most common morphosyntactic features of Swedish. This leads to poorer performance than what has been reported for, for example, an equally-sized English news corpus and a much smaller German news corpus, tagged with the statistical TnT tagger (Brants, 2000). Both the English and German models show an accuracy of 96.7%, while the same tagger trained on SUC only has an accuracy of 95.5%. As SUC obviously is too small to be used alone as training data for any higher-accuracy tagger, we have used it as a seed corpus to bootstrap a much larger, unannotated, corpus, that can be added as training data. The bootstrapped corpus could represent another modality, domain or genre, if we are looking for adaptation. This sort of bootstrapping process has proved to be a viable approach (cf. Forsbom, 2006; Merialdo, 1994; Nivre and Grönqvist, 2001; Sjöbergh, 2003). Here, we are interested in seeing the effect the genre balance of bootstrapped corpus has on the performance, when drilling down by SUC genres.