Experiments on sentence segmentation in Old Swedish editions

We present experiments on automatic segmentation of electronic Old Swedish editions into sentence-like units. Our target material is characterized by a great variation in the type of boundaries that are marked orthographically, the extent of boundary marking, and the means of boundary marking. We begin with an exploration of boundary marking in a large, unannotated corpus of Old Swedish texts. Then we show that we are able to improve upon a simple but effective segmenting baseline, using a conditional random field model trained on a manually annotated corpus. A more valuable lesson the modelling teaches us, however, is that we need to address the boundary marking variation explicitly.