Do POS Tags Help to Learn Better Morphological Segmentations?

Do POS Tags Help to Learn Better Morphological Segmentations? Kairit Sirts Department of Computing Macquarie University Australia kairit.sirts@gmail.com Mark Johnson Department of Computing Macquarie University Australia mark.johnson@mq.edu.au Abstract The utility of using morphological features in part-of-speech (POS) tagging is well es- tablished in the literature. However, the usefulness of exploiting information about POS tags for morphological segmentation is less clear. In this paper we study the POS-dependent morphological segmenta- tion in the Adaptor Grammars framework. We experiment with three different scenar- ios: without POS tags, with gold-standard tags and with automatically induced tags, and show that the segmentation F1-score improves when the tags are used. We show that the gold-standard tags lead to the biggest improvement as expected. How- ever, using automatically induced tags also brings some improvement over the tag- independent baseline. 1 Introduction Linguistially, part-of-speech (POS) tagging and morphology are closely related and this relation has been heavily exploited in both supervised and unsupervised POS tagging. For instance, the su- pervised Stanford tagger (Toutanova et al., 2003) as well as some unsupervised POS taggers (Berg- Kirkpatrick et al., 2010; Lee et al., 2010) use char- acter prefix and/or suffix features, while the model by Christodoulopoulos et al. (2011) makes use of suffixes learned with an unsupervised morphologi- cal segmentation model. There have been some attempts to exploit the relation in the opposite direction to learn the seg- mentations dependent on POS tags. For instance, the segmentation procedures described by Freitag (2005) and Can and Manandhar (2009) find the syntactic clusters of words and then perform mor- phology learning using those clusters. Few works have included a small number of syntactic classes directly into the segmentation model (Goldwater et al., 2006; Lee et al., 2011). However, Goldwa- ter et al. (2006) only trains the model on verbs, which means that the classes model different verb paradigms rather than POS tags. Secondly, the model is never evaluated in a single class configura- tion and thus it is not known whether incorporating those classes gives any actual improvement. The results of Lee et al. (2011) show small improve- ments when the POS-word component (a bigram HMM) is incorporated into the model. However, the number of syntactic categories they learn is only 5, which is smaller than the number of main POS categories in most annotated corpora. More- over, the main gain in the segmentation F-score is obtained by modeling the agreements between adjacent words, rather than exploiting the relation to syntactic classes. Another line of previous work has attempted to model the POS tags and morphological segmenta- tions jointly in an unsupervised model (Can, 2011; Sirts and Alum¨ae, 2012; Frank et al., 2013). How- ever, the results presented in those papers fail to demonstrate clearly the utility of using the tag infor- mation in segmentation learning over the scenario where the tags are missing. The goal of this paper is to explore the relation between POS tags and morphological segmenta- tions and in particular, to study if and how much the …

Content truncated to 500 words