Morfessor FlatCat: An HMM-Based Method for Unsupervised and Semi-Supervised Learning of Morphology

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1177–1185, Dublin, Ireland, August 23-29 2014. Morfessor FlatCat: An HMM-Based Method for Unsupervised and Semi-Supervised Learning of Morphology Stig-Arne Gr¨onroos1 stig-arne.gronroos@aalto.fi Sami Virpioja2 sami.virpioja@aalto.fi Peter Smit1 peter.smit@aalto.fi Mikko Kurimo1 mikko.kurimo@aalto.fi 1Department of Signal Processing and Acoustics, Aalto University 2Department of Information and Computer Science, Aalto University Abstract Morfessor is a family of methods for learning morphological segmentations of words based on unannotated data. We introduce a new variant of Morfessor, FlatCat, that applies a hid- den Markov model structure. It builds on previous work on Morfessor, sharing model compo- nents with the popular Morfessor Baseline and Categories-MAP variants. Our experiments show that while unsupervised FlatCat does not reach the accuracy of Categories-MAP, with semi- supervised learning it provides state-of-the-art results in the Morpho Challenge 2010 tasks for English, Finnish, and Turkish. 1 Introduction Morphological analysis is essential for automatic processing of compounding and highly-inflecting lan- guages, for which the number of unique word forms may be very large. Apart from rule-based analyzers, the task has been approached by machine learning methodology. Especially unsupervised methods that require no linguistic resources have been studied widely (Hammarstr¨om and Borin, 2011). Typically these methods focus on morphological segmentation, i.e., finding morphs, the surface forms of the mor- phemes. For language processing applications, unsupervised learning of morphology can provide decent- quality analyses without resources produced by human experts. However, while morphological ana- lyzers and large annotated corpora may be expensive to obtain, a small amount of linguistic expertise is more easily available. A well-informed native speaker of a language can often identify the different pre- fixes, stems, and suffixes of words. Then the question is how many annotated words makes a difference. One answer was provided by Kohonen et al. (2010), who showed that already one hundred manually segmented words provide significant improvements to the quality of the output when comparing to a linguistic gold standard. The semi-supervised approach by Kohonen et al. (2010) was based on Morfessor Baseline, the sim- plest of the Morfessor methods by Creutz and Lagus (2002; 2007). The statistical model of Morfessor Baseline is simply a categorical distribution of morphs—a unigram model in the terms of statistical lan- guage modeling. As the semi-supervised Morfessor Baseline outperformed all unsupervised and semi- supervised methods evaluated in the Morpho Challenge competitions (Kurimo et al., 2010a) so far, the next question is how the approach works for more complex models. Another popular variant of Morfessor, Categories-MAP (CatMAP) (Creutz and Lagus, 2005), models word formation using a hidden Markov model (HMM). The context-sensitivity of the model improves the precision of the segmentation. For example, it can prevent splitting a single s, a common English suffix, from the beginning of a word. Moreover, it can disambiguate between identical morphs that are actually surface forms of different morphemes. Finally, separation of stems and affixes in the output makes it simple to use the method as a stemmer. In contrast to Morfessor Baseline, the …

Content truncated to 500 words