Enhanced Word Decomposition by Calibrating the Decision Threshold of Probabilistic Models and Using a Model Ensemble

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 375–383, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Enhanced word decomposition by calibrating the decision threshold of probabilistic models and using a model ensemble Sebastian Spiegler Intelligent Systems Laboratory, University of Bristol, U.K. spiegler@cs.bris.ac.uk Peter A. Flach Intelligent Systems Laboratory, University of Bristol, U.K. peter.flach@bristol.ac.uk Abstract This paper demonstrates that the use of ensemble methods and carefully calibrat- ing the decision threshold can signifi- cantly improve the performance of ma- chine learning methods for morphologi- cal word decomposition. We employ two algorithms which come from a family of generative probabilistic models. The mod- els consider segment boundaries as hidden variables and include probabilities for let- ter transitions within segments. The ad- vantage of this model family is that it can learn from small datasets and easily gen- eralises to larger datasets. The first algo- rithm PROMODES, which participated in the Morpho Challenge 2009 (an interna- tional competition for unsupervised mor- phological analysis) employs a lower or- der model whereas the second algorithm PROMODES-H is a novel development of the first using a higher order model. We present the mathematical description for both algorithms, conduct experiments on the morphologically rich language Zulu and compare characteristics of both algo- rithms based on the experimental results. 1 Introduction Words are often considered as the smallest unit of a language when examining the grammatical structure or the meaning of sentences, referred to as syntax and semantics, however, words them- selves possess an internal structure denominated by the term word morphology. It is worthwhile studying this internal structure since a language description using its morphological formation is more compact and complete than listing all pos- sible words. ical analysis.This study is called morpholog- According to Goldsmith (2009) four tasks are assigned to morphological analy- sis: word decomposition into morphemes, build- ing morpheme dictionaries, defining morphosyn- tactical rules which state how morphemes can be combined to valid words and defining mor- phophonological rules that specify phonological changes morphemes undergo when they are com- bined to words. Results of morphological analy- sis are applied in speech synthesis (Sproat, 1996) and recognition (Hirsimaki et al., 2006), machine translation (Amtrup, 2003) and information re- trieval (Kettunen, 2009). 1.1 Background In the past years, there has been a lot of inter- est and activity in the development of algorithms for morphological analysis. All these approaches have in common that they build a morphologi- cal model which is then applied to analyse words. Models are constructed using rule-based meth- ods (Mooney and Califf, 1996; Muggleton and Bain, 1999), connectionist methods (Rumelhart and McClelland, 1986; Gasser, 1994) or statisti- cal or probabilistic methods (Harris, 1955; Hafer and Weiss, 1974). Another way of classifying ap- proaches is based on the learning aspect during the construction of the morphological model. If the data for training the model has the same struc- ture as the desired output of the morphological analysis, in other words, if a morphological model is …

Content truncated to 500 words