Unsupervised Morphology Learning with Statistical Paradigms

Proceedings of the 27th International Conference on Computational Linguistics, pages 44–54 Santa Fe, New Mexico, USA, August 20-26, 2018. 44 Unsupervised Morphology Learning with Statistical Paradigms Hongzhi Xu1, Mitch Marcus1, Charles Yang2, Lyle Ungar1 1 Computer and Information Science, University of Pennsylvania 2 Linguistics Department, University of Pennsylvania Philadelphia, PA 19104, USA 1 {xh, mitch, ungar}@cis.upenn.edu 2 charles.yang@ling.upenn.edu Abstract This paper describes an unsupervised model for morphological segmentation that exploits the notion of paradigms, which are sets of morphological categories (e.g., suffixes) that can be applied to a homogeneous set of words (e.g., nouns or verbs). Our algorithm identifies statistically reliable paradigms from the morphological segmentation result of a probabilistic model, and chooses reliable suffixes from them. The new suffixes can be fed back iteratively to improve the accuracy of the probabilistic model. Finally, the unreliable paradigms are subjected to pruning to eliminate unreliable morphological relations between words. The paradigm-based algorithm significantly improves segmentation accuracy. Our method achieves state-of-the- art results on experiments using the Morpho-Challenge data, including English, Turkish, and Finnish. 1 1 Introduction Morphological learning aims to automatically uncover constitutive units of words. It is an especially important task for many NLP applications such as language generation, information retrieval etc. (Sproat, 1992). Morphology analyzing is non-trivial especially for morphologically rich languages such as Turkish where the word formation process is extremely productive and can create in principle tens of billions of word forms. The identification of morphological relations between words provides a basis for uncovering their syntactic and semantic relations, which in turn can be exploited by downstream NLP applications. Most unsupervised models of morphological segmentation (Virpioja et al., 2013; Goldwater and Johnson, 2004; Creutz and Lagus, 2005; Creutz and Lagus, 2007; Lignos, 2010; Poon et al., 2009; Snyder and Barzilay, 2008) treat words as concatenation of morphemes. In some models, the dependencies between morphemes (e.g., the English suffix -es often follows a verbal stem with y changed to i, as in carries) are recognized (Narasimhan et al., 2015), making use of transformations akin to rewrite rules (Goldwater and Johnson, 2004; Lignos et al., 2010). In all these approaches, the dependency between morphemes is generally local, and the overall distribution of the underlying paradigms implied by the segmentation result is not explored. In this paper, we propose to exploit the notion of the paradigm, a global property of morphological systems, for the task of unsupervised morphological segmentation (Parkes et al., 1998; Goldsmith, 2001; Chan, 2006). The idea of using paradigm to describe the morphological structure of a language can be traced back to a long time ago, and has been widely adopted in modern linguistic studies, starting from Ferdinand de Saussure. A paradigm refers to a set of morphological categories such as suffixes that can be can be applied to a homogeneous class of words. For instance, the paradigm (NULL, -er, -est, -ly) is defined over English adjectives (e.g., high, higher, highest, highly), the paradigm (NULL, -ing, -ed, -s, -er) is defined over English verbs (e.g, walk, walking, …

Content truncated to 500 words