Adaptor Grammars for Learning Non-Concatenative Morphology

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 345–356, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Adaptor Grammars for Learning Non-Concatenative Morphology Jan A. Botha and Phil Blunsom Department of Computer Science University of Oxford Oxford, OX1 3QD, UK {jan.botha,phil.blunsom}@cs.ox.ac.uk Abstract This paper contributes an approach for expressing non-concatenative morphological phenomena, such as stem derivation Semitic languages, in terms of a mildlyin context-sensitive grammar formalism. This offers a convenient level of modelling ab- straction while remaining computationally tractable. The nonparametric Bayesian frame- work of adaptor grammars is extended to this richer grammar formalism to propose a prob- abilistic model that can learn word segmenta- tion and morpheme lexicons, including ones with discontiguous strings as elements, from unannotated data. Our experiments on He- brew and three variants of Arabic data find that the additional expressiveness to capture roots and templates as atomic units improves the quality of concatenative segmentation and stem identification. We obtain 74% accuracy in identifying triliteral Hebrew roots, while performing morphological segmentation with an F1-score of 78.1. 1 Introduction Unsupervised learning of morphology is the task of acquiring, from unannotated data, the intra-word building blocks of a language and the rules by which they combine to form words. This task is of interest both as a gateway for studying language acquisition in humans and as a way of producing morphological analyses that are of practical use in a variety of nat- ural language processing tasks, including machine translation, parsing and information retrieval. A particularly interesting version of the morphol- ogy learning problem comes from languages that use templatic morphology, such as Arabic, He- brew and Amharic. These Semitic languages de- rive verb and noun stems by interspersing abstract root morphemes into templatic structures in a non- concatenative way. For example, the Arabic root k·t·b can combine with the template (i-a) to derive the noun stem kitab (book). Established morpho- logical analysers typically ignore this process and simply view the derived stems as elementary units (Buckwalter, 2002), or their account of it coincides with a requirement for extensive linguistic knowl- edge and hand-crafting of rules (Finkel and Stump, 2002; Schneider, 2010; Altantawy et al., 2010). The former approach is bound to suffer from vocabu- lary coverage issues, while the latter clearly does not transfer easily across languages. The practical appeal of unsupervised learning of templatic mor- phology is that it can overcome these shortcomings. Unsupervised learning of concatenative morphol- ogy has received extensive attention, partly driven by the MorphoChallenge (Kurimo et al., 2010) in re- cent years, but that is not the case for root-templatic morphology (Hammarstr¨om and Borin, 2011). In this paper we present a model-based method that learns concatenative and root-templatic mor- phology in a unified framework. We build on two disparate strands of work from the literature: Firstly, we apply simple Range Concatenating Grammars (SRCGs) (Boullier, 2000) to parse contiguous and discontiguous morphemes from an input string. These grammars are mildly-context sensitive (Joshi, 1985), a superset of …

Content truncated to 500 words