Supervised Learning of Complete Morphological Paradigms

Proceedings of NAACL-HLT 2013, pages 1185–1195, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Supervised Learning of Complete Morphological Paradigms Greg Durrett∗ Computer Science Division University of California, Berkeley gdurrett@cs.berkeley.edu John DeNero Google, Inc. denero@google.com Abstract We describe a supervised approach to predict- ing the set of all inflected forms of a lexical item. Our system automatically acquires the orthographic transformation rules of morpho- logical paradigms from labeled examples, and then learns the contexts in which those trans- formations apply using a discriminative se- quence model. Because our approach is com- pletely data-driven and the model is trained on examples extracted from Wiktionary, our method can extend to new languages without change. Our end-to-end system is able to pre- dict complete paradigms with 86.1% accuracy and individual inflected forms with 94.9% ac- curacy, averaged across three languages and two parts of speech. 1 Introduction For natural languages with rich morphology, knowl- edge of how to inflect base forms is critical for both text generation and analysis. Hand-engineered, rule- based methods for predicting inflections can offer extremely high accuracy, but they are laborious to construct and do not exist with full lexical cover- age in all languages. By contrast, a large number of example inflections are freely available in a semi- structured format on the Web. The English Wik- tionary1 is a crowd-sourced lexical resource that in- cludes complete inflection tables for many lexical items in many languages. We present a supervised ∗Research conducted during an internship at Google. 1http://en.wiktionary.org system that, given only data from Wiktionary, au- tomatically discovers and learns to apply the ortho- graphic transformations governing a language’s in- flectional morphology.2 Our data-driven approach is designed to extend to any language for which we have a substantial num- ber of example inflection tables. The design of our model is guided by three structural assumptions: 1. The inflections of many lexical items are governed by a few repeated morphological paradigms. 2. A morphological paradigm can be decom- posed into independent orthographic transfor- mation rules (including prefix, suffix, and stem changes), which are triggered by orthographic context. 3. A base form is transformed in consistent, cor- related ways to produce its inflected variants. Learning proceeds in two stages that both utilize the same training set of labeled inflection tables. First, an inventory of interpretable transformation rules is generated by aligning each base form to all of its inflected forms. Second, a semi-Markov con- ditional random field (CRF) (Sarawagi and Cohen, 2004) is trained to apply these rules correctly to un- seen base forms. As we demonstrate experimentally, the CRF is most effective when jointly predicting all inflected forms of a lexical item together, forcing the system to adopt a single consistent analysis of each base form. 2See http://eecs.berkeley.edu/~gdurrett for our datasets and code. 1185 Previous work has also described supervised and semi-supervised approaches to predicting inflec- tional morphology (Yarowsky and Wicentowski, 2000; Wicentowski, 2004; Dreyer and Eisner, 2011). Our approach differs primarily in its use of auto- matically …

Content truncated to 500 words