A Hybrid Approach to the Induction of Underlying Morphology

A Hybrid Approach to the Induction of Underlying Morphology Michael Tepper Department of Linguistics University of Washington Seattle, WA 98195 mtepper@u.washington.edu Fei Xia Department of Linguistics University of Washington Seattle, WA 98195 fxia@u.washington.edu Abstract We present a technique for refining a base- line segmentation and generating a plausible underlying morpheme segmentation by inte- grating hand-written rewrite rules into an ex- isting state-of-the-art unsupervised morpho- logical induction procedure. Performance on measures which consider surface-boundary accuracy and underlying morpheme consis- tency indicates this technique leads to im- provements over baseline segmentations for English and Turkish word lists. 1 Introduction 1.1 Unsupervised Morphological Induction The primary goal of unsupervised morphological in- duction (UMI) is the simultaneous induction of a reasonable morphological lexicon as well as an op- timal segmentation of a corpus of words, given that lexicon. The majority of existing approaches em- ploy statistical modeling towards this goal, but dif- fer with respect to how they learn or refine the mor- phological lexicon. While some approaches involve lexical priors, either internally motivated or moti- vated by the minimal description length (MDL) cri- terion, some utilize heuristics. Pure maximum like- lihood (ML) approaches may refine the lexicon with heuristics in lieu of explicit priors (Creutz and La- gus, 2004), or not make categorical refinements at all concerning which morphs are included, only proba- bilistic refinements through a hierarchical EM pro- cedure (Peng and Schuurmans, 2001). Approaches that optimize the lexicon with respect to priors come in several flavors. There are basic maximum a priori (MAP) approaches that try to maximize the proba- bility of the lexicon against linguistically motivated priors (Deligne and Bimbot, 1997; Snover and Brent, 2001; Creutz and Lagus, 2005). An alternative to MAP, MDL approaches use their own set of pri- ors motivated by complexity theory. These studies attempt to minimize lexicon complexity (bit-length in crude MDL) while simultaneously minimizing the complexity (by maximizing the probability) of the corpus given the lexicon (de Marcken, 1996; Gold- smith, 2001; Creutz and Lagus, 2002). Many of the approaches mentioned above utilize a simplistic unigram model of morphology to produce the segmentation of the corpus given the lexicon. Substrings in the lexicon are proposed as morphs within a word based on frequency alone, indepen- dently of phrase-, word- and morph-surroundings (de Marcken, 1996; Peng and Schuurmans, 2001; Creutz and Lagus, 2002). There are many approaches, however, which further constrain the segmentation procedure. The work by Creutz and Lagus (2004; 2005; 2006) constrains segmentation by accounting for morphotactics, first assigning mophotactic cate- gories (prefix, suffix, and stem) to baseline morphs, and then seeding and refining an HMM using those category assignments. Other more structured mod- els include Goldsmith’s (2001) work which, instead of inducing morphemes, induces morphological sig- natures like {ø, s, ed, ing} for English regular verbs. Some techniques constrain possible analyses by em- ploying approximations for morphological meaning or usage to prevent false derivations (like singed = sing + ed). There is work by Schone and Juraf- sky (2000; 2001) where meaning …

Content truncated to 500 words