An Unsupervised Method for Uncovering Morphological Chains

An Unsupervised Method for Uncovering Morphological Chains Karthik Narasimhan, Regina Barzilay and Tommi Jaakkola CSAIL, Massachusetts Institute of Technology {karthikn, regina, tommi}@csail.mit.edu Abstract Most state-of-the-art systems today produce morphological analysis based only on ortho- graphic patterns. In contrast, we propose a model for unsupervised morphological anal- ysis that integrates orthographic and seman- tic views of words. We model word forma- tion in terms of morphological chains, from base words to the observed words, breaking the chains into parent-child relations. We use log-linear models with morpheme and word- level features to predict possible parents, in- cluding their modifications, for each word. The limited set of candidate parents for each word render contrastive estimation feasible. Our model consistently matches or outper- forms five state-of-the-art systems on Arabic, English and Turkish.1 1 Introduction Morphologically related words exhibit connections at multiple levels, ranging from orthographical pat- terns to semantic proximity. For instance, the words playing and played share the same stem, but also carry similar meaning. Ideally, all these comple- mentary sources of information would be taken into account when learning morphological structures. Most state-of-the-art unsupervised approaches to morphological analysis are built primarily around orthographic patterns in morphologically-related words (Goldwater and Johnson, 2004; Creutz and Lagus, 2007; Snyder and Barzilay, 2008; Poon et al., 2009). In these approaches, words are com- monly modeled as concatenations of morphemes. 1Code is available at https://github.com/ karthikncode/MorphoChain. This morpheme-centric view is well-suited for un- covering distributional properties of stems and af- fixes. But it is not well-equipped to capture semantic relatedness at the word level. In contrast, earlier approaches that capture se- mantic similarity in morphological variants oper- ate solely at the word level (Schone and Juraf- sky, 2000; Baroni et al., 2002). Given two candi- date words, the proximity is assessed using standard word-distributional measures such as mutual infor- mation. However, the fact that these models do not model morphemes directly greatly limits their per- formance. In this paper, we propose a model to integrate or- thographic and semantic views. Our goal is to build a chain of derivations for a current word from its base form. For instance, given a word playfully, the corresponding chain is play → playful → playfully. The word play is a base form of this derivation as it cannot be reduced any further. Individual deriva- tions are obtained by adding a morpheme (ex. -ful) to a parent word (ex. play). This addition may be implemented via a simple concatenation, or it may involve transformations. At every step of the chain, the model aims to find a parent-child pair (ex. play- playful) such that the parent also constitutes a valid entry in the lexicon. This allows the model to di- rectly compare the semantic similarity of the parent- child pair, while also considering the orthographic properties of the morphemic combination. We model each step of a morphological chain by means of a log-linear model that enables us to in- corporate a wide range of features. At the seman- tic level, we consider …

Content truncated to 500 words