Modeling Syntactic Context Improves Morphological Segmentation

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 1–9, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Modeling Syntactic Context Improves Morphological Segmentation Yoong Keok Lee Aria Haghighi Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {yklee, aria42, regina}@csail.mit.edu Abstract The connection between part-of-speech (POS) categories and morphological properties is well-documented in linguistics but underuti- lized in text processing systems. This pa- per proposes a novel model for morphologi- cal segmentation that is driven by this connec- tion. Our model learns that words with com- mon affixes are likely to be in the same syn- tactic category and uses learned syntactic cat- egories to refine the segmentation boundaries of words. Our results demonstrate that incor- porating POS categorization yields substantial performance gains on morphological segmen- tation of Arabic. 1 1 Introduction A tight connection between morphology and syntax is well-documented in linguistic literature. In many languages, morphology plays a central role in mark- ing syntactic structure, while syntactic relations help to reduce morphological ambiguity (Harley and Phillips, 1994). Therefore, in an unsupervised lin- guistic setting which is rife with ambiguity, model- ing this connection can be particularly beneficial. However, existing unsupervised morphological analyzers take little advantage of this linguistic property. In fact, most of them operate at the vo- cabulary level, completely ignoring sentence con- text. This design is not surprising: a typical mor- phological analyzer does not have access to syntac- 1The source code for the work presented in this paper is available at http://groups.csail.mit.edu/rbg/code/morphsyn/. tic information, because morphological segmenta- tion precedes other forms of sentence analysis. In this paper, we demonstrate that morphological analysis can utilize this connection without assum- ing access to full-fledged syntactic information. In particular, we focus on two aspects of the morpho- syntactic connection: • Morphological consistency within POS cat- egories. Words within the same syntactic cat- egory tend to select similar affixes. This lin- guistic property significantly reduces the space of possible morphological analyses, ruling out assignments that are incompatible with a syn- tactic category. • Morphological realization of grammatical agreement. In many morphologically rich lan- guages, agreement between syntactic depen- dents is expressed via correlated morphological markers. For instance, in Semitic languages, gender and number agreement between nouns and adjectives is expressed using matching suf- fixes. Enforcing mutually consistent segmen- tations can greatly reduce ambiguity of word- level analysis. In both cases, we do not assume that the relevant syntactic information is provided, but instead jointly induce it as part of morphological analysis. We capture morpho-syntactic relations in a Bayesian model that grounds intra-word decisions in sentence-level context. Like traditional unsuper- vised models, we generate morphological structure from a latent lexicon of prefixes, stems, and suffixes. 1 In addition, morphological analysis is guided by a latent variable that clusters together words with sim- ilar affixes, acting as a proxy for POS tags. More- over, a sequence-level component further refines the analysis by correlating segmentation decisions be- tween adjacent words that exhibit morphological …

Content truncated to 500 words