Getting More from Morphology in Multilingual Dependency Parsing

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 315–326, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Getting More from Morphology in Multilingual Dependency Parsing Matt Hohensee and Emily M. Bender University of Washington Department of Linguistics Box 354340 Seattle WA 98195-4340, USA {hohensee, ebender}@uw.edu Abstract We propose a linguistically motivated set of features to capture morphological agreement and add them to the MSTParser dependency parser. Compared to the built-in morphologi- cal feature set, ours is both much smaller and more accurate across a sample of 20 morpho- logically annotated treebanks. We find in- creases in accuracy of up to 5.3% absolute. While some of this results from the feature set capturing information unrelated to morphol- ogy, there is still significant improvement, up to 4.6% absolute, due to the agreement model. 1 Introduction Most data-driven dependency parsers are meant to be language-independent. They do not use any information that is specific to the language being parsed, and they often rely heavily on n-grams, or sequences of words and POS tags, to make parsing decisions. However, designing a parser without in- corporating any specific linguistic details does not guarantee its language-independence; even linguis- tically na¨ıve systems can involve design decisions which in fact bias the system towards languages with certain properties (Bender, 2011). It is often taken for granted that using linguistic information necessarily makes a system language- dependent. But it is possible to design a linguisti- cally intelligent parser without tuning it to a specific language, by modeling at a high level phenomena which appear cross-linguistically. Such a system is still language-independent; it does not require any knowledge or modeling of specific languages, but it does use linguistic knowledge to make the most of the available data. We present modifications to an existing system, MSTParser (McDonald et al., 2006), to incorporate a very simple model of mor- phological agreement. These modifications improve parsing performance across a variety of languages by making better use of morphological annotations. 2 Background and related work 2.1 Morphological marking of agreement Most languages show some morphological agree- ment via inflected noun, adjective, verb, and deter- miner forms, although the degree to which this hap- pens varies. At one end of the spectrum are analytic, or “morphologically impoverished”, languages. An extreme example is Chinese, which shows no inflec- tion at all; words do not take different forms de- pending on features such as person or gender. En- glish has some inflection, but is relatively morpho- logically poor. At the other end are synthetic or “morphologi- cally rich” languages such as Czech, which has, inter alia, four genders and seven cases. In synthetic lan- guages, words which are syntactically related in cer- tain ways must agree: e.g., subject-verb agreement for gender or determiner-noun agreement for case (Corbett, 2006). Words participating in agreement may be marked explicitly for the property in ques- tion (via affixing or other morphological changes), or may possess it inherently (with no specific …

Content truncated to 500 words