Special Techniques for Constituent Parsing of Morphologically Rich Languages

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 135–144, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Special Techniques for Constituent Parsing of Morphologically Rich Languages Zsolt Sz´ant´o, Rich´ard Farkas University of Szeged Department of Informatics {szanto,rfarkas}@inf.u-szeged.hu Abstract We introduce three techniques for improv- ing constituent parsing for morphologi- cally rich languages. We propose a novel approach to automatically find an optimal preterminal set by clustering morphologi- cal feature values and we conduct exper- iments with enhanced lexical models and feature engineering for rerankers. These techniques are specially designed for mor- phologically rich languages (but they are language-agnostic). We report empirical results on the treebanks of five morpho- logically rich languages and show a con- siderable improvement in accuracy and in parsing speed as well. 1 Introduction From the viewpoint of syntactic parsing, the languages of the world are usually categorized according to their level of morphological rich- ness (which is negatively correlated with config- urationality). At one end, there is English, a strongly configurational language while there is Hungarian at the other end of the spectrum with rich morphology and free word order (Fraser et al., 2013). A large part of the methodology for syn- tactic parsing has been developed for English but many other languages of the world are fundamen- tally different from English. In particular, mor- phologically rich languages – the other end of the configurational spectrum – convey most sentence- level syntactic information by morphology (i.e. at the word level), not by configuration. Because of these differences the parsing of morphologically rich languages requires techniques that differ from or extend the methodology developed for English (Tsarfaty et al., 2013). In this study, we present three techniques to improve constituent parsing and these special techniques are dedicated to han- dle the challenges of morphologically rich lan- guages. Constituency parsers have advanced consider- ably in the last two decades (Charniak, 2000; Charniak and Johnson, 2005; Petrov et al., 2006; Huang, 2008) boosted by the availability of the Penn Treebank (Marcus et al., 1993). While there is a progress on parsing English (especially the Penn Treebank), the treebanks of morphologi- cally rich languages have been attracted much less attention. For example, a big constituent treebank has been available for Hungarian for almost 10 years (Csendes et al., 2005) and to the best of our knowledge our work is the first one report- ing results on this treebank. One reason for the moderate level of interest in constituent parsing of morphologically rich languages is the widely held belief that dependency structures are better suited for representing syntactic analyses for morpho- logically rich languages than constituent represen- tations because they allow non-projective struc- tures (i.e. discontinuous constituents). From a theoretical point of view, Tsarfaty et al. (2010) point out, however, this is not the same as prov- ing that dependency parsers function better than constituency parsers for parsing morphologically rich languages. For a detailed discussion, please see Fraser et al. (2013). …

Content truncated to 500 words