A Transition-Based System for Joint Part-of-Speech Tagging and Labeled Non-Projective Dependency Parsing

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1455–1465, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics A Transition-Based System for Joint Part-of-Speech Tagging and Labeled Non-Projective Dependency Parsing Bernd Bohnet Institute for Natural Language Processing University Stuttgart bohnet@ims.uni-stuttgart.de Joakim Nivre Department of Linguistics and Philology Uppsala University joakim.nivre@lingfil.uu.se Abstract Most current dependency parsers presuppose that input words have been morphologically disambiguated using a part-of-speech tagger before parsing begins. We present a transition- based system for joint part-of-speech tagging and labeled dependency parsing with non- projective trees. Experimental evaluation on Chinese, Czech, English and German shows consistent improvements in both tagging and parsing accuracy when compared to a pipeline system, which lead to improved state-of-the- art results for all languages. 1 Introduction Dependency-based syntactic parsing has been the focus of intense research efforts during the last decade, and the state of the art today is represent- ed by globally normalized discriminative models that are induced using structured learning. Graph- based models parameterize the parsing problem by the structure of the dependency graph and normally use dynamic programming for inference (McDonald et al., 2005; McDonald and Pereira, 2006; Carreras, 2007; Koo and Collins, 2010; Bohnet, 2010), but other inference methods have been explored espe- cially for non-projective parsing (Riedel and Clarke, 2006; Smith and Eisner, 2008; Martins et al., 2009; Martins et al., 2010; Koo et al., 2010). Transition- based models parameterize the problem by elemen- tary parsing actions and typically use incremental beam search (Titov and Henderson, 2007; Zhang and Clark, 2008; Zhang and Clark, 2011). Despite notable differences in model structure, graph-based and transition-based parsers both give state-of-the- art accuracy with proper feature selection and opti- mization (Koo and Collins, 2010; Zhang and Nivre, 2011; Bohnet, 2011). It is noteworthy, however, that almost all depen- dency parsers presuppose that the words of an input sentence have been morphologically disambiguated using (at least) a part-of-speech tagger. This is in s- tark contrast to the best parsers based on PCFG mod- els, such as the Brown parser (Charniak and John- son, 2005) and the Berkeley parser (Petrov et al., 2006; Petrov and Klein, 2007), which not only can perform their own part-of-speech tagging but nor- mally give better parsing accuracy when they are al- lowed to do so. This suggests that joint models for tagging and parsing might improve accuracy also in the case of dependency parsing. It has been argued that joint morphological and syntactic disambiguation is especially important for richly inflected languages, where there is consid- erable interaction between morphology and syntax such that neither can be fully disambiguated with- out considering the other. Thus, Lee et al. (2011) show that a discriminative model for joint morpho- logical disambiguation and dependency parsing out- performs a pipeline model in experiments on Latin, Ancient Greek, Czech and Hungarian. However, Li et al. (2011) and Hatori et al. (2011) report improve- ments with a joint model also for …

Content truncated to 500 words