Joint Morphological and Syntactic Analysis for Richly Inflected Languages

Joint Morphological and Syntactic Analysis for Richly Inflected Languages Bernd Bohnet∗ Joakim Nivre⋆ Igor Boguslavsky•◦ Rich´ard Farkas⋄ Filip Ginter† Jan Hajiˇc‡ ∗University of Birmingham, School of Computer Science ⋆Uppsala University, Department of Linguistics and Philology •Universidad Polit´ecnica de Madrid, Departamento de Inteligencia Artificial ◦Russian Academy of Sciences, Institute for Information Transmission Problems ⋄University of Szeged, Institute of Informatics †University of Turku, Department of Information Technology ‡Charles University in Prague, Institute of Formal and Applied Linguistics Abstract Joint morphological and syntactic analysis has been proposed as a way of improving parsing accuracy for richly inflected languages. Start- ing from a transition-based model for joint part-of-speech tagging and dependency pars- ing, we explore different ways of integrating morphological features into the model. We also investigate the use of rule-based mor- phological analyzers to provide hard or soft lexical constraints and the use of word clus- ters to tackle the sparsity of lexical features. Evaluation on five morphologically rich lan- guages (Czech, Finnish, German, Hungarian, and Russian) shows consistent improvements in both morphological and syntactic accuracy for joint prediction over a pipeline model, with further improvements thanks to lexical con- straints and word clusters. The final results improve the state of the art in dependency parsing for all languages. 1 Introduction Syntactic parsing of natural language has witnessed a tremendous development during the last twenty years, especially through the use of statistical mod- els for robust and accurate broad-coverage parsing. However, as statistical parsing techniques have been applied to more and more languages, it has also been observed that typological differences between languages lead to new challenges. In particular, it has been found over and over again that languages exhibiting rich morphological structure, often to- gether with a relatively free word order, usually ob- tain lower parsing accuracy, especially in compar- ison to English. One striking demonstration of this tendency can be found in the CoNLL shared tasks on multilingual dependency parsing, organized in 2006 and 2007, where richly inflected languages clustered at the lower end of the scale with respect to pars- ing accuracy (Buchholz and Marsi, 2006; Nivre et al., 2007). These and similar observations have led to an increased interest in the special challenges posed by parsing morphologically rich languages, as evidenced most clearly by a new series of work- shops devoted to this topic (Tsarfaty et al., 2010), as well as a special issue in Computational Linguistics (Tsarfaty et al., 2013) and a shared task on parsing morphologically rich languages.1 One hypothesized explanation for the lower pars- ing accuracy observed for richly inflected languages is the strict separation of morphological and syn- tactic analysis assumed in many parsing frame- works (Tsarfaty et al., 2010; Tsarfaty et al., 2013). This is true in particular for data-driven dependency parsers, which tend to assume that all morphological disambiguation has been performed before syntactic analysis begins. However, as argued by Lee et al. (2011), in morphologically rich languages there is often considerable interaction between morphology and syntax, such that neither can be disambiguated …

Content truncated to 500 words