Exploiting Synergies Between Open Resources for German Dependency Parsing, POS-tagging, and Morphological Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 601–609, Hissar, Bulgaria, 7-13 September 2013. Exploiting Synergies Between Open Resources for German Dependency Parsing, POS-tagging, and Morphological Analysis Rico Sennrich, Martin Volk and Gerold Schneider Institute of Computational Linguistics University of Zurich Binzmühlestr. 14 CH-8050 Zürich {sennrich,volk,gschneid}@cl.uzh.ch Abstract We report on the recent development of ParZu, a German dependency parser. We discuss the effect of POS tagging and morphological analysis on parsing perfor- mance, and present novel ways of improv- ing performance of the components, in- cluding the use of morphological features for POS-tagging, the use of syntactic in- formation to select good POS sequences from an n-best list, and using parsed text as training data for POS tagging and statisti- cal parsing. We also describe our efforts towards reducing the dependency on re- strictively licensed and closed-source NLP resources. 1 Introduction German NLP tools such as part-of-speech taggers, morphology tools, and syntactic parsers often re- quire licensing and suffer from usage restrictions, which makes the deployment of an NLP pipeline that combines several components cumbersome at best, impossible at worst (if no license can be ob- tained). Some restrictions are rooted in the copy- right and/or licenses of the annotated corpora on which statistical taggers or parsers can be trained for German, such as TIGER (Brants et al., 2002) or Tüba-D/Z (Telljohann et al., 2004). There have been attempts to bypass these restrictions through corpus masking (Rehm et al., 2007), but for statis- tical models that require lexical information, this is not an option. We discuss ParZu, a German dependency parser that relies on external tools for POS tagging and morphological analysis, and combines a hand- written grammar and a statistical disambiguation module that is trained on a treebank. We describe attempts to move towards components with freer licensing. We also discuss techniques to improve parsing performance by better exploiting the vari- ous resources, specifically by using morphological information in POS tagging, and through n-best POS tagging. 2 Parser Architecture ParZu, first described in (Sennrich et al., 2009), is a hybrid dependency parser for German, which implements the grammar described by Foth (2005). It combines a hand-written grammar with a statistical disambiguation module, building on the same architecture as the English Pro3Gres parser (Schneider, 2008). The hand-written gram- mar is mostly unlexicalized and operates on the level of parts-of-speech.1 To give the subject rela- tion as example, the grammar constrains the pos- sible parts-of-speech of the head (a finite verb) and the dependent (typically a noun or pronoun, but some other classes such as numbers are also allowed). The dependent must be in nomina- tive case, and either agree with the verb in per- son and number, or be a coordinated structure. Since each word form may have multiple possible morphological analyses, the morphological con- straints are unification-based and allow for under- specified representations. Also, at most one sub- ject is allowed per finite verb, and some topologi- cal restrictions must be met, such as only one con- stituent …

Content truncated to 500 words