Down-stream effects of tree-to-dependency conversions

Proceedings of NAACL-HLT 2013, pages 617–626, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Down-stream effects of tree-to-dependency conversions Jakob Elming, Anders Johannsen, Sigrid Klerke, Emanuele Lapponi†, Hector Martinez, Anders Søgaard Center for Language Technology, University of Copenhagen †Institute for Informatics, University of Oslo Abstract Dependency analysis relies on morphosyntac- tic evidence, as well as semantic evidence. In some cases, however, morphosyntactic ev- idence seems to be in conflict with seman- tic evidence. For this reason dependency grammar theories, annotation guidelines and tree-to-dependency conversion schemes often differ in how they analyze various syntactic constructions. Most experiments for which constituent-based treebanks such as the Penn Treebank are converted into dependency tree- banks rely blindly on one of four-five widely used tree-to-dependency conversion schemes. This paper evaluates the down-stream effect of choice of conversion scheme, showing that it has dramatic impact on end results. 1 Introduction Annotation guidelines used in modern depen- dency treebanks and tree-to-dependency conversion schemes for converting constituent-based treebanks into dependency treebanks are typically based on a specific dependency grammar theory, such as the Prague School’s Functional Generative Description, Meaning-Text Theory, or Hudson’s Word Grammar. In practice most parsers constrain dependency struc- tures to be tree-like structures such that each word has a single syntactic head, limiting diversity be- tween annotation a bit; but while many dependency treebanks taking this format agree on how to an- alyze many syntactic constructions, there are still many constructions these treebanks analyze differ- ently. See Figure 1 for a standard overview of clear and more difficult cases. The difficult cases in Figure 1 are difficult for the following reason. In the easy cases morphosyn- tactic and semantic evidence cohere. Verbs gov- ern subjects morpho-syntactically and seem seman- tically more important. In the difficult cases, how- ever, morpho-syntactic evidence is in conflict with the semantic evidence. While auxiliary verbs have the same distribution as finite verbs in head position and share morpho-syntactic properties with them, and govern the infinite main verbs, main verbs seem semantically superior, expressing the main predi- cate. There may be distributional evidence that com- plementizers head verbs syntactically, but the verbs seem more important from a semantic point of view. toTree-to-dependency conversion schemes used convert constituent-based treebanks into dependency-based ones also take different stands on the difficult cases. In this paper we consider four dif- ferent conversion schemes: the Yamada-Matsumoto conversion scheme yamada,1 the CoNLL 2007 format conll07,2 the conversion scheme ewt used in the English Web Treebank (Petrov and McDonald, 2012),3 and the lth conversion scheme (Johansson 1The Yamada-Matsumoto scheme can be replicated by running penn2malt.jar available at http://w3.msi.vxu.se/∼nivre/research/Penn2Malt.html. We used Malt dependency labels (see website). The Yamada- Matsumoto scheme is an elaboration of the Collins scheme (Collins, 1999), which is not included in our experiments. 2The CoNLL 2007 conversion scheme can be obtained by running pennconverter.jar available at http://nlp.cs.lth.se/software/treebank converter/with the ’conll07’ flag set. 3The EWT conversion scheme can be repli- cated using the Stanford converter available at http://nlp.stanford.edu/software/stanford-dependencies.shtml 617 Clear cases Difficult cases Head Dependent …

Content truncated to 500 words