Language Independent Dependency to Constituent Tree Conversion

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 421–428, Osaka, Japan, December 11-17 2016. Language Independent Dependency to Constituent Tree Conversion Young-Suk Lee IBM T. J. Watson Research Center Yorktown Heights, NY 10598 USA ysuklee@us.ibm.com Zhiguo Wang IBM T. J. Watson Research Center Yorktown Heights, NY 10598 USA zhigwang@us.ibm.com Abstract We present a dependency to constituent tree conversion technique that aims to improve con- stituent parsing accuracies by leveraging dependency treebanks available in a wide variety in many languages. The technique works in two steps. First, a partial constituent tree is derived from a dependency tree with a very simple deterministic algorithm that is both language and de- pendency type independent. Second, a complete high accuracy constituent tree is derived with a constraint-based parser, which uses the partial constituent tree as external constraints. Evaluated on Section 22 of the WSJ Treebank, the technique achieves the state-of-the-art conversion F- score 95.6. When applied to English Universal Dependency treebank and German CoNLL2006 treebank, the converted treebanks added to the human-annotated constituent parser training cor- pus improve parsing F-scores significantly for both languages. 1 Introduction State-of-the-art parsers require human annotation of a training corpus in a specific representation, e.g. constituent structure in Penn Treebank (Charniak and Johnson, 2005; Petrov and Klein, 2007) or de- pendency relations in a dependency treebank (Yamada and Matsumoto, 2003; McDonald et al., 2005) . Creation of human-annotated treebanks, however, is knowledge and labor intensive and it is desired that one can improve parsing performance by leveraging treebanks annotated in representations of a wide variety. While there have been quite a few papers on automatic conversion from dependency to constituent trees and vice versa (Wang et al., 1994; Collins et al., 1999; Forst, 2003; de Marneffe et al., 2006; Jo- hansson and Nugues, 2007; Xia et al., 2008; Hall and Nivre, 2008; Rambow, 2010; Wang and Zong, 2010; Zhang et al., 2013; Simk´o et al., 2014; Kong et al., 2015) , very few papers address the issue of whether or not the converted treebank actually improves the performance of the target parser when added to the human-annotated gold treebanks for parser training. In addition, much of the work on de- pendency to constituency conversion relies on dependency trees automatically derived from the Penn Treebank (Marcus et al., 1993) via head rules and assumes that the head-modifier definitions are consis- tent between the constituent and dependency trees (Xia et al., 2008). However, such techniques cannot easily generalize to dependencies that diverge from the Penn Treebank in head-modifier definitions and dependency labels, e.g. Universal Dependency (Nivre et al., 2015) in Figure 1(b), and the dependencies of a wide variety available in CoNLL shared tasks. In this paper, we propose a very simple dependency to constituent tree conversion technique which is applicable to any languages and any dependencies, e.g. Universal Dependency (UD), CoNLL de- pendencies (CoNLL), Stanford dependencies (Stanford), while achieving the state-of-the-art conversion accuracy. The technique works in two steps. We first derive a partial …

Content truncated to 500 words