Using Supertags and Encoded Annotation Principles for Improved Dependency to Phrase Structure Conversion

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 305–314, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Using Supertags and Encoded Annotation Principles for Improved Dependency to Phrase Structure Conversion Seth Kulick and Ann Bies and Justin Mott Linguistic Data Consortium University of Pennsylvania Philadelphia, PA 19104 {skulick,bies,jmott}@ldc.upenn.edu Abstract We investigate the problem of automatically converting from a dependency representa- tion to a phrase structure representation, a key aspect of understanding the relationship between these two representations for NLP work. We implement a new approach to this problem, based on a small number of su- pertags, along with an encoding of some of the underlying principles of the Penn Tree- bank guidelines. The resulting system signifi- cantly outperforms previous work in such au- tomatic conversion. We also achieve compara- ble results to a system using a phrase-structure parser for the conversion. A comparison with our system using either the part-of-speech tags or the supertags provides some indication of what the parser is contributing. 1 Introduction and Motivation Recent years have seen a significant increase in interest in dependency treebanks and dependency parsing. Since the standard training and test set for English parsing is a phrase structure (PS) treebank, the Penn Treebank (PTB) (Marcus et al., 1993; Mar- cus et al., 1994), the usual approach is to convert this to a dependency structure (DS) treebank, by means of various heuristics for identifying heads in a PS tree. The resulting DS representation is then used for training and parsing, with results reported on the DS representation. Our goal in this paper is to go in the reverse di- rection, from the DS to PS representation, by find- ing a minimal DS representation from which we can use an approximate version of the principles of the PTB guidelines to reconstruct the PS. Work in this conversion direction is somewhat less studied (Xia et al., 2009; Xia and Palmer, 2001), but it is still an important topic for a number of reasons. First, because both DS and PS treebanks are of current in- terest, there is an increasing effort made to create multi-representational treebank resources with both DS and PS available from the beginning, without a loss of information in either direction (Xia et al., 2009). Second, it is sometimes the case that it is convenient to do annotation in a dependency repre- sentation (e.g., if the annotators are already famil- iar with such a representation), though the treebank will in final form be either phrase-structure or multi- representational (Xia et al., 2009). However, our concern is somewhat different. We are specifically interested in experimenting with de- pendency parsing of Arabic as a step in the annota- tion of the Arabic Treebank, which is a phrase struc- ture treebank (Maamouri et al., 2011). Although we currently use a phrase structure parser in this annota- tion pipeline, there are advantages to the flexibility of being able to experiment with advances in pars- ing technology …

Content truncated to 500 words