Linguistic Phenomena, Analyses, and Representations: Understanding Conversion between Treebanks

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1234–1242, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Linguistic Phenomena, Analyses, and Representations: Understanding Conversion between Treebanks Rajesh Bhatt Univ. of Massachusetts Amherst, MA 01003, USA bhatt@linguist.umass.edu Owen Rambow CCLS, Columbia University New York, NY 10115, USA rambow@ccls.columbia.edu Fei Xia Univ. of Washington Seattle, WA 98195, USA fxia@uw.edu Abstract Treebanks are valuable resources for nat- ural language processing (NLP). There is much work in NLP which converts treebanks from one representation (e.g., phrase structure) to another (e.g., depen- dency) before applying machine learning. This paper provides a framework in which to think about the question of when such a conversion is possible. 1 Introduction There has been much interest in converting tree- banks from one representation to another; for in- stance, from phrase structure to dependency struc- ture (e.g., motivated by the recent surge in interest in dependency parsing), or from phrase structure to other grammatical frameworks such as LTAG, HPSG, CCG, or LFG. While there has been much work on converting between treebank representa- tions (Collins et al., 1999; Xia and Palmer, 2001; Cahill et al., 2002; Nivre, 2003; Hockenmaier and Steedman, 2007), there has not been a general yet precise discussion of what conditions are neces- sary for such conversion to happen. In this paper, we provide an analytical frame- work for determining how difficult it would be to convert representations under one set of annota- tion guidelines M1 to representations under an- other set of guidelines M2. We are only interested in cases where annotation guidelines are available for both levels of representation, since it is not clear how one would interpret an undocumented representation, and thus it would not be clear how to evaluate the conversion results. Given two sets of guidelines and a particular linguistic phe- nomenon, there are three possible scenarios: (1) the phenomenon is represented only on one side; (2) the phenomenon is represented on both sides with incompatible analyses; (3) the phenomenon is represented on both sides with compatible anal- yses. We give a formal definition of compatibility and a procedure for distinguishing these three sce- narios. We also discuss how each scenario will af- fect automatic conversion. Using this framework, researchers can determine the difficulty of a con- version task between existing guidelines, or they can design guidelines for new treebanks so auto- matic conversion to other representations can be as smooth as possible. Note that we are not addressing general ques- tions such as “In general, is it easier to convert de- pendency to phrase structure or vice versa?” We believe that such general questions cannot be an- swered. One needs to examine what information is being represented before the issue of conver- sion can be addressed, i.e., we must first study the guidelines of the two levels of representation. While we propose a general approach to an- alyzing syntactic representations, throughout the paper we will use examples based on convert- ing dependency structures to …

Content truncated to 500 words