Unsupervised Parse Selection for HPSG

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 694–704, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Unsupervised Parse Selection for HPSG Rebecca Dridan and Timothy Baldwin Dept. of Computer Science and Software Engineering University of Melbourne, Australia rdridan@csse.unimelb.edu.au, tb@ldwin.net Abstract Parser disambiguation with precision gram- mars generally takes place via statistical rank- ing of the parse yield of the grammar using a supervised parse selection model. In the standard process, the parse selection model is trained over a hand-disambiguated treebank, meaning that without a significant investment of effort to produce the treebank, parse selec- tion is not possible. Furthermore, as treebank- ing is generally streamlined with parse selec- tion models, creating the initial treebank with- out a model requires more resources than sub- sequent treebanks. In this work, we show that, by taking advantage of the constrained nature of these HPSG grammars, we can learn a dis- criminative parse selection model from raw text in a purely unsupervised fashion. This al- lows us to bootstrap the treebanking process and provide better parsers faster, and with less resources. 1 Introduction Parsing with precision grammars is generally a two- stage process: (1) the full parse yield of the preci- sion grammar is calculated for a given item, often in the form of a packed forest for efficiency (Oepen and Carroll, 2000; Zhang et al., 2007); and (2) the individual analyses in the parse forest are ranked us- ing a statistical model (“parse selection”). In the do- main of treebank parsing, the Charniak and Johnson (2005) reranking parser adopts an analogous strat- egy, except that ranking and pruning are incorpo- rated into the first stage, and the second stage is based on only the top-ranked parses from the first stage. For both styles of parsing, however, parse se- lection is based on a statistical model learned from a pre-existing treebank associated with the grammar. Our interest in this paper is in completely remov- ing this requirement of parse selection on explicitly treebanked data, ie the development of fully unsu- pervised parse selection models. The particular style of precision grammar we ex- periment with in this paper is HPSG (Pollard and Sag, 1994), in the form of the DELPH-IN suite of grammars (http://www.delph-in.net/). One of the main focuses of the DELPH-IN collab- oration effort is multilinguality. To this end, the Grammar Matrix project (Bender et al., 2002) has been developed which, through a set of question- naires, allows grammar engineers to quickly pro- duce a core grammar for a language of their choice. Bender (2008) showed that by using and expanding on this core grammar, she was able to produce a broad-coverage precision grammar of Wambaya in a very short amount of time. However, the Gram- mar Matrix can only help with the first stage of pars- ing. The statistical model used in the second stage of parsing (ie parse selection) requires a treebank to learn the features, but as we explain in Section 2, …

Content truncated to 500 words