Hybrid Ways to Improve Domain Independence in an ML Dependency Parser

Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 1119–1123, Prague, June 2007. c⃝2007 Association for Computational Linguistics Hybrid Ways to Improve Domain Independence in an ML Dependency Parser Eckhard Bick Institute of Language and Communication University of Southern Denmark 5230 Odense M, Denmark eckhard.bick@mail.dk Abstract This paper reports a hybridization experi­ ment, where a baseline ML dependency pars­ er, LingPars, was allowed access to Con­ straint Grammar analyses provided by a rule­ based parser (EngGram) for the same data. Descriptive compatibility issues and their in­ fluence on performance are discussed. The hybrid system performed considerably better than its ML baseline, and proved more ro­ bust than the latter in the domain adaptation task, where it was the best­scoring system in the open class for the chemical test data, and the best overall system for the CHILDES test data. 1 Introduction LingPars, a language­independent treebank­learner developed in the context of the CoNLL­X 2006 shared task (http://nextens.uvt.nl/~conll/), was in­ spired by the Constraint Grammar (CG) parsing approach (Karlsson et al. 1995) in the sense that it prioritized the identification of syntactic function over syntactic form, basing the dependency poten­ tial of a word on "edge" labels like subject, object etc. rather than the other way around. The system also used other features typical of CG systems, such as BARRIER conditions, tag chains of vari­ able length, implicit clause boundaries and tag sets (Bick 2006). For the 2007 task only one such fea­ ture was newly introduced ­ a directedness marker for a few major functions, splitting subject, adver­ bial and adnominal labels into pairs of left­ and right­attaching labels (e.g. SBJ­L, SBJ­R, NMOD­ L, NMOD­R). Even this small addition, however, increased the memory space requirements of the model to such a degree that only runs with 50­75% of the training data were possible on the available hardware. The main purpose of the LingPars architecture changes for CoNLL2007 (Nivre et al. 2007), how­ ever, was to test two core hypotheses: ● Can an independent, rule­based parser be made to conform to different, data­imposed descriptive conventions without too great a loss in accuracy? ● Does a rules­based dependency parser have a better chance than a machine­learned one to identify long­distance relations and global sentence structure, thus providing valuable arbiter information to the latter? Obviously, both points rule out a test involving many languages with the same parser (CoNLL task 1). The domain adaptation task (task 2), however, satisfied the single­language condition and also adressed the descriptive adaptation problem (sec­ ond hypothesis), involving three English treebanks ­ Wall Street Journal data from the Penn treebank (PTB, Marcus et al. 1993) for training, and the Pchem (Kulick et al. 2004) and CHILDES (Brown 1973 and MacWhinney 2000) treebanks with biomedical and spoken language data, respectively. 2 Developing and adapting EngGram A parser with hand­written rules pays a high "labour price" to arrive at deep, linguistically pre­ 1119 dictable and versatile analyses. For CG systems as employed by the author, the cost, from lexicon …

Content truncated to 500 words