Dependency Parsing and Domain Adaptation with LR Models and Parser Ensembles

Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 1044–1050, Prague, June 2007. c⃝2007 Association for Computational Linguistics Dependency Parsing and Domain Adaptation with LR Models and Parser Ensembles Kenji Sagae1 and Jun’ichi Tsujii1,2,3 1Department of Computer Science University of Tokyo Hongo 7-3-1, Bunkyo-ku, Tokyo, Japan 2School of Computer Science, University of Manchester 3National Center for Text Mining {sagae,tsujii}@is.s.u-tokyo.ac.jp Abstract We present a data-driven variant of the LR algorithm for dependency parsing, and ex- tend it with a best-first search for probabil- istic generalized LR dependency parsing. Parser actions are determined by a classifi- er, based on features that represent the cur- rent state of the parser. We apply this pars- ing framework to both tracks of the CoNLL 2007 shared task, in each case taking ad- vantage of multiple models trained with different learners. In the multilingual track, we train three LR models for each of the ten languages, and combine the analyses obtained with each individual model with a maximum spanning tree voting scheme. In the domain adaptation track, we use two models to parse unlabeled data in the target domain to supplement the labeled out-of- domain training set, in a scheme similar to one iteration of co-training. 1 Introduction There are now several approaches for multilingual dependency parsing, as demonstrated in the CoNLL 2006 shared task (Buchholz and Marsi, 2006). The dependency parsing approach pre- sented here extends the existing body of work mainly in four ways: 1. Although stepwise 1 dependency parsing has commonly been performed using parsing algo- 1 Stepwise parsing considers each step in a parsing algo- rithm separately, while all-pairs parsing considers entire rithms designed specifically for this task, such as those described by Nivre (2003) and Yamada and Matsumoto (2003), we show that this can also be done using the well known LR parsing algorithm (Knuth, 1965), providing a connec- tion between current research on shift-reduce dependency parsing and previous parsing work using LR and GLR models; 2. We generalize the standard deterministic step- wise framework to probabilistic parsing, with the use of a best-first search strategy similar to the one employed in constituent parsing by Rat- naparkhi (1997) and later by Sagae and Lavie (2006); 3. We provide additional evidence that the parser ensemble approach proposed by Sagae and La- vie (2006a) can be used to improve parsing ac- curacy, even when only a single parsing algo- rithm is used, as long as variation can be ob- tained, for example, by using different learning techniques or changing parsing direction from forward to backward (of course, even greater gains may be achieved when different algo- rithms are used, although this is not pursued here); and, finally, 4. We present a straightforward way to perform parser domain adaptation using unlabeled data in the target domain. We entered a system based on the approach de- scribed in this paper in the CoNLL 2007 shared trees. For a more complete definition, see the CoNLL- X shared task description paper (Buchholz and Marsi, 2006). …

Content truncated to 500 words