Multilingual Dependency Parsing and Domain Adaptation using DeSR

Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 1112–1118, Prague, June 2007. c⃝2007 Association for Computational Linguistics Multilingual Dependency Parsing and Domain Adaptation using DeSR Giuseppe Attardi Felice Dell’Orletta Maria Simi Dipartimento di Informatica largo B. Pontecorvo 3 I-56127 Pisa, Italy attardi@di.unipi.it felice.dellorletta@ ilc.cnr.it simi@di.unipi.it Atanas Chanev Università di Trento via Matteo del Ben 5 I-38068 Rovereto, Italy Fondazione Bruno Kessler-irst via Sommarive 18 I-38050 Povo, Italy chanev@form.unitn.it Massimiliano Ciaramita Yahoo! Research Barcelona Ocata 1 S-08003 Barcelona, Spain massi@yahoo-inc.com Abstract We describe our experiments using the DeSR parser in the multilingual and do- main adaptation tracks of the CoNLL 2007 shared task. DeSR implements an incre- mental deterministic Shift/Reduce parsing algorithm, using specific rules to handle non-projective dependencies. For the multi- lingual track we adopted a second order averaged perceptron and performed feature selection to tune a feature model for each language. For the domain adaptation track we applied a tree revision method which learns how to correct the mistakes made by the base parser on the adaptation domain. 1 Introduction Classifier-based dependency parsers (Yamada and Matsumoto, 2003; Nivre and Scholz, 2004) learn from an annotated corpus how to select an appropriate sequence of Shift/Reduce actions to construct the dependency tree for a sentence. Learning is based on techniques such as SVM (Vapnik 1998) or Memory Based Learning (Daelemans 2003), which provide high accuracy but are often computationally expensive. For the multilingual track in the CoNLL 2007 Shared Task, we employed a Shift/Reduce parser which uses a perceptron algorithm with second-order feature maps, in order to verify whether a simpler and faster algorithm can still achieve comparable accuracy. For the domain adaptation track we wished to explore the use of tree revisions in order to incorporate language knowledge from a new domain. 2 Multilingual Track The overall parsing algorithm is a deterministic classifier-based statistical parser, which extends the approach by Yamada and Matsumoto (2003), by using different reduction rules that ensure deterministic incremental processing of the input sentence and by adding specific rules for handling non-projective dependencies. The parser also performs dependency labeling within a single processing step. The parser is modular and can use several learning algorithms. The submitted runs used a second order Average Perceptron, derived from the multiclass perceptron of Crammer and Singer (2003). No additional resources were used. No pre- processing or post-processing was used, except stemming for English, by means of the Snowball stemmer (Porter 2001). 3 Deterministic Classifier-based Parsing DeSR (Attardi, 2006) is an incremental determinis- tic classifier-based parser. The parser constructs dependency trees employing a deterministic bot- tom-up algorithm which performs Shift/Reduce actions while analyzing input sentences in left-to- right order. Using a notation similar to (Nivre and Scholz, 2003), the state of the parser is represented by a 1112 quadruple 〈S, I, T, A〉, where S is the stack of past tokens, I is the list of (remaining) input tokens, T is a stack of temporary tokens and A is the arc rela- tion for the dependency graph. Given …

Content truncated to 500 words