A Pipeline Approach for Syntactic and Semantic Dependency Parsing

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 228–232 Manchester, August 2008 A Pipeline Approach for Syntactic and Semantic Dependency Parsing Yotaro Watanabe and Masakazu Iwatate and Masayuki Asahara and Yuji Matsumoto Nara Institute of Science and Technology, Japan 8916-5, Takayama, Ikoma, Nara, Japan, 630-0192 {yotaro-w, masakazu-i, masayu-a, matsu}@is.naist.jp Abstract This paper describes our system for syn- tactic and semantic dependency parsing to participate the shared task of CoNLL- 2008. We use a pipeline approach, in which syntactic dependency parsing, word sense disambiguation, and semantic role labeling are performed separately: Syn- tactic dependency parsing is performed by a tournament model with a support vector machine; word sense disambigua- tion is performed by a nearest neighbour method in a compressed feature space by probabilistic latent semantic indexing; and semantic role labeling is performed by a an online passive-aggressive algorithm. The submitted result was 79.10 macro- average F1 for the joint task, 87.18% syn- tactic dependencies LAS, and 70.84 se- mantic dependencies F1. After the dead- line, we constructed the other configura- tion, which achieved 80.89 F1 for the joint task, and 74.53 semantic dependencies F1. The result shows that the configuration of pipeline is a crucial issue in the task. 1 Introduction This paper presents the description of our system in CoNLL-2008 shared task. We split the shared task into five sub-problems – syntactic dependency parsing, syntactic dependency label classification, predicate identification, word sense disambigua- tion, and semantic role labeling. The overview of our system is illustrated in Figure 1. Our de- c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Figure 1: Overview of the System pendency parsing module is based on a tourna- ment model (Iida et al., 2003), in which a depen- dency attachment is estimated in step-ladder tour- nament matches. The relative preference of the at- tachment is modeled by one-on-one match in the tournament. Iwatate et al. (Iwatate et al., 2008) initially proposed the method for Japanese depen- dency parsing, and we applied it to other languages by relaxing some constraints (Section 2.1). Depen- dency label classification is performed by a linear- chain sequential labeling on the dependency sib- lings like McDonald’s schemata (McDonald et al., 2006). We use an online passive-aggressive al- gorithm (Crammer et al., 2006) for linear-chain sequential labeling (Section 2.2). We also use the other linear-chain sequential labeling method to annotate whether each word is a predicate or not (Section 2.3). If an identified predicate has more than one sense, a nearest neighbour classifier disambiguates the word sense candidates (Section 2.4). We use an online passive-aggressive algo- rithm again for the semantic role labeling (Section 2.5). The machine learning algorithms used in sep- arated modules are diverse due to role sharing.1 1Unlabeled dependency parsing was done by Iwatate, de- pendency label classification and semantic role labeling was done by Watanabe, predicate identification and word sense 228 We attempt to construct a framework in which each module passes k-best solutions …

Content truncated to 500 words