Dependency Tree-based SRL with Proper Pruning and Extensive Feature Engineering

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 253–257 Manchester, August 2008 Dependency Tree-based SRL with Proper Pruning and Extensive Feature Engineering Hongling Wang Honglin Wang Guodong Zhou Qiaoming Zhu JiangSu Provincial Key Lab for Computer Information Processing Technology School of Computer Science and Technology, Soochow University, Suzhou, China 215006 {redleaf, 064227065055,gdzhou, qmzhu}@suda.edu.cn Abstract This paper proposes a dependency tree- based SRL system with proper pruning and extensive feature engineering. Official evaluation on the CoNLL 2008 shared task shows that our system achieves 76.19 in la- beled macro F1 for the overall task, 84.56 in labeled attachment score for syntactic dependencies, and 67.12 in labeled F1 for semantic dependencies on combined test set, using the standalone MaltParser. Be- sides, this paper also presents our unofficial system by 1) applying a new effective pruning algorithm; 2) including additional features; and 3) adopting a better depend- ency parser, MSTParser. Unofficial evalua- tion on the shared task shows that our sys- tem achieves 82.53 in labeled macro F1, 86.39 in labeled attachment score, and 78.64 in labeled F1, using MSTParser on combined test set. This suggests that proper pruning and extensive feature engineering contributes much in dependency tree-based SRL. 1 Introduction Although CoNLL 2008 shared task mainly evaluates joint learning of syntactic and semantic parsing, we focus on dependency tree-based se- mantic role labeling (SRL). SRL refers to label the semantic roles of predicates (either verbs or nouns) in a sentence. Most of previous SRL sys- tems (Gildea and Jurafsky, 2002; Gildea and Palmer, 2002; Punyakanok et al., 2005; Pradhan © 2008. Licensed under the Creative Commons Attribution- Noncommercial-Share Alike 3.0 Unported (http://creativecommons.org/licenses/by-nc-sa/3.0/). license Some rights reserved. et al., 2004, 2005) work on constituent structure trees and has shown to achieve remarkable re- sults. For example, Punyakanok et al. (2005) achieved the best performance in the CoNLL 2005 shared task with 79.44 in F-measure on the WSJ test set and 77.92 on the combined test set (WSJ +Brown). With rapid development of dependency pars- ing in the last few years, more and more re- searchers turn to dependency tree-based SRL with hope to advance SRL from viewpoint of dependency parsing. Hacioglu (2004) pioneered this work by formulating SRL as a classification problem of mapping various dependency rela- tions into semantic roles. Compared with previ- ous researches on constituent structure tree-based SRL which adopts constituents as labeling units, dependency tree-based SRL adopts dependency relations as labeling units. Due to the difference between constituent structure trees and depend- ency trees, their feature spaces are expected to be somewhat different. In the CoNLL 2008 shared task, we extend the framework by Hacioglu (2004) with maximum entropy as our classifier. For evaluation, we will mainly report our official SRL performance us- ing MaltParser (Nivre and Nilsson, 2005). Be- sides, we will also present our unofficial system by 1) applying a new effective pruning algorithm; 2) including additional features; and 3) adopting a better dependency parser, MSTParser (McDon- ald, 2005). In the remainder …

Content truncated to 500 words