Semantic Dependency Parsing using N-best Semantic Role Sequences and Roleset Information

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 233–237 Manchester, August 2008 Semantic Dependency Parsing using N-best Semantic Role Sequences and Roleset Information Joo-Young Lee, Han-Cheol Cho, and Hae-Chang Rim Natural Language Processing Lab. Korea University Seoul, South Korea {jylee,hccho,rim}@nlp.korea.ac.kr Abstract In this paper, we describe a syntactic and semantic dependency parsing system sub- mitted to the shared task of CoNLL 2008. The proposed system consists of five mod- ules: syntactic dependency parser, predi- cate identifier, local semantic role labeler, global role sequence candidate generator, and role sequence selector. The syntac- tic dependency parser is based on Malt Parser and the sequence candidate gen- erator is based on CKY style algorithm. The remaining three modules are imple- mented by using maximum entropy classi- fiers. The proposed system achieves 76.90 of labeled F1 for the overall task, 84.82 of labeled attachment, and 68.71 of labeled F1 on the WSJ+Brown test set. 1 Introduction In the framework of the CoNLL08 shared task (Surdeanu et al., 2008), a system takes POS tagged sentences as input and produces sentences parsed for syntactic and semantic dependencies as output. A syntactic dependency is represented by an ID of head word and a dependency relation between the head word and its modifier in a sentence. A Semantic dependency is represented by predicate rolesets and semantic arguments for each predi- cate. The task combines two sub-tasks: syntactic dependency parsing and semantic role labeling. Among the sub-tasks, we mainly focus on the se- mantic role labeling task. Compared to previous 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. CoNLL 2004 and 2005 shared tasks (Carreras and M`arquez, 2004; Carreras and M`arquez, 2005) and other semantic role labeling research, major dif- ferences of our semantic role labeling task are 1) considering nominal predicates and 2) identify- ing roleset of predicates. Based on our observa- tion that verbal predicate and nominal predicate have have different characteristics, we decide to build diffent classification modeles for each pred- icate types. The modeles use same features but, their statistical parameters are different. In this paper, maximum entropy1 is used as the classifi- cation model, but any other classification models such as Naive Bayse, SVM, etc. also can be used. To identify roleset, we investigate a roleset match scoring method which evaluate how likely a roleset is matched with the given predicate. 2 System Description The proposed system sequentially performs syn- tactic dependency parsing, predicate identification, local semantic role classification, global sequence generation, and roleset information based selec- tion. 2.1 Syntactic Dependency Parsing In the proposed system, Malt Parser (Nivre et al., 2007) is adopted as the syntactic dependency parser. Although the training and test set of CoNLL08 use non-projective dependency gram- mar, we decide to use projective parsing algorithm, Nivre arc-standard, and projective/non- projective conversion functions that Malt Parser provides. The reason is that non-projective parsing shows worse performance than projective parsing with conversion in our preliminary experiment. 1We use …

Content truncated to 500 words