What Information is Helpful for Dependency Based Semantic Role Labeling

International Joint Conference on Natural Language Processing, pages 781–787, Nagoya, Japan, 14-18 October 2013. What information is helpful for dependency based Semantic Role Labeling Yanyan Luo Kevin Duh Yuji Matsumoto Computational Linguistics, Nara Institute of Science and Technology Takayama, Ikoma, Nara 630-0192, Japan {yanyan-l; kevinduh; matsu}@is.naist.jp Abstract Semantic Role Labeling (SRL) is an im- portant task since it benefits a wide range of natural language processing applica- tions. Given a sentence, the task of SRL is to identify arguments for a pred- icate (target verb or noun) and assign semantically meaningful labels to them. Dependency parsing based methods have achieved much success in SRL. How- ever, due to errors in dependency pars- ing, there remains a large performance gap between SRL based on oracle parses and SRL based on automatic parses in prac- tice. In light of this, this paper investi- gates what additional information is nec- essary to close this gap. Is it worthwhile to introduce additional dependency informa- tion in the form of N-best parse features, or is it better to incorporate orthogonal non- dependency information (base chunk con- stituents)? We compare the above features in a SRL system that achieves state-of-the- art results on the CoNLL 2009 Chinese task corpus. Our findings suggest that or- thogonal information in the form of con- stituents is much more helpful in improv- ing dependency based SRL in practice. 1 Introduction In recent years, SRL has become an impor- tant component in many kinds of deep natural language processing applications, such as ques- tion answering (Narayanan and Harabagiu, 2004), event extraction (Riedel and McCallum, 2011), document categorization (Persson et al., 2009). SRL aims at identifying the semantic relations be- tween predicates in a sentence and their associ- ated arguments, with these relations drawn from a pre-specific list of possible semantic roles for corresponding predicates. Syntax information is essential in SRL systems. To date, both con- stituent parsing and dependency parsing based SRL have been investigated (Xue, 2008; Johans- son and Nugues, 2008), with dependency based systems giving superior results in CoNLL 2008 (Surdeanu et al., 2008) and CoNLL 2009 shared tasks ( Hajiˇc et al., 2009). However, the performance gap is still quite large between SRL systems using oracle ”perfect” dependency parses and SRL systems using auto- matic dependency parses. We observe as much as 10% F-score difference in our experiments. Clearly, errors in the 1-best dependency parse af- fects SRL prediction. This leaves an open ques- tion: in order to improve dependency based SRL, is it more worthwhile to incorporate more depen- dency information (in the form of N-best parse), or to incorporate an entirely separate source of infor- mation, such as base phrase chunks? We perform such an analysis in this paper, using a state-of-the- art Chinese SRL system. Our findings suggest that constituent informa- tion such as chunking nicely complements depen- dency based SRL, achieving more improvements compared to N-best dependency information. Fi- nally, we also report the best results to date on the CoNLL 2009 Chinese shared …

Content truncated to 500 words