Experimental Evaluation of LTAG-Based Features for Semantic Role Labeling

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 590–599, Prague, June 2007. c⃝2007 Association for Computational Linguistics Experimental Evaluation of LTAG-based Features for Semantic Role Labeling Yudong Liu and Anoop Sarkar Simon Fraser University Burnaby, BC, Canada {yudongl,anoop}@cs.sfu.ca Abstract This paper proposes the use of Lexical- ized Tree-Adjoining Grammar (LTAG) for- malism as an important additional source of features for the Semantic Role Labeling (SRL) task. Using a set of one-vs-all Sup- port Vector Machines (SVMs), we evalu- ate these LTAG-based features. Our exper- iments show that LTAG-based features can improve SRL accuracy significantly. When compared with the best known set of fea- tures that are used in state of the art SRL sys- tems we obtain an improvement in F-score from 82.34% to 85.25%. 1 Introduction Semantic Role Labeling (SRL) aims to identify and label all the arguments for each predicate occurring in a sentence. It involves identifying constituents in the sentence that represent the predicate’s arguments and assigning pre-specified semantic roles to them. [A0seller Ports of Call Inc.] reached agreements to [Vverb sell] [A1thing its remaining seven aircraft] [A2buyer to buyers that weren’t disclosed] . is an example of SRL annotation from the PropBank corpus (Palmer et al., 2005), where the subscripted information maps the semantic roles A0, A1, A2 to arguments for the predicate sell as defined in the PropBank Frame Scheme. For SRL, high accuracy has been achieved by: (i) proposing new types of features (see Table 1 in Section 3 for previously proposed features), (ii) modeling the predicate frameset by capturing de- pendencies between arguments (Gildea and Juraf- sky, 2002; Pradhan et al., 2004; Toutanova et al., 2005; Punyakanok et al., 2005a), (iii) dealing with incorrect parser output by using more than one parser (Pradhan et al., 2005b). Our work in this paper falls into category (i). We propose several novel features based on Lexicalized Tree Adjoining Grammar (LTAG) derivation trees in order to improve SRL performance. To show the usefulness of these features, we provide an ex- perimental study comparing LTAG-based features with the standard set of features and kernel meth- ods used in state-of-the-art SRL systems. The LTAG formalism provides an extended domain of locality in which to specify predicate-argument relationships and also provides the notion of a derivation tree. These two properties of LTAG make it well suited to address the SRL task. SRL feature extraction has relied on various syn- tactic representations of input sentences, such as syntactic chunks (Hacioglu et al., 2004) and full syntactic parses (Gildea and Jurafsky, 2002). In contrast with features from shallow parsing, previ- ous work (Gildea and Palmer, 2002; Punyakanok et al., 2005b) has shown the necessity of full syntactic parsing for SRL. In order to generalize the path fea- ture (see Table 1 in Section 3) which is probably the most salient (while being the most data sparse) fea- ture for SRL, previous work has extracted features from other syntactic representations, such as …

Content truncated to 500 words