A Grammar-driven Convolution Tree Kernel for Semantic Role Classification

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 200–207, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics A Grammar-driven Convolution Tree Kernel for Se- mantic Role Classification Min ZHANG1 Wanxiang CHE2 Ai Ti AW1 Chew Lim TAN3 Guodong ZHOU1,4 Ting LIU2 Sheng LI2 1Institute for Infocomm Research {mzhang, aaiti}@i2r.a-star.edu.sg 2Harbin Institute of Technology {car, tliu}@ir.hit.edu.cn lisheng@hit.edu.cn 3National University of Singapore tancl@comp.nus.edu.sg 4 Soochow Univ., China 215006 gdzhou@suda.edu.cn Abstract Convolution tree kernel has shown promis- ing results in semantic role classification. However, it only carries out hard matching, which may lead to over-fitting and less ac- curate similarity measure. To remove the constraint, this paper proposes a grammar- driven convolution tree kernel for semantic role classification by introducing more lin- guistic knowledge into the standard tree kernel. The proposed grammar-driven tree kernel displays two advantages over the pre- vious one: 1) grammar-driven approximate substructure matching and 2) grammar- driven approximate tree node matching. The two improvements enable the grammar- driven tree kernel explore more linguistically motivated structure features than the previ- ous one. Experiments on the CoNLL-2005 SRL shared task show that the grammar- driven tree kernel significantly outperforms the previous non-grammar-driven one in SRL. Moreover, we present a composite kernel to integrate feature-based and tree kernel-based methods. Experimental results show that the composite kernel outperforms the previously best-reported methods. 1 Introduction Given a sentence, the task of Semantic Role Label- ing (SRL) consists of analyzing the logical forms expressed by some target verbs or nouns and some constituents of the sentence. In particular, for each predicate (target verb or noun) all the constituents in the sentence which fill semantic arguments (roles) of the predicate have to be recognized. Typical se- mantic roles include Agent, Patient, Instrument, etc. and also adjuncts such as Locative, Temporal, Manner, and Cause, etc. Generally, semantic role identification and classification are regarded as two key steps in semantic role labeling. Semantic role identification involves classifying each syntactic element in a sentence into either a semantic argu- ment or a non-argument while semantic role classi- fication involves classifying each semantic argument identified into a specific semantic role. This paper focuses on semantic role classification task with the assumption that the semantic arguments have been identified correctly. Both feature-based and kernel-based learning methods have been studied for semantic role classi- fication (Carreras and Màrquez, 2004; Carreras and Màrquez, 2005). In feature-based methods, a flat feature vector is used to represent a predicate- argument structure while, in kernel-based methods, a kernel function is used to measure directly the similarity between two predicate-argument struc- tures. As we know, kernel methods are more effec- tive in capturing structured features. Moschitti (2004) and Che et al. (2006) used a convolution tree kernel (Collins and Duffy, 2001) for semantic role classification. The convolution tree kernel takes sub-tree as its feature and counts the number of common sub-trees as the similarity between two predicate-arguments. This kernel has shown very 200 promising results in SRL. However, as a …

Content truncated to 500 words