Hierarchical Semantic Role Labeling

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 201–204, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Hierarchical Semantic Role Labeling Alessandro Moschitti⋄ moschitti@info.uniroma2.it ⋄ DISP - University of Rome “Tor Vergata”, Rome, Italy † ITC-Irst, ‡ DIT - University of Trento, Povo, Trento, Italy Ana-Maria Giuglea⋄ ana-maria.giuglea@topex.ro Bonaventura Coppola†‡ coppolab@itc.it Roberto Basili⋄ basili@info.uniroma2.it Abstract We present a four-step hierarchical SRL strategy which generalizes the classical two-level approach (boundary detection and classification). To achieve this, we have split the classification step by group- ing together roles which share linguistic properties (e.g. Core Roles versus Ad- juncts). The results show that the non- optimized hierarchical approach is com- putationally more efficient than the tradi- tional systems and it preserves their accu- racy. 1 Introduction For accomplishing the CoNLL 2005 Shared Task on Semantic Role Labeling (Carreras and M`arquez, 2005), we capitalized on our experience on the se- mantic shallow parsing by extending our system, widely experimented on PropBank and FrameNet (Giuglea and Moschitti, 2004) data, with a two- step boundary detection and a hierarchical argument classification strategy. Currently, the system can work in both basic and enhanced configuration. Given the parse tree of an input sentence, the basic system applies (1) a bound- ary classifier to select the nodes associated with cor- rect arguments and (2) a multi-class labeler to assign the role type. For such models, we used some of the linear (e.g. (Gildea and Jurasfky, 2002; Pradhan et al., 2005)) and structural (Moschitti, 2004) features developed in previous studies. In the enhanced configuration, the boundary an- notation is subdivided in two steps: a first pass in which we label argument boundary and a second pass in which we apply a simple heuristic to elimi- nate the argument overlaps. We have also tried some strategies to learn such heuristics automatically. In order to do this we used a tree kernel to classify the subtrees associated with correct predicate argument structures (see (Moschitti et al., 2005)). The ratio- nale behind such an attempt was to exploit the cor- relation among potential arguments. Also, the role labeler is divided into two steps: (1) we assign to the arguments one out of four possi- ble class labels: Core Roles, Adjuncts, Continuation Arguments and Co-referring Arguments, and (2) in each of the above class we apply the set of its spe- cific classifiers, e.g. A0,..,A5 within the Core Role class. As such grouping is relatively new, the tradi- tional features may not be sufficient to characterize each class. Thus, to generate a large set of features automatically, we again applied tree kernels. Since our SRL system exploits the PropBank for- malism for internal data representation, we devel- oped ad-hoc procedures to convert back and forth to the CoNLL Shared Task format. This conversion step gave us useful information about the amount and the nature of the parsing errors. Also, we could measure the frequency of the mismatches between syntax and role annotation. In the remainder of this paper, Section …

Content truncated to 500 words