A Joint Model for Semantic Role Labeling

Proceedings of the 9th Conference on Computational Natural Language Learning (CoNLL), pages 173–176, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics A Joint Model for Semantic Role Labeling Aria Haghighi Dept of Computer Science Stanford University Stanford, CA, 94305 aria42@stanford.edu Kristina Toutanova Dept of Computer Science Stanford University Stanford, CA, 94305 kristina@cs.stanford.edu Christopher D. Manning Dept of Computer Science Stanford University Stanford, CA, 94305 manning@cs.stanford.edu Abstract We present a semantic role labeling sys- tem submitted to the closed track of the CoNLL-2005 shared task. The system, in- troduced in (Toutanova et al., 2005), im- plements a joint model that captures de- pendencies among arguments of a predi- cate using log-linear models in a discrimi- native re-ranking framework. We also de- scribe experiments aimed at increasing the robustness of the system in the presence of syntactic parse errors. Our final system achieves F1-Measures of 76.68 and 78.45 on the development and the WSJ portion of the test set, respectively. 1 Introduction It is evident that there are strong statistical patterns in the syntactic realization and ordering of the argu- ments of verbs; for instance, if an active predicate has an A0 argument it is very likely to come before an A1 argument. Our model aims to capture such de- pendencies among the labels of nodes in a syntactic parse tree. However, building such a model is computation- ally expensive. Since the space of possible joint la- belings is exponential in the number of parse tree nodes, a model cannot exhaustively consider these labelings unless it makes strong independence as- sumptions. To overcome this problem, we adopt a discriminative re-ranking approach reminiscent of (Collins, 2000). We use a local model, which la- bels arguments independently, to generate a smaller number of likely joint labelings. These candidate la- belings are in turn input to a joint model which can use global features and re-score the candidates. Both the local and global re-ranking models are log-linear (maximum entropy) models. In the following sections, we briefly describe our local and joint models and the system architecture for combining them. We list the features used by our models, with an emphasis on new features, and com- pare the performance of a local and a joint model on the CoNLL shared task. We also study an approach to increasing the robustness of the semantic role la- beling system to syntactic parser errors, by consid- ering multiple parse trees generated by a statistical parser. 2 Local Models Our local model labels nodes in a parse tree inde- pendently. We decompose the probability over la- bels (all argument labels plus NONE), into a product of the probability over ARG and NONE, and a prob- ability over argument labels given that a node is an ARG. This can be seen as chaining an identification and a classification model. The identification model classifies each phrase as either an argument or non- argument and our classification model labels each potential argument with a specific argument label. The two models use the …

Content truncated to 500 words