Probabilistic Model for Syntactic and Semantic Dependency Parsing

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 263–267 Manchester, August 2008 Probabilistic Model for Syntactic and Semantic Dependency Parsing Enhong Chen Department of Computer Science, University of Sci- ence and Technology of China, Hefei, China cheneh@ustc.edu.cn Liu Shi Department of Computer Science, University of Sci- ence and Technology of China, Hefei, China shiliu@ustc.edu Dawei Hu Department of Computer Science, University of Sci- ence and Technology of China, Hefei, China dwhu@mail.ustc.edu.cn Abstract This paper proposes a novel method to analyze syntactic dependencies and label semantic dependencies around both the verbal predicates and the nouns. In this method, a probabilistic model is designed to obtain a global optimal result. More- over, a predicate identification model and a disambiguation model are proposed to label predicates and their senses. The ex- perimental results obtained on the wsj and brown test sets show that our system obtains 77% of labeled macro F1 score for the whole task, 84.47% of labeled at- tachment score for syntactic dependency task, and 69.45% of labeled F1 score for semantic dependency task. 1 Introduction There are two difficulties in the CoNLL 2008 shared task. One is how to label semantic role on a dependency-based representation and how to label verbal predicates and nouns. The other one is how to combine the syntactic task with the semantic task together. On the basis of statistical analysis of labeling results, we optimize the traditional approaches of syntactic dependency parsing and semantic role labeling. Moreover, we design a predicate identification model and a disambiguation model, which will be described in section 2.3, for labeling predicates and their senses. In the disambiguation model, an exhaustion method is used to find the best sense which is corresponding to a frame of predicate. In order to obtain a global optimization result for every © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. sentence, a probabilistic model is designed to combine all subtasks. The rest of this paper is organized as follows: our system is described in section 2; and section 3 reports our results on development and test sets; at last we conclude the paper in section 4. 2 A Probabilistic Model for Syntactic and Semantic Dependency Labeling Compared with previous tasks, this shared task is more complex. It aims to merge both syntactic and semantic dependencies under a unified representation. Obviously, it can be divided into two subtasks: syntactic dependency parsing and semantic dependency labeling. For the second subtask, predicates and their senses should be labeled before semantic arguments for predicates are labeled. Since many predicates have only one sense, it is inefficient to build a multi-label classifier to classify each predicate. When a classification approach is used, it is mandatory to consider multiple senses for those predicates with only one or two senses. To prevent assigning irrelevant senses to predicates, we do not adopt classification approach. Instead, two more subtasks, i.e., predicate identification and predicate sense labeling, are introduced in …

Content truncated to 500 words