Dependency Parsing and Semantic Role Labeling as a Single Task

International Conference RANLP 2009 - Borovets, Bulgaria, pages 275–280 Dependency Parsing and Semantic Role Labeling as a Single Task Roser Morante, Vincent Van Asch CLiPS - Computational Linguistics University of Antwerp Prinsstraat 13 B-2000 Antwerpen, Belgium Roser.Morante,Vincent.VanAsch@ua.ac.be Antal van den Bosch Tilburg Centre for Creative Computing Tilburg University P.O. Box 90153 NL-5000 LE Tilburg, The Netherlands Antal.vdnBosch@uvt.nl Abstract We present a comparison between two systems for establishing syntactic and semantic depen- dencies: one that performs dependency parsing and semantic role labeling as a single task, and another that performs the two tasks in isola- tion. The systems are based on local memory- based classifiers predicting syntactic and seman- tic dependency relations between pairs of words. In a second global phase, the systems perform a deterministic ranking procedure in which the output of the local classifiers is combined per sentence into a dependency graph and seman- tic role labeling assignments for all predicates. The comparison shows that in the learning phase a joint approach produces better-scoring classi- fiers, while after the ranking phase the isolated approach produces the most accurate syntactic dependencies, while the joint approach yields the most accurate semantic role assignments. Keywords Joint learning, dependency parsing, semantic role labeling 1 Introduction In their currently popular definitions, dependency parsing and semantic role labeling are partly overlap- ping tasks. In their standard definitions they map to differently structured output spaces: dependency graphs span over sentences, while semantic role assign- ments center around individual predicates. Yet, the spaces overlap; in a dependency graph verbal predi- cates will tend to have dependency relations with the same modifiers that have a semantic role as argument of that predicate. In general, even though the la- bels are different, syntactic dependencies between two words often co-occur with the existence of certain se- mantic roles. Although they do not signify the same, the “subject” dependency relation, for example, often co-occurs with the “A0” label that denotes the agent role in the PropBank annotation scheme [14]. Over- laps such as these naturally suggest the possibility of jointly learning the two labeling tasks as if they were one. In this paper we present a system that performs de- pendency parsing and semantic role labeling jointly, which we submitted to the CoNLL Shared Task 2009 [8]. The task combines the identification and label- ing of syntactic dependencies and semantic roles for seven languages. Details about the task setting and the data sets used can be found in the web page of the task1. Additionally, we present a comparison of the joint system with another version of the system (“isolated” system) that processes semantic and syn- tactic dependencies separately. In this way, we are able to evaluate whether and where the joint learning approach is more efficient and successful than the iso- lated approach. As far as we know, this is the first time that such a comparison is performed. In the joint system, the two labeling tasks are learned jointly by merging the syntactic and seman- tic dependencies, which implies …

Content truncated to 500 words