Hybrid Learning of Dependency Structures from Heterogeneous Linguistic Resources

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 198–202 Manchester, August 2008 Hybrid Learning of Dependency Structures from Heterogeneous Linguistic Resources Yi Zhang Language Technology Lab DFKI GmbH yzhang@coli.uni-sb.de Rui Wang Computational Linguistics Saarland University, Germany rwang@coli.uni-sb.de Hans Uszkoreit Language Technology Lab DFKI GmbH uszkoreit@dfki.de Abstract In this paper we present our syntactic and semantic dependency parsing system par- ticipated in both closed and open compe- titions of the CoNLL 2008 Shared Task. By combining the outcome of two state-of- the-art syntactic dependency parsers, we achieved high accuracy in syntactic de- pendencies (87.32%). With MRSes from grammar-based HPSG parsers, we achieved significant performance improvement on semantic role labeling (from 71.31% to 71.89%), especially in the out-domain evaluation (from 60.16% to 62.11%). 1 Introduction The CoNLL 2008 shared task (Surdeanu et al., 2008) provides a unique chance of comparing dif- ferent syntactic and semantic parsing techniques in one unified open competition. Our contribution in this joint exercise focuses on the combination of different algorithms and resources, aiming not only for state-of-the-art performance in the com- petition, but also for the dissemination of the learnt lessons to related sub-fields in computational lin- guistics. The so-called hybrid approach we take has two folds of meaning. For syntactic dependency pars- ing, we build our system based on state-of-the art algorithms. Past CoNLL share task results have shown that transition-based and graph-based algo- rithms started from radically different ideas, yet achieved largely comparable results. One of the question we would like investigate is whether the c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. combination of the two approach on the output level leads to even better results. For the semantic role labeling (SRL) task, we would like to build a system that allows us to test the contribution of different linguistic resources. To our special interest is to examine the deep linguistic parsing systems based on hand-crafted grammars. During the past decades, various large scale linguistic grammars have been built, some of which achieved both broad coverage and high precision. In combination with other advances in deep linguistic processing, e.g. efficient pars- ing algorithms, statistical disambiguation models and robust processing techniques, several systems have reached mature stage to be deployed in ap- plications. Unfortunately, due to the difficulties in cross-framework evaluation, fair comparison of these systems with state-of-the-art data-driven sta- tistical parsers is still hard to achieve. More impor- tantly, it is not even clear whether deep linguistic analysis is necessary at all for tasks such as shallow semantic parsing (also known as SRL). Drawing a conclusion on this latter point with experiments using latest deep parsing techniques is one of our objective. The remainder of the paper is structure as fol- lows. Section 2 introduces the overall system ar- chitecture. Section 3 explains the voting mecha- nism used in the syntactic parser. Section 4 de- scribes in detail the semantic role labeling com- ponent. Section 5 presents evaluation results and …

Content truncated to 500 words