Using String-Kernels for Learning Semantic Parsers

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 913–920, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Using String-Kernels for Learning Semantic Parsers Rohit J. Kate Department of Computer Sciences The University of Texas at Austin 1 University Station C0500 Austin, TX 78712-0233, USA rjkate@cs.utexas.edu Raymond J. Mooney Department of Computer Sciences The University of Texas at Austin 1 University Station C0500 Austin, TX 78712-0233, USA mooney@cs.utexas.edu Abstract We present a new approach for mapping natural language sentences to their for- mal meaning representations using string- kernel-based classifiers. Our system learns these classifiers for every production in the formal language grammar. Meaning repre- sentations for novel natural language sen- tences are obtained by finding the most probable semantic parse using these string classifiers. Our experiments on two real- world data sets show that this approach compares favorably to other existing sys- tems and is particularly robust to noise. 1 Introduction Computational systems that learn to transform nat- ural language sentences into formal meaning rep- resentations have important practical applications in enabling user-friendly natural language com- munication with computers. However, most of the research in natural language processing (NLP) has been focused on lower-level tasks like syntactic parsing, word-sense disambiguation, information extraction etc. In this paper, we have considered the important task of doing deep semantic parsing to map sentences into their computer-executable meaning representations. Previous work on learning semantic parsers either employ rule-based algorithms (Tang and Mooney, 2001; Kate et al., 2005), or use sta- tistical feature-based methods (Ge and Mooney, 2005; Zettlemoyer and Collins, 2005; Wong and Mooney, 2006). In this paper, we present a novel kernel-based statistical method for learn- ing semantic parsers. Kernel methods (Cristianini and Shawe-Taylor, 2000) are particularly suitable for semantic parsing because it involves mapping phrases of natural language (NL) sentences to se- mantic concepts in a meaning representation lan- guage (MRL). Given that natural languages are so flexible, there are various ways in which one can express the same semantic concept. It is difficult for rule-based methods or even statistical feature- based methods to capture the full range of NL con- texts which map to a semantic concept because they tend to enumerate these contexts. In contrast, kernel methods allow a convenient mechanism to implicitly work with a potentially infinite number of features which can robustly capture these range of contexts even when the data is noisy. Our system, KRISP (Kernel-based Robust In- terpretation for Semantic Parsing), takes NL sen- tences paired with their formal meaning represen- tations as training data. The productions of the for- mal MRL grammar are treated like semantic con- cepts. For each of these productions, a Support- Vector Machine (SVM) (Cristianini and Shawe- Taylor, 2000) classifier is trained using string sim- ilarity as the kernel (Lodhi et al., 2002). Each classifier then estimates the probability of the pro- duction covering different substrings of the sen- tence. This information is used to compositionally build a complete meaning representation (MR) of …

Content truncated to 500 words