An algorithm of Identifying Semantic Arguments of a Verb From Structured Data

Proceedings of Recent Advances in Natural Language Processing, pages 568–573, Hissar, Bulgaria, 12-14 September 2011. An Algorithm of Identifying Semantic Arguments of a Verb from Structured Data Minhua Huang Department of Computer Science Graduate School and University Center City University of New York New York, U.S.A mhuang@gc.cuny.edu Robert M. Haralick Department of Computer Science Graduate School and University Center City University of New York New York, U.S.A haralick@aim.com Abstract We discuss a method for identifying se- mantic arguments of a verb from a sen- tence. It differs from existing methods by an unique feature that represents all se- mantic arguments of a verb in a syntactic parse tree. The feature is a path in which at least one of the children of a node is a root of a subtree that associates with a semantic argument. Experiments on WSJ data from Penn TreeBank and PropBank show that our method achieves an average of preci- sion 92.3% and an average of recall 94.2% on identifying semantic arguments of over six hundred verbs. 1 Introduction Semantic argument identification is one of the sub- tasks of semantic role labeling (Gildea and Juraf- sky, 2002) (Chen and Rambow, 2002) (Hacioglu, 2004b) which classifies a sequence of words as- sociated with a semantic argument of a verb but does not assign its role. It is the most difficult task in semantic role labeling. Moreover, it is one of the core techniques for a machine to understand the semantics of a sentence. For instance, in the sentences Lisa cut the ribbon with a pair of scis- sors. and The ribbon was cut by Lisa with a pair of scissors, cut is the verb. Semantic arguments of cut will be Lisa, the ribbon, and a pair of scis- sors, where Lisa is the one who performs the ac- tion of cutting, the ribbon is the material to be cut by Lisa, and a pair of scissors is the tool used for cutting. For semantic role labeling, arguments of cut need to be determined. Then, each argument will be assigned to a label, such as agent, theme, and instrument in the example. In this report, we presents an algorithm for finding a semantic argu- ment of a verb which is the first task required for assigning a role for the verb. Over the years, two approaches have been dis- cussed by researchers, such as methods developed based on hierarchical trees (Gildea and Jurafsky, 2002) (Hacioglu, 2004b) (Hacioglu, 2004c) and methods developed based on flat chunks (Hacioglu and Ward, 2003) (Hacioglu, 2004a). In almost all the methods of the first approach, a syntactic tree is transformed into a sequence of constituents. Each semantic argument of a verb is represented by a set of constituents. Each constituent is rep- resented by a set of features. These features are extracted based on linguistic knowledge and local knowledge of the tree structure. Finally, sophisti- cated classifiers such as support vector machines or maximum entropy modeling classifiers are em- ployed to identify semantic …

Content truncated to 500 words