Identifying Semantic Roles Using Combinatory Categorial Grammar

Identifying Semantic Roles Using Combinatory Categorial Grammar Daniel Gildea and Julia Hockenmaier University of Pennsylvania {dgildea,juliahr}@cis.upenn.edu Abstract We present a system for automatically identifying PropBank-style semantic roles based on the output of a statistical parser for Combinatory Categorial Grammar. This system performs at least as well as a system based on a traditional Treebank parser, and outperforms it on core argu- ment roles. 1 Introduction Correctly identifying the semantic roles of sentence constituents is a crucial part of interpreting text, and in addition to forming an important part of the infor- mation extraction problem, can serve as an interme- diate step in machine translation or automatic sum- marization. Even for a single predicate, semantic arguments can have multiple syntactic realizations, as shown by the following paraphrases: (1) John will meet with Mary. John will meet Mary. John and Mary will meet. (2) The door opened. Mary opened the door. Recently, attention has turned to creating cor- pora annotated with argument structures. The PropBank (Kingsbury and Palmer, 2002) and the FrameNet (Baker et al., 1998) projects both doc- ument the variation in syntactic realization of the arguments of predicates in general English text. Gildea and Palmer (2002) developed a system to predict semantic roles (as defined in PropBank) from sentences and their parse trees as determined by the statistical parser of Collins (1999). In this paper, we examine how the syntactic representations used by different statistical parsers affect the performance of such a system. We compare a parser based on Combinatory Categorial Grammar (CCG) (Hocken- maier and Steedman, 2002b) with the Collins parser. As the CCG parser is trained and tested on a cor- pus of CCG derivations that have been obtained by automatic conversion from the Penn Treebank, we are able to compare performance using both gold- standard and automatic parses for both CCG and the traditional Treebank representation. The Treebank- parser returns skeletal phrase-structure trees with- out the traces or functional tags in the original Penn Treebank, whereas the CCG parser returns word- word dependencies that correspond to the under- lying predicate-argument structure, including long- range dependencies arising through control, raising, extraction and coordination. 2 Predicate-argument relations in PropBank The Proposition Bank (Kingsbury and Palmer, 2002) provides a human-annotated corpus of semantic verb-argument relations. For each verb appearing in the corpus, a set of semantic roles is defined. Roles for each verb are simply numbered Arg0, Arg1, Arg2, etc. As an example, the entry- specific roles for the verb offer are given below: Arg0 entity offering Arg1 commodity Arg2 price Arg3 benefactive or entity offered to These roles are then annotated for every instance of the verb appearing in the corpus, including the following examples: (3) [ARG0 the company] to offer [ARG1 a 15% stake] to [ARG2 the public]. (4) [ARG0 Sotheby’s] ... offered [ARG2 the Dorrance heirs] [ARG1 a money-back guarantee] (5) [ARG1 an amendment] offered by [ARG0 Rep. Peter DeFazio] (6) [ARG2 Subcontractors] will be offered [ARG1 a settlement] A variety of additional roles are assumed to …

Content truncated to 500 words