Can Semantic Roles Generalize Across Genres?

Proceedings of NAACL HLT 2007, pages 548–555, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Can Semantic Roles Generalize Across Genres? Szu-ting Yi Dept of Computer Science University of Pennsylvania Philadelphia, PA 19104 Edward Loper Dept of Computer Science University of Pennsylvania Philadelphia, PA 19104 Martha Palmer Dept of Computer Science University of Colorado at Boulder Boulder, CO 80309 Abstract PropBank has been widely used as train- ing data for Semantic Role Labeling. However, because this training data is taken from the WSJ, the resulting machine learning models tend to overfit on idiosyn- crasies of that text’s style, and do not port well to other genres. In addition, since PropBank was designed on a verb-by-verb basis, the argument labels Arg2 - Arg5 get used for very diverse argument roles with inconsistent training instances. For exam- ple, the verb “make” uses Arg2 for the “Material” argument; but the verb “multi- ply” uses Arg2 for the “Extent” argument. As a result, it can be difficult for auto- matic classifiers to learn to distinguish ar- guments Arg2-Arg5. We have created a mapping between PropBank and VerbNet that provides a VerbNet thematic role la- bel for each verb-specific PropBank label. Since VerbNet uses argument labels that are more consistent across verbs, we are able to demonstrate that these new labels are easier to learn. 1 Introduction Correctly identifying semantic entities and success- fully disambiguating the relations between them and their predicates is an important and necessary step for successful natural language processing applica- tions, such as text summarization, question answer- ing, and machine translation. For example, in or- der to determine that question (1a) is answered by sentence (1b), but not by sentence (1c), we must de- termine the relationships between the relevant verbs (eat and feed) and their arguments. (1) a. What do lobsters like to eat? b. Recent studies have shown that lobsters pri- marily feed on live fish, dig for clams, sea urchins, and feed on algae and eel-grass. c. In the early 20th century, Mainers would only eat lobsters because the fish they caught was too valuable to eat themselves. An important part of this task is Semantic Role Labeling (SRL), where the goal is to locate the con- stituents which are arguments of a given verb, and to assign them appropriate semantic roles that describe how they relate to the verb. Many researchers have investigated applying machine learning to corpus specifically annotated with this task in mind, Prop- Bank, since 2000 (Chen and Rambow, 2003; Gildea and Hockenmaier, 2003; Hacioglu et al., 2003; Mos- chitti, 2004; Yi and Palmer, 2004; Pradhan et al., 2005b; Punyakanok et al., 2005; Toutanova et al., 2005). For two years, the CoNLL workshop has made this problem the shared task (Carreras and M´arquez, 2005). However, there is still little con- sensus in the linguistic and NLP communities about what set of role labels are most appropriate. The Proposition Bank (PropBank) corpus (Palmer et al., 2005) avoids this issue by using theory-agnostic la- bels …

Content truncated to 500 words