Use of Deep Linguistic Features for the Recognition and Labeling of Semantic Arguments

Use of Deep Linguistic Features for the Recognition and Labeling of Semantic Arguments John Chen Department of Computer Science Columbia University New York, NY 10027 jchen@cs.columbia.edu Owen Rambow Department of Computer Science Columbia University New York, NY 10027 rambow@cs.columbia.edu Abstract We use deep linguistic features to predict semantic roles on syntactic arguments, and show that these perform considerably better than surface-oriented features. We also show that predicting labels from a “lightweight” parser that generates deep syntactic features performs comparably to using a full parser that generates only sur- face syntactic features. 1 Introduction Syntax mediates between surface word order and meaning. The goal of parsing (syntactic analysis) is ultimately to provide the first step towards giv- ing a semantic interpretation of a string of words. So far, attention has focused on parsing, because the semantically annotated corpora required for learn- ing semantic interpretation have not been available. The completion of the first phase of the PropBank (Kingsbury et al., 2002) represents an important step. The PropBank superimposes an annotation of semantic predicate-argument structures on top of the Penn Treebank (PTB) (Marcus et al., 1993; Marcus et al., 1994). The arc labels chosen for the arguments are specific to the predicate, not universal. In this paper, we find that the use of deep lin- guistic representations to predict these semantic la- bels are more effective than the generally more surface-syntax representations previously employed (Gildea and Palmer (2002)). Specifically, we show that the syntactic dependency structure that results load arg0 John arg2 hay arg1 truck Figure 1: PropBank-style semantic representation for both John loaded the truck with hay and John loaded hay into the truck from the extraction of a Tree Adjoining Grammar (TAG) from the PTB, and the features that accom- pany this structure, form a better basis for determin- ing semantic role labels. Crucially, the same struc- ture is also produced when parsing with TAG. We suggest that the syntactic representation chosen in the PTB is less well suited for semantic process- ing than the other, deeper syntactic representations. In fact, this deeper representation expresses syntac- tic notions that have achieved a wide acceptance across linguistic frameworks, unlike the very partic- ular surface-syntactic choices made by the linguists who created the PTB syntactic annotation rules. The outline of this paper is as follows. In Sec- tion 2 we introduce the PropBank and describe the problem of predicting semantic tags. Section 3 presents an overview of our work and distinguishes it from previous work. Section 4 describes the method used to produce the TAGs that are the basis of our experiments. Section 5 specifies how train- ing and test data that are used in our experiments are derived from the PropBank. Next, we give re- sults on two sets of experiments. Those that predict semantic tags given gold-standard linguistic infor- mation are described in Section 6. Those that do prediction from raw text are described in Section 7. Finally, in Section 8 we present concluding remarks. 2 The PropBank and the …

Content truncated to 500 words