FrameNet-based Semantic Parsing using Maximum Entropy Models

FrameNet-based Semantic Parsing using Maximum Entropy Models Namhee Kwon Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 nkwon@isi.edu Michael Fleischman Messachusetts Institute of Technology, 77 Massachusetts Ave Cambridge, MA 02139 mbf@mit.edu Eduard Hovy Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 hovy@isi.edu Abstract As part of its description of lexico-semantic predicate frames or conceptual structures, the FrameNet project defines a set of semantic roles specific to the core predicate of a sentence. Recently, researchers have tried to automatically produce semantic interpretations of sentences using this information. Building on prior work, we describe a new method to perform such interpretations. We define sentence segmentation first and show how Maximum Entropy re-ranking helps achieve a level of 76.2% F-score (answer among top- five candidates) or 61.5% (correct answer). 1 Introduction To produce a semantic analysis has long been a goal of Computational Linguistics. To do so, however, requires a representation of the semantics of each predicate. Since each predicate may have a particular collection of semantic roles (agent, theme, etc.) the first priority is to build a collection of predicate senses with their associated role frames. This task is being performed in the FrameNet project based on frame semantics (Fillmore, 1976). Each frame contains a principal lexical item as the target predicate and associated frame-specific roles, such as offender and buyer, called frame elements. FrameNet I contains 1,462 distinct predicates (927 verbs, 339 nouns, 175 adjectives) in 49,000 annotated sentences with 99,000 annotated frame elements. Given these, it would be interesting to attempt an automatic sentence interpretation. We build semantic parsing based on FrameNet, treating it as a classification problem. We split the problem into three parts: sentence segmentation, frame element identification for each segment, and semantic role tagging for each frame element. In this paper, we provide a pipeline framework of these three phases, followed by a step of re-ranking from n-best lists of every phase for the final output. All classification and re-ranking are performed by Maximum Entropy. The top-five final outputs provide an F-score of 76.2% for the correct frame element identification and semantic role tagging. The performance of the single best output is 61.5% F-score. The rest of the paper is organized as follows: we review related work in Section 2, explain Maximum Entropy in Section 3, describe the detailed method in Section 4, show the re-ranking process in Section 5, and conclude in Section 6. 2 Related Work The first work using FrameNet for semantic parsing was done by Gildea and Jurafsky (G & J, 2002) using conditional probabilistic models. They divide the problem into two sub-tasks: frame element identification and frame element classification. Frame element identification identifies the frame element boundaries in a sentence, and frame element classification classifies each frame element into its appropriate semantic role. The basic assumption is that the frame element (FE) boundaries match the parse constituents, and both identification and classification are then done for each constituent1. …

Content truncated to 500 words