Maximum Entropy Models for FrameNet Classification

Maximum Entropy Models for FrameNet Classification Michael Fleischman, Namhee Kwon and Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 {fleisch, nkwon, hovy }@ISI.edu Abstract The development of FrameNet, a large database of semantically annotated sen- tences, has primed research into statistical methods for semantic tagging. We ad- vance previous work by adopting a Maximum Entropy approach and by using previous tag information to find the high- est probability tag sequence for a given sentence. Further we examine the use of sentence level syntactic pattern features to increase performance. We analyze our strategy on both human annotated and automatically identified frame elements, and compare performance to previous work on identical test data. Experiments indicate a statistically significant im- provement (p<0.01) of over 6%. 1 Introduction Recent work in the development of FrameNet, a large database of semantically annotated sentences, has laid the foundation for statistical approaches to the task of automatic semantic classification. The FrameNet project seeks to annotate a large subset of the British National Corpus with seman- tic information. Annotations are based on Frame Semantics (Fillmore, 1976), in which frames are defined as schematic representations of situations involving various frame elements such as partici- pants, props, and other conceptual roles. In each FrameNet sentence, a single target predicate is identified and all of its relevant frame elements are tagged with their semantic role (e.g., Agent, Judge), their syntactic phrase type (e.g., NP, PP), and their grammatical function (e.g., ex- ternal argument, object argument). Figure 1 shows an example of an annotated sentence and its appro- priate semantic frame. She clapped her hands in inspiration. Frame: Body-Movement Frame Elements: Agent Body Part Cause -NP -NP -PP -Ext -Obj -Comp Figure 1. Frame for lemma “clap” shown with three core frame elements and a sentence annotated with ele- ment type, phrase type, and grammatical function. As of its first release in June 2002, FrameNet has made available 49,000 annotated sentences. The release contains 99,000 annotated frame ele- ments for 1462 distinct lexical predicates (927 verbs, 339 nouns, and 175 adjectives). While considerable in scale, the FrameNet da- tabase does not yet approach the magnitude of re- sources available for other NLP tasks. Each target predicate, for example, has on average only 30 sen- tences tagged. This data sparsity makes the task of learning a semantic classifier formidable, and in- creases the importance of the modeling framework that is employed. 2 Related Work To our knowledge, Gildea and Jurafsky (2002) is the only work to use FrameNet to build a statis- tically based semantic classifier. They split the problem into two distinct sub-tasks: frame element identification and frame element classification. In the identification phase, syntactic information is extracted from a parse tree to learn the boundaries of the frame elements in a sentence. In the classi- fication phase, similar syntactic information is used to classify those elements into their semantic roles. In both phases Gildea and Jurafsky (2002) build a model of the conditional probabilities of …

Content truncated to 500 words