Identifying FrameNet Frames for Verbs from a Real-Text Corpus

Proceedings of the Australasian Language Technology Workshop 2005, pages 200–206, Sydney, Australia, December 2005. Identifying FrameNet Frames for Verbs from a Real-Text Corpus Matthew HONNIBAL and Tobias HAWKER Language Technology Research Group School of Information Technologies Madsen Building (F09) University of Sydney NSW 2006, Australia {mhonn,toby}@it.usyd.edu.au Abstract Previous systems that automatically tag text with FrameNet labels have been trained from the FrameNet example data, as there is no FrameNet annotated corpus. The FrameNet data is systemat- ically biased by the criteria for the examples’ se- lection, as annotators attempt to select simple sen- tences that include the target word. Instead of using the FrameNet examples, we train a maximum entropy model classifier to identify verb frames on text from the Penn Treebank. We use ex- amples of verbs with only one entry in FrameNet as training data, and evaluate the system on human annotated text from the Wall Street Journal. We ac- curately identify the frame used by 76% of finite verbs. We also investigate how well the system performs on verbs it has not encountered before. This task examines the feasibility of using the system to auto- matically extend the coverage of FrameNet by clas- sifying verbs with no FrameNet entries. The classi- fier accurately assigns a frame to 55% of instances of verbs it has not been trained on. 1 Introduction FrameNet (Ruppenhofer et al., 2005) is a lexical se- mantic database that categorises words into frames, and gives extensive examples of their use. A frame records the semantic type of a predicate and the se- mantic roles of its arguments. Usually, the predicate will be a verb, and its roles will be realised by con- stituents of its clause. Recently, Senseval-3 (Litowski, 2004) used FrameNet as the basis of a semantic role labelling shared task. Participants were asked to replicate the role labelling of frame elements in the FrameNet example data given the frame. However, a system which is trained and evaluated on the FrameNet ex- amples may not perform comparably on other text. FrameNet examples are selected by the annotators to illustrate the semantic and syntactic combinatory possibilities of each word, with minimal confusion from irrelevant performance variables and com- plicated syntactic constructions. They are there- fore a systematically skewed sample, with much of the complexity of natural text under-represented or missing entirely. The problem is that there is currently no FrameNet annotated corpus — only four manually annotated Wall Street Journal texts, recently re- leased on the FrameNet website. There is, how- ever, still a way to train a system for a subset of the FrameNet annotation task on real text without one. Frame selection is often unambiguous given a particular verb: currently, 73% of verbs entered in FrameNet are associated with only one frame. These verbs head 40% of the finite clauses in the Penn Treebank (Marcus et al., 1993). The verb senses associated with a particu- lar FrameNet frame share similar semantics and argument structures. For instance, the Activ- ity finish frame includes …

Content truncated to 500 words