Exploiting unlabeled data and lexical/ ontological structure for frame-semantic parsing

Semantic parsing is the challenge of taking a sentence as input and generating a structured output, ideally a lexically and syntactically neutral representation of the semantics of the sentence. One common computational representation of semantics uses predicate logic; another makes use of predicate-argument structures defined with respect to some lexicon. The latter formulation of the problem is addressed here. English FrameNet (http://framenet.icsi.berkeley.edu) is a lexical resource prepared in the course of a major linguistic annotation project at Berkeley. FrameNet encompasses three forms of data: • A structured lexicon of frames, which are complex concepts (typically events); associated lexical units, linguistic words and phrases that evoke a particular frame; and frame elements or roles, which constitute the participants or attributes constituting the internal structure of a frame. Frames and their roles exist in an inheritance hierarchy and are marked as related to each other in several different ways (similar to entries in WordNet). For the most part, the frames which were selected for inclusion in the current version of FrameNet are frequent and domain-general in English. • Hand-selected exemplar sentences from a corpus which illustrate a particular use of a frame; these are annotated to indicate the frame-evoking word and any arguments to its roles. These are generally annotated for just a single frame, though they may also involve concepts represented by other