ParsingSpeechRepairwithoutSpecializedGrammarSymbols.PDFVIP

Parsing Speech Repair without Specialized Grammar Symbols? Tim Miller Luan Nguyen William Schuler University of Minnesota University of Minnesota University of Minnesota tmill@cs.umn.edu lnguyen@cs.umn.edu schuler@cs.umn.edu Abstract same problem, by completely separating the ?uent grammar from the operations of the parser. The This paper describes a parsing model for grammar thus has no representation of dis?uency speech with repairs that makes a clear sep- or speech repair, such as the “EDITED” category aration between linguistically meaningful used to represent a reparandum in the Switchboard symbols in the grammar and operations corpus, as such categories are seemingly at odds speci?c to speech repair in the operation of with the typical nature of a linguistic constituent. the parser. This system builds a model of Rather, the approach presented here uses a how un?nished constituents in speech re- grammar that explicitly represents incomplete pairs are likely to ?nish, and ?nishes them constituents being processed, and repair is rep- probabilistically with placeholder struc- resented by rules which allow incomplete con- ture. These modi?ed repair constituents stituents to be prematurely merged with existing and the restarted replacement constituent structure. While this model is interesting for its are th