Wide-coverage parsing of speech transcripts

Abstract

This paper discusses the performance difference of wide-coverage parsers on small-domain speech transcripts. Two parsers (C&C CCG and RASP) are tested on the speech transcripts of two different domains (parent-child language, and picture descriptions). The performance difference between the domain-independent parsers and two domain-trained parsers (MSTParser and MEGRASP) is substantial, with a difference of at least 30 percent point in accuracy. Despite this gap, some of the grammatical relations can still be recovered reliably.