Making the most of multiplicity: a multi-parser multi-strategy architecture for the robust processing of spoken language

This paper describes ongoing research on robust spoken language understanding in the context of the Verbmobil speech-to-speech machine translation project. We focus on recent developments in the processing steps which map a word lattice to a semantic representations. The approach described firstly applies speech repair correction to word lattices. Four analysis methods of varying depth are then applied in parallel to the normalized word lattices, producing output for sub-portions of the lattice in the same semantic description language, the VIT format. These fragmentary analyses are stored and combined by a further processing component, which finally selects a sequence of semantic representations as a result.