8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Estimation of Semantic Confidences on Lattice Hierarchies

Robert Lieb, Tibor Fabian, Guenther Ruske, Matthias Thomae

Technische Universität München, Germany

Inspired by the well-known method for confidence measure calculation via estimation of word posterior probabilities on the word graph, we devised a technique to estimate confidences on all levels of the hierarchically structured output of our one-stage decoder for interpretation of natural speech (ODINS). By constructing a nested lattice hierarchy, the generalized counterpart of the word graph, we estimate posterior probabilities for all nodes in the decoded semantic tree, namely for all contained semantic units and words. The obtained experimental results show that the tree node confidence measure performs significantly better than the confidence error base line, no matter if the evaluation is carried out on tree nodes representing semantic concepts, word classes, or words. Furthermore, the paper proposes possible applications of the tree node confidences to improve the grounding strategy of spoken dialog systems.

Full Paper

Bibliographic reference.  Lieb, Robert / Fabian, Tibor / Ruske, Guenther / Thomae, Matthias (2004): "Estimation of semantic confidences on lattice hierarchies", In INTERSPEECH-2004, 569-572.