A very important issue for the Natural Language Processing (NLP) in the Speech Understanding perspective is the processing of input utterances that deviate from their correct expectations. It is possible to find the Ill-Formedness problem at every level of the Speech Processing: in the phonetic transcription of the signal, in the lexical interpretation, in the syntactical and semantical parsing. The main way to represent the Ill-Formedness is creating sets of hypotheses which give temporal and grammatical interpretations of the input: segment lattice, word lattice, syntactical chart. At the properly syntactical level, three main kinds of action to process the uncertainty and the incompleteness are possible: hypothesis scoring, island driven chart parsing, top-down prediction. In this paper, we present simulation test results in terms of word accuracy - 85% of word accuracy if the simulated phonemic transcription contains 10% of substitutions, deletions and insertions.
Bibliographic reference. Carlo, Andrea Di / Falcone, Rino (1991): "Ill-formedness problem in the spoken language processing", In EUROSPEECH-1991, 749-752.