Learning of dialogue states and language model of spoken information system

摘要

In this invention dialogue states for a dialogue model are created using a training corpus of example human-human dialogues. Dialogue states are modelled at the turn level rather than at the move level, and the dialogue states are derived from the training corpus. ;The range of operator dialogue utterances is actually quite small in many services and therefore may be categorized into a set of predetermined meanings. This is an important assumption which is not true of general conversation, but is often true of conversations between telephone operators and people. ;Phrases are specified which have specific substitution and deletion penalties, for example the two phrases ‘I would like to’ and ‘can I’ may be specified as a possible substitution with low or zero penalty. Thus allows common equivalent phrases are given low substitution penalties. Insignificant phrases such as ‘erm’ are given low or zero deletion penalties.