Using phone features to improve dialogue state tracking generalisation to unseen states.

The generalisation of dialogue state trackingto unseen dialogue states can be verychallenging. In a slot-based dialogue system,dialogue states lie in discrete spacewhere distances between states cannot becomputed. Therefore, the model parametersto track states unseen in the trainingdata can only be estimated from more generalstatistics, under the assumption thatevery dialogue state will have the same underlyingstate tracking behaviour. However,this assumption is not valid. For example,two values, whose associated conceptshave different ASR accuracy, mayhave different state tracking performance.Therefore, if the ASR performance of theconcepts related to each value can be estimated,such estimates can be used as generalfeatures. The features will help to relateunseen dialogue states to states seenin the training data with similar ASR performance.Furthermore, if two phoneticallysimilar concepts have similar ASRperformance, the features extracted fromthe phonetic structure of the concepts canbe used to improve generalisation. Inthis paper, ASR and phonetic structurerelatedfeatures are used to improve thedialogue state tracking generalisation tounseen states of an environmental controlsystem developed for dysarthric speakers.