The generalisation of dialogue state trackingudto unseen dialogue states can be veryudchallenging. In a slot-based dialogue system,uddialogue states lie in discrete spaceudwhere distances between states cannot beudcomputed. Therefore, the model parametersudto track states unseen in the traininguddata can only be estimated from more generaludstatistics, under the assumption thatudevery dialogue state will have the same underlyingudstate tracking behaviour. However,udthis assumption is not valid. For example,udtwo values, whose associated conceptsudhave different ASR accuracy, mayudhave different state tracking performance.udTherefore, if the ASR performance of theudconcepts related to each value can be estimated,udsuch estimates can be used as generaludfeatures. The features will help to relateudunseen dialogue states to states seenudin the training data with similar ASR performance.udFurthermore, if two phoneticallyudsimilar concepts have similar ASRudperformance, the features extracted fromudthe phonetic structure of the concepts canudbe used to improve generalisation. Inudthis paper, ASR and phonetic structurerelatedudfeatures are used to improve theuddialogue state tracking generalisation toudunseen states of an environmental controludsystem developed for dysarthric speakers.ud
展开▼