An information gain and grammar complexity based approach to attribute selection in speech enabled information retrieval dialogs

An effective dialog driven method is required for today´s speech enabled information retrieval systems, such as name dialers. Similar to the dynamic sales dialog for electronic commerce scenarios, information gain measure based approaches are widely used for attribute selection and dialog length reduction. However, for speech enabled information retrieval systems, another important factor influencing attribute selection is speech recognition accuracy. Too low accuracy results in a failed dialog. Recognition accuracy varies with many issues, including acoustic model performance and grammar complexity. The acoustic model is fixed for a whole dialog, while grammar is different for each interaction round, thereby grammar complexity influences the attribute selected for the next question. An approach combining both information gain measurement and grammar complexity is presented for a dynamic dialog driven system. Offline evaluations show that this approach can give a trade-off between the target of faster discrimination of the candidates for retrieval and higher recognition accuracy.