Hybrid vector space model for flexible voice search

This paper addresses incorporation of semantic analysis into information retrieval (IR) based on the vector space model (VSM) for flexible matching of spontaneous queries in a voice search system. Information of semantic slots or concepts that correspond to database fields is expected to help enhancing IR, but the semantic analyzer often fails or needs a large amount of training data. We propose a hybrid model which combines dedicated VSMs for concept slots with a general VSM as a back-off. The model has been evaluated in a book search task and shown to be effective and robust against ASR and SLU errors.