The thoughtful elephant: strategies for spoken dialog systems

摘要: We present technology used in spoken dialog systems for applications of a wide range. They include tasks from the travel domain and automatic switchboards as well large scale directory assistance. The overall goal developing is to allow natural flexible flow similar human-human interaction. This imposes challenging task recognize interpret user input, where he/she allowed choose an unrestricted vocabulary infinite set possible formulations. therefore put emphasis on strategies that make system more robust while still maintaining high level naturalness flexibility. In view this paradigm, we found two fundamental principles characterize many proposed methods: consider available sources information early possible; keep alternative hypotheses delay decision single option long possible. describe how our architecture caters incorporating application specific knowledge, including, example, database constraints, determination best sentence hypothesis turn. On next higher level, use history assess plausibility by applying consistency checks with items previous turns. particular, demonstrate combination decisions over several turns can be exploited boost recognition performance system.