Statistical methods in natural language understanding and spoken dialogue systems

摘要: Modern automatic spoken dialogue systems cover a wide range of applications. There are for hotel reservations, restaurant guides, travel and timetable information, as well telephone-banking services. Building the different components system combining them in an optimal way such that reasonable becomes possible is complex task because during course dialogue, has to deal with uncertain information. In this thesis, we use statistical methods model combine system’s components. Statistical provide well-founded theory modeling where decisions have be made under uncertainty. Starting from Bayes’ decision rule, define evaluate various models these components, which comprise speech recognition, natural language understanding, management. The problem understanding described special machine translation source sentence translated into formal target consisting concepts. For this, two models. first generative based on source-channel paradigm. Because word context plays important role tasks, phrasebased order take local dependencies account. second direct maximum entropy framework works similar tagger. model, several feature functions capture between words Both advantage only source-target pairs form input-output sentences must provided training. Thus, there no need generate grammars manually, significantly reduces costs building new domains. Furthermore, propose investigate minimum error rate training results tighter coupling recognition understanding. This allows easy integration multiple knowledge sources by minimizing overall criterion. it add features thus minimize rate, or slot rate. Finally, develop task-independent manager using trees fundamental data structure. Based cost function, chooses next action minimal costs. design task-independence leads strict separation given application operations performed manager, simplifies porting existing domain. We report field test was able choose 90% decisions. techniques handling confidence measures defined performance when incorporated strategy. Experiments been carried out TelDir database, German in-house telephone directory assistance corpus, Taba train time scheduling task.