Recognition of topics from utterances in a spoken dialog system

Recognition of the topic from utterances is part of the discourse analysis in a spoken dialog system. At present, many spoken dialog systems take the key point on semantic surface, ignoring the context dependent interpretation of utterances. We present an approach to recognition of user´s intentions and topics through bottom-up and top-down analyses. It uses a transition network of user intentions and a tree of topics to describe the dialog history. And the experimental results show that the correct recognition rates were 83.8% for the topic and 81.3% for the user intention. It can provide more knowledge for system and therefore enhance the system performance greatly.