Studies on Robust Language and Dialogue Processing for Spoken Dialogue Systems

In spoken dialogue systems, robust language processing for spontaneous speech understanding and robust dialogue processing for achieving user goal are inevitable. Previously, research of speech recognition and research of natural language understanding were done independently. At first glance, it seems to be no problem to combine these two technologies, because the purpose of speech recognition is to identify spoken object (word or sentence) from the input speech, and one of the purposes of the natural language understanding is to make some semantic representation from the sentence. However, the demonstration dialogue system which simply concatenate these two technologies did not work well if the input is not correct utterance, or the recognition error occurs. In this thesis, we describe robust language processing and robust dialogue processing for spoken dialogue systeJ;llS. First, we explain the robust language processing method, that is, keyword-driven parser, which uses path analysis of the semantic network. The keyword-driven parser can generate partial semantic representation toward the noisy input. Next, we propose a cognitive process model of spoken dialogue, which specifies the cognitive process of whole dialogue understanding process, the interaction between understanding process and dialogue management process, and the recovering method from input errors into this cognitive process. Such robustness should be evaluated in pseudo real environment for spoken dialogue systems, that is, interactive environment under communication errors. We implemented system-to-system dialogue evaluation environment with linguistic noise, and showed the effectiveness of proposed dialogue model. In chapter 1, we clarify our position with describing general framework of spoken dialogue systems and describing the previous approach to robust language / dialogue processing. In chapter 2, we present a keyword-driven speech parser as a robust language processing. Generally speaking, previous robust parsing methods are divided into two groups: grammar-based approach (it is a kind of theory-based technique) that generates all possible