A Maximum Entropy model for modelling context in a spoken dialogue system

Rob Koeling (SRI-International Cambridge)


Last year I reported on work in progress on a dialogue system
developed within the NWO priority program OVIS. At that time I could
not present any proper results of the proposed approach to include
contextual knowledge, in this talk I will elaborate on it a bit
further. 

In the OVIS project a dialogue system is developed with which information about the railway schedule can be obtained via the telephone in spoken Dutch. A speech recogniser analyses the user utterance and returns a (possibly big) number of hypotheses to the natural language processing component. It is the task of the NLP component to find the best hypothesis. To reach this goal we try to exploit as many information sources as possible (e.g. linguistic, accoustic and statistical (n-gram)) An extra source of information that is investigated in this work is the (linguistic) context in which the user uttered the sentence.

In this talk I discuss experiments to model the relation between words in a wordgraph and contextual information by means of a statistical model (Maximum Entropy). I try to model dependencies between words in the user utterance and information like the type of the corresponding system question. I will describe the most successful experiments carried out, report on a significant improvement of the performance of the parser and discuss some pros and cons of the suggested approach.