We propose in this paper to use a novel multiview boosting- like algorithm called Mumbo for processing human-human con- versations recorded in a call-senter setting. We present how dia- log classification can be seen as a multiview classification prob- lem and we compare the performance of Mumbo and the one of a standard boosting algorithm. The first results obtained on a subset of the DECODA corpus show that a significant improve- ment in classification performance can be achieved, especially on high Word Error Rate transcriptions.

Index Terms: human-human conversation, multiview learning, boosting, spoken language understanding