Projects per year
Understanding user utterances in human-computer spoken dialogue systems involves a multi-level pragmatic-semantic analysis of noisy natural language input streams. These analyses are heavily dependent on the dialogue context, and are complex due to the inherent ambiguity of language use, and to the errors induced by the intermediate speech recognition system. We review work on applying k-nearest-neighbour classification to this multi-level task split into (1) dialogue act classification, (2) slot filling identification, and (3) communication problem signalling, showing that co-learning some of these tasks produces superior results over learning them in isolation. We also show that additional feature selection can produce succinct feature sets, illustrating the viability of simple keyword-based shallow understanding.
|Title of host publication||Workshop Proceedings of the 6th International Conference on Case-Based Reasoning, August 2005|
|Place of Publication||Chicago, IL|
|Number of pages||10|
|Publication status||Published - 2005|