A tool to build a treebank for conversational Chinese

Yves Lepage, Nicolas Auclerc, Satoshi Shirai

Abstract: N-grams have been extensively used with phonemes or words as basic units in speech recognition. Recently, it has been proposed to use n-grams with phrase tree structures as units to increase speech recognition quality. In order to test this idea on Chinese, a treebank of Chinese hotel reservation con- versation utterances is needed. Because no such treebank is yet available, we have to build it. We propose to see the process of building a tree-bank as a sequence of edition and search operations: input or copy a new utterance (edit a text); search for similar existing utterances to get their corresponding structures and adapt them to the new utterance; adapt the structure (edit a tree) ; earch for similar structures to ensure representaion and coding consistency. This way of doing will have a benefic "snow-ball" effect: the bigger the treebank, the faster and the more consistent its extension.