Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Natural Utterance Segmentation and Discourse Label Assignment

Mutsuko Tomokiyo

ATR Interpreting Telecommunications Research Laboratories, Kyoto, Japan

This paper proposes an approach based on pragmatics of spontaneously-spoken Japanese dialogue. Input to the system presented here appears as an unbroken stream of "utterance" i.e, a sequence of clauses or clause fragments separated by particles and other transitional elements, uttered by one speaker. To cope with data of this sort, a three-step procedure is employed. The aim of the present paper is motivate and discribe step 1 and 2 of this three-step process. As step 1, the input stream is automatically segmented into discourse units and in step 2, discourse labels are assigned. Test results of 1742 utterances show that the procedures presented here effected discourse segmentation correctly on average of 95.32% of the time, and they assigned discourse labels correctly on average of 85.12% of the time.

Full Paper

Bibliographic reference.  Tomokiyo, Mutsuko (1994): "Natural utterance segmentation and discourse label assignment", In ICSLP-1994, 1671-1674.