An example-based method for estimating attributes of out-of-vocabulary words for spoken dialogue systems

摘要

This paper proposes a new example-based method for estimating attributes of out-of-vocabulary (OOV) words in user utterances through natural conversation without explicit Q/A. This method is advantageous because there are many cases in which the system can continue a dialogue if only attributes of OOV words are acquired. We evaluated the proposed method using examples collected from dialogues between the weather information system and the user, and making the user utterance including OOV words by assuming a part of the system's understanding of the user utterance to be OOV words. The proposed method achieved attribute estimation accuracy of 69 %.