Dialogue Breakdown Detection with Long Short Term Memory
Tittaya Mairittha, Tsuyoshi Okita, Sozo Inoue,
EAI International Conference on Mobile Computing, Applications and Services (MobiCASE) poster
This paper aims to detect the utterance which can be categorized as the breakdown of the dialogue flow. We propose a logistic regression-based and a Long Short-Term Memory (LSTM)-based methods. Using the input with utterance-response pairs the performance of the LSTM-based method is superior to that of the logistic regression-based method in 36% measured with F-measure. We also measured the performance using the performance with utterance-response pairs: the performance with the input only with responses is unexpectedly inferior to those with responses in 6% to 23% measured with F-measure.