Intention Understanding in Small Training Data Sets by Using Transfer Learning

Hideaki Joko, Yusuke Koji, Hayato Uchide, Takahiro Otsuka

Abstract: This research proposes an intention understanding method that uses transfer learning from Japanese-English translation data. It was found that the proposed method improved performance over the baseline method for small training data, with intention understanding accuracy improving by a maximum of 10.8 points when the number of data for each intention label was 1.