Intention Understanding with Small Training Data Sets by Utilizing Multi-Task Transfer Learning

Abstract

In this research, we propose intention understanding method utilizing multi-task transfer learning. Our method improves intention understanding accuracy using data of different kind of domain as source domain. As source domain's training data, we use Japanese-English translation data (translation task) and Japanese Wikipedia data (sentence prediction task). As target domain's training data, we use transcribed utterance data of voice control of equipment. In this data, each utterance has one intention label. As an experimental result, we found that proposed method provides a performance improvement over previous transfer learning method in the case of small training data (the number of data for each intention label are 1, 3, 5, 10 and 30).