Childbot: A Conversational Assistant for Childcare

Abstract

In this paper, we used the RNN Encoder-Decoder (seq2seq) model in Korean conversational robot (Konvbot). As a first step of being a general conversational model, we restricted our domain to childcare situation which includes six circumstances: 1) waking up, 2) morning exercise, 3) having breakfast, 4) taking a shower, 5) wearing clothes, and 6) going to school. We collected about 10,000 dialogue pairs for this scenario from more than 30 people. With the data, we implemented the base conversational model, which is used for collecting more dialogues in real environments (e.g. lab tour, conference demo). We present experiments in our expected scenario as well as general conversations which are out-of-script, and finally, real conversation with children. The result showed that our model could catch slight different expressions in the similar context, but it can cover only specific domain and has low vocabulary due to the small amount of training data. With more real experiments, we can collect data from experimenters, the better conversation the model will generate. Keywords—chatbot; conversational model; seq2seq;

1 Figure or Table

Cite this paper

@inproceedings{Jo2016ChildbotAC, title={Childbot: A Conversational Assistant for Childcare}, author={Hwiyeol Jo and Dong-Sig Han and Woo-Young Kang and Byoung-Tak Zhang}, year={2016} }