| 06 |
Real dialogue data tells how people get closeChat collection for developing long-term chat model 
|
|---|
Demand is growing for long-term chat data with which chatbots can be developed that gradually become acquainted with a user in a manner that resembles human interaction. In this study, we collected long-term text chat data over eight weeks by recording text chats between speakers who met for the first time. These pairs gradually became more acquainted with each other through repeated interactions. Conventional long-term chat data are acting-based data (i.e. simulated data) crafted by workers in virtual speaker settings. Our comparative analysis showed that the actual data collected in this study are more natural in terms of speech level and dialogue acts than conventional acting-based chat data. Our future challenge is developing a long-term chat dialogue model using these data to maintain consistency with a dialogue's history and to reflect the establishment of ongoing speaker relationships.
[1] T. Arimoto, H. Sugiyama, H. Narimatsu, M. Mizukami, “Comparison of the Intimacy Process between Real and Acting-based Long-term Text Chats,” in Proc. The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024.
Tsunehiro Arimoto
Interaction Research Group, Innovative Communication Laboratory