Japanese / English Ishida Lab. > Learning Conversational Agent

Learning Conversational Agent


We are developing conversational agents which guide the visitors of tour spots. There are the following two research themes for these agents:

Dialogue Model
Figure 1 Dialogue Model

1. Incremental Learning Algorithms for Dialogue Models

We can not acquire a full dialogue model at first because it is hard to collect dialogues on a specific task. Therefore, a dialogue model needs to be learned every time the number of example dialogues increases. However, re-learning for a little increase is very costful. Incremental learning algorithms are desirable for this situation.
In this theme, we use probabilistic DFAs (PDFAs) for dialogue models. We propose and evaluate incremental algorithms for learning PDFAs.
WOZ Interface
Figure 2 WOZ Interface

2. Tour Guide Agent on Digital City by Wizard of Oz Method

We develop a tour-guide agent which aquires the ability of conversation gradually through collecting example dialogues and learning dialogue models. The agent is constructed through the cycle of the following two steps:

  • Collecting example dialogues with Wizard of Oz method
  • Learning dialogue models from example dialogues

In this theme, we construct the Kyoto tour-guide agent with the method.


Masayuki OKAMOTO (Kyoto University) okamoto@kuis.kyoto-u.ac.jp

Nobutoshi YAMANAKA (Kyoto University) nobutosi@kuis.kyoto-u.ac.jp





Ishida Laboratory, Department of Social Informatics, Kyoto University

Digital City Research Center, JST

Internal Pages (password required)