A robot that can engage in both task-oriented and non-task-oriented dialogues

This paper presents a new type of conversational humanoid robot, which can engage in both task-oriented dialogues and non-task-oriented dialogues. Most previously built conversational robots can engage in either task-oriented dialogues for accurately understanding human requests or non-task-oriented dialogues to allow humans to enjoy conversations. Since both are beneficial functionalities for a humanoid robot as a human partner, it is desirable for one humanoid robot to be able to engage in both types of dialogues. Our model is based on a multiexpert model, which features control modules called experts each or which is specialized to perform certain kinds of tasks through performing physical actions and engaging in dialogues. One of the experts takes charge in understanding human utterances and deciding robot utterances or actions. Non-task-oriented dialogue functionality is incorporated into this model by building an expert called the chat expert which is dedicated to non-task-oriented dialogues. The chat expert utilizes the outputs of a large-vocabulary speech recognizer, while other task-oriented experts utilize the outputs of a small-vocabulary speech recognizer. By selecting an appropriate expert according to the speech recognition result and dialogue context, we can alleviate degradation in speech recognition accuracy in spite of incorporating a large-vocabulary speech recognizer. The chat expert is dealt with as the default expert that has the responsibility to reply to human utterances. If a human utterance is considered to be a request for a task with a high plausibility, the expert for understanding the request is selected. The implemented system, which has been combined with Honda ASIMO, demonstrates that it can dynamically change dialogue strategy based on speech recognition results.