Acquisition of Word-meaning and Dialogue Strategies Aiming at the Communication in a Symbiotic Society of Human and Agents

在不久的将来，人类和智能体的共生社会将使他们能够进行真实的互动。为了加速相关研究，口语对话互动的设施和开发工具是非常重要的。我们首先研究了：（a）图像对象的词义获取，（B）智能体的对话策略获取：在词义获取中，提出了一种基于Online-EM算法的方法，该算法能够使智能体找出由口语单词表示的对象的特征，但是该算法需要大量的例子来确定词义或正确的分布。为了克服这个问题，我们应用了两种类型的学习偏见，形状偏见和互斥偏见，观察儿童的语言发展。实验结果表明，改进后的方法可以有效地获取词义。在对话策略获取中，我们提出了一种基于Q学习的算法，智能体根据自己的角色（提问或教学），通过对方的面部表情和话语来估计对方的理解水平，从而获得不同的策略。实验结果表明，教与问两种策略都是有效的。我们还开发了一个具有上述词义习得和对话策略习得功能的工具。合作作业的实验结果表明，获得的行动策略，以快速完成任务的有效性。

The symbiotic society of human and agents that enables them real interaction in the near future is expected. To accelerate related studies, facilities of spoken dialogue interaction and development tools are very important. We have studied first: (a) word-meaning acquisition of image objects, (b) dialogue strategy acquisition for the agents.In the word-meaning acquisition, we propose a method based on an Online-EM algorithm that enables agents to find out the features of objects represented by spoken words, however, the algorithm needs a lot of examples to determine the word-meaning, or correct distributions. To overcome the problem, we have applied two types of learning biases, shape bias and mutual exclusivity bias, observed in children's language development. Experimental results showed that the improved method with biases could efficiently acquire word meanings.In the dialogue strategy acquisition, we propose a Q-learning based algorithm in which agents acquire different strategies according with their roles, asking or teaching, by estimating counterpart's comprehension level from its facial expression and utterance. Experimental results showed the effectiveness of both strategies of teaching and asking. We have also developed a tool with the abovementioned functionalities of word-meaning acquisition and dialogue strategy acquisition. The experimental results on cooperative works showed the effectiveness of acquired action strategies to complete tasks quickly.