Connectives Acquisition in a Humanoid Robot Based on an Inductive Learning Language Acquisition Model

Abstract

1.1 Background Humanoid robots, like the Honda’s ASIMO1, have been developed and such machines will be expected to help people to do a variety of tasks in every day life. Then robots that exist in the same dynamic environment as humans should be able to interact with humans using natural language. However, when it comes to developing robots that understand language and work in dynamic environment, there are two problems to be solved in usual methods of the natural language processing. One problem is that of the language grounding which was also pointed out by Roy (Roy, 2003). In natural language processing systems, meanings of words are defined by other words circularly. That is, the words are not connected to objects, movements, colors or other physical features in the real world. However, connecting words to the real world is needed when a robot is to perform a concrete action following a human’s utterance. For example, when a person orders a robot to “kick the red ball”, if the word “red” is not connected to a specific color representation in accordance with the real world then the robot cannot realize the order. Another problem is that the top-down approach, which is very common in natural language processing systems, cannot deal with a dynamic environment. For example, SHRDLU (Winograd, 1972) is a system that connects words to a virtual world where some blocks exist. This system can understand utterances that refer to the virtual world such as “put the red cylinder on the green box” and can perform such actions when told, because the designer describes knowledge about the limited and static virtual world to the system in advance. On the other hand, SHRDLU cannot reply to utterances that refer to things out side of the virtual world such as “is the weather fine today?”. Hence, the top-down approach is not appropriate to design a robot that has to understand language and work in an open and dynamic environment. Therefore, the bottom-up approach is needed. In the bottom-up approach, a robot dynamically learns the connection between words and the real world and