Efficient learning of word meanings by agents using biases observed in language development of children

Abstract

Recently, studies on learning of word meanings by agents have begun. In these studies, a human shows objects to an agent and utters words such as "red" or "box". The agent finds out object's feature represented by each spoken word. In our method, firstly, the agent learns probability distribution p(x) and conditional probability distribution p(x|w), where x is an object, feature and w is a word. If a word w does not represent a feature x, p(x) and p(x|w) will be almost same distribution because x is independent of w. This fact enables the agent to use distance between p(x) and p(x|w) when inferring which feature the word represents. Previous works also employ similar stochastic approaches to detect the feature. However, such approaches need a lot of examples to learn correct distributions. To overcome this problem, we apply two types of biases: shape bias and mutual exclusivity bias that are observed in children's language development. When a child hears a novel word about an object, he/she often applies the word to other objects similar in shape. This tendency is called the shape bias. This bias works effectively for learning word meanings, because not a few words in the real world do not represent its color and material but represent its shape. In order to implement this bias into the agent, we formulate it as the variable that reduces the distances of non-shape features. Therefore, when a novel word is taught, the agent with the bias decides that the word represents shape feature and it can apply the word to other objects similar in shape. On the other hand, if a child already knows some words about an object, he/she often seeks the meaning of a novel word outside the meanings of known words. This tendency is called the mutual exclusivity bias. This bias is formulated as the variable that reduces the distances associated with features represented by known words. For example, if the agent already knows a word representing a shape, it decides that other words do not represent the same shape by reducing the distance associated with the shape. Experimental results show that proposed method with biases can acquire word meanings more efficiently than the traditional stochastic only approach.