Estimating concept embeddings from their child concepts

We estimate the concept embeddings, which are distributed representations of the concepts, in a concept dictionary from the concept embeddings of their child concepts. The concept dictionaries represent the systematic classification of the concepts of nouns and verbs in a way and the concept embeddings represent the meanings of the concepts. This paper investigates what kind of composition calculation can represent the relation between the concepts and their child concepts in a concept dictionary, which is a hierarchical relationship that humans assume. We examined four methods to estimate the concept embeddings and investigated three size of dimensions. The experiments revealed that the best method was the simple summation of concept embeddings of the child concepts and the similarities increased when the vector size decreased. We also examined that whether the similarities between the actual and estimated concept embeddings will increase when we restricted the concepts to calculate the similarities by the minimum number or percentage of their children's concept embeddings. However, the experiments revealed that they decreased if the concepts were restricted.