Objective Hierarchy of Abstract Concepts : Organization of Abstract Nouns via Distribution of Adjectives

Abstract

Our purpose in this research is to find hypernimic concepts of words experimentally by using a large corpora and a neural network model. At first we treat adjectives. We focused on semantic relations between abstract nouns and adjectives. We made linguistic data by extracting semantic relations between abstract nouns and adjectives from large corpora, and we use them as an input data for the Self-Organizing Semantic Map (SOM), which is a neural network model (Kohonen 1995). On the Semantic Map, words are located near to or far from each other depending on their similarities. In some previous research, word meanings were classified from linguistic data based on syntactic information using a statistical method. Hindle (1990) used syntactic relations between nouns and verbs, and Hatzivassiloglou and McKeown (1993) used semantic relations between adjective-adjective pairs, both of which are modifiers for a head noun. Their methods are useful for an organization of words in a subordinate layer, however, if we consider it in a superordinate layer, their methods seem not to be enough to solve the problem. Previous researches on the classification of words by a SOM treat a small amount of data (Kohonen et al., 1995), however, we made an experiment with a large amount of data, such as 42 years worth of newspaper articles.