This paper describes a primitive-based representation for adjective meaning. The meanings of words are roughly divided into two parts : core meaning and connotative meaning. The core meaning is independent of context, situation, and the speaker's emotions. On the other hand, the connotative meaning is strongly dependent on the speaker's emotions and feelings. However, the boundary of two parts for adjectives is less clear than for nouns or verbs. This is why most research in natural language processing has centered on the analysis of the meanings of nouns and verbs. But when natural language is regarded as a method for human communication, the analysis of adjective meaning is indispensable for natural language processing because adjectives represent human emotions, perception, and attributes of objects. We have analyzed about 700 Japanese adjectives and decomposed the meanings of these adjectives into approximately 400 semantic primitives. The essential points of this primitive-based representation are the following. 1) Core meaning is represented by one semantic primitive and connotative meaning is represented by additional primitives. Moreover, in order to clarify the differences among adjectives which have similar meanings, additional information (e.g. subjective/objective, positive/negative evaluation) is used. 2) Semantic relations between primitives are also used for representing the meaning relations among adjectives. 3) Selectional restrictions are made available for semantic disambiguation. This representation can provide much information useful for natural language processing.