A frame-based association mechanism is proposed aiming at advanced speech recognition. The mechanism is based on two knowledge bases. One is a hierarchical dictionary and the other, an association index file. In the association index file "Association Confidence Factor (ACF)" is introduced to express a directional weight of each link from a source concept to a target concept using a membership function. As it is impossible to assign ACFs on every links among all concepts, the number of links to which ACFs are assigned is limited, so it is required to provide means to give appropriate ACF values for the rest of the links on demand employing available ACFs. A set of formulae is proposed for that purpose. In addition, psychological experiments were conducted to investigate human association activity.