Taiwanese tone sandhi was argued by-generative phonologists as productive and reducible to a rule-type operation. This paper attempts to explore into the nature of the productivity of the phenomenon by conducting an experiment. Twenty nonsense words were made up with four words for each of the five tones. Eighteen native speakers were asked to perform a task in which the nonsense words assumed the meaning of color adjectives. The subjects were instructed to use the words in sentences in a tone sandhi environment. The items were run through five times. The results showed that different tones demonstrated different degrees of productivity, and that productivity increased over five trials. This seems to indicate that familiarity of lexical items does contribute to the productivity of the phenomenon. A non-generative, organizational account is proposed to accomodate the results. It is argued that phonological forms coexist in the mental structure, and linguistic generalizations serve as organizational principles rather than as derivational/deterministic procedures for relevant forms.