Associative connections between letter-digit pairs were established in three separate experiments using a probabilistic paired-associate procedure. In each experiment, three or more letters were each paired with the same three digits. The S"s were then given an association task, which required a digit response to each of the stimulus letters, and a "speaker" communication task, which required a digit response which would discriminate one letter (referent) from another (nonreferent). In general, the association data indicated that relative frequency of emission was monotonically related to relative frequency of occurrence in training; in one experiment, however, certain systematic discrepancies suggested a "nonassociative" factor. "Speaker" response choices were related to induced associative strength by qualitative and quantitative application of a two-stage model. Results from all three experiments indicated that the two-stage model successfully predicted speaker response choices and was superior to a simpler one-stage model.