A large number of word forms in natural language are polysemous, that is, associated with several related senses (e.g., line, run, tight, etc.). While such polysemy appears to cause little difficulty in verbal communication, it poses a range of theoretical and descriptive problems. One concerns its very existence: What is it about our language systems that make them so susceptible to polysemy? In this paper I discuss two approaches to polysemy with different answers to this question: (i) A code-based approach that treats polysemy in terms of the operation of lexicon-internal generative rules, and (ii) an inference-based approach that takes polysemy to be governed by pragmatic inferential processes applying at the level of individual words. After evaluating how each of these accounts fares with respect to some empirical data, I look more broadly at their implications for the emergence and development of polysemy. I conclude that, overall, the pragmatic approach provides the most promising basis for a unified account of the role of polysemy in several domains, and for explaining what motivates its proliferation natural language.