This paper is about the status of compositionality in our linguistic communication and its theorizing. The principle of compositionality has been regarded as an essential precondition under which semantic theories are built since the principle seems to properly explain some important facts about natural language, such as learnability and novelty. According to recent discussions, however, the precise understanding of compositionality casts doubt on its necessity. In this paper, I take a closer look at how we have started and improved our semantic investigations, and present a new argument for compositionality which is based on our linguistic intuition about the notion of un-synonymy and substitution. My point of the argument is that compositionality can serve as a useful tool for theorizing natural language semantics even if semantic rules themselves are described non-compositionally.