The notion of words is obscure and yet is supposedly one of the basic notions in theory of linguistics. When we ask what words are, we cannot get a definite answer like that a chemist could provide for the question “what is water”. We talk about words in various contexts of linguistic theories: in semantics we ask questions like “what does the word ‘definite’ mean?”, in acquisition we say the baby learned ten new words today, theorists in speech perception make claims like “in speech, beginnings of words arrive temporally prior to middles and ends, and word length cannot be initially apparent”, and etc. Some people might argue that a theory about what words are is not necessary: we can talk about words in different ways in different linguistic theories and yet don’t have any problem understanding the theoretical talk, nor would the child have any difficulty learning words without really knowing what words are. In this talk I argue that a metaphysical theory of words, i.e. a theory about what the child tacitly takes words to be, is very important both empirically and theoretically. Such a theory is closely related to our hypotheses about what kind of learning mechanisms a child has. I shall also critically review the existing theories of words (Jackendoff (2002), Kaplan (1990) and Lepore & Hawthorne (2011)) and their implications for what cognitive architecture a child must have in order to successfully acquire words and how such implications might lead to learning problems due to their theoretical limitations.