Diary observations of two-year-olds' over-extended word use have been interpreted as arising from the word's underlying semantic feature structure, retrieval error, or limited vocabulary. The first interpretation was tested by analysing whether the same few features systematically characterized most over-extensions of a word (picture-naming task), and whether the same word was over-extended in comprehension (picture-pointing task). Four instances of over-extended use were studied in each of 5 children aged 1; 9 to 2; 3. A semantic feature interpretation is rejected, as is the view that overextensions are used by the child deliberately, to refer to objects which he knows are not properly named by the words. The need to construct models of early word meaning which reflect the early instability of conceptual organization, the probability of retrieval error in recall tasks, and the developmental history of individual words – particularly parental acceptance or rejection of use – is discussed.