The word substitution errors from a corpus of 2,400 French slips of the tongue were grouped into several categories: contaminational, semantic, formal, and mixed cases; substitutions of syntagmatic codependents also occurred. Semantic and formal substitutions involved a resemblance between target and error. In addition, all substitutions exhibited a strong degree of word class and gender identity. The various types of resemblance were analyzed with reference to three-layer models of lexicalization. They did not make a lemma layer necessary, but stronger evidence came from another error category – semantic blends.