How are lexical representations retrieved during sign production? Similar to spoken languages, lexical representation in sign language must be accessed through semantics when naming pictures. However, it remains an open issue whether lexical representations in sign language can be accessed via routes that bypass semantics when retrieval is elicited by written words. Here we address this issue by exploring under which circumstances sign retrieval is sensitive to semantic context. To this end we replicate in sign language production the cumulative semantic cost: The observation that naming latencies increase monotonically with each additional within-category item that is named in a sequence of pictures. In the experiment reported here, deaf participants signed sequences of pictures or signed sequences of Italian written words using Italian Sign Language. The results showed a cumulative semantic cost in picture naming but, strikingly, not in word naming. This suggests that only picture naming required access to semantics, whereas deaf signers accessed the sign language lexicon directly (i.e., bypassing semantics) when naming written words. The implications of these findings for the architecture of the sign production system are discussed in the context of current models of lexical access in spoken language production.