Recent masked priming experiments have brought to light a morphological level of analysis that is exclusively based on the orthographic appearance of words, so that it breaks down corner into corn- and -er, as well as dealer into deal- and -er (Rastle, Davis, & New, 2004). Being insensitive to semantic factors, this morpho-orthographic segmentation process cannot capture the morphological relationship between irregularly inflected words and their base forms (e.g., fell-fall, bought-buy); hence, the prediction follows that these words should not facilitate each other in masked priming experiments. However, the first experiment described in the present work demonstrates that fell does facilitate fall more than orthographically matched (e.g., fill) and unrelated control words (e.g., hope). Experiments 2 and 3 also show that this effect cannot be explained through orthographic sub-regularities that characterize many irregular inflections, as no priming arose when unrelated words showing the same orthographic patterns were tested (e.g., tell-tall vs. toll-tall). These results highlight the existence of a second higher-level source of masked morphological priming; we propose that this second source of priming is located at the lemma level, where inflected words (but not derived words) share their representation irrespective of orthographic regularity.