This paper revisits the question raised by Kishimoto (1996): what determines the unergative-unaccusative split in the behavior of the V1 in a Japanese deverbal nominal construction (V1-V2-gen N) (e.g., toke-kake-no yuki “the snow, almost melted”). While previous accounts (e.g., Tsujimura & Iida 1999) relied on purely semantic treatments, this paper takes a fresh look at the construction, turning to syntax. This analysis crucially points out that the modifier subsumes two distinct units, “adjectival” and “reduced clausal,” and that the V1s split differently in each type on the basis of lexical aspect (i.e., the former requires the V1 to be telic and the latter, non-state). Construction-specific requirements which further affect the distribution of the elements are also delineated.