What makes one item in memory more accessible than another in language processing? While evidence is steadily mounting in favor of a cue-based parsing system (e.g., Lewis & Vasishth, 2005), in which items are directly accessed in parallel and compared against the retrieval cue in an associative cue-matching procedure, numerous choice points in the retrieval architecture remain open. One choice point is the extent to which linguistic structure and grammatical constraints influence retrieval; another is how discourse status of items in memory affects accessibility.

I report two studies that address these issues via sluicing ellipsis with d-linked ‘which’ phrases, as in “The student still needs to take a linguistics class, but I don't know which one”. The studies exploit several key properties of these constructions, allowing us to compare how the processor uses structural preferences and discourse constraints to pair the remnant (which one) with a correlate (a linguistics class). In particular, I explore how a structural preference favoring the closest possible correlate for the remnant (Frazier & Clifton, 1998) is moderated by (i) the specificity of cues at the remnant site, and (ii) the discourse status of potential correlates.

In keeping with previous research, the results show an immediate penalty for distant correlates. However, the size and time course of the penalty appears to depend on the specificity of the retrieval cues provided by the remnant. Results also indicate that while discourse status guides correlate selection, it is not sufficient to completely override structural preferences. Implications for current memory models will be discussed.