Many dependencies show retrieval interference, and ellipsis does too! Dan Parker, Hayley Shankle, & Irene Williams (William & Mary) dparker@wm.edu<mailto:dparker@wm.edu> It is well-established that memory retrieval for linguistic dependency formation relies on an associative retrieval procedure [1]. However, it remains unclear what information (“cues”) at the retrieval site are used to recover an antecedent in memory. In this study, we test dependencies involving verb phrase (VP) ellipsis to better understand what types of non-structural information are used to recover an antecedent in memory. Previous research has focused primarily on how a noun phrase antecedent is recovered, but VP-ellipsis is an important test case to examine what cues are used to recover a VP. We present the results of two self-paced reading experiments testing whether voice features (e.g., active vs. passive) are used to recover a VP antecedent. Results based on susceptibility to retrieval interference revealed that passive and active voice features are used differently in antecedent retrieval. Motivation. VP-ellipsis involves a missing VP that requires retrieval of a previously-processed VP for interpretation, e.g., (1) [2]. A constraint on VP-ellipsis is that the antecedent and ellipsis must match in voice, e.g., (2) [3]. A recent ERP study [4] showed that voice features can trigger interference from a grammatically-irrelevant but feature-matching antecedent at retrieval. However, this study confounded voice with grammaticality (grammatical sentences = active, ungrammatical sentences = passive), and recent computational modeling work suggests that active and passive ellipsis might behave differently at retrieval resulting in different RT profiles [5]. To address this issue, we tested active and passive sentences independently using self-paced reading. Experiment 1: Passive voice (n = 120). Experiment 1 manipulated grammaticality (grammatical vs. ungrammatical for passive ellipsis) and interference (voice features of active vs. passive on a relative clause VP that cannot serve as an antecedent) (see Table 1). RTs at the region immediately following the ellipsis marker (did too) showed a main effect of grammaticality (t = 2.76) and an interaction of grammaticality × interference (t = 2.23) carried by interference in the ungrammatical conditions (t = 2.78). Sensitivity to the passive relative clause VP suggests that passive voice is used to guide retrieval for ellipsis. Experiment 2: Active voice (n = 120). Experiment 2 used the same paradigm as in Experiment 1, but tested active ellipsis (Table 2). RTs following the ellipsis marker showed an effect of grammaticality (t = 5.46), but no interference (ts < 1.5). These results suggest that active voice is not used in retrieval for ellipsis or at least is not a highly selective cue to the antecedent. This proposal is supported by an interference × experiment interaction (t = 2.65) showing that active and passive voice behave differently with respect to interference effects. Discussion. These results are surprising because it seems that a cue for active voice would be just as useful in recovering an antecedent as a cue for passive voice. However, the results are consistent with the recent claim based on other dependencies that not all cues at the retrieval site are used in the same way [6-8]. The selective profile with respect to voice interference likely reflects a markedness effect, like that observed subject-verb agreement (“agreement attraction”), where only marked structures (e.g., plural verbs) show interference. For ellipsis, the marked passive ellipsis would deploy a voice cue, whereas the unmarked active counterpart would not, leading to the observed profiles. These results point to a uniform account of markedness effects across dependencies with respect to retrieval interference.