The Predictive Value of Tagalog Voice Morphology in Filler-Gap Dependency Formation

Research has shown that when processing filler-gap dependencies, comprehenders do not wait until they encounter all of the bottom-up information in the input. Instead, they use various types of linguistic information to predictively posit a gap that would allow the dependency to be resolved. They can use syntactic (Traxler and Pickering, 1996), lexical (Trueswell et al., 1994), morphological (Kamide et al., 2003), and prosodic (Nagel et al., 1994) information. Here we examine whether Tagalog comprehenders use the language's voice morphology to guide their incremental interpretations. We hypothesized that voice allows comprehenders to commit to an interpretation upon encountering the verb, since they have information about the event structure at this point in time and by virtue of the voice morphology, the thematic role of the filler. In experiment 1, using an acceptability judgment study, we found that comprehenders differed in how they used the different voices in different filler-gap contexts to detect the licitness of displacements. These differences may have consequences for how voice is used in real-time. In experiments 2 and 3, using the stops-making-sense paradigm (Boland et al., 1990), we found that comprehenders used voice as a cue to actively associate the filler with the gap. However, in experiment 3, the way in which they used voice varied by type and varied across types of filler-gap dependencies. We argue that comprehenders were using construction-specific cue validities when processing filler-gap dependencies. However, they also engaged with other classes of linguistic information, including (but not limited to) information about the structural similarities and the thematic complexity of the dependencies involved, and the relative frequency of the different types of voices in the language. These interactions resulted in processing asymmetries.

One key difference between the cumulative d' and unrejected RTs is the degree of commitment to an interpretation that a comprehender must have given what she has encountered so far. When she chose to reject at a region in any given trial, she had committed to a particular interpretation: what continuations were possible and what continuations were not. In contrast, when she chose to continue reading but displayed slower readings times, she might have experienced processing difficulty-due to a plausibility mismatch effect, for example-but she may have been unable or unwilling to fully commit to an interpretation just yet (Mauner and Koenig, 2000). Together, these measures provide a more nuanced picture of the time courses of FGD-processing.
In Table S1, we report the mean unrejected reading times in ms by VOICE and PLAUSIBILITY at the critical and disambiguating regions (i.e., the verb-, XP-, and co-argument-regions). In Table S2, we present the estimated models. In the text that follows, we only report significant effects on their unrejected RTs.

Experiment 3A: Comparing AV and PV in wh-questions
At the verb-region, their unrejected RTs indicated that on trials where they chose to continue reading, they took longer when the verb exhibited PV (M = 830 ms) than when it exhibited AV (M = 781 ms). This effect was qualified by a significant interaction, such that they read implausible sentences significantly slower than plausible ones when it exhibited PV (∆RT = 89 ms) than when it exhibited AV (∆RT = -56 ms).
At the XP-region, their RTs were longer in the implausible conditions (M = 1168 ms) compared to the plausible conditions (M = 978 ms). At the co-argument-region, their RTs were again longer in the implausible conditions (M = 1343 ms) compared to the plausible conditions (M = 969 ms). This effect was qualified by a significant interaction, such that the difference between plausible and implausible sentences was greater when the verb exhibited PV (∆RT = 447 ms) than when it exhibited AV (∆RT = 299 ms). They were significantly faster when reading plausible sentences with PV.

Experiment 3B: Comparing AV and PV in relative clauses
At the verb-region, their unrejected RTs suggest that on trials where they chose to continue reading, they still detected the plausibility mismatches. Their RTs were longer in the implausible conditions (M = 902 ms) compared to the plausible ones (M = 748 ms).
At the XP-region, their RTs were again longer in the implausible conditions (M = 1047 ms) than in the plausible ones (M = 902 ms). This effect was qualified by a significant interaction, such that the difference between plausible and implausible sentences was greater when the verb exhibited AV (∆RT = 222 ms) than when it exhibited PV (∆RT = 78 ms).
At the co-argument-region, their RTs were again longer in the implausible conditions (M = 1180 ms) compared to the plausible ones (M = 980 ms). This effect was again qualified by a significant interaction, such that the difference between plausible and implausible sentences was greater when the verb exhibited AV (∆RT = 291 ms) than when it exhibited PV (∆RT = 125 ms).

Experiment 3C: Comparing AV and PV in ay-inverted sentences
At the verb-region, their unrejected RTs suggest that on trials where they chose to continue reading, they took longer when the verb exhibited PV (M = 831 ms) compared to when it exhibited AV (M = 754 ms).
At the XP-region, their RTs were longer in the implausible conditions (M = 1017 ms) compared to the plausible ones (M = 955 ms).
At the co-argument region, their RTs were longer in the implausible conditions (M = 1392 ms) compared to the plausible ones (M = 1168 ms).

Discussion
The goal of experiment 3 was to directly compare the time courses of dependency formation when the verb exhibited AV and PV in three different types of filler-gap dependencies. We found the following. First, we replicated the finding from experiment 2 that comprehenders actively associated the filler with the gap even before the fully disambiguating co-argument. Second, we found that the way in which voice was used varied across different voice types and across different dependencies.
As we saw, the participants' d's at the verb-and XP-regions provide evidence that voice was used a cue when interpreting FGDs. Participants were correctly rejecting implausible sentences as early as the verb-and XP-regions, even before they encountered the disambiguating co-argument. We add that their unrejected RTs provided further evidence that they were using voice as a cue. On trials where participants decided to continue reading, their longer reading times for implausible sentences suggest that they did detect the implausibility of the sentence even before the co-argument was encountered.
When we also consider their unrejected RTs, a more nuanced picture emerges. There was a PV-advantage at the co-argument-region in wh-questions and an AV-advantage at the XP-region in relative clauses. We allude to these asymmetries in greater detail in our general discussion. Table S1.  Table S2. Summary of linear mixed-effects models in experiment 3. The models included the log-transformed unrejected RTs as the dependent measure, and sum-coded VOICE, PLAUSIBILITY and their interaction into the models as fixed effects. Random effects included the maximal structure that converged.