Glossa Psycholinguistics

1.1 Word order and case clitics in Hindi

The grammatical system of Hindi has two particularly important properties for the current study. First, Hindi has Subject–Object–Verb (SOV) word order. In consequence, while a Hindi sentence’s tense and aspect only appears at the end, case marking on the subject (ergative in (1)) appears already within the first syntactic phrase:

The second important property of Hindi for the purpose of this article is that it is a split-ergative language (Anand and Nevins, 2006; Coon and Parker, 2019; Kachru and Pandharipande, 1978), showing different patterns of case marking that are conditional on two factors: perfectivity and transitivity (Kachru, 1980; Mohanan, 1994). In this pattern, the agent argument in a transitive sentence is assigned ergative case when the verb is inflected for the perfective (e.g., (1)). When the verb does not have perfective morphology, the agent argument will occur with the nominative (zero) case as seen in (2). Intransitives, on the other hand, will typically occur with the nominative (zero) case on the single argument, regardless of perfective morphology (3), giving rise to a split-ergative system.²

When viewed from the point of view of real-time language comprehension, the case marking on the subject (mohan in (1) and (2)) can serve as a very early cue for the grammatical aspect marked on the verb, and trigger inferences about whether an event is culminated or not. As mentioned, this is different from English and many other languages, where this aspectual distinction is only marked morphologically on the verb describing the event (Dahl and Velupillai, 2013).

A consequence of this property is that ergative markers should be strong signals to indicate culminated events (Arunachalam and Kothari, 2011; Butt, 2017). Indeed, a corpus search reveals that the association between ergative marking and perfective tense is just as strong as commonly assumed.³ Ergative case is an extremely reliable cue for perfectivity, while nominative case is not (Figure 1).

Figure 1: Corpus data: Distribution of case marking in the HindiTreeBank, by tense.

Despite the strength of the association between ergative case and perfective aspect, Arunachalam and Kothari (2011) showed that Hindi perfective sentences are not necessarily always interpreted as culminated events, especially when those sentences describe events with incremental themes (e.g., peeling an orange) as opposed to non-incremental themes (e.g., dropping an orange; see also Wittek (2002) and Pederson (2007) for similar results in Tamil and German). However, Arunachalam and Kothari (2011) did not study the role of the ergative marker in contrast to null-marked subjects in the nominative; and the main interest in these studies was the defeasibility of the culmination inference, as opposed to the incremental unfolding of predictions about an event interpretation based on grammatical properties.

From an incremental processing perspective, Hindi remains understudied, but several recent studies exploit its link between ergative/nominative case and perfective/imperfective aspect to gain insight into the underlying parsing process. Choudhary et al. (2009) conducted an ERP study of case marking violations in Hindi and found that a mismatch of ergative case with imperfective aspect triggered a stronger N400 than a mismatch of nominative case with perfective aspect, and a late positivity (P600).

This P600 effect was replicated by Dillon et al. (2012) for ungrammatical combinations of ergative markers with future tense; in contrast, a nominative-perfective mismatch did not result in a late positivity, which the authors attribute to a small class of transitive perfective verbs that obligatorily take nominative marking on the agent (Mohanan, 1994), and to the fact that omission of ergative marking is quite common in some Hindi dialects (Butt and King, 2004). This asymmetry between violations of ergative vs. nominative case was, however, not found by Bickel et al. (2015).

That said, the nature of the signaling function of ergative markers in perfective sentences is still subject to debate. An orthogonal hypothesis about what, if anything, ergative case markers may contribute beyond their grammatical necessity claims that ergativity also signals volitionality, or an action done with a certain degree of intention: An event that is described with the ergative marker on the subject is said to convey a sense of ‘conscious choice’ (Butt, 1995; Longenbaugh and Polinsky, 2017; Mohanan, 1994; also see Maitreyee et al., 2023 for evidence that both children and adults are sensitive to the degree of volitionality in the context of ergative subjects). We will address this hypothesis in Section 2.

1.2 Previous studies of event structure inferences

Our article is one in a growing psycholinguistic tradition asking whether and at which point during incremental sentence comprehension people draw culmination inferences. So far, the evidence has been mixed: While there is some evidence that comprehenders do not commit to culmination inferences as a sentence unfolds incrementally (Misersky, 2022; Pickering et al., 2006; Proctor et al., 2004), other work has found that aspectual information affects ongoing sentence interpretation (Baggio et al., 2010; Kaplan et al., 2021; Misersky et al., 2021; Paczynski et al., 2014).

Several studies have shown that the rapid calculation of culmination inferences can be detected using the Visual World Paradigm. For instance, Altmann and Kamide (2007) showed that when people heard sentences containing verbs in the past perfect tense (The man has drunk…), they rapidly look to an object compatible with a culminated action, e.g., an empty glass. However, whenever people heard sentences containing verbs in future tense (The man will drink…), they were more likely to look to an object compatible with a non-culminated action, e.g., a full bottle.

These rapid inferences of event culmination based on aspectual information are documented in different populations, and across languages. For instance, Zhou et al. (2014) found that young Mandarin-speaking children used aspectual verbal morphology to rapidly recognize event structure: when they heard the durative suffix -zhe, they looked towards a picture of an ongoing version of the event, and when they heard the perfective suffix -le, they looked towards the completed version – and these children did so as rapidly as adults. A similar study in Russian, which has a highly grammaticized distinction between imperfective and perfective aspect, found that adult Russian listeners used aspectual morphology to identify a target picture even before verb offset (Minor et al., 2022). Finally, Turkish children can draw culmination inferences based on partitive case marking before the verb is heard (Özge et al., 2017).

Our present studies are built on a recent article that used a visual world paradigm in Italian to ask when a verb’s morphosyntactic telicity⁴ cues (perfective aspect) are used to make inferences about event structure (Foppolo et al., 2021). Importantly, Foppolo et al. (2021) separated telicity and culmination by using sentences containing perfective verbs that were either durative (e.g., has peeled the apple) or punctual (e.g., has dropped the ball): Unlike durative verbs, the culmination, or end point, is lexically encoded in punctual⁵ verbs. Thus, the authors were able to ask not only whether people draw telicity inferences based on the verb, but also whether the lexical information they glean from the verb is used to draw culmination inferences during incremental sentence processing, and at which point.

In their studies, Foppolo et al. (2021) showed two pictures containing different objects, one with a completed and one with a non-completed action (e.g., an open newspaper and a closed newspaper). Participants listened to sentences like Valeria has opened the newspaper while their eye-movements were measured. The results show that when listening to sentences containing perfective tense, participants launched anticipatory looks towards the picture of the event in its completed state before the offset of the predicate. Crucially, participants were equally fast to look towards the target picture in both the punctual and the durative verb conditions. In other words, perfective aspectual morphology on the durative verbs was enough to trigger a culmination inference.

What is an open question, however, is whether any morphological cue could be a potential trigger for successful culmination inferences: the previous literature has relied entirely on morphosyntactic marking within the verb phrase, that is, case or aspectual marking either on the verb or the object. Across the world’s languages, however, there are many ways to grammatically convey information about properties of an event’s structure. One cross-linguistically frequent way is ergative marking on subjects, specifically, split ergativity, as found in Modern Standard Hindi. As reviewed above, several studies have shown that Hindi speakers do use case marking on the subject to predict upcoming verbal morphology (Arunachalam and Kothari, 2011; Choudhary et al., 2009; Dillon et al., 2012; but see Bickel et al., 2015). But do speakers also use this abstract information to predict event structure, such as an event’s degree of culmination?

1.3 Current studies and predictions

In this article, we ask whether Hindi speakers draw predictive inferences about event culmination based on case marking on the grammatical subject. In four studies, we ask whether Hindi native speakers draw semantic inferences about events, and when they do so, based on the presence or absence of an ergative marker on the subject.

One prediction for the current studies is that if Hindi speakers indeed use case marking as a predictive cue, they should theoretically be able to do so at the very first noun in a sentence, given Hindi’s SOV structure. This prediction is supported by data from Choudhary et al. (2009) and Dillon et al. (2012): comprehenders appear to generate expectations about upcoming input based on subject case marking, and mismatches produce an ERP signature that has been interpreted as indexing surprisal or reanalysis.

However, anticipation of grammatically licensed input is not equivalent to inferencing event properties before the verb has been encountered: although there is robust evidence (Foppolo et al., 2021; Minor et al., 2022; Zhou et al., 2014) that grammatical aspect can trigger rapid inferences of event culmination, an alternative prediction is that Hindi case marking does not function as a predictive cue for event structure at all.

This prediction is rooted within a theoretical linguistic literature in which the verb phrase is the domain of aspectual interpretation: under this view, aspectual interpretations are crucially dependent on features within the verb phrase (e.g., Husband and Stockall, 2015; Kratzer, 1996).⁶ Located outside of the event-encoding predicate, the subject’s case marking may be superfluous to aspectual composition in the verb phrase, or unavailable for integration with the event semantics until the verb phrase constituent is fully parsed. Under this account, ergative case mostly fulfills a structural requirement, and the semantic reading, be it of event culmination or conscious choice, or both, emerges from the complex interaction between the lexical affordances of the main verb, the semantic contribution of the light verb (if present),⁷ and the ergative marker (Bashir, 1999; Butt, 1995; Ramchand, 1990; Ranjan, 2016).

A third possibility is one of partial prediction: that, since perfective marking on the verb does not reliably entail event culmination, the presence of an ergative marker does not serve as a predictive cue to a completed event – but that the absence of one, i.e., nominative case, does allow comprehenders to predict an incomplete event. Such an absence could be interpreted and integrated as rapidly as a realized affix (Minor et al., 2022). A weaker version of this scenario is one where, in Hindi, neither case marker is a predictive cue, but there is an overall preference for interpreting an event as incomplete unless and until there is evidence not just of an event telos, but also of a result state. This preference could be a cross-linguistic strategy, as found in Foppolo et al. (2021). However, partial prediction could also be driven by language-specific affordances changing participants’ processing strategies, which are adapted to the contrasts available in particular languages (e.g., Bott and Gattnar, 2015; Bott and Hamm, 2014; Flecken et al., 2014; von Stutterheim et al., 2012).

These predictions about culmination inferences are somewhat orthogonal to another question: whether Hindi ergative case marking is primarily associated with “conscious choice” or intentionality (Butt, 2017; Butt and King, 2004; Maitreyee et al., 2023; Mohanan, 1994; also see Demirdache and Martin, 2015; Mathis and Papafragou, 2022; van Hout et al., 2017 for arguments that knowledge about the goals of an agent, or the degree of an agent’s control over an event, influences inferences about event construal). If this is the case, then culmination and intentionality may interact in a way to mask any culmination inferences.

Thus, our first experiment here will address the question of the role of intentionality inferences, before we move to the question of event culmination inferences in Experiment II–IV.

2.1 Participants

We paid 48 participants on Amazon Mechanical Turk to participate in this experiment. All participants were self-reported Hindi native speakers and had an IP address in India. As an attention check, and to ensure that they were indeed Hindi native speakers, they had to solve two riddles in Hindi. 17 participants failed this test, so 31 participants were included in the final analysis.⁸

2.2 Methods and Materials

In this and the following study, we presented 31 sentences without the verb, containing only the subject and an inanimate direct object (e.g., Mohan(-ne) orange-ko …). In order to see whether subject case marking alone triggers inferences on event completeness, the sentence fragments either contained an ergative-marked subject (Mohan-ne) or a nominative unmarked subject (Mohan), as a within-participants manipulation. The direct object was generally definite, singular, and inanimate, and marked with the accusative marker -ko,⁹ ensuring that the fragment could not be interpreted as passive and nominative marking could not be confused for an unmarked patient role. Participants then answered a forced-choice task about whether they thought this fragment described an action that was done on purpose (see Figure 2 and Table 1).

Figure 2: Example of a forced choice task. The task was the same for Experiments I and II, and differed only in whether it asked for intuitions on intentionality (Experiment I) or event culmination (Experiment II).

In addition to the critical sentences, we also included 24 fillers, in which random portions of the sentence were eliminated.

2.3 Results

All our data, with the exception of Experiment IV, were analyzed using mixed-effects models with maximal effects structure on contrast-coded conditions, with random slopes and intercepts for participants and items. In Experiment I, the fixed effect was case marking.

As Table 2 and Figure 3a show, people drew intentionality inferences for fragments with nominative and ergative-marked subjects at the roughly the same rate: Fragments containing an ergative subject were rated to describe an intentional action around 74% of the time, and fragments containing a nominative subject, 76% of the time.

Figure 3: Results of Experiments I and II: Completeness and intentionality inferences, based on subject and object information alone.

The two conditions of the fixed effect, case marking, were contrast-coded as ergative = 1 and nominative = –1. Within this model, the effect of case was statistically non-significant and negative (beta = –0.05, p = 0.575; see Table 3), meaning that the likelihood of participants judging an item to be intentional was very slightly, but not meaningfully, higher in the nominative condition.

2.4 Discussion of Experiment I

This binary choice task served to determine whether ergative case marking by itself, i.e., in the absence of further semantic information about the predicate, carries a meaning of intentionality with it, a question that has repeatedly arisen throughout the previous literature (e.g., Butt, 1995; Longenbaugh and Polinsky, 2017; Maitreyee et al., 2023; Mohanan, 1994). However, our data show a similar proportion of intentionality inferences for both types of case marking, nominative and ergative.

One possibility for this lack of difference could be that the wording of our task was ill-chosen, given that similar notions are employed throughout the literature, from intentionality to volitionality to “conscious choice”. To the extent that all of these notions draw from a closely related conceptual space, however, these data show that if there is any intentional meaning that arises from a sentence containing ergative marking, this meaning is likely to be inferred from either the predicate alone (which was absent in our study), or by the combined effect of ergativity and predicative meaning.¹⁰

Another possible explanation is that the semantic contrast of a volitional vs. involuntary agent conferred by ergative vs. nominative marking only exists in a (small) subset of intransitive verbs (De Hoop and Narasimhan, 2005; Maitreyee et al., 2023; Mohanan, 1994). Given a sentence fragment containing first an animate noun, followed by an inanimate (and accusative-marked) noun, a comprehender has no reason to assume that circumstance applies; they may infer that the nouns form two arguments of a transitive verb, of which the first is judged to have a higher likelihood of being volitional.

For the purposes of this article, however, we lay the intentionality hypothesis aside: We have no reason to assume that subject case marking alone plays any role in inferences on volition or “conscious choice” in event structure. But does it for event culmination?

3.1 Participants

We paid 48 participants on Amazon Mechanical Turk to participate in this experiment. All participants were self-reported Hindi native speakers and had an IP address in India. As an attention check, and to ensure that they were indeed Hindi native speakers, they had to solve two riddles in Hindi. 18 participants failed this test, so 30 participants were included in the final analysis.

3.2 Methods and Materials

We used the same materials as in Experiment I, but instead of asking people to rate whether they thought the event described by the sentence fragment was intentional, we asked whether they thought the action described by the event was completed.

3.3 Results

As Table 4 and Figure 3b show, people had clear intuitions about whether the event expressed by the fragment was completed or not, depending on case marking: Fragments containing an ergative subject were rated to be incomplete only 26% of the time, but fragments containing a nominative subject, 76% of the time.

As in Experiment I, the two conditions of the fixed effect were contrast-coded as ergative = 1 and nominative = –1. Within this model, the effect of case was statistically significant and negative (beta = –1.10, p < .001; see Table 5), meaning that the likelihood of participants judging an item to be incomplete was much higher in the nominative condition.

3.4 Discussion of Experiment II

This binary choice task established that ergative case marking by itself, without any lexical or contextual information about the predicate, gives rise to event completion or culmination inferences: The recorded proportion of culmination inferences is significantly higher than in the nominative-marking condition.

This means that Hindi speakers privilege an event culmination inference when they encounter ergative subject marking, even in the absence of any further information about the kind of event itself. Taking these results as a starting point, the following studies ask whether we can trace when during sentence comprehension these inferences arise: From the earliest point on, that is, right after case marking on the subject is heard? Or do people hold off on their interpretation until they have all available information, as has been argued for other phenomena (e.g., Pickering et al., 2006)?

4.1 Method & materials

We created 24 sets of stimuli like (4–7). Each set described two events involving the same object (i.e., an orange) in a Patient role. One event was durative (‘peel’), and the other event was punctual (‘drop’).¹¹ Each event either described a completed event like (4) or (6), or an event in progress, like (5) or (7). Each sentence was preceded by the command hame ye bataein (‘Tell us…’).

We paired the sentences above with a pair of pictures showing an event either in its completed or its incomplete state. Each participant saw each object (for instance, an orange) only paired with one event; so, for a given participant, Example (4) would be paired with the two pictures in Figure 4a, and for the next participant, Example (5) would be paired with the two pictures in Figure 4a. A third and fourth set of participants saw Examples (7) and (6), respectively, paired with Figure 4b, such that each participant saw 24 trials, six of each condition.

Figure 4: Pictures of stages of events associated with the same object (orange): durative (a) and punctual (b).

4.2 Procedure

As with Experiments I and II, all studies were deployed on Amazon Mechanical Turk. The norming studies (see below) were also built using the survey tools of Mechanical Turk; the main experiment was built using jsPsych (De Leeuw, 2015).

In preparation for this experiment, we conducted a series of norming studies to make sure that any potential effects would not be due to confounds introduced by the stimuli.

4.2.2 Norming Study I: Sentence Stimuli

4.2.2.1 Participants

We paid 96 workers for their participation on Amazon Mechanical Turk. In all experiments, all participants were self-reported Hindi native speakers and had an IP address in India. 33 workers participated twice, so we only included their first submitted dataset in the analysis. We excluded 28 workers based on their performance on the riddles, leaving us with 35 individual participants’ datasets.

4.2.2.2 Methods and results

Participants rated all sentences on a scale from 1 (ungrammatical) to 7 (grammatical). All factors were contrast-coded and the z-scored responses were analyzed using generalized linear mixed-effects regression with maximal effects structure. To determine significance, we used model comparisons.

As Table 6 and Figure 5 show, there were no significant differences between conditions: Neither the main effect of event type, nor the main effect of case, nor their interactions were significant (main effects χ²s (2, 25) < 0.33, ps > .36; interaction χ² (1, 25) = 0.8092, ps > 0.37). These results make it more likely that any difference we find in the sentence-picture matching task is not due to the varying grammaticality of the sentence conditions.

Table 6: Average grammaticality ratings of sentences in each condition, with Standard Deviations, Standard Errors, and Confidence Intervals. N indicates number of measurements per condition.

Case	Event	N	rating	SD	SE
ergative	durative	369	5.39	1.63	0.09
ergative	punctual	374	5.28	1.66	0.09
nominative	durative	363	5.39	1.65	0.09
nominative	punctual	371	5.40	1.59	0.08

Figure 5: Norming Study I: Sentence ratings on a scale from 1 (bad) to 7 (good). Error bars show standard errors.

4.2.3 Norming Study II: Matching sentences to pictures

This norming study was carried out to ensure that our sentences and pictures were well matched, that is, that the pictures represented the sentences accurately.

4.2.3.1 Participants

We paid 96 workers for their participation. This time, 50 workers participated twice, so we only included their first submitted dataset in the analysis. We excluded an additional 21 workers based on their performance on the riddles, leaving us with 25 individual participants’ datasets.

4.2.3.2 Methods and Results

In each trial, participants saw one written sentence and two pictures: the incomplete and completed versions of the event described by the sentence. We manipulated case/aspect as a within-subjects factor: for each object in the set of stimuli (e.g., an orange), participants saw either the two progressive sentences with nominative case (i.e., ‘peeling’ and ‘dropping’), or the two perfective sentences with ergative case (i.e., ‘peeled’ and ‘dropped’). Participants were asked to choose which picture of the two best matched the sentence. We coded the responses as ‘correct’ when a completed event picture was matched to a sentence with ergative case, and when an incomplete picture was matched to a sentence with nominative case.

We used logistic mixed models to determine whether the log-odds of selecting the correct picture were above chance level in each of the case/event type conditions, with random intercepts for items and participants.

As Table 7 and Figure 6a show, people reliably matched pictures to the correct sentences (>50% correct in all conditions). The log-odds of selecting the incomplete picture to match the nominative sentence were significantly higher than 0, i.e. chance level (Intercept = 2.97, SE = 0.58, Z = 5.07, p < 0.001), and likewise for selecting the complete picture to match the ergative sentence (Intercept = 1.8, SE = 0.42, Z = 4.3, p < 0.001).

Table 7: Average correct matching of sentences to pictures in each condition, with standard error. N indicates number of measurements per condition.

Case	Event	N	% correct	SE
ergative	durative	147	78%	0.03
ergative	punctual	146	75%	0.04
nominative	durative	152	89%	0.02
nominative	punctual	155	86%	0.03

Figure 6: Results of Norming Studies II and III.

We then compared the conditions using binomial mixed-effects models, with sum-coded case and event type as fixed effects. We only used random intercepts for items and participants, because the model failed to converge with a maximal random-effects structure. To determine significance, we used model comparisons.

The percentage of correct matches did not depend on event type: There was no difference between punctual and durative events (χ² (2, 6) < 1.84, p > .40), and there was no interaction between event type and case, either (χ ² (1, 6) < 0.52, p > .47). However, there was a main effect of case: Sentences were more likely to be matched to the correct picture in the nominative condition than in the ergative condition (χ² (2, 6) < 14.02, p < .001). So even though people demonstrated an above-chance preference both for matching ergative sentences with complete pictures, and nominative sentences with incomplete pictures, that preference appears to be more flexible for ergative sentences – despite the (in)completeness inference tested in Experiment II being about equally strong for both cases.

4.3 Timed picture-matching experiment

In summary, to prepare for this experiment, we ran a visual complexity analysis, a grammaticality norming task, and two sentence/picture norming tasks. We found that the visual complexity of the picture stimuli was comparable across conditions; that the grammaticality of the sentence stimuli was, likewise, rated highly and very similarly across conditions; and that participants reliably matched sentences to the correct picture, and vice versa.

In this timed picture-matching experiment, we paired four kinds of sentences (crossing ergative or nominative case with durative or punctual events) with a pair of pictures showing an event either in its completed or its incomplete state. In each trial, the two pictures were presented side-by-side on the screen, without preview; and the sentence (created using Google’s speech-to-text API) was presented auditorily. The experiment included no fillers.¹² Participants were instructed to choose the picture that best matched what they heard, and to try to respond as quickly as possible. We analysed participants’ accuracy rates (i.e., how often they matched the sentence with the correct picture) and response times. As before, we coded matches of ergative sentences with completed-state pictures, and nominative sentences with incomplete-state pictures, as ‘correct’.

4.4 Discussion of Experiment III

In our data (see Figure 7), even though Case and Event Type do interact in terms of the accuracy of participants’ responses, they do not in terms of speed. Sentences in the nominative condition elicited longer response times regardless of event type – but it was only when the verb was durative that this longer response time also resulted in higher accuracy.

The strong prediction hypothesis, wherein ergative-marked sentences should be matched to completed-event pictures as soon as the ergative marker is processed, was, therefore, not borne out: ergative case marking may have resulted in faster response times, but with accuracy rates hovering just above chance level, that speed did not translate to the kinds of high accuracy rates expected, based on Experiment II, and on the rates of correct sentence-to-picture matching in Norming Study II. The two chief differences between that study and this one are the presentation mode of the sentence (audio rather than written) and the instruction to respond as quickly as possible. However, despite that prompt, participants waited to respond until they heard the verb: the average response time for ergative sentences was roughly 700 ms after verb onset, and roughly 1000 ms after verb onset for nominative sentences. What may explain this discrepancy between the response times of the ergative and nominative conditions is the object case marking: since the direct object in the nominative condition is the first constituent with overt case marking, the parser needs to disambiguate thematic roles and forecast a transitive predicate, resulting in a slower reaction. A corpus check in the Hindi Dependency Treebank (Bhat et al., 2017) further supports this: the likelihood of encountering a transitive verb after hearing ergative -ne is very high, compared to other verb valencies.¹³ The sentence type that was, by some margin, most reliably matched to the correct picture was nominative and durative – in other words, an in-progress event with null subject case marking, where the matching picture was a prototypical representation of the verb content.

Figure 7: Average proportion of correct responses (a) and average response time in milliseconds after subject offset (b).

There is a possibility that the timing data is somewhat unreliable, because the experiment was run in the browser with remote participants each using different hardware set-ups, internet speeds, etc. Bridges et al. (2020), who ran an exhaustive study cross-testing the precision and accuracy of all the most popular browsers, operating systems, and behavioural experiment software libraries, note that though there are unavoidable timing lags (in the range of tens of milliseconds) that are variable across different (combinations of) hardware and software equipment, whatever lag exists does appear to be generally consistent within subjects, which is what matters. In any case, we were able to control for this factor by running Experiment IV in the lab, using the same hardware and software configuration for all participants.

In addition, reaction times alone, as in Experiment III, do not provide sufficient evidence for the unfolding of incremental language comprehension: While reaction times can reflect the speed at which an individual responds to a stimulus, they do not capture the underlying cognitive processes that occur during the decision-making and response selection stages. To gain a comprehensive understanding of incremental comprehension, Experiment IV replicated Experiment III using an eye-tracking paradigm.

5.1 Method & materials

We used the same materials and method as in Experiment III, collecting data via an EyeTribe portable eye-tracker and custom-made software based on jsPsych, using a localhost server to record participants’ eye-movements. Data were recorded at 30 Hz on a 15” MacBook Air, with participants sitting about 60 cm away, and the eye-tracker mounted on the bottom of the laptop screen (see Figure 8).

Figure 8: Participant seated in front of portable EyeTribe eyetracker setup (on bottom of the screen).

5.2 Participants

We paid 46 students of Indian Institute of Technology (IIT) Delhi to participate in this experiment, recruited by an IIT Delhi listserv. All participants were Hindi native speakers. We excluded two participants from the analyses because of calibration issues.

5.3 Results

Picture selection data. As before, we used logistic mixed models to determine whether the log-odds of selecting the correct picture were above chance level in each of the case/event type conditions; with random intercepts for items and participants. Results are presented in Table 11.

We then fitted a binomial mixed model with contrast-coded event type and case as fixed effects, and random intercepts for item and participant; using pairwise comparisons to determine significance. As in Experiment III, both the main effects of event type (χ² (2, 7) = 36.1, p < 0.0001) and case (χ² (2, 7) = 111.6, p < 0.0001) were significant, and so was their interaction (χ² (1, 7) = 36.1, p < 0.0001).

Eye-tracking data. To analyze eye-tracking behaviour, we excluded trials whose track-loss proportion was greater than 25%, resulting in a removal of 95 trials. We registered a track loss of 3.6% of all observations, leading to no exclusions of either participants or items. The gaze data are presented in Figure 9.

Figure 9: A. Experiment IV, percentage of looks to complete picture, B. by event type, C. by case. Grey shaded area indicates significant difference; color shaded areas mark confidence intervals. The green vertical line marks the average verb onset (1597 ms after subject offset).

In order to identify in which time windows participants’ looking behaviour diverged significantly between conditions, we performed a cluster-based permutation analysis (Maris and Oostenveld, 2007) using the R package eyetrackingR (Dink and Ferguson, 2015). In this analysis, gaze data was binned into 50 ms bins, and each bin was t-tested (with Bonferroni correction) to identify clusters of time bins during which looking preference was significantly above chance.¹⁴ To estimate the likelihood that these clusters represent genuine effects, this analysis was then repeated 1000 times with permuted predictor variables, resulting in a sum z-score distribution against which the original cluster sum z-scores were compared.

We performed separate analyses for the predictor variables Case and Event type. No significant clusters were found for Event type; for Case, we found one significant cluster, between 2460 and 3150 ms (sum Z = 80.78, p < 0.0001).¹⁵ In other words, between 2490 ms and 3150 ms after the subject was heard, 1500 ms before average sentence end, people looked more towards the incomplete picture in the nominative case.

However, ergative case marking did not lead to more looks to completed pictures (Figure 10), in line with the results from Experiment III, and the picture selection data. To better understand how participants’ picture selection speed related to their picture choice, we plotted by-participant mean response time (x-axis) against their mean accuracy rate, in a scatterplot with Spearman correlations for the ergative and nominative conditions. By doing so, we can see a bimodal distribution in the participants: some that respond early, before they have fully heard the verb phrase, and a larger group that responds later. The first group’s accuracy rate is around chance level. In the latter group, there is a clear divergence between the two Case conditions: a relatively tight cluster of high-accuracy responses in the nominative condition, and a wide range of responses in the ergative condition, with some participants almost never matching ergative-marked sentences with their completed-event picture; some preferentially matching ergative sentences with the completed picture; and the majority falling somewhere in the middle.

Figure 10: Experiment IV: (a) mean accuracy by Case and Event Type, and (b) scatterplot of mean accuracy vs. mean response time, by subject, with two Spearman correlations (by ergative and nominative condition) overlaid.

5.4 Discussion of Experiment IV

This experiment confirmed the timed picture-matching data from Experiment III: Hindi speakers do not use ergative case marking for drawing inferences about event structures at the time the case marking would allow them to. Instead, we observe a delay in interpretative commitment until the lexical verb is heard.

In plotting by-subject mean offline accuracy against mean reaction time, we could see that only a handful of participants chose a picture early (i.e. before the verb was fully processed), and that their accuracy was at chance in both the ergative and nominative conditions. The majority of participants chose a picture well after verb offset, where a cluster of above-chance mean accuracy could be observed in the nominative condition, and a much more dispersed scattering of high-to-low accuracy in the ergative condition.

Case marking alone is clearly not a predictive cue, but the time window in which looks towards the completed event diverge significantly between conditions is found before verb offset. Assuming about 200 ms processing time to plan and execute an eye movement (Huettig et al., 2011; Saslow, 1967), that suggests that comprehenders in the nominative condition heard, at most, the onset of the verb before looking towards the incomplete picture. This looking preference is brief: immediately after verb offset, looks return to chance level. But it is notable for the fact that it emerges before the comprehender encounters any potential aspectual morphology (or lack thereof).

The preferential looks towards the incomplete event in the nominative condition align with higher offline accuracy in the picture choice (89% for durative verbs, and 74% for punctual verbs). The comparatively low average offline accuracy scores in the ergative condition confirm that, (as observed by Arunachalam and Kothari, 2011) perfective simple verb predicates do not entail event culmination. Event type does appear to have an effect, with choice of the completed event picture marginally above chance in the ergative punctual condition in both Experiments III and IV – again, consistent with Arunachalam and Kothari (2011), who found generally lower acceptability ratings for partially completed punctual events. However, we found no evidence for any effect of Event type in sentence processing: this is likewise consistent with Foppolo et al.’s, 2021 results. Their reason for including the durative/punctual event contrast was that for punctual events, culmination is lexically encoded in the verb, and should, therefore, lead to faster, more decisive culmination inferences during incremental processing. They, like us, found no difference between the two event types – perhaps that is because this experimental manipulation is awkward to realize with punctual events. It requires that an event that, by its lexical aspectual definition, should not have an in-progress stage, or be split into its incomplete and completed components. The combination of static pictures with the sentence preamble (‘show me where…’), implying a frozen snapshot of an event the participant can peruse at leisure, may have coerced the punctual events towards a more progressive reading; or the combination of lexical and grammatical aspect may simply be more ‘elastic’ (in the words of Pederson, 2007: about Tamil) in Hindi, as suggested by the fact that punctual events with nominative case (i.e., imperfective aspect) normed just as high as the other three sentence types in Norming Study I.

All this suggests that, while there might not be any aggregated looking preference towards either picture in the ergative condition, there is considerable variation among individual participants in their ultimate interpretation of perfective sentences. What could explain the degree of freedom Hindi speakers have to decide whether an ergatively-marked subject has finished doing something or not?