1.1 Clausal ellipsis structures

Clausal ellipsis is ellipsis with a clausal source in which only a single element, the remnant, is overtly expressed. Cases of sluicing (2a; Ross 1969; Merchant 2001; van Craenenbroek 2010, 2012), fragment answers (2b; Merchant 2004), and stripping ellipsis (2c; Lobeck 1995; Merchant 2003; Nakao 2009; Wurmbrand 2017) illustrate:

The literature on ellipsis has spawned a wealth of approaches, which can be divided roughly into syntactic (Sag 1976; Williams 1977; Fiengo and May 1994; Chung et al. 1995; Fox 2000; among others) and non-syntactic (Sag and Hankamer 1984; Merchant 2001; van Craenenbroeck 2010; Aelbrecht 2010; among others) accounts. While our study does not crucially depend on a particular analysis, we will adopt a syntactic “move and delete” approach (Jayaseelan 2001; Merchant 2001, 2003, 2004; Depiante 2000; Ortega-Santos et al. 2014; Yoshida et al. 2015; among others), in which the remnant has moved from TP to the CP domain, before TP ellipsis takes place.⁴ For instance, the clausal source for the respective ellipsis structures in (2) would be spelled out roughly as (3) below. In these cases, the remnant has moved to the Spec of FocP, out of the material targeted for deletion (represented with a strikethrough):

In languages such as German, it has been proposed that the remnant of stripping ellipsis can function as a contrastive topic (4B) or contrastive focus (4C) (Konietzko & Winkler 2010). One of the diagnostics Konietzko and Winkler use to determine the discourse status of the remnant in stripping ellipsis is the relative placement of a sentential adverb and the remnant. In particular, contrastive topic remnants occur above (4B) but not below a sentential adverb (4B’), unlike contrastive focus remnants (4C).

The question in (4A) asks about a set of individuals, which can then be presupposed in the context, allowing for an answer that contains contrastive topic, as in (4B). On the other hand, in (4C) as an answer to (4A’), the remnant Hans represents new information; therefore, it functions as a focalized element (see Konietzko & Winkler 2010 for further discussion on how to differentiate topic and focus in German). In contrastive-topic ellipsis (4B), the remnant is assumed to move to TopP in the left-periphery while in contrastive-focus ellipsis (4C), the remnant is assumed to move to FocP, below the negation and sentential adverb (Konietzko & Winkler 2010).

In this paper, we will be concerned with related cases of Polarity stripping in Persian (5), in which the negative marker na follows the remnant, and a coordinator like but introduces the elided clause in which the polarity is reversed (Toosarvandani 2013, 2014; Rasekhi 2018, 2020).⁵ In (5), the contrastive topic remnant moāven contrasts with a correlate in the antecedent clause (see Rasekhi 2018, 2020 for evidence). However, the correlate-remnant pairing is formally ambiguous, as the remnant could contrast with either the object (5a) or the subject (5b) of the antecedent clause.

The underlying structure of the ellipsis for object (5a) and subject (5b) contrastive-topic remnants is assumed to be as in (6a) and (6b), respectively (see section 1.3.5. for further discussion on polarity stripping in Persian).

As illustrated in (6), the structure and the interpretation of the ellipsis depend on the choice of correlate-remnant pairing. An Object contrast places the remnant moāven “assistant” in object position within the ellipsis, whereas a Subject contrast places the remnant in subject position.

Correlate-remnant pairing itself appears to be sensitive to several factors, including the position of the correlate in the antecedent clause and a general preference for similarity between the remnant and its correlate. These factors will now be discussed in the context of how such decisions might be made during sentence processing.

1.2 Processing clausal ellipsis structures

There is a large and rapidly expanding literature on ellipsis in sentence processing, most of which will not be reviewed here (see, e.g., Phillips & Parker 2014; Frazier 2018; or Yoshida 2018). As a starting point, we assume that the sentence processor must engage in at least three tasks when processing clausal elliptical structures (Harris & Carlson 2018):

The steps in (7) are assumed to follow in the order presented. Assuming that the correlate and remnant match in syntactic category, the category of the remnant must first be parsed (Step 1) before the appropriate correlate can be found (Step 2). We further assume that the correlate must be located prior to constructing the ellipsis (Step 3), as the choice of correlate will determine the meaning of the ellipsis. We concentrate here on two factors that have been implicated in Step 2: Locality and Parallelism. Note that many of the studies reviewed below report reaction time measures. While our study is concerned with offline measures that probe the final interpretation, similar factors are likely to affect online and offline processing. This issue is discussed in more depth in the General Discussion.

1.3 Resolving Ambiguity in Persian Ellipsis

1.4 Current Study

The central questions that are addressed in this paper are as follows: Does Persian exhibit a general preference for Locality or Morphological Parallelism in the resolution of ambiguous stripping structures? When these two principles are placed in conflict, which one determines how the structure will be interpreted? In two experimental studies, we study how Locality and Parallelism interact and which, if either, Persian speakers prioritize in offline language comprehension. We explore three options.

First, if Locality is prioritized, we would expect the remnant to contrast with the structurally/linearly most local correlate regardless of Morphological Parallelism between the remnant and correlate. In this case, Object contrasts should be preferred except when they appear in sentence initial position via scrambling. Second, if Morphological Parallelism is prioritized, we would expect a strong preference for the correlate-remnant pair to share a similar morphological/semantic shape regardless of the correlate’s location.¹⁴ It should be noted that the test items in this study were designed so that both Subject and Object contrast interpretations were equally plausible pragmatically.

As a third option, it is possible that one of the Locality or Morphological Parallelism principles is the primary factor in locating a correlate for the remnant, but that the other, non-primary principle could still have a secondary impact in certain contexts – e.g., when the primary factor is insufficient to resolve the pairing. As discussed, English tends to prefer local correlates in clausal ellipsis, but this preference is reduced, but not entirely overturned, when a non-local noun receives pitch accent (Frazier & Clifton 1998; Carlson, Dickey, Frazier & Clifton 2009; Harris & Carlson 2018). In contrast, Persian speakers might in general rely on Morphological Parallelism, given the possibility of -rā marking the object, but nonetheless show sensitivity to Locality when the object is scrambled, in which case, the subject is the structurally or linearly more accessible correlate.

To determine which option holds in Persian, we used three kinds of antecedent clauses in our experiments (canonical SOV (34), canonical marked SO-rāV (35), and scrambled O-rāSV (36)) and two types of remnants (unmarked and -rā marked).¹⁵ It should be noted that -rā marked nouns are unambiguously objects, and thus remnants with rā marking are expected to contrast with objects in the antecedent clause. The -rā marked remnants were included in this study for two reasons. First, they provide a morphologically unambiguous baseline in a full factorial design. Second, previous studies have shown that although pragmatic and prosodic indicators of contrast overturn Locality in English, they nonetheless elicit online processing costs, which might manifest in offline ratings judgments (Harris & Carlson 2016, 2018). Even in German, where the case of the determiner disambiguates the structure, prosodic mismatches elicited processing difficulty (Stolterfoht & Bader 2004; Stolterfoht et al. 2007). Thus, there is good evidence that morphological disambiguation does not eliminate the influence of other factors in ellipsis resolution. In the rest of this section, we present our predictions regarding the role of Locality and Morphological Parallelism in interpreting ambiguous ellipsis structures in Persian.

Regarding canonical SOV antecedent clauses (34), if Persian speakers use Locality to match the remnant with its correlate, we expect an Object contrast interpretation to be preferred, in cases where the remnant is either -rā marked or not. However, if Persian speakers depend primarily on Morphological Parallelism for disambiguation, then the Object and Subject contrast interpretation should be roughly matched on balance when the remnant is not -rā marked. In this case, the remnant (“assistant”) and the object (“secretary”) and subject (“manager”) in the antecedent clause have a similar morphological form and meaning. With respect to Morphological Parallelism, an unmarked remnant matches with both subject and object nouns, whereas a -rā marked remnant matches with neither.

In canonical marked SO-rāV antecedent clauses (35), the object is -rā marked. If Locality is the dominant strategy in correlate-remnant pairing, then an Object contrast should again be preferred in this structure, regardless of the form of the remnant. Parallelism, however, predicts that the choice of correlate will entirely depend on whether the remnant is -rā marked or not.

In the third type of antecedent clause (36), the object has been scrambled and surfaces in the sentence-initial position. The -rā marking on the fronted object monshi-rā (“secretary”) is required by the grammar; see section 1.3.4. Here, Locality predicts a bias towards a Subject, rather than an Object, contrast. Morphological Parallelism again predicts that the choice of correlate depends on the morphological shape of the remnant.

This final case is especially important in determining whether Locality and Morphological Parallelism interact. If, for example, comprehenders largely prefer to match remnants with similar correlates but are still sensitive to the location of the correlate, then scrambling should weaken the bias for Parallelism.

The predictions of Morphological Parallelism and Locality are summarized in Table 1. While -rā marked remnants must unambiguously be interpreted as object remnants, we do not assume that the presence of -rā necessarily obviates the Locality bias, as discussed above. These cells are marked by an asterisk ‘*’ in Table 1. Although it is likely that -rā marking will have a strong effect on judgments, it is an empirical question whether other factors continue to influence sentence interpretation, as shown in previous studies on English and German.

The cases where Locality and Morphological Parallelism conflict are crucial for determining which principle, if either, is dominant in Persian. There are two such cases. In the first, Canonical -rā marked matrix clauses paired with an unmarked remnant should yield a Subject contrast preference under Morphological Parallelism, but an Object contrast preference under Locality. In the second, Scrambled -rā marked matrix clauses paired with a -rā marked remnant should produce a preference for the scrambled Object under Morphological Parallelism, and a bias towards Subject contrast under Locality. A methodological advantage of the design is that different interpretations are favored under different configurations, so that the patterns cannot be explained as a result of a more general bias towards object or subject correlates.

To summarize, Locality predicts that the structurally or linearly closest noun is the preferred correlate. In contrast, Morphological Parallelism predicts that the noun that shares the most semantic or morphological features is the preferred correlate. If the nouns are equally similar, they should be equally accessible as correlates. However, it is also possible that Persian, like English, does not employ one strategy exclusively in pairing a remnant with a correlate, but may instead be influenced by each factor. In that case, we would expect, say, that a general preference for Morphological Parallelism might be mitigated when a non-parallel noun is found in a more local position, as in the case of Scrambling. We address these possibilities in two experiments reported below: a naturalness rating study and a sentence completion study.

2.1 Participants

Sixty self-reported native speakers of Persian completed the experiment voluntarily without receiving payment.¹⁶ Subjects were recruited via email using snowballing sampling techniques to recruit additional participants. To identify inattentive readers, six catch items were included, but no participant was removed on this basis. All participants signed an informed consent online before starting the experiment and were told they could leave the experiment at any time they wanted.

2.2 Materials

Twenty-four sextets as in Table 2 were constructed in a 2 × 3 design, crossing Remnant type (No -rā in remnant and -rā in remnant,) and antecedent clause word order (Canonical SOV, Canonical -rā marked SO-rāV, and Scrambled O-rāSV). In the ellipsis clause, the remnant without -rā marking (a–c) is ambiguous between a Subject and an Object contrast interpretation, as the English translations at the bottom of the table show. However, adding the -rā marking on the remnant grammatically disambiguates the remnant as an object (d–f). As the Subject and Object contrast interpretations illustrate, the sentences in (d–f) do not have a Subject contrast interpretation as the remnant with -rā marking can only be interpreted as an object.

In addition to manipulating the -rā marking on the remnant, we also varied the word order in the antecedent clause. In the Canonical (SOV) word order, there is no -rā marking on the object, while in Canonical -rā marked (SO-rāV) and Scrambled (O-rāSV) word orders, the object is -rā marked. Test items are provided in Appendix A.

The items were carefully constructed to ensure that the remnant could pragmatically contrast equally well with both the subject and object nouns in the matrix clause. Interpreting the ellipsis with a Subject contrast interpretation (“a/the penguin did not catch fish/the fish”) and an Object contrast interpretation (“a/the shark did not catch a/the penguin”) are equally plausible.¹⁷

2.3 Procedure

The questionnaire was conducted on Ibex Farm (Drummond 2013). Materials were presented in a counterbalanced and randomized order so that the subjects saw only one sentence from each sextet for any experimental item. The 24 experimental sentences were presented with 10 sentences from an unrelated experiment and 20 non-experimental filler sentences, and 6 catch sentences. Each participant saw a total of 60 sentences. Completing the questionnaire took approximately 15 minutes on average.

Participants had two tasks to perform. First, they were instructed to rate naturalness of items like in Table 2 on a Likert scale of 1 (completely unnatural) to 7 (completely natural). Second, they were asked to select an interpretation of the item by choosing one of four options (37), which indicated (a) a Subject contrast, (b) an Object contrast, (c) both, or (d) “I am not sure”. Half of the test items had Subject contrast answers followed by the Object contrast answer, while in the other half the order was reversed.

We first present results of the naturalness rating task, followed by the interpretation question task. Our predictions for the naturalness ratings portion were as follows. Both Locality and Morphological Parallelism play a role in rating how natural the sentences are; however, if Locality takes priority over Morphological Parallelism, then sentences with a remnant that contrasts with a local correlate should receive higher naturalness ratings. Specifically, Locality predicts an interaction between conditions, in which a -rā marked remnant should receive higher naturalness ratings in the Canonical -rā marked condition, as the -rā marked object is the most local noun, but lower naturalness ratings in the Scrambled -rā marked condition, as the -rā marked object is non-local. However, if Morphological Parallelism is prioritized over Locality, then -rā marking should improve naturalness ratings whenever there is -rā marking on both the remnant and a potential correlate in the antecedent clause (Canonical -rā marked and Scrambled clause types).

Pairing interpretation questions with naturalness ratings allows us to determine how the sentences were interpreted. To review, we predicted that ambiguous remnants, without -rā marking (cells a–c in Table 2), can have both Subject and Object contrast interpretations since the remnant can match with both the subject and object in the antecedent clause. However, Morphological Parallelism would predict a preference for Subject contrasts for unmarked remnants (cells b and c), as the object DP in the antecedent clause is -rā marked. However, Locality predicts a general bias towards local correlates (an Object contrast preference in (a–b) and a Subject contrast preference in (c)) regardless of morphology. In contrast, in remnants with -rā marking (conditions d–f in Table 2), we expected to observe a strong bias towards the Object contrast interpretation as -rā marking occurs only with object DPs. The Scrambled condition is crucial in that it places the two principles in conflict, and generates a scenario in which we might observe a preference to pair -rā marked remnants in interpretation, but a cost associated with violating Locality in the naturalness ratings.

2.4 Naturalness ratings results

The median naturalness rating for each condition is provided in Table 3. Although we only analyze the median values statistically, we also report means and standard errors, shown also in Figure 1B, for comparison with studies that have not used ordinal regression models.

Figure 1

Experiment 1: Probabilities of responses obtained from the first ordinal fixed-effect regression model (Panel A, top). Naturalness ratings means for each condition (Panel B, bottom).

Values on the Likert scale were treated as inherently categorical, instead of numeric or metric. We analyzed the results as an ordinal linear regression (cumulative link logit) with fixed and random effects using the ordinal package (Christensen 2019) in R (R Core Team 2020). These models include threshold coefficients, in addition to fixed effects. While we will be interested in fixed effects exclusively, the threshold coefficients serve a similar role as the Intercept in linear models, and are included for replicability. Briefly, a threshold coefficient x|y corresponds to the log odds increase of a value x compared to all the higher ordered categories y, provided that all other variables in the model are held constant. For example, the coefficient of 4|5 corresponds to the log odds of any particular item receiving a rating of 4 or lower, rather than a higher rating of 5, 6, or 7 on a 7-point scale. Threshold coefficients provide probability thresholds between ordered categories, ignoring any effects of predictor variables. The probabilities of responses from the model are shown in Figure 1A.

Figure 2

Experiment 1: Percent response selected for Subject, Object, and Both responses in interpretation questions. “I am not sure” responses were excluded.

Treatment coding was used to explore the effect of each factor and their interactions against the baseline condition of a Canonical antecedent with an Ambiguous Remnant (Table 2, cell a). The model is presented in Table 4. Compared to the baseline, there was no effect of -rā marking on the remnant (cell a vs. d). There was also no effect of a -rā marked object in the Canonical Marked with an Ambiguous remnant condition (cell a vs. b). However, Scrambling (cell a vs. c) elicited overall lower naturalness ratings in the Ambiguous remnant condition, z = –3.63, p < .001.

In investigating the interactions between Clause Type and Remnant Type, we compared the effect of a -rā marked remnant in Canonical Marked (the difference between cells b and e) and Scrambled (the difference between cells c and f) against the baseline effect in the Canonical antecedent order (the difference between cells a and d). Both interactions were found to be significant; whereas adding -rā marking to the remnant after matrix clauses in sentences with the baseline Canonical order reduced naturalness ratings by 1 point, -rā marking on the remnant increased ratings by 1 point following Canonical Marked (b vs. e), z = 5.84, p < .001, and Scrambled (c vs. f), z = 5.54, p < .001, antecedent clauses.

An analysis of the model’s marginal means was conducted to further investigate the interactions using emmeans (Lenth 2021). In Canonical clauses, an Ambiguous remnant was rated no differently than rā-marked remnants (95% CI [–0.28, 0.41], p = 0.72). However, Ambiguous rated lower than rā-marked remnants following both Canonical Marked (95% CI [–1.81, –1.08], $\widehat{\beta }=-1.44$ , SE = 0.19, z = –7.72, p < .001) and Scrambled (95% CI [–1.68, 0.98], $\widehat{\beta }=-1.33$ , SE = 0.18 z = –7.48, p < .001) antecedent clause orders.¹⁸

Since both Canonical Marked and Scrambled conditions contained -rā marked DPs in the matrix clause, we constructed a second ordinal regression model to determine whether Scrambled word orders (Table 2, cells c and f combined) had an effect over and above -rā marking in the Canonical Marked condition (cells b and e combined).¹⁹ In this model, Canonical Marked and Scrambled conditions were first collapsed to test for a general effect of -rā marking. The conditions were then compared against each other, removing the Canonical condition (cells a and d). User-defined contrasts were accordingly specified to compare interactions of theoretical interest, in particular (i) the Combined effect of -rā marking in Canonical Marked and Scrambling conditions (contrasts: Canonical = –0.66, Canonical Marked = .33, Scrambled = .33), and (ii) the effect of Scrambled over Canonical Marked conditions (contrasts: Canonical = 0, Canonical Marked = –.5, Scrambled = .5).

In this second model, naturalness ratings increased when the remnant was -rā marked, $\widehat{\beta }=0.45$ , SE = 0.05, z = 8.62, p < .001. Compared to Canonical Marked antecedents (cells b and e), Scrambling was associated with decreased ratings (cells c and f), $\widehat{\beta }=-0.64$ , SE = 0.13, z = –4.98, p < .001. In addition, there was a general naturalness ratings advantage for the combination of Canonical Marked and Scrambled conditions, i.e., items with -rā marking in the antecedent clause, over the Canonical word order, $\widehat{\beta }=0.40$ , SE = 0.11, z = 3.66, p < .001.

Interactions indicated that conditions with a -rā marked matrix object patterned together with respect to Remnant Type. While -rā marking on the remnant had no effect on structures with Canonical word order (the difference between cells a and d), it improved naturalness ratings when the object in the matrix clause was also -rā marked (cells b and c compared to e and f), $\widehat{\beta }=0.73$ , SE = 0.11, z = 6.57, p < .001. However, the size of the effect of -rā marking was no different for Canonical Marked antecedents (the difference between cells b and e) compared to Scrambled word orders (the difference between cells c and f), z = –0.43, p = 0.67, resulting in a 1-point increase for both conditions.

2.5 Interpretation question results

The “I am not sure” responses accounted for less than 4% of responses, and were removed from the data prior to analysis for convenience. The mean percent of interpretation responses for each condition is provided in Table 5, along with χ²-tests on the distribution of Subject, Object, and Both answers in each of the conditions. The distribution of responses is presented in Figure 2.

The fact that the dependent variable is an unordered nominal response makes both linear mixed-effects and cumulative link ordinal models inappropriate for analysis. We therefore conducted a series of χ²-tests over the distribution of raw counts for each response within each condition. Sentences were effectively disambiguated by -rā marking on the remnant, garnering approximately 85% of Object-contrast interpretations overall, regardless of word order and -rā marking in the matrix clause. Morphologically ambiguous remnants, i.e., those that were not -rā marked, generated a more evenly distributed set of responses. Morphologically ambiguous remnants showed a bias towards the Subject contrast interpretation in both Canonical Marked and Scrambled, but not Canonical, orders.

2.6 Discussion

Results from the naturalness ratings task were compatible with the predictions made by Morphological Parallelism. First, remnants with and without -rā marking were rated as equally acceptable in Canonical clauses. Second, while there was an advantage when -rā marking appeared on both the remnant and the object, the size of the effect was not decreased in Scrambled word orders. The patterns observed in the interpretation question results support the claim that Morphological Parallelism plays a crucial role in remnant-correlate pairing when resolving stripping ellipsis in Persian.

The perception of acceptability and naturalness is influenced by a great number of factors, including complexity, plausibility, frequency, and so on (e.g., Chomsky 1965; Newmeyer 1983; Schütze 2011; Sprouse 2018). For example, interpreting the remnant as a specific object may have been more plausible in the particular experimental sentences provided, making these sentences generally easier to interpret. In addition, our choice of nouns might have implicitly introduced confounding bias into the results. In Table 2, for example, subjects could have relied on extra-grammatical knowledge in interpreting the remnant penguin as contrasting with the matrix object fish rather than the subject shark, even though great care was taken to avoid such bias during stimulus design. Another possibility is that participants may have generally preferred sentences with less ambiguity, resulting in a general ratings advantage for sentences with a -rā marked remnant and correlate. Although we think such confounds are unlikely given the plausibility pre-test and the results (especially as the Canonical ambiguous form showed no difference between remnant types), we conducted a follow up study in a method that avoids imposing a specific remnant onto participants, and that allows us to further confirm the patterns in a more open-ended fashion.

3.1 Participants

Twenty-two self-reported native speakers of Persian participated in the experiment without compensation, and were recruited via the same method as Experiment 1.²⁰ Six catch items were included to identify inattentive participants, but no one was removed on this basis.

3.2 Materials

Twenty-four sentence fragments with Canonical, Canonical -rā marked, and Scrambled antecedents were constructed in which the remnant noun was left blank; see Table 6. Items were designed to be equally compatible with animate/agent (Subject) and inanimate/non-agent (Object) nouns.²¹ This was designed to help us determine whether the participants preferred a Subject or Object contrast interpretation. All items are provided in Appendix B.

3.3 Procedure

The study was administered to participants over the Internet and responses were collected with Ibex Farm. In addition to the 24 experimental sentence fragments, the questionnaire included 43 non-experimental filler fragments and 6 catch fragments. Items were presented in a counterbalanced and individually randomized order so that the subjects responded to only one condition from each triplet. Each participant saw a total of 73 sentences. Completing the questionnaire took less than 20 minutes on average.

Participants were instructed to fill in the blank with the first appropriate word that occurred to them. The animacy/agency of the remnant noun disambiguated the sentence. When participants supplied an animate/agent noun, the remnant was taken to contrast with the subject. However, if they completed the blank with an inanimate/non-agent noun, the remnant was assumed to contrast with the object.

The predictions of the completion study largely follow the previous experiment. A preference for Local correlates would result in a propensity to provide a remnant that contrasted with the closest noun (the object in Canonical sentences and the subject in Scrambled sentences). A preference for Morphological Parallelism would elicit remnant completions that match the correlate, regardless of position.

3.4 Results

Completion responses were categorized according to whether the noun provided was animate/agent or inanimate/non-agent and whether it was -rā marked or not. By hypothesis, animate/agent responses indicated a Subject contrast, whereas inanimate/non-agent responses indicated an Object contrast. Results are provided in Table 7.

As subject remnants were never -rā marked, the data were categorically skewed, and a logistic linear mixed-effect regression model of the data could not be fit due to complete lack of variance. Rather than remove the subject remnant level, the relevant comparisons were made with exact one-sided binomial tests on the observed count data. The 95% confidence interval (CI) is reported along with the significance level of the test for significant comparisons. We first report an analysis that ignores -rā marking on the remnant (shown in the Total column in Table 7), followed by an analysis that distinguishes responses on the basis of -rā marking (shown in the final two columns in Table 7).

Ignoring -rā marking on the remnant, approximately half (49%) of the remnant completions were inanimate/non-agent, indicating an Object contrast. Object contrasts were numerically preferred in Canonical SOV (M = 59%) and Canonical SOV Marked (M = 56%) conditions, but did not differ significantly from Subject contrasts. However, Scrambled OSV conditions showed significantly fewer Object contrast completions (M = 32%), 95% CI [.23, .43], p < .001, in favor of Subject contrasts (68%).

Distinguishing completions by -rā marking on the remnant, there were no cases of -rā marking on animate/agent completions (Subject contrasts) in any condition, which is likely due to the fact that -rā marking is reserved for objects. In addition, there were no cases of -rā marked (object) remnants following a Canonical matrix clause. In conditions with -rā marked objects in the matrix clause, inanimate/non-agent remnants (Object contrasts) tended to also receive -rā marking; Object contrasts were -rā marked more often than not in Canonical Marked (35 of 54 cases) and Scrambled (29 of 32 cases) word orders, 95% CI [.51, .77], p < .05, and 95% CI [.75, .98], p < .001, respectively.²²

Conversely, contrasts elicited with unmarked remnants also showed a strong effect of Antecedent clause type. While unmarked remnants were essentially evenly distributed between Subject and Object contrasts (56 vs. 49 cases), they were more likely to associate with Subject contrasts in Canonical -rā marked (42 of 61 cases) word orders, 95% CI [.20 .44], p < .01, and Scrambled -rā marked (65 of 68 cases) word orders, 95% CI [.88 .99], p < .001.

3.5 Discussion

Remnant completions provided by participants showed a strong sensitivity to clause type and the presence of -rā marked objects in the antecedent. In the Canonical clause condition, responses appeared to be guided by Parallelism, as no remnant was marked with -rā, even though participants rated -rā marked remnants on par with unmarked remnants in Experiment 1. Although there was a numerical bias towards local correlates in Canonical unmarked and Canonical -rā marked clauses, Object and Subject contrasts were provided at statistically equivalent rates (though see footnote 22). Completions supplied in the Canonical -rā marked clause condition indicate that Morphological Parallelism had a non-categorical influence on responses. Participants provided significantly more -rā marked objects in this condition, though bare objects were also provided.

However, the pattern of completions provided for Scrambled antecedents showed a sensitivity to both Locality and Morphological Parallelism. Compared to the baseline condition, participants provided significantly more Subject contrast completions in Scrambled clauses, indicating an increased preference for local correlates in this condition. Participants also provided fewer unmarked Object contrasts, indicating a stronger preference for parallel contrasts.

In addition to corroborating the central findings of the interpretation question portion of Experiment 1, the results provide evidence that Persian speakers are sensitive to Locality in Scrambled conditions. However, having a -rā marked remnant match with the -rā marked object in the antecedent clause (in order to maintain Morphological Parallelism) would violate Locality. While these preferences may not be encoded in the grammar, they do appear to strongly affect how comprehenders disambiguate ellipsis structures, interacting in seemingly complex ways.