2.1 Linear approach: Marušič et al. (2015); Willer-Gold et al. (2016)

Marušič et al. (2015) focused on gender agreement with conjoined subjects in Slovenian and they conducted written and spoken elicitation experiments. Their results reveal that part can show agreement with C1, C2, or masculine agreement in the subject-part order as shown in (4a). However, part can only show agreement with C1 and masculine agreement but not C2 agreement in the part-subject order (4b). Willer-Gold et al. (2016) conducted similar experiments in both BCS and Slovenian and the patterns observed are largely the same as those in Marušič et al. (2015), see (1)–(2) for examples in BCS.

To capture these patterns, Marušič et al. (2015) follow Bhatt & Walkow (2013) and propose an analysis where linear order can condition agreement. Although the analysis in Marušič et al. (2015) is proposed for Slovenian, Willer-Gold et al. (2016) largely adapt it for BCS.

In this approach, the Agree operation proceeds in two steps: Agree-Link (establishing a dependency between the controller and the target) and Agree-Copy (copying the feature from the controller to the target). Furthermore, they assume that the ConjP has a number feature, but lacks a gender feature. In conjunction agreement, part first Agree-Links with the ConjP, but cannot get any gender feature. In this situation, three grammars are available in Slovenian/BCS. In Grammar I, the probing head never looks inside the ConjP (‘No-Peeking’ in their terms). Instead, a default masculine feature is inserted on ConjP. part will then Agree-Copy this M feature, resulting in masculine agreement. On the other hand, in the absence of a gender feature on ConjP, the probing head in Grammars II and III looks into the ConjP (‘Peeking’or ‘No-Default’) and searches for a gender feature on one of the conjuncts. In this case, two locality considerations are relevant in choosing which conjunct to agree with. In Grammar II, Agree-Copy takes place in PF before linearization, part thus copies the gender feature of the hierarchically higher conjunct, i.e. the first/highest conjunct. In Grammar III, Agree-Copy takes place in PF after linearization. As a result, part would copy the gender feature from the conjunct that is linearly closest to it, i.e. the first conjunct in the part-subject order and the last conjunct in subject-part order.

This linear approach correctly predicts the patterns in BCS and Slovenian mentioned above. In sentences with the subject-part order, Grammar I will result in the default masculine agreement, Grammar II will result in agreement with the higher conjunct, i.e. C1, while Grammar III will result in agreement with linearly closest conjunct, i.e C2. The schematic illustration of Grammar II and Grammar III is shown in (5).

In sentences with the part-subject order, Grammar I will again result in the default masculine agreement, Grammar II results in agreement with C1 as it is the higher conjunct, Grammar III also results in C1 agreement as C1 is the linearly closest conjunct to part in this word order as is shown in (6). Note that there is no way for C2 to control agreement on part in this order as it is neither the highest nor the closest conjunct.

2.2 Non-linear approach: Murphy & Puškar (2018)

Unlike the linear approach, the analyses within the non-linear approach argue that the linear effect in conjunction agreement results from purely hierarchical relations. In other words, linear order plays no role in deriving the agreement patterns. These analyses include van Koppen (2005) for Dutch dialects and Bošković (2009); Murphy & Puškar (2018) for BCS. In this paper, I focus on the analysis proposed by Murphy & Puškar (2018) as their empirical claims are more in line with the experimental results from Willer-Gold et al. (2016).

Murphy & Puškar (2018) argue that the agreement patterns observed in BCS result from different orders of operations in narrow syntax including merge of the conjuncts, upward agree (↑agr↑), and downward agree (↓agr↓). The agreement process takes place in two cycles: inside the ConjP, the Conj head agrees with the conjuncts and projects its value onto the ConjP; external to the ConjP, part agrees with the ConjP. They assume that i. the order of the operations within each cycle is in principle free, ii. that the order of ↑agr↑ and ↓agr↓ is constant inside and outside the ConjP in one derivation, and iii. that EPP movement of the agreement controller to the Spec,PartP position is only driven to feed upward agreement. The authors argue that this analysis generates all the attested patterns in BCS and rules out the unattested pattern, i.e. C2 agreement in the part-subject order, is only compatible with a derivation in which movement to Spec,PartP takes place. Different orders and the generated agreement patterns are summarized in Table 1.

The readers are referred to Murphy & Puškar (2018) for the detailed derivations of all the possibilities. In this paper, I use C2 agreement in subject-part order in (7) to illustrate how the ‘closest conjunct agreement’ (CCA) is derived in this system. In (7), the part prodata ‘sold’ shows neuter agreement with the second conjunct sva odela ‘all suits’.

According to Murphy & Puškar (2018), the agreement pattern in (7) is generated with the order ↑agr↑ ≫ merge ≫ ↓agr↓. First, as shown in (8), the Conj head agrees upward (dashed line) and does not find a DP, since merge of the conjuncts has not occurred yet. After that, both conjuncts merge with the Conj head. The Conj head then agrees down to get the feature from the second/lower conjunct, shown in (9). The resulting ConjP projects the feature of the second conjunct i.e. npl. External to the ConjP, since the order of ↑agr↑ and ↓agr↓ is constant inside and outside the ConjP in one derivation, the part head agrees upward first and requires the movement of the ConjP to the Spec,PartP position (solid line), as shown in (10). After the movement, part gets the npl feature from the ConjP. The result sentence (7) is one where part shows agreement with the second conjunct while the ConjP is in the preverbal position. On the surface, it is a CCA pattern; however, linear order plays no role in deriving the pattern.

As shown, both the linear and the non-linear approach can account for the agreement options available in BCS (and Slovenian). The advantage of one approach over the other is argued for based on conceptual reasons. In what follows, I will provide an empirical argument that teases the approaches apart.

2.3 Coordinate Structure Constraint violating movement in BCS

Having established the basic patterns of conjunction agreement and the two approaches under discussion, this section introduces an important piece of the argument: Coordinate Structure Constraint violating movement in BCS. Ross (1967) observes that moving one conjunct out of the conjunction is banned and proposes the Coordinate Structure Constraint (CSC) formulated in (11).

CSC contains two components: the Conjunct Constraint, a ban on movement of the conjuncts, and the Element Constraint, a ban on movement out of the conjuncts, following terminology from Grosu 1973; Pollard & Sag 1994. Since Grosu 1973, the two components have been argued to be independent from each other (see Oda 2021 for a recent review). In this paper, we will focus on the Conjunct Constraint and leave the Element Constraint aside. Thus I will use the label Coordinate Structure Constraint violating movement (hereinafter CSCV) to refer specifically to movement of the first conjunct out of the conjunction.

While the ban on movement of conjuncts is observed in many languages, there are cases of CSCV in selected languages. The sentence in (12) shows such a case in BCS: the first conjunct is moved to the sentence initial position, out of the conjunction. See Bošković 2009; Stjepanović 2014; 2017; Oda 2017; Arano & Oda 2019; Arsenijević et al. 2020a; Gračanin-Yuksek & Arsenijević 2017; Oda 2021 for discussion of CSCV movement and similar constructions.

CSCV like (12) has been reported but the judgments are subject to speaker variation.

2.4 Predictions on conjunction agreement with CSCV

The non-linear and linear approaches to conjunction agreement make different predictions for sentences with CSCV.⁴ As a reminder, sentences in the part-subject order are repeated below: first conjunct agreement (C1) and default masculine agreement are both allowed and second conjunct agreement (C2) is not.

The first conjunct moves to the sentence initial position through scrambling/topicalization in sentences in (14).

Let’s now look at the predictions the three grammars of the linear approach make. Grammar I does not peek inside the ConjP but inserts a default M feature in the ConjP in narrow syntax. In this grammar, the inserted default M feature controls agreement in sentences with CSCV just like the sentences without CSCV, as the ConjP is assumed to not compute a gender feature. Thus, Grammar I predicts masculine agreement in (14).

In Grammars II and III of the linear approach, no default feature is inserted, the probing head part looks inside the ConjP. In Grammar II, Agree-Copy is postponed to PF and occurs before linearization. In sentences without CSCV, this results in part agreement with the highest/first conjunct. In sentences with CSCV, if we follow the standard assumption that the lower copy or trace of the first conjunct is not present in PF, the second conjunct would be the only conjunct in the conjunction and thus the highest conjunct. In this case, C2 agreement is predicted.

The reviewers suggest another possibility: in Grammar II with the ‘Peeking’ preference, part Agree-Links with both conjuncts prior to the movement of C1, and this dependency between part and the conjuncts persists even when C1 moves away. Since the moved C1 is hierarchically higher than C2, in Grammar II, part could choose C1 to copy the features from. This interpretation of Grammar II would additionally predict C1 agreement. It is worth noting that this interpretation differs from the original proposal by Marušič et al. (2015) that the current paper is engaging with (see Section 4.2 in Marušič et al. 2015 for the specific implementation of Grammar II and III). In particular, the Agree-Link would have to be between part and each conjunct, rather than the ConjP.⁵

However, as will be discussed, the non-linear approach also predicts CSCV sentences to show C1 agreement. What is important is that the linear approach predicts multiple possibilities including C1, C2, and M agreement, while the non-linear approach predicts C1 agreement only. Thus, including the prediction of C1 agreement in Grammar II does not change the argument that will be developed.

In Grammar III, Agree-Copy is postponed to PF and occurs after linearization. In this case, the part looks inside ConjP and agrees with the conjunct that is linearly closest to part. Since in (14), the first conjunct is moved away in narrow syntax, the second conjunct would become the linearly closest conjunct in the conjunction. It is thus predicted that the second conjunct can control agreement in (14), i.e. C2 agreement. The predictions from Grammars II and III are shown in (15).

In sum, the linear approach predicts sentences with CSCV to show second conjunct agreement and masculine agreement. First/highest conjunct agreement is also predicted under a certain interpretation of Grammar II discussed above.

Now we turn to the non-linear approach. Since Move and Agree are connected in the non-linear approach, it is necessary to accommodate CSCV in the system. Following Murphy & Puškar (2018)’s assumption that movement of the subject that results in subject-part order is required by Upward Agree (↑agr↑), I assume that when part probes upward and does not find a goal, either the entire conjunction or the first conjunct moves to Spec,PartP in BCS to feed ↑agr↑. Moving the entire conjunction results in the subject-part order, and moving the first conjunct results in a sentence with CSCV.

It then follows that sentences with CSCV can only be generated in derivations where ↑agr↑ is ordered before ↓agr↓. In derivations where ↓agr↓ occurs first, CSCV (or movement of the entire conjunction) would not be required and thus does not occur. This restricts the relevant orders to the three in (16).

Since movement is required by ↑agr↑ and it is the moved element that controls agreement, all three orders in (16) predict that the CSCVed first conjunct would control agreement. Masculine agreement or second conjunct agreement cannot be generated.

Take the order in (16a) for example, within the ConjP, the Conj first probes up and does not find a controller because the conjuncts have not been merged yet, as shown in (17). The conjuncts then merge with the Conj head, and the Conj head probes down and gets the feature from the second conjunct. The ConjP will end up with the feature of the second conjunct as is shown in (18). Outside the ConjP in (19), part probes up (dashed line) and requires movement of the first conjunct (solid line). The moved first conjunct thus controls the agreement on part.

The result for the other two orders is similar. In the order in (16b), within the ConjP, both the conjuncts merge with the Conj head and the Conj head gets the features from both conjuncts. Outside the ConjP, part probes upward first, like the previous derivation, and the CSCVed first conjunct controls the agreement on part. In the order in (16c), within the ConjP, the Conj probes up and down first but does not find a controller. The conjuncts then merge with the Conj. No features end up on the Conj head or the ConjP. Outside the ConjP, like the previous two derivations, part probes up and the first conjunct moves up to feed it, which ends up controlling the agreement on part.

As shown, all three orders of the non-linear approach predict the moved first conjunct to control agreement on part. In other words, for CSCV sentences, the non-linear approach predicts only first conjunct agreement to be available, and masculine agreement or second conjunct agreement to be unavailable.⁶ With the predictions worked out, we are ready to see the empirical patterns of agreement in sentences with CSCV.

3.2 Status of CSCV

First we look at the status of CSCV in BCS. The sentence in (20) is a base sentence with post-verbal coordinated subjects and no CSCV. On the other hand, sentences in (21) and (22) involve the first conjunct moving out of the conjunction to the sentence initial position, (21) being scrambling and (22) being wh-movement.⁷ These sentences were included in Survey 1. Based on the responses from 30 participants, the mean rating for the base sentence is 4.6 out of 5, which indicates that the sentence is acceptable. On the other hand, (21) gets a mean rating of 2.6, below the midpoint of the scale, and (22) a mean rating of 1.37, close to the lower end of the scale.

The rating in the middle of the scale could result from a uniform mid-point rating by most of the speakers, or a divide among the speakers, some accepting the sentence while others rejecting it. Thus we look at the distribution of the judgments among the speakers. Table 2 shows the distribution of individual ratings as well as the mean and median ratings. The numbers in the columns labeled 1–5 indicate the number of speakers who chose that point of the scale for the corresponding sentence.

As is shown, (20) and (22) show uniform distributions. 27 speakers rated (20) as 4 or 5, and 2 rated it as 1 or 2, meanwhile 27 speakers rated (22) as 1 or 2 and only 1 rated it as 4 or 5. We can thus conclude that (20) is generally accepted and (22) is not accepted. In contrast, the ratings for (21) show more speaker variation: 16 out of 30 speakers rated it as 1 or 2, and 9 speakers gave it a rating of 4 or 5. In other words, the relatively low mean rating of (21) does not result from a degraded status of CSCV in general (like for (22)), but rather from speaker variation. From the mean and the distribution, I conclude that CSCV is accepted by some speakers when it is scrambling, but it is generally unacceptable when it involves wh-movement. This finding is compatible with the fact that the existing CSCV examples in the literature involve scrambling and not wh-movement. Although this asymmetry between wh-movement and scrambling is an interesting fact worth investigating, we will focus on CSCV-scrambling in this paper.

Another property of CSCV that will be crucial for the argument is that CSCV feeds variable binding, which puts CSCV in the narrow syntax, rather than PF.⁸ Following the standard assumption, structures in narrow syntax are spelled out to PF and LF. Movement that occurs in narrow syntax is thus able to affect interpretation in LF. On the other hand, movement in PF cannot affect the interpretation. If CSCV has interpretive effects, e.g. making a reading available, it cannot be a PF movement.

To test this, the sentences in (23) were also included in Survey 1. The sentence in (23a) is the base sentence without CSCV and in (23b), the first conjunct ‘every general’ is moved to the sentence initial position, c-commanding the subject ‘his soldiers’.

The relevant reading is one where ‘every’ binds ‘his’ in ‘his soldiers’, which can be paraphrased as ‘for every general, the general’s soldiers love the general and the soldier’s own country.’ Since in (23a), ‘every general’ does not c-command ‘his soldiers’, the bound variable reading is predicted to be unavailable. On the other hand, ‘every general’ is moved to a c-commanding position for ‘his soldiers’ via CSCV in (23b). If CSCV occurs in the PF, no interpretive difference is predicted between (23a) and (23b). If CSCV occurs in narrow syntax, (23b) is predicted to make the bound variable reading available.

To probe the pattern, the participants were asked ‘Whose soldiers are they?’ after each sentence in (23) and were given two options in (24). Choosing (24a) ‘the generals” would indicate the bound variable reading and (24b) would indicate a non-bound variable reading where ‘his’ is not bound by ‘every’.

For (23a), only 5 speakers out of 30 chose (24a) and 25 speakers chose (24b), indicating that the bound variable reading is not available in (23a) as is predicted by the standard assumption. Crucially, (23b) shows the opposite pattern: 25 participants chose (24a) and 5 chose (24b), indicating that CSCV made the bound variable reading available. I thus conclude that CSCV does not occur in PF but rather is a movement in narrow syntax.⁹

3.3 Conjunction agreement in CSCV

Turning to conjunction agreement, Survey 1 also includes the sentences in (25) and (26). The sentences in (25) (repeated from (13)) are baseline sentences in part-subject order without CSCV. part shows first conjunct agreement (C1 agreement) in (25a), masculine agreement in (25b), and second conjunct agreement (C2 agreement) in (25c).

Based on the responses from the 30 participants, first conjunct agreement has a mean rating of 4.1. Masculine agreement in (25b) has a mean of 4.27. In contrast, second conjunct agreement has a rating of 2.03, notably lower than the other two agreement options. The mean ratings replicated existing claims in the literature, as well as experimental results, indicating that for part-subject sentences, both first conjunct agreement and masculine agreement are available while second conjunct agreement is not. The distribution, mean, and median of these sentences are shown in Table 3. There was no major variation among the participants.

The sentences in (26) are counterparts of (25) with CSCV. The first conjunct moves to the sentence initial position in all three sentences, with part showing first conjunct agreement in (26a), masculine agreement in (26b), and second conjunct agreement in (26c).

Based on responses from 30 participants, the mean rating for C1 agreement is 2.3 out of 5; the mean rating for masculine agreement is 1.3 and the mean for C2 agreement is also 1.3. The distribution, mean, and median for each option are shown in Table 4. Ratings of (21), which also involves CSCV scrambling but with no agreement, are also included for comparison.

The mean rating of C1 agreement with CSCV (26a) is similar to that of CSCV with no agreement (21) at around 2.3–2.6. C2 agreement and masculine agreement, on the other hand, are rated lower at 1.3. C1 agreement with CSCV also shows similar distribution profile as (21): 20 speakers gave it a rating of 1 or 2, while 8 gave it 4 or 5. In contrast, 28 out of 30 speakers gave masuline agreement a rating of 1–2 and only 2 rated it as 4–5; 29 speakers gave C2 agreement a rating of 1–2 and only one rated it as 4–5. It is clear that the speakers uniformly rejected C2 agreement and masculine agreement in sentences with CSCV, while C1 agreement in CSCV is rated on par with CSCV with no agreement.

Although there is a difference between C1 agreement and the other two agreement options in CSCV, none of the agreement options are rated above the midpoint of the scale. I propose two reasons for the low rating: 1. as is discussed above, CSCV in BCS shows speaker variation independently from agreement patterns, the mean ratings also include the low ratings of speakers who do not allow CSCV in general; 2. even for the speakers who do allow CSCV, sentences with CSCV are not rated as high as other more ‘vanilla’ constructions in (20) and (25), likely due to the low frequency or the processing difficulty of CSCV constructions.¹⁰ It is thus important to keep in mind that it is the contrast between the agreement options within each construction that the approaches to agreement make different predictions about.

The distinction between C1 agreement in CSCV and the other two agreement options is more obvious if responses from the 16 speakers who gave CSCV scrambling (21) a rating of 1–2 are set aside. These speakers are likely to not allow CSCV in BCS in general. Distribution, mean ratings, and medians of the sentences discussed so far from the remaining 14 speakers are summarized in Table 5. Note that the acceptability patterns for sentences with no CSCV remain the same: C1 agreement and masculine agreement in (25a) and (25b) are rated at 4.2 and C2 agreement in (25c) is rated as 2. In sentences with CSCV, the wh-question in (22) remains unacceptable, so are masculine agreement (26b) and C2 agreement (26c). C1 agreement, on the other hand, has a mean of 3.2, notably higher than the other agreement options.

Factoring the speaker variation and the lower rating of CSCV sentences in general, the data above show that first conjunct agreement is the only acceptable option in sentences with CSCV. Masculine agreement and C2 agreement are not available. These results are compatible with the predictions made by the non-linear approach as laid out in the previous section. The predictions made by the linear approach were not supported: neither masculine agreement nor C2 agreement is available. The agreement patterns in sentences with CSCV thus argue for the non-linear approach and against the linear approach.

3.4 Evaluating the ellipsis analysis

This section addresses an alternative ellipsis analysis of the CSCV data from the last section. So far I have been assuming the sentences in (26) to be base-generated with a conjunction subject where the two NPs are coordinated. It is the CSCV movement that derives the word order in the sentence as is shown in (27).

The alternative analysis is one where (26) is derived from a conjunction of two clauses with most elements in the second clause elided. As is shown in (28), two clauses with the subject-part order are coordinated. Within each conjunct, part shows agreement with its own local subject. part of the second conjunct is elided along with other elements. This derivation does not involve CSCV at all. Agreement with sablje is expected since it is the true subject of the clause where the part is. Masculine agreement or agreement with koplja would not be generated. Thus if (28) is the right analysis of (26), the agreement pattern would not tease the two approaches to conjunction agreement apart (since there is no conjunction agreement to begin with).

Using three tests, I will argue that the bi-clausal ellipsis structure cannot be the only source of sentences with CSCV discussed above. The tests are drawn from Arsenijević et al. (2020b) in which the authors argue against a sole ellipsis analysis for sentences with closest conjunct agreement (without CSCV) in BCS.

If the bi-clausal ellipsis analysis is the only way to generate the sentences with CSCV, the sentence with ellipsis should have the same interpretation as the non-elliptical form with two full clauses. On the other hand, the structure with coordinated subjects predicts that sentences with CSCV can have different interpretations from the coordinated clauses. We will focus on three interpretations that can tease these structures apart: the collective reading, the cumulative reading, and the internal reading of different.

The sentence in (29a) has two interpretations: the distributive reading where sabers collided with sabers and spears collided with spears, and the collective reading where sabers collided with spears. On the other hand, the sentence with two coordinated clauses in (29b) only allows the distributive reading and not the collective reading. If the ellipsis structure is the only source, the sentence in (26a) is predicted not to allow the collective reading, just like the two coordinated clauses. On the other hand, the coordinated subject + CSCV analysis predicts the collective reading to be available.

The same logic goes with the cumulative reading. The sentence in (30a) can mean that sabers and spears were put in three boxes altogether, while (30b) is the most natural in a scenario where sabers were put in three boxes and spears were put three different boxes. The ellipsis analysis thus predicts the CSCV constructions with C1 agreement to be judged false in a scenario where a mixture of sabers and spears are put in three boxes. The coordinated subject + CSCV analysis would predict the construction to be true.

Lastly, (31a) is compatible with the scenario where sabers are put in one box and spears were put in a different box. This reading is labeled as the internal reading. The sentence in (31b) does not allow the internal reading but is judged true in a scenario where the sabers are put in multiple boxes and spears are put in multiple boxes. The ellipsis analysis then predicts the relevant sentence in (26a) to disallow the internal reading.

To probe these interpretations, Survey 1 also included the following three sentences in (32) along with the pictures in Figures 1, 2, and 3 respectively. The sentence in (32a) with Figure 1 is for the collective reading; (32b) with Figure 2 for the cumulative reading; and (32c) with Figure 3 for the internal reading of different. The participants were asked to judge whether the sentences are true given the scenarios depicted in the pictures by choosing from Da ‘Yes’ and Ne ‘No’. All the sentences are with CSCV and C1 agreement. As discussed above, the bi-clausal ellipsis analysis is not compatible with the readings, thus the sentences are predicted to be rejected. On the other hand, the coordinated subject + CSCV analysis predicts the sentences to be judged true.

Figure 1

collective (Figure 2 in Arsenijević et al. 2020b).

Figure 2

cumulative.

Figure 3

internal.

Out of the 30 participants, 19 chose ‘Da.’ and 11 chose ‘Ne.’ for (32a); 20 chose ‘Da’ and 10 chose ‘Ne’ for (32b); and 19 chose ‘Da’ and 11 chose ‘Ne.’ for (32c). For each reading, the number of speakers who accept the reading is twice as many as the ones that rejected it. This indicates that 19–20 speakers accepted the readings that are predicted to be bad if the bi-clausal ellipsis structure is the only source for sentences with CSCV.

One may wonder about the 10–11 speakers who rejected the readings. Note that the sentences in (32) involve CSCV with C1 agreement. The responses from all 30 speakers include ones that are from speakers who do not allow CSCV in the first place. In order to more specifically probe how speakers who allow CSCV interpret the sentences, we can look at responses only from speakers who rated the sentence with CSCV scrambling in (21) as 3 and above. As mentioned in the last section, 16 participants rated (21) as 1 or 2. Among the remaining 14, 13 speakers chose ‘Da.’ for (32a) and (32b) and 12 chose ‘Da.’ for (32c). In other words, when looking at responses only from speakers who allow CSCV, the readings in (32) predicted by the coordinated subject + CSCV analysis are uniformly accepted. I take this as evidence that the bi-clausal ellipsis analysis cannot be the only source for the sentences with CSCV, thus the argument for the non-linear approach made in the last section stands.

4.1 Order: Aux C1 Part and C2

The patterns observed in the previous section established a generalization: when C1 moves out of the conjunction, it must control agreement. We have shown that this generalization is predicted by the non-linear approach and not the linear approach. The sentences we have looked at are of the shape in (33), i.e. the first conjunct moves to the sentence initial position before the auxiliary while the rest of the conjunction stays after the part. This section looks at the order in (34) where the first conjunct moves to the position between Auxiliary and part.

Before discussing the agreement patterns of (34), we discuss the derivation of the word order in (34). Following the standard assumption, the subject ConjP is base-generated in a post-verbal position as in (35a). When the entire ConjP moves to Spec,PartP, the order in (35b) is derived with the subject between Aux and part. Sentences with sentence-initial subjects (e.g. (1)) result from moving the subject from Spec,PartP to the sentence initial position (e.g. Spec,TP) as in (35c).

Having covered the orders without CSCV, we move on to the order with CSCV in (33) and (34). Starting again from the post-verbal subject in (36a), the first conjunct can undergo CSCV to Spec,PartP in BCS, resulting in the first conjunct between Aux and Part and the rest of the ConjP at the post-verbal position as in (36b). This is how (34) is derived. Then the first conjunct can further move to the sentence initial position resulting in (36c), the order in (33) (which is a schematic representation of the crucial data (26) in the last section).

Sentences in the order in (34) were tested using sentences in (37) in Survey 2. The predictions made by the two approaches should be the same as those for (33): the linear approach predicts first conjunct agreement (37a), masculine agreement (37b), and second conjunct agreement (37c) while the non-linear approach predicts first conjunct agreement (37a) only.

Based on responses from 18 participants, first conjunct agreement has a mean rating of 2.8 out of 5; masculine agreement has a mean of 1.2 and C2 agreement has a mean of 1.4. The pattern of agreement options in this order is the same as the order in (36c) reported in (3): first conjunct agreement is preferred over the other two options.

Distribution of the ratings of (37) as well as the mean and median are reported in Table 6. 8 speakers rated C1 agreement as 1 or 2 while 7 rated it as 4 or 5, an even divide which indicates speaker variation. In contrast, all 18 speakers gave (37b) a rating of 1 or 2 and 17 out of the 18 speakers gave (37c) a rating of 1 or 2. Masculine agreement and C2 agreement in this order are uniformly rejected.

This observed pattern is parallel to the previous data with CSCV in (26): first conjunct agreement is accepted while second conjunct agreement and masculine agreement are not. The identical pattern observed in (37) and in (26) is expected in the non-linear approach.

4.2 C1 does not always control agreement: C1 Aux and C2 Part

In the patterns we have seen so far, the CSCVed C1 always controls agreement on part. This section will show that this is not always the case. Rather, C1 controls agreement when the non-linear approach predicts it to do so, and the pattern changes when the approach predicts otherwise.

The non-linear approach has been successful in accounting for the CSCV data because it connects movement to Spec,PartP and agreement on part. In all the CSCV sentences we have discussed so far, C1 is moved to Spec,PartP, thus it is predicted to control the agreement. On the other hand, if the CSCV does not involve the movement of C1 to Spec,PartP, the non-linear approach would not straight-forwardly predict that agreement on part is controlled by only C1. The order in (38) is such a case. Following the standard assumption, I assume that the order in (38) is derived from the structure in (35b) repeated here as (39a), where the ConjP moves to Spec,PartP. In (38), instead of moving the entire ConjP further to the sentence initial position (as in (35c)), the first conjunct moves out of the ConjP to the sentence initial position as in (39b).

Since the entire ConjP is in Spec,PartP position, the non-linear approach predicts sentences with the order in (38) to show the same agreement patterns as any sentences with pre-verbal subjects: first conjunct agreement, masculine agreement, and second conjunct agreement. The derivation has been worked out in detail by Murphy & Puškar (2018) (section 4.4). In particular, the following three orders would generate the three agreement patterns in sentences with pre-verbal subjects. Crucially, the further movement of C1 from Spec,PartP to the sentence initial position is predicted not to affect the agreement patterns on part.

The order in (38) was tested using the sentences in (41): first conjunct agreement (41a), masculine agreement (41b), and second conjunct agreement (41c). These sentences were included in Survey 2 with 18 participants.¹¹ Based on the responses, first conjunct agreement in (41a) has a mean rating of 2.1. Masculine agreement (41b) has a rating of 3. Second conjunct agreement (41c) has a mean of 2.7.

The distribution, the mean and the median of the judgments are shown in Table 7. Table 7 also includes previous cases of acceptable CSCV sentences for comparison.

Focusing on masculine agreement and C2 agreement first, both options have a mean rating close to 3, the midpoint of the scale, and show speaker variation: 7 speakers out of 18 rated (41b) as 1 or 2 and 9 speakers rated it as 4 or 5; whereas 10 speakers rated (41c) as 1 or 2 and 7 rated it as 4 or 5. This pattern is similar to the CSCV sentence without agreement in (21) as well as the CSCV sentences with C1 agreement in (26a) and (37a). Thus I conclude that in the order where the conjunction precedes the part, both masculine agreement and C2 agreement are available even with CSCV.

On the other hand, C1 agreement in (41a) shows a lower mean rating of 2.1, and 12 out of 18 speakers gave it a rating of 1 or 2. I propose that this pattern results from two factors: 1) the lower rating of CSCV constructions in general discussed above, and 2) a dispreference of highest conjunct agreement in subject-part order in BCS. Although C1 agreement in subject-part order has been acknowledged as a legitimate agreement strategy in multiple elicitation experiments (Marušič et al. (2015); Willer-Gold et al. (2016)), it is consistently less preferred compared to masculine agreement and C2 agreement.¹² In Exp 1 in Willer-Gold et al. (2016), the participants were asked to produce part based on coordinated subjects of different gender combinations. When the first conjunct is feminine and the second is neuter (the same combination in Survey 1 and 2 in (41)), the highest conjunct agreement only takes 11% of all responses, compared to 36% for masculine agreement and 53% for neuter agreement.¹³ This dispreference has been acknowledged and discussed in Marušič et al. 2015: fn. 4, Willer-Gold et al. 2016: Section 4.1. Taking these two factors into consideration, I conclude that C1 agreement, C2 agreement and masculine agreement are all viable agreement options in the subject-part order with C1 moved away. The parallel between (41) with CSCV and the subject-part order without CSCV also confirms the predictions made by the non-linear approach.

The availability of C1, C2 and masucline agreement in (41) also argues against a variation of the linear approach suggested by a reviewer. In this suggested variation, both the moved C1 and the stranded ConjP are potential goals for Agree-Link in sentences with CSCV movement. If one assumes that in this configuration, part always Agree-Links with the higher goal, C1, and ignores the ConjP, the linear approach would also predict only C1 agreement in sentences with CSCV movement. However, this variation would also predict a C1-agreement-only pattern in (41), contrary to the fact.

Table 8 shows the mean ratings of each agreement strategy in the three orders we have discussed so far. Note that the linear approach makes identical predictions with the non-linear approach regarding the last order. The sentences in (41) are thus not on their own a supporting piece of evidence for the non-linear approach. However, what’s crucial here is that the non-linear approach correctly predicts patterns observed in all three orders in Table 8, including the differences among them, while the linear approach makes wrong predictions for the first two orders.

4.3 Sentences with three conjuncts

In all the sentences we have seen so far, part cannot target the second conjunct for agreement, which I have argued to support the non-linear approach. A potential alternative explanation is that second conjunct agreement is in principle available. However, when the speakers parse these sentences, the overt conjunction head ‘and’ provides a strong cue to stop the search for the agreement controller. After all, in no sentences with the part-subject word order would part agree with anything that comes after the coordinator ‘and’. Note that this alternative refers to the processing/parsing of the sentences rather than the syntactic structure. If this idea is on the right track, the degraded C2 agreement reported above would not directly argue against the linear approach. This is shown schematically in (42a).

In this section, I show that this possibility is untenable using data from sentences with three conjuncts. In the construction with three conjuncts shown in (42b), there is no overt Conj head between part and C2, thus C2 agreement would not be affected by the processing factor mentioned above. If Agree-Copy can be postponed to PF after linearization according to the linear approach, part is predicted to be able to agree with C2.

To test this alternative, the sentences in (43) and (44) were included in Survey 2 with 18 participants. First, sentences without CSCV with different agreement patterns in (43) were tested to establish a baseline of different agreement strategies in sentences with three conjuncts. The first conjunct is of neuter gender while the following two conjuncts are of feminine gender. First conjunct agreement has a mean rating of 4.7, masculine agreement has a mean rating of 4.3, second conjunct agreement, on the other hand, has a meaning rating of 2.7. In line with sentences with two conjuncts in (25), first conjunct agreement (43a) and masculine agreement (43b) are acceptable while C2 agreement (43c) is degraded.

The distribution, mean, and median for the three sentences are shown in Table 9. C1 agreement and masculine agreement are uniformly accepted. On the other hand, C2 agreement shows speaker variation: 9 speakers gave (43c) a rating of 1 or 2 while 8 gave it a rating of 4 or 5. This is surprising for both the linear and the non-linear approach as both predict (43c) to be unacceptable. I will leave this data point for future research and tentatively conclude that the C2 agreement is degraded in (43c).¹⁴

In the sentences in (44), the first conjuncts moved to the pre-part position which I will assume to be Spec,PartP. If the unavailability of C2 agreement in sentences with CSCV reported in previous sections results from the overt Conj head blocking the search for the agreement controller, it is predicted that agreeing with the second conjunct is acceptable here as there is no overt ‘and’ between part and the second conjunct. As is shown, based on responses of 18 speakers, the first conjunct agreement in (44a) has a mean rating of 2.3, the resolved agreement in (44b) has a mean rating of 1.3, and the second conjunct agreement in (44c) has a mean rating of 1.4.

The distribution, mean, and median for these sentences are shown in Table 10. The data here is similar to agreement with two conjuncts in sentences in (37) and Table 6.

The results above show that, contrary to the prediction made by the alternative analysis, second conjunct agreement remains unavailable even when there is no overt Conj head between part and the second conjunct. Similarly, default masculine agreement remains unavailable. Agreement with the moved conjunct is still the best option out of the three.¹⁵¹⁶ The results rule out the hypothesis that the overt Conj head blocks the agreement.