On the movement analysis of null subjects in Brazilian Portuguese: Experimental results from extraction of embedded subjects*

There is an on-going debate on the empirical adequacy of the movement approach to definite null subjects in Brazilian Portuguese, BP (cf. Ferreira 2000, Rodrigues 2004, and Nunes 2009). On the one hand, Modesto (2000) observes that the null subject of finite embedded clauses associated with object control verbs like convencer ‘to convince’ is subject-oriented, rather than object-oriented and this is unexpected from a movement perspective. On the other hand, Rodrigues (2004) claims that these embedded clauses are actually adjuncts in BP and the observed subject orientation can be accounted for in terms of Hornstein’s (1999, 2001) movement analysis of adjunct control. This paper aims to contribute to this debate by presenting the results of an experiment on grammaticality judgments by BP speakers on the extraction of embedded subjects out of complement and adjunct clauses, as well as finite embedded clauses associated with convencer. The results show that when a distinctive pattern could be observed, finite clauses associated with convencer behaved like adjunct clauses rather than complement clauses. The experiment thus provides confirming evidence for Rodrigues’s (2004) adjunct analysis, invalidating Modesto’s (2000) argument against the movement approach to definite null subjects in BP.


Introduction
The considerably large literature on null subjects in Brazilian Portuguese (henceforth BP) converges on the conclusion that they do not behave like the null subjects of canonical pro-drop languages 2 . (1a) below, for instance, shows that a definite (i.e. nonexpletive, nonarbitrary) third person null subject in BP is banned from matrix environments that exclude topic drop. This amounts to saying that definite null subjects in BP typically appear in embedded clauses. (1b) further shows that the embedded null subject requires an antecedent and that this antecedent must be the most local c-commanding DP.
On the movement analysis of null subjects goes be promoted 'João said that [Pedro's father]k thinks that hek is going to be promoted' Facts like the ones illustrated in (1) have led Kato (1999) to propose that null subject constructions in BP involve PRO rather than pro. Reinterpreting the gist of Kato's proposal in terms of Hornstein's (1999Hornstein's ( , 2001) Movement Theory of Control (henceforth MTC), Ferreira (2000Ferreira ( , 2009 and Rodrigues (2002Rodrigues ( , 2004 have argued that definite null subjects in BP are traces of A-movement. Under this approach, contrasts such as the one illustrated in (2) below, for instance, are accounted for in terms of minimality. The null subject in both (2a) and (2b), being a trace of A-movement, requires that its antecedent be the closest c-commanding DP. In other words, the null subject must be interpreted as a Maria 'Maria' in (2a) and o médico 'the (male) doctor' in (2b). However, the lexical meaning of grávida 'pregnant' as well as its feminine agreement morphology are only compatible with the former; hence the contrast between (2a) and (2b).
(2) a. O médico disse que a Maria acha que t está grávida the doctor.MASC said that the Maria thinks that is pregnant.FEM 'The doctor said that Maria thinks that she is pregnant' b. *A Maria disse que o médico acha que t está grávida the Maria said that the doctor.MASC thinks that is pregnant.FEM 'Maria said that the doctor thinks that she is pregnant' In this paper we examine an empirical challenge raised by Modesto (2000Modesto ( , 2011 to approaches that analyze null subjects in BP as traces of A-movement (cf. Ferreira 2000, Rodrigues 2002, and Nunes 2008). The challenge is based on his empirical observation that the null subject of finite embedded clauses associated with object control verbs like convencer 'to convince' are subject oriented, rather than object oriented. Assuming that the matrix object of these constructions c-commands the embedded clause, Modesto concludes that it should block movement of the embedded subject to the matrix subject position; in other words, the subject orientation reading should be prohibited, contrary to fact. Rodrigues (2004) objects to this reasoning, claiming that these embedded clauses are actually adjuncts in BP and that the observed subject orientation can be accounted for in terms of Hornstein's (1999Hornstein's ( , 2001) movement analysis of adjunct control. Extraction of embedded subjects in these constructions should in principle suffice to settle this debate, but the reported judgments in this regard are conflicting (fully acceptable according to Modesto 2011 and marked as ?? according to Rodrigues 2004). Since there is no agreement on judgments regarding crucial pieces of data, we designed a grammaticality judgment experiment with the goal of shedding more light on the discussion. The results to be presented below show that when a distinctive pattern could be observed, finite clauses associated with convencer behaved like adjunct clauses rather than complement clauses. The paper is organized as follows. In section 2, we present the details of Modesto's (2000) challenge to the movement approach to null subjects in BP and Rodrigues's (2004) alternative analysis of the apparently problematic data. We show that A'-extraction of embedded subjects should in principle tease the two approaches apart, but a proper assessment of its relevance becomes clouded as the judgments presented by these authors go in opposite directions. In section 3, we then present the experiment on judgments on such extractions that we carried out. In sections 4 and 5 we present and discuss the results of the experiment. Finally, section 6 concludes the paper. Hornstein (1999Hornstein ( , 2001 has argued within the MTC that object control follows from minimality. Under a Larsonian shell analysis of ditransitive structures, the matrix object of a sentence like (3a), for instance, should c-command into the infinitival clause, as sketched in (3b). If obligatorily controlled PRO is a trace of A-movement, as defended by Hornstein, the subject position of the infinitival clause in (3b) Modesto (2000: 20) takes the contrast between (3a) and (4) below to provide evidence against the movement approach to null subjects in BP finite clauses outlined in (2) (cf. Ferreira 2000, Rodrigues 2002, Nunes 2008. He points out that if the embedded null subject of (4) is an A-trace, it should also find its antecedent in the matrix object position, contrary to fact. His conclusion is that the null subject of the infinitival clause in (3a) may be prone to a movement analysis, but not the null subject of the finite clause in (4).

A challenge to the movement analysis of null subjects in BP
(4) [O Paulo]1 convenceu [a Maria]2 que ec1/*2 tinha que ir embora the Paulo convinced the Maria that had that go away 'Paulo convinced Maria that he had to go away' Modesto further makes the important observation that if the matrix object of sentences such as (4) undergoes A'-movement, it may then be interpreted as the antecedent for the embedded null subject, as shown in (5).
(5) Quem1 que [o Pedro]2 convenceu t1 que ec1/*?2 tinha que ir embora? who that the Pedro convinced that had that go away 'Who did Pedro convince that he had to go away?' (Modesto 2000:85) Modesto takes the contrast between (4) and (5) as evidence for his claim that null subjects in BP are pros that must be licensed via A'-binding. He proposes that subjects of finite clauses in BP may move to Spec,AgrP, which is taken to be On the movement analysis of null subjects Isogloss 2018, 4/1 in Brazilian Portuguese 89 an A'-position, and from this position they can license an embedded null subject. That would be the case of (4). By contrast, in (5) the matrix object passes through Spec,AgrP on its way to Spec,CP and the matrix subject stays put in its Case checking position; hence, only the matrix object should be able to function as an antecedent for the embedded null subject.
Before we discuss the contrast between (3a) and (4), it should be observed that for purposes of exposition, we have maintained Modesto's original judgments in (5), according to which only the object reading is licit when the object undergoes A'-movement. Although there may be a preference for the object reading for some speakers, Rodrigues (2004: 217) notes that the sentence in (6) below 3 clearly shows that the subject reading is available, as well.
Quemi [a Maria]k convenceu ti que eck estava grávida? who the Maria convinced that was pregnant 'Who did Maria convince that she was pregnant?' Going back to the contrast between (3a) and (4), the tacit assumption in Modesto's argument is that they have the same structural configuration (see (3b)), differing only in terms of finiteness. There are reasons to believe that this is not the case, though. Ferreira (2000) shows that the matrix object does not ccommand into the finite embedded clause, as indicated by the lack of a Principle C effect in a sentence such as (7) below. Besides, Rodrigues (2004) shows that the embedded clause of (4) does not behave as a typical complement, for it does not allow extraction from within it, as shown in (8) (cf. Nunes 2009 for further discussion) 4 . Accordingly, we find a clear contrast between (7)-(8) and sentences with infinitivals like (3a) (cf. Nunes 2013). As (9) shows, in the case of infinitivals, the matrix object c-commands into the infinitival clause, inducing a Principle C effect in (9a), and the embedded clause is transparent for extraction (see (9b)).  Even more interesting, Rodrigues (2004) shows that the additional object reading available in sentences like (5) when the object undergoes A'-movement is also observed in BP with respect to adjunct finite clauses. The pair of sentences in (10), for instance, shows that the null subject of the finite adjunct clause obligatorily takes the matrix subject as its antecedent if the matrix object remains in situ (see (10a)), but may be interpreted as either the matrix subject or the matrix object if the latter undergoes A'-movement (see (10b) who the Maria visited when went to Brasília 'Which person did Maria visit when she/that person went to Brasília?' (Rodrigues 2004: 228) In the face of these facts, Rodrigues (2004) proposes that the finite embedded clause associated with verbs like convencer 'to convince' in BP is an adjunct of sorts 5 . This would account for why the object does not induce a Principle C effect in sentences such as (7) and why extraction out of the embedded clause as in (8) does not yield acceptable outcomes. As for the additional reading in (5), it should in principle be subject to whatever explanation is offered for the pattern in (10b) (cf. Rodrigues 2004and Nunes 2013, 2014 for specific proposals).
Importantly, Rodrigues shows that if the finite embedded clause of sentences like (4) is an adjunct, its subject orientation can be properly accounted for under the movement approach. Hornstein (1999Hornstein ( , 2001 has argued that adjunct 5 Nunes (2009) has argued that the finite embedded clause associated with convencer may display a complement behavior when preceded by the preposition de 'of'. In particular, embedded subject extraction is allowed, as illustrated in (i) below. In this paper, we will restrict our discussion to constructions without de. See Coelho (in progress) for a grammaticality judgment experiment testing the role of de in these constructions.

(i)
Quemi o João convenceu a Maria de [que ti vem amanhã]? (Nunes 2009: 257) who the J. convinced the M. of that comes tomorrow 'Who did João convince Maria will come tomorrow?' On the movement analysis of null subjects Isogloss 2018, 4/1 in Brazilian Portuguese 91 control is to be derived in terms of sideward movement (in the sense of Nunes 2001Nunes , 2004 and that subject orientation in adjunct control follows from Mergeover-Move economy computations (cf. Chomsky 1995). Extending Hornstein's adjunct control analysis to finite adjunct control in BP, Ferreira (2000) and Rodrigues (2004) argue that the null subject of a finite adjunct such as the one in (10a) is a trace of the matrix subject, which has undergone sideward movement before the embedded clause becomes an adjunct. From this perspective, the derivation of (10a) proceeds along the (simplified) lines of (11). Given the derivational step in (11a), the verb visitou can have its selectional requirements satisfied either via selection and merger of quem or via sideward movement of a Maria. Assuming that Merge is more economical than Move (cf. Chomsky 1995), the computational system chooses merger of quem (see (11b)) before a Maria undergoes sideward movement to external argument position of visitou (see (11c)). Notice that movement of a Maria in (11c) crucially takes place while L is a root syntactic object -in other words, before L becomes an island (cf. Hornstein 2001 for detailed discussion). Finally, a Maria moves to the matrix Spec,TP, yielding the subject reading for the embedded subject.
Applying this analysis to constructions involving finite clauses associated with verbs like convencer, Rodrigues argues that the subject orientation observed in (4), for instance, is the outcome of a derivation along the lines sketched in (12) below. Given the derivational step in (12a), Merge-over-Move triggers merger of a Maria before o Paulo undergoes sideward movement to the external argument position of convencer. Thus, the embedded null subject of (4) is interpreted as the matrix subject for it is a trace of the matrix subject 6 . In a later paper, Modesto (2011) questions Rodrigues's (2004) judgments on subject extraction in sentences like (8) (which are the same as ours). According to him, the sentence (13a), which is parallel to (8), is "perfectly grammatical", whereas the sentence in (13b), which involves subject extraction out of a bona fide adjunct, is "grossly ungrammatical" (p. 16).
(13) a. Quem1 (que) o Pedro convenceu a Cilene que t1 vai viajar? who that the Pedro convinced the Cilene that goes travel 'Who is the person that Pedro convinced Cilene that he will travel?' b. *Quem1 o Pedro viu a Cilene enquanto t1 fazia compras? who the Pedro saw the Cilene while did shopping *'Who did Pedro see Cilene while was shopping?' (Modesto 2011: 15-16) Establishing the grammaticality status of sentences such as (8) and (13a) therefore becomes a central issue on the debate on grammatical nature of null subjects in BP. Bearing that in mind, in this paper we subject the contradicting judgments reported by Rodrigues (2004) and Modesto (2011) to a closer scrutiny. We developed an acceptability experiment to test judgments by BP speakers on extraction of subjects of finite clauses, comparing their judgments on standard complement and adjunct clauses with their judgments on parallel cases involving convencer 7 . The prediction is clear: extraction out of cases involving convencer should pattern like extraction out of complements if Modesto's (2000) analysis of these cases is correct, but like extraction out of adjuncts if Rodrigues's (2004) analysis is correct. In the next sections, we describe the experiment and present the results found.
On the movement analysis of null subjects Isogloss 2018, 4/1 in Brazilian Portuguese 93

The experiment
In order to investigate the judgments on the types of sentences discussed above, we designed an acceptability test and applied it to 20 participants (10 males and 10 females) with 20 to 30 years of age, all native speakers of BP who were born and raised in the state of São Paulo 8 . The participants were told that we were evaluating the performance of a computer software designed to formulate complex questions in Portuguese and in order to do that, we needed Portuguese speakers to judge which questions built up by the software were well formed and which were not. The participants listened to audio recorded sentences read by a BP speaker and pressed YES or NO in the computer to indicate if the sentence listened to was well formed or not. The participants could also press REPETE in case they wanted to listen to the sentence once again before pressing YES or NO.
Before the presentation of the relevant sentences of the experiment, there was a familiarization period where the participants were presented with 8 questions unrelated to our research topic (4 with expected YES-answers and 4 with expected NO-answers). This allowed us to check if the participants had correctly understood the instructions of the experiment.
The actual experiment consisted of 10 sentences with subject extraction out of a finite complement clause, 10 sentences with subject extraction out of a finite adjunct clause, 10 sentences with subject extraction out of a finite clause associated with the verb convencer, and 8 distractors (4 with expected YES answers and 4 with expected NO answers). (14) below shows an example of each type of sentence. All the sentences with complement clauses involved a ditransitive verb (see (14c)) and all the sentences with adjunct clauses involved a transitive verb in the matrix clause (see (14d)). This ensured that an eventual distinct behavior of sentences with convencer should not be attributed to the complexity induced by the presence of an object in the matrix clause. We also computed if the extracted wh-subject was syntactically simple (quem 'who' or o que 'what') or complex (que pessoa 'which person' or que prato 'which dish', for instance). The sentences of the different paradigms were mixed and all participants were exposed to the same order of presentation of the sentences. There was a total of 920 answers. The data were collected and categorized with the free software TP from Worken (http://www.worken.com.br/tp_regfree.php), developed for speech perception experiments.

The actual sample for the analysis
The answers of one participant were excluded, for she pressed YES for all the sentences, which suggests that she did not understand her task. Out of the set of 46 sentences, the three sentences in (15) below were also excluded, for they exhibited an atypical behavior with respect to both the global picture (see Figure  1) and the individual participants (see Graphic 1) 9 .
On the movement analysis of null subjects

Graphic 1. Proportion of NO-answers with respect to convencer-sentences by individual participants
With the exclusion of the participant mentioned above and the outliers sentences in (15), the actual sample to be examined consisted of 817 answers (43 sentences times 19 participants), distributed between simple and complex whphrases as depicted in Table 1 Distribution of complex and simplex wh-phrases Sentence type Simplex whphrases Complex whphrases

Homogeneity and reliability of the data
The remaining 19 participants behaved homogeneously and no participant was identified as an outlier, as we can see in Figure 2.     The dispersion Graphics 2 and 3 further show that there is no correlation between the correct answers provided by individual participants with respect to familiarization and distractor sentences and their answers to convencer-sentences (the coefficient of linear correlation between x and y axes is 0,24 in the former and 0,11 in the latter). In other words, it is not the case that a given participant is providing more YES or NO answers to convencer-sentences because (s)he did not understand the task of the experiment or was not paying sufficient attention. Familiarization Distractors Correct answer

Repetition and syntactic complexity of the extracted wh-phrase
We examined whether there was a correlation between the syntactic complexity of the wh-phrase and repetition. One could expect that simple wh-phrases should trigger repetition more often than complex wh-phrases, given that the latter is more prone to a D-linking interpretation (in the sense of Pesetsky 1987). This expectation was not fulfilled, though. Figure 9 shows that the type of wh-phrase is not a determining factor in triggering repetition (p = 0.406910).  Figure 10 below shows that familiarization sentences are the ones that display the larger proportion of repetition (0.48/0.52). This is an expected result, as it indicates that the participants wanted to make sure they got familiarized with the type of sentences they were supposed to evaluate. This is further confirmed in Graphic 4, which shows that the number of repetitions decreases as the experiment unfolds 10 .

Graphic 4. Repetition along the experiment
Recall that only the familiarization sentences were presented together at the beginning of the experiment (the first 8 sentences); the sentences of the other paradigms were intermingled (and presented in the same order to all participants). 10 Recall (see section 4.1) that out of the 46 original sentences, we have excluded three outliers (s21, s31, and s32). For easiness of reference, we maintained the original identification of the other sentences, though. So, s46 in Graphic 4, for instance, actually refers to the 43 rd sentence of the sample (the last one). The directionality seen in Graphic 4 suggests that as the experiment unfolds, the participants become more and more secure with respect to the task itself, pressing REPETE only when in doubt with respect to specific sentences. In this regard, note that the proportion of repetition for convencer-sentences (0.35/0.65) in Figure 10 above is much more similar to the proportion for adjunct sentences (0.36/0.64; p = 0.726557) than the proportion for complement sentences (0.26/0.74; p = 0.119698). In fact, the proportion of repetition for complement sentences is much closer to the proportion for distractors (0.3/0.7; p = 0.418481).

Tendency for YES-and NO-answers during the experiment
We also examined whether the order of presentation of the sentences affected the participants' answers, that is, whether there was a tendency for participants to give more YES-or NO-answers for each type of sentence in the beginning of the experiment than at the end.
Graphics 5 and 6 below respectively depict the proportion of expected answers for complement and adjunct sentences as the experiment unfolds. As extraction is allowed out of complement clauses but not out of adjunct clauses, YES was the expected answer for complement sentences and NO for adjunct sentences. Given that the aim of the experiment was to determine the behavior of convencer-sentences, there was no a priori expected answer in this case. However, in order to make the appropriate comparison with complement and adjunct sentences, it was necessary to arbitrarily choose between YES and NO. For concreteness, we computed NO as the expected answer for convencer-sentences and the result is shown in Graphic 7 below. We submitted the results depicted in Graphics 5-7 as well as the repetition results shown in Graphic 4 to two tendency tests (Cox-Stuart test and Mann-Kendall test) in order to determine whether the proportion of YES-and NOanswers, as well as repetitions, showed a statistically significant tendency as the experiment unfolded. Table 2 below shows that under both tendency tests, only the decrease of repetition in the course of the experiment turned out to be statistically significant (p = 0.006 under Cox-Stuart test and p = 0.001 under Mann-Kendall test). The p-values above 0.05 for each sentence type in both tests show that there was no correlation between expected answers and order of presentation. In other words, the reliability of the answers for the different sentence types was not affected by the order in which the sentences were presented.   Figure 11 below shows the overall pattern of YES-and NO-answers for sentences involving complements, adjuncts, and convencer. There is no statistically significant difference between answers for convencer-sentences and complement sentences (p-value for YES-answers = 0.39723; p-value for NO-answers = 0.311338) or between convencer-sentences and adjunct sentences (p-value for YES-answers = 0.401755; p-value for NO-answers = 0.421641). Figure 11 by itself is not very illuminating. Recall that Rodrigues's (2004) took extraction of the subject of the finite clause associated with convencer to be quite marginal (see (8)), whereas Modesto (2011) took it to be completely acceptable (see (13a)). However, the results depicted in Figure 11 place convencer-sentences in a middle position between clear complements and adjuncts. In this regard, the overall performance of individual participants is not very helpful either: out of the 19 participants, 10 answered YES with respect to convencer-sentences most of the times and 9 answered NO. Figure 12 below depicts the proportion between YES and NO answers for the combined set of complement, adjunct and convencer-sentences when the type of the wh-phrase is taken into account. It shows that in the overall pattern, the    Figure 13 shows that the effect of the complexity of the wh-phrase on extraction out of complement clauses is not statistically significant (p = 0.5706).

YES-and NO-answers and syntactic complexity of the extracted wh-phrase
In turn, Figure 14 shows that the complexity of the wh-phrase is statistically significant in the case of adjunct sentences (p = 0.01801), with complex whphrases favoring YES answers. The relevant result for our purposes is shown in Figure 15: subject extraction out of the finite clause associated with convencer is sensitive to the complexity of the wh-phrase (p = 6.801e-5), patterning like subject extraction out of adjunct clauses (see Figure 14) and not like subject extraction out of complement clauses (see Figure 13). This result gets even more robust when we examine if there are correlations among the answers for different sentence types. We have already seen in section 4.2 that the answers for familiarization and distractor sentences did not correlate with the answers for convencer-sentences (see Graphics 2 and 3). In fact, the only correlation found among specific answers with respect to different sentence types involved answers for adjunct sentences and answers for convencersentences. As Graphic 8 below shows, the larger the number of YES answers a given participant offers to adjunct sentences, the larger the number of YES answers (s)he provides to convencer-sentences (p = 0.03619).

Discussion
At first sight, the overall pattern of YES-and NO-answers per sentence type depicted in Figure 11 and the description of the individual behavior by the participants (10 favoring YES-answers and 9 favoring NO-answers with respect to convencer-sentences) seem to suggest that we are simply facing a gardenvariety idiolectal variation, with some speakers allowing subject extraction out of embedded clauses associated with convencer and some other speakers disallowing it. However, it could in principle be the case that the inconclusive picture seen in Figure 11 actually results from the interaction of each type of sentence with independent conditions. We have seen in Figure 9 that the complexity of the wh-phrase was not a relevant factor for repetition. However, the descriptive content available in complex wh-phrases is arguably more prone to licensing a D-linked interpretation and D-linking attenuates island effects (cf. Pesetsky 1987). Thus, one could expect the complexity of the moved wh-phrase to have a greater impact on YES answers in adjunct sentences than in complement sentences. This expectation was indeed borne out: the complexity of the wh-phrase was statistically significant for extraction out of adjuncts (see Figure 14), but not for extraction out of complements (see Figure 13). Once this contrast was established, the next step was to consider the behavior of convencer-sentences in this regard and the result was that the complexity of the wh-phrase was also statistically significant (see Figure 15). In other words, this indicates that the picture seen in Figure 11 is misleading and that convencer-sentences pattern like adjunct sentences and not like complement sentences.
This conclusion becomes even clearer when we take the overall contrast between Figure 13, on the one hand, and Figures 14 and 15, on the other, in the light of the results displayed in Graphic 8. Graphic 8 shows that the tolerance regarding subject extraction out of finite clauses associated with convencer exhibited by individual participants correlates with their tolerance regarding subject extraction out of adjunct clauses. The more tolerant a given speaker is with respect to subject extraction out of adjunct clauses, the more tolerant s/he will be with respect to convencer-sentences.
Another revealing result was provided by the behavior of convencersentences regarding repetition. Recall that participants were requested to say if the questions formulated by a computer software were well formed or not (see section 3). Given that they had the possibility of pressing REPETE to listen to the sentence once again before passing on their judgment, it is reasonable to think that they would take advantage of this option in cases that were harder to process or harder to judge. Furthermore, given that complements are transparent domains for extraction as opposed to adjuncts, the expectation was that adjunct sentences should be harder to process and evaluate. Therefore, the proportion of repetition was expected to be larger for adjunct sentences than for complement sentences. This expectation was indeed borne out, as seen in Figure 10. Importantly, the proportion of repetition for convencer-sentences (0.35/0.65) in Figure 10 is much more similar to the proportion for adjunct sentences (0.36/0.64) than complement sentences (0.26/0.74). Again, this corroborates Rodrigues's (2004) proposal that the finite clause associated with convencer has the behavior of an adjunct clause.

Conclusion
As discussed in section 2, subject extraction out of finite clauses associated with verbs like convencer in BP may provide a crucial basis for us to evaluate Modesto's (2000Modesto's ( , 2011 empirical argument against analyzing definite null subjects in BP in terms of Hornstein's (1999Hornstein's ( , 2001) MTC, as advocated by Ferreira 2000, Rodrigues 2002and Nunes 2008, among others. If the embedded clause of a sentence such as (4), for instance, repeated below in (16), turns out to function as a complement, Modesto's argument is a valid one, for movement of the embedded subject to the matrix subject position should be blocked by the intervening object. On the other hand, if it turns out to function as an adjunct, as proposed by Rodrigues (2004), Modesto's point becomes mute, for the MTC takes adjunct control to be derived via sideward movement of the subject of the will-be adjunct to the subject position of the subordinating clause (see Hornstein 1999Hornstein , 2001; in this case, an object in the subordinating clause does not count as an intervener, for it does not c-command the trace of the moved subject (see (11) and (12)). The results of our experiment help us understand the disagreement in the judgments regarding subject extraction in convencer-sentences reported in the literature (see (8) vs. (11a)). Taken in isolation, these sentences appear to fluctuate between complement and adjunct sentences as far as subject extraction is concerned and this fluctuation is also observed with respect to the participants individually, with some speakers being more tolerant than others. We have shown that this conundrum can nonetheless be unveiled if we take into account factors that may independently affect speakers' judgments regarding subject extraction.
One such factor was repetition. Under the assumption that island violations are harder to process, the sentences of the experiment that involved island violations should trigger more instances of repetition. As we see in Figure  16 (see Figure 10), convencer-sentences patterned like adjunct sentences with respect to repetition. Another factor that independently affects the acceptability of constructions involving movement is D-linking. In particular, island effects get weakened if the moved element involves a D-linked wh-phrase (cf. Pesetsky 1987). In the experiment, we indirectly controlled for D-linking by investigating the behavior of simple and complex wh-phrases for each sentence type. Figure 17 below, which portrays the results for NO-answers in the Figures 13, 14, and 15, shows that complex wh-phrases have a greater impact in reducing the proportion of NO answers in adjunct and convencer-sentences than in complement sentences. In other words, the amelioration effect that D-linking produces with respect to adjunct island violations is also observed in convencer-sentences. Finally, Graphic 8, repeated below, shows that the proportion of YES answers to adjunct sentences was correlated with the proportion of YES answers to convencer-sentences. This indicates that the individual variation detected among the participants regarding the acceptability of subject extraction out of convencer-sentences is actually a reflex of how tolerant a given individual is with respect to subject extraction out of adjunct islands.

Graphic 8. Correlation between YES answers for adjunct and convencer sentences
These results support Rodrigues's (2004) proposal that the finite embedded clause associated with convencer in BP is not a real complement but some sort of adjunct, thus invalidating Modesto's (2000Modesto's ( , 2011 argument. The conclusion is that subject orientation in sentences like (16) does not provide empirical evidence against the movement approach to definite null subjects in BP,