Constraining Free Merge

3.1 Labeling-Based Derivations According to Chomsky (2013, 2015)

Following Chomsky (2013, 2015), I assume that a form of Labeling is at work with respect to language generation. Labeling is necessary for interpreting phrases. Labeling refers to a process of finding a prominent feature of an SO via the search process involved in language, Minimal Search (Chomsky 2013, 2015).³ Chomsky (2013, p. 43) writes that “[t]he simplest assumption is that LA [Labeling Algorithm] is just minimal search, presumably appropriating a third factor principle, as in Agree and other operations.” In this way, Labeling is really just a form of Minimal Search, which finds prominent features that can function as labels.

Labeling via Minimal Search works as follows. Assume that a head X and a phrase YP Merge, forming {X, YP}. In this case, the label is X, assuming that X has prominent features that are capable of Labeling. If an XP and a YP, both phrases, Merge to form {XP, YP}, then shared features can label. For example, assume that XP (specifically the head X) has phi-features and YP (the head Y) has unchecked phi-features. Minimal Search results in the XP and YP forming an Agree relation so that the uPhi on YP are checked by the iPhi (interpretable phi-features) on X, and then the shared phi-features on XP and YP can label. Chomsky also takes the position that the English T is too weak to label on its own—this accounts for the requirement that a clause have a subject (the traditional EPP effect of Chomsky (1981)). Given the structure {T, YP}, T alone cannot label. However, given {XP, TP} where TP and XP Agree in terms of phi-features, the shared phi-features label. This is accounted for as follows. T inherits uPhi from C. Given an {XP, TP} structure, the uPhi on T Agree with the phi-features on X and the shared phi-features label. Crucially, in an {XP, YP} structure in which XP and YP do not Agree, Labeling is not possible. Also, given a {T, YP} structure, Labeling is not possible, assuming that T is too weak to label. These are summarized in (1). Furthermore, consider a root that has Merged with a functional head (e.g., a categorizer) to form what is essentially a head-head structure; for example, the root walk Merges with a categorizer n. In this case, the root is too weak to label by itself, but the functional head can label. This is the position that Chomsky (2013, p. 47) takes, following Marantz (1997), Embick and Marantz (2008), and Borer (2005a, 2005b, 2013). I assume that a root can label after it Merges with a categorizer.

The convergent derivation of Tom read books,⁴ following Chomsky (2013, 2015), proceeds as shown in Figure 2. The lines below each terminal node represent the frontier of the derivation—the portion of the derivation that is sent to Spell-Out to be pronounced. The root books set-Merges with the functional categorizer n, and n labels, as the root books is not capable of Labeling. Chomsky (2015) claims that a verbal root undergoes internal pair-Merge (head movement) with v to form <v*, read>, resulting in v* being dephased (also see Epstein et al., 2016). The <v*, read> pair-Merged structure is represented with a dotted arc. Dephasing is a process in which an element that would typically function as a phase head, thus being a point of transfer, no longer functions as phase head.⁵ In this case, Chomsky proposes that phasehood is passed onto the complement of v*. Thus, the complement of v* will function as a phase and be transferred. A phase head passes uFs to its complement, so the uPhi (uninterpretable Phi) of v* are passed onto the verbal root read, which being a root, is unable to label by itself. The object a book undergoes IM (Internal Merge) with read to form an {XP, YP} structure. In the matrix clause, the subject is initially-Merged with vP, and then it internally-Merges with the TP. The uPhi of C are inherited by T. Minimal Search results in phi-feature agreement in the {NP, {Tpast…}} and {NP, {read…}} structures, and these shared phi-features are able to label, where the label is indicated as <ɸ, ɸ>.

Click to enlarge

Figure 2

Structure of “Tom Read Books”

Note. Adapted from Chomsky (2015, p. 10).

In the following subsections, I explain my assumptions about Labeling Theory. Note that, for reasons discussed in the following sections, I do away with some of the operations utilized in the type of derivation shown in Figure 2.

3.2 Phases

Labeling Theory follows the view that the structures of sentences are constructed hierarchically in a bottom-up fashion, and sentences consist of phases, which are portions of sentences that essentially become inaccessible after construction. The core phases are generally assumed to be a transitive Verb Phrase (v*P) and a Complementizer Phrase (CP), following Chomsky (2000, 2001). Both (2)a–b are well-formed and both crucially are formed from the same set of lexical items. These examples differ, however, with respect to the ordering of lexical items. The embedded CP in (2)a is a phase that is constructed from a lexical array that does not contain there. As a result, a man raises to subject position of the CP. The expletive there is associated with the higher phase of the matrix clause. In (2)b, on the other hand, the expletive there is available in the embedded CP phase. As a result, there is inserted in subject position of the CP and a man does not need to move.

Once a phase is complete, the complement of the phase head becomes inaccessible to further operations, which is proposed to reduce memory burden—the mind can essentially put a completed phase to the side and compute the next phase. Note that when a phase head is Merged, there are differing views about which portions of the phase become inaccessible in accord with the Phase Impenetrability Condition (Chomsky, 2000, 2001; Müller, 2004; Richards, 2011). Under one version of the Phase Impenetrability Condition, the complement of the phase head becomes inaccessible and is transferred, but in another version, the complement (if present) of the lower phase head becomes inaccessible and is transferred. As noted by Boeckx and Grohmann (2007, p. 206),⁶ referring to Chomsky (2000), “[c]omputation cost reduction is the prime conceptual advantage and motivation for phases.” This means that all feature checking operations within the phase must be complete, and any elements that need to move out of the phase must have moved to the edge of the phase before completion. Since phases are thought to be complete (in some sense), they ideally should be of some advantage when accounting for island effects, although whether or not this is the case is open to debate (e.g., see Chomsky, 2008; Gallego, 2010). I incorporate the notion of phases into this model, since they are utilized in Labeling theory. I assume that the phases, following Chomsky (2001), are transitive VP (v*P) and CP. Note that, essentially following Chomsky (2021), I will assume that when v* or C is Merged, the v*P/CP is transferred. Thus, the head of the phase is transferred together with its complement but the specifier, if present, remains outside of the transferred phase.

3.3 Feature Inheritance and Agreement

Feature inheritance is an operation in which a phase head passes features onto a complement. The notion of feature inheritance was proposed by Chomsky (2008), based on work (to the best of my knowledge) by Carstens (2003) and Miyagawa (2005), among others (also see references in Carstens, 2003, and Miyagawa, 2005). Chomsky (2008, pp. 143–144) writes:

Feature inheritance can be useful for accounting for Exceptional Case Marking (ECM) constructions. In the ECM (3)a, the embedded T, pronounced as to, occurs without C. In this case, T lacks agreement features and him Agrees with the matrix verb expect. In (3)b, on the other hand, T, pronounced as past tense on the verb win (resulting in won), occurs with C. T has agreement features and Agrees with the subject, resulting in the nominative pronoun he. These types of simple examples demonstrate how T, in the presence of C, has agreement features, which it lacks in the absence of C. While feature inheritance is useful for accounting for the ECM data, it isn’t necessarily clear if it is required. If non-finite T simply lacks a full set of unchecked/uninterpretable phi-features (uPhi), and tensed T has a full set of uPhi, the same facts can be accounted for, without recourse to feature inheritance.

Feature inheritance notably is a complex operation that involves copying agreement features from C onto T, or the passing of features from C to T. Chomsky notes that this violates the No-Tampering Condition (Chomsky, 2000, pp. 136–137; Chomsky, 2008, p. 138), as it requires altering an already formed syntactic structure. The question then arises of whether or not it is conceptually necessary.

Complementizer agreement is found in a variety of languages such as Frisian, some Dutch and Germanic dialects, and Bantu languages (Koppen, 2017). Note that both C and a verb can show agreement with a subject, as in (4)a–b, in which the complementizer and the verb show agreement with the subject. Assuming that verbal agreement indicates agreement on T, then both C and the verb Agree with the subject in these examples.

Although the existence of complementizer agreement as in (4) has been given as evidence for feature inheritance (Chomsky, 2008; Miyagawa, 2005), complementizer agreement tends to be less common and less complete than agreement with T. Matasović (2018, p. 9) writes “that the most common agreement pattern within the domain of the clause is verbal agreement.” Koppen (2017, p. 7) writes “The CA [Complementizer Agreement] paradigm is usually defective, however, in the sense that not all person/number combinations of the subject lead to an overt agreement reflex on the complementizer.” Koppen (2005, p. 35) points out this defectivity in a variety of Germanic/Dutch languages/dialects. In Frisian, a complementizer shows agreement only with a second person singular embedded subject, whereas a verb shows agreement with all types of subjects, as shown in Table 1. Koppen (2005) discusses similar paradigms in Tegelen Dutch, Bavarian, and Lapscheure Dutch. In all of these languages/dialects, there are variations with respect to the extent of complementizer agreement, but there are fewer complementizer agreement suffixes than verbal suffixes, thus providing further evidence for the notion that complementizer agreement tends to be defective.

If it truly is the case that verbal agreement is more common than complementizer agreement and that verbal agreement tends to be more complete than complementizer agreement, then this may be an indication that a complementizer is not always the origin of agreement features. If C were the locus of agreement features, then one might expect agreement with C to be more common than it is, and for agreement with C to tend to be more, not less, complete than agreement with T.

Feature inheritance also raises technical problems. If uPhi are inherited by T, one possibility is that all of the uPhi of C are passed from C onto T and no longer remain on C. This would be the case when agreement only shows up on T (usually visible on the verb). This does not appear to be the case in the examples in (4) in which there is agreement between a subject and both C and the verb (assuming the verbal agreement is the result of agreement on T). Another possibility is that the uPhi of C are copied onto T, so that they appear on both C and T. This could account for the data in (4). Again, copying of features from one element onto another seriously alters an already formed SO, again violating the No-Tampering Condition. It would be simpler if C and T come with their necessary agreement features.

Richards (2007) provides arguments for feature-inheritance, proposing that feature transmission is a “conceptual necessity” in order to avoid transfer of uninterpretable features. Uninterpretable features, by definition, cannot be processed by the semantic component of a derivation. If uninterpretable features are checked but are not transferred immediately, then they should stay around and cause a derivation to crash, according to Richards, since they can’t be interpreted. Thus, uninterpretable features must be transferred as soon as they are checked. The assumption seems to be that when checked, uninterpretable features are transferred with the phase. They are “deleted” so that they are no longer visible to the semantic component. Assume that T has uPhi that are checked via Agree with a subject, before the phase head C is Merged. When these features are checked, they cannot be transferred until after the phase head C is Merged. Thus, these uPhi cannot be deleted as soon as they are checked. These checked uPhi, according to Richards, then become indistinguishable from interpretable features and interpretable phi-features on T, presumably, will cause a derivation to crash. The idea seems to be that since T is not an argument, phi-features (which are associated with arguments) cannot be interpreted on T. On the other hand, if uPhi are inherited from C by T, then as soon as they are inherited, they are checked, and since the phase level has been reached, the checked uPhi are instantly transferred, so that they no longer remain for the semantic component. Assuming that uninterpretable features originate on a phase head predicts phase-level operations of inheritance of features, Agree (e.g., checking of uPhi on T by phi-features on a subject), and transfer of the relevant portion of the phase.

However, the idea that uninterpretable features need to be deleted as soon as they are checked is not necessarily a given. Since these features are uninterpretable, by definition, they could cause a derivation to crash if they are transferred, but as long as they are deleted before transfer, it isn’t clear why they need to be deleted immediately – this seems to be a stipulation. Furthermore, some recent work takes the position that the complement of a phase head is not transferred immediately. Chomsky (2015) argues that phasehood can be transferred to the complement of a phase head, based on ECM constructions and the that-trace effect.⁷ As noted by Goto (2017), the motivation for feature-inheritance based on the need to delete uninterpretable features as soon as they are checked may not necessarily hold.

Another issue with feature inheritance involves probe-goal agreement. Since Chomsky (2001), agreement has typically been assumed to involve a probe-goal relation. For example, assume that T in (5) has uPhi that probe for and Agree with the phi-features on a subject. Similarly, v* has uPhi that probe for and Agree with phi-features on an object. The relations Agree(T_[uPhi], Mary_[iPhi]) and Agree(v*_[uPhi], books_[iPhi]) check the uPhi on T and Mary via probe-goal agreement.

Now consider how probe-goal agreement of this sort works given feature-inheritance. If T must inherit its uPhi from C, then the uPhi on T cannot probe until after C is Merged. Thus, probing is counter-cyclic, not from a root node, which is contrary to the original notion of probe-goal in which probing occurs from the root node (Richards, 2006).

Another problem, pointed out by Epstein, Kitahara, and Seely (2022), hereafter EKS (2022), is that given feature inheritance, there are cases in which agreement must occur with a goal that is no longer visible. Assuming feature inheritance, in (6), the uPhi features on T are inherited from C. Thus, T does not obtain its uPhi features until after C is Merged, and also after the subject has internally Merged with the TP (assuming that a subject raises to the specifier of TP). Then, following Chomsky’s (2013) view that only the highest copy of a syntactic object (SO) is visible to probing, the lower copy of the subject is not visible to probing. This means that the probe cannot find the subject in its base position. The higher copy of the subject is in the specifier of the TP, so that the past tense T does not c-command it. See Figure 3. EKS (2022) propose a solution based on Minimal Search (Agreement occurs between T and the subject in the TP). However, none of this is necessary if there is no feature inheritance. If T simply comes with its relevant set of uPhi features, then it can probe as soon as it is Merged. There is no need for counter-cyclic Agree relations, and the problem of agreement with an invisible copy of an SO does not arise.

Click to enlarge

Figure 3

Agreement Given Feature Inheritance

The facts regarding feature inheritance are far from settled, but I will assume that from the perspective of the Strong Minimalist Thesis, it is best to do without it.⁸ Feature inheritance is best eliminated from the current theory from the perspective of simplicity; it is a complex operation, and a complex operation such as feature inheritance requires extraordinary justification.

3.4 Agreement and Case

Case is subject to a great deal of cross-linguistic and language-internal variation. As is well known, Case morphology in English is basically “phonologically zero” (Pesetsky & Torrego, 2011, p. 55), except for pronouns. Case shows up on nouns in Latin, as in (7). In Russian, there are declinable and indeclinable nouns (Pesetsky & Torrego, 2011, p. 55), so that whether or not Case appears overtly on a noun can depend on the particular noun, as in (8). In Icelandic, the verb luku ‘finished’ occurs with a dative object and the verb vitjuðum ‘visited’ occurs with a genitive object, as in (9)a–b. Icelandic is also well-known for constructions in which a subject appears with dative Case and an object with nominative Case, as in (9)c. Furthermore, Bobaljik (2008) points out that nominative-accusative case systems and ergative case systems (which typically mark the subject of an intransitive verb and the object of a transitive verb with ergative case) assign Case differently, but arguments seems to be treated syntactically in the same way in both types of systems, suggesting that Case is not truly a syntactic relation.

These Case facts can be accounted for if Case is primarily a Spell-Out phenomenon. Marantz (2000, p. 20) argues that “case and agreement morphemes are inserted only after SS [Sentence Structure] at a level we could call “MS” or morphological structure.” Bobaljik (2008), following Marantz, writes that “the proper place of the rules of m-case [morphological-case] assignment is thus the Morphological complement, a part of the PF interpretation of syntactic structure (Bobaljik, 2008, p. 300).” Chomsky (2021, p. 23) suggests that “Case is part of externalization” further writing that “there seems to be no general semantic reason” for Case systems and “[p]erhaps establishing relations among elements facilitates perception/parsing.”

In my model, I take a Spell-Out-based approach to Case. I assume that Case appears at Spell-Out, following Chomsky’s view that Case is a reflex of phi-feature agreement (Chomsky, 2000, 2001). This approach can account, at least to a certain extent, for some of the language-internal and cross-linguistic idiosyncrasies that occur with Case. I assume that unchecked phi-features, uPhi, must be checked for a derivation to converge. The result of phi-feature agreement can lead to an argument being pronounced with overt Case morphology. Case, however, is a Spell-Out phenomenon. The exact form of Case can be subject to language internal and cross-linguistic variation, but the actual form of Case on an argument does not have an influence on syntax. Note that if an argument is unable to be pronounced with Case, a derivation can crash at Spell-Out (see Section 3.6).

3.5 Head Movement

Head movement is a controversial topic. It appears to be ubiquitous. However, it isn’t clear how exactly it works. Consider the basic examples in (10) and (11) which show typical head-movement of T to C. Assume that C is selected and Merged with TP. Then assume that T raises and undergoes IM (undergoes head-movement) with C, as shown in (10)b and in (11)b (assume that will is in T). These head-movement operations appear to violate the No Tampering Condition because an already formed CP is altered. Head movement also violates the Extension Condition (Chomsky, 1993, 1995) which is the requirement that operations extend the SO. Movement of a head to a position within the clause does not extend the size of the SO.

Chomsky has proposed differing views on head movement. Chomsky (1993, p. 23) argues that the Extension Condition does not apply to adjunction operations, which means that it does not apply to head movement, assuming that head movement is adjunction (also see Dékány, 2018, p. 5). However, making head-movement an exception to the Extension Condition is not exactly ideal. From the perspective of Minimalism, the notion that all Merge operations target the root of an SO is optimal. Thus, head-movement seems to violate this requirement, and this alone is enough to make head-movement suspect.

In recent work in Labeling Theory, a version which I assume here, Chomsky (2015) makes use of head movement. Chomsky (2015, p. 12) proposes that a verbal root R raises to v*, resulting in dephasing of v* (see Figure 2 above). This raising operation forms “the amalgam [R-v*]”, which Epstein, Kitahara, and Seely (2016) (hereafter EKS, 2016) specifically describe as being a case of internal pair-Merge. EKS (2016, p. 90) write “pair-Merge internally forms <R, v*> (=R with v* affixed).” Internal pair-Merge, unless it is implemented via sidewards movement, requires R to internally pair-Merge with v* after v* has Merged with the SO. This head-movement again violates the No Tampering Condition and the Extension Condition.

Despite making use of head movement in some works such as those discussed above, a variety of issues, including the violation of the Extension Condition/counter-cyclicity, lead Chomsky (2001) to propose that head-movement is a phonological (PF) operation. Chomsky (2001, p. 37) writes “[t]here are some reasons to suspect that a substantial core of head-raising processes, excluding incorporation in the sense of Baker (1988), may fall within the phonological component.” Some other reasons from Chomsky (2001, pp. 37–38) why head movement is problematic are as follows (see Roberts, 2011, for a summary of these arguments). There are no clear interpretation differences in languages such as French and Icelandic, in which the verb appears in a position generally considered to be T (possibly a result of head movement), compared with languages such as English, in which the verb remains below T. If head movement were responsible for verb movement, and if head movement influences interpretation, then the expectation is that semantic differences would arise (Roberts, 2011, p. 99). Another issue is that if a head raises and adjoins to a higher head within an SO, then “the raised head does not c-command its trace (Chomsky, 2001, p. 38).” Also, a phrase can undergo successive-cyclic movement whereby it moves from the specifier of one phrase to the specifier of another phrase, but this doesn’t appear to occur with a head. Rather, “it always involves ‘roll-up’ (i.e. movement of the entire derived constituent … iterated head movement always forms a successively more complex head (Roberts, 2011, p. 201).” For example, in an English interrogative, an auxiliary (Aux) moves to T (assuming that the auxiliary is not base generated in T), and then Aux-T raises to C. An auxiliary cannot move to C and leave T behind.

There are a variety of proposals related to head movement in the literature which make use of syntactic movement and/or post-syntactic PF operations. Embick and Noyer (2001) propose that there are postsyntactic lowering operations in which a head can lower and combine with another head. Matushansky (2006) argues that typical cases of head movement can result from a syntactic operation of movement of a head to a specifier position followed by a morphological (non-syntactic) operation that creates adjoined heads. Harizanov and Gribanova (2019) argue that some cases of head movement occur in the syntax via IM and other cases involve a morphological process that occurs outside of the syntax via an operation of “postsyntactic amalgamation” which can involve postsyntactic lowering or raising of a head to form a head-head adjunction structure. Harley (2004) argues for a non-syntactic operation of conflation (following Hale & Keyser, 2002) in which the defective phonological features of one head combine with the phonological features of a complement. Platzack (2013) also develops a purely phonological approach to head movement which permits pronunciation of heads to occur in a position that is different from in the syntax. See Roberts (2011) and Dékány (2018) for in-depth discussion of problems with head movement, as well as discussion of potential ways to account for head-movement, both in narrow syntax and at PF. Also see Roberts (2010) for an attempt to account for some components of head-movement in the syntax.

Notably, there are two types of proposals in the literature involving head-movement that occurs in the syntax without violating the Extension Condition.⁹

In some accounts (e.g., Harizanov & Gribanova, 2019; Matushansky, 2006; among others), a head can undergo IM to a specifier position. Thus, a head essentially functions as a specifier. The head moves to the root of the SO, so this movement is not counter-cyclic and does not violate the Extension Condition. However, the result of movement is an {_Y X, {YP…X}} structure in which the X does not label, so X functions as a specifier. This is problematic given the core notions of Labeling Theory, which assume that given an {X, YP} structure, the head X labels.

Another approach is consistent with Labeling Theory. In this approach, a head undergoes IM with the core SO, and then the head relabels. Thus, IM of X with YP, forms an {_X X, {YP…X}} structure in which X functions as the label. Presumably, X also functions as a label in its base position, and thus, this is sometimes referred to as “relabeling” as well as “reprojection”. For example, this type of head movement might be plausible in relative clauses, if one assumes that a nominal head raises and relabels, and if relabeling by a simplex head does not go against the core assumptions of Labeling Theory. This type of relabeling approach can be found in Georgi and Müller (2010), Donati and Cecchetto (2011), Cecchetto and Donati (2015), and Fong and Ginsburg (2023), among others. Note that this type of head movement might be compatible with my computer model, although it is not utilized for the derivations discussed in this paper (as it is not necessary).¹⁰

Although the issues regarding head-movement are far from settled, I adopt a model in which there is no counter-cyclic head movement in the narrow syntax. I take an affix hopping approach (Chomsky, 1957) in which a set of Spell-Out rules are applied to the output of the syntax.¹¹ Crucially, if X (an affix) and Y are adjacent at PF, then it is possible for Y to be linearized before X.

The basic rules that I implemented are given in (12) below. Note that only a small set of rules is sufficient to account for the basic English constructions produced by my model. The computer model generates a syntactic structure. If a derivation does not crash, the nodes of the tree are sent to Spell-Out. Then the basic PF (Phonological Form) rules apply when necessary. Examples of rule applications for particular Spell-Out forms are shown in Table 2. Note that for these rules to apply, the PF component of the derivation has to have access to some syntactic category information. Thus, if the model finds T adjacent to v* which is adjacent to a root R, or an auxiliary, then T attaches onto the root or auxiliary, and v* is eliminated from PF, since it has no pronunciation. As shown in Table 2, for Tom saw Fred, the adjacent SOs T(Past,3rd,sg) v* see are converted into T(Past,3rd,sg)+see which is pronounced as saw, where T ends up suffixing onto the adjacent see; with a regular verb, past tense is generally pronounced as -ed. The verbal head v* is not pronounced. For Will Tom read a book, the interrogative C and T combine to form C_Q+T(Pres,3rd,sg). Note that this requires T to attach onto C_Q by moving over the subject at Spell-Out. Furthermore, the auxiliary will must move over the subject and combine with T to form T(Pres,3rd,sg)+will, which ends up being pronounced as will. In cases in which T combines with C, but there is no overt element in T, then the appropriate form of do (depending on Tense and agreement) is pronounced. The appropriate forms of do as well as irregular verb forms are listed in a lexicon, which the model consults.¹²

3.6 FormCopy

Chomsky (2021) proposes a FormCopy operation which accounts for how nominals are interpreted as Copies. Chomsky (2021) describes FormCopy as a rule that assigns “the relation Copy to certain identical inscriptions” (p. 17) that are in a c-command relation, and that presumably are in the same phase. Identical inscriptions refers to arguments that are identical in form. Consider how this accounts for the Control construction (Chomsky, 1981, 1986) in (13)a, as shown in (13)b. Here, I assume that the Control construction is a TP.¹³ The NP many people₁ is externally-Merged in the vP theta-position and it undergoes IM to the non-finite TP specifier position. The NP many people₂ is separately externally-Merged in the matrix vP theta-position, and internally Merged with the matrix T. FormCopy applies and all inscriptions of many people, except for the highest many people₂ in the matrix TP, are interpreted as Copies, which are not pronounced. If FormCopy were not to apply in (13), the lower many people₁ would be treated as a separate NP (a repetition, not a Copy) from the higher many people₂. This is fine, as it has a separate theta-role from the matrix many people. However, the construction would then crash. Why exactly it would crash is then an issue. I assume that it would crash at Spell-Out due to Case reasons, not due to issues with the syntax. The relevant argument many people₁, not being a Copy, would need to be pronounced, but it would not be in a position in which it could obtain Case, in the edge of a non-finite TP, and the matrix try is not the kind of verb that assigns accusative Case.

Note that FormCopy has the advantage of eliminating the need for memory of movement operations. In (14), assuming the standard view of the VP-internal subject hypothesis, the subject John is externally-Merged in the v*P. Then it internally-Merges with T in subject position. With FormCopy, there is no need to retain memory of the movement of John. When the construction is generated, John undergoes IM. Then, at the phase level, FormCopy applies and the lower inscription of John is interpreted as a Copy of the higher John.¹⁴

A number of issues arise regarding the FormCopy operation and its formulation in the literature. First, the question arises of whether or not FormCopy can apply freely. Chomsky (2021, p. 25) writes:

Chomsky indicates that FormCopy is optional, but I don’t take this to mean that it applies freely. Rather, it applies in some, but not all, cases in which there are multiple inscriptions with the same form.¹⁶ Specifically, it applies at the phase level. Chomsky et al. (2023, p. 25) write that “[i]n technical terms, the point at which FC [FormCopy] applies is referred to as the phase level.” Limiting FormCopy to the phase level accounts clearly for why FormCopy cannot convert the lower John₁ into a Copy in John₂ saw John₁. As shown in (15), FormCopy should apply when the lower phase head v* is Merged, before John₂ is Merged. After John₂ is Merged, the lower v* and its complement are no longer accessible to FormCopy. Thus, FormCopy cannot apply between John₂ and John₁.

In my model, I adopt this phase-level approach to FormCopy. Once a phase-head is Merged, FormCopy applies if possible. If X and Y are identical inscriptions of arguments in the same phase, then FormCopy converts Y into a Copy of X. Furthermore, FormCopy is beneficial as it enables us to do away with the need for the language faculty to memorize movement operations.¹⁷

3.7 Strength and Labeling

Chomsky (2013, 2015) utilizes a notion of strength, combined with the need for Labeling, to account for the traditional EPP effects in languages such as English. T is weak so that it must be strengthened. In order to be strengthened it must form an {XP, YP} structure with an argument. Chomsky (2015, p. 7) suggests that a lack of “rich agreement” may be a reason for the weakness of T in English, unlike in null subject languages like Italian that have rich agreement. Chomsky suggests that in a null subject language, T may be able to label by itself.

The Labeling-based analysis of TP in English, in one sense, enables us to do away with the EPP. But on the other hand, the notion of weakness is not entirely clear. If richness of agreement is at play, then questions arise around languages (e.g., Japanese) which do not appear to show any agreement, but allow null subjects.

If certain heads like the English T can be weak, then the question also arises of what to do with non-finite T. Assume that finite T is weak and requires agreement to be strengthened for Labeling. But non-finite T does not show agreement. In (16), non-finite T, pronounced as to, and the embedded subject Mary are adjacent. One possibility is that the embedded TP has the structure in (16)b with Mary and the non-finite T forming an {XP, YP} structure. In (17), there is no overt argument in the embedded clause, but there should be a Copy of John (or PRO) in the embedded clause, as in (17)b. In this case, since the lower Copy of John is not pronounced, it should be invisible to Labeling, and thus non-finite T cannot be strengthened. Presumably, non-finite T can label, and thus, it does not require strengthening.

Chomsky (2015) assumes that in an ECM construction such as (16), the embedded subject raises to the matrix object position. This follows work by Postal (1974) and Lasnik and Saito (1991), among others. For example, the embedded subject in an ECM construction can be passivized and it behaves like it is in the matrix clause with respect to binding effects. Chomsky argues that the Root expect is weak and thus must be labeled by an {XP, YP} structure. On this account, the embedded subject raises and forms an {XP, YP} structure with the matrix verbal root. For example, in (16), the structure {Mary, {expect Mary to arrive}} is formed. Problems with this approach are that the root expect must inherit uPhi from the higher v* (if feature inheritance is assumed) and that expect has to undergo some form of head movement over Mary. If head movement is a Spell-Out phenomenon, then this would happen at Spell-Out.

Due to the complexities involved, I take a simpler approach—Mary is in the embedded non-finite TP in (16), where it Agrees with the higher v*, resulting in accusative Case appearing on Mary at Spell-Out. The evidence that the subject of a non-finite clause can behave like the object of the matrix clause is strong, but whether or not this requires the embedded subject to actually undergo IM with the matrix verbal root is not clear. I will simply assume that, due to the lack of an intervening phase boundary, an ECM subject can behave like it is a matrix object even if it is in the embedded clause.¹⁸

There is evidence that an overt subject can appear in a non-finite clause. In examples such as (18), the subject him appears to be in the embedded clause. It seems to be a fairly standard assumption that the embedded subject is in the non-finite T (e.g., see Chomsky & Lasnik, 1977). If Labeling Theory is correct, then there must be some type of agreement relation between non-finite T to and him.

If the embedded subject of a non-finite clause remains in the non-finite T, as in (16) and (18), there is a potential Labeling issue. If non-finite T and the subject do not Agree, then there should be a Labeling failure. To get around this problem, I assume that non-finite T contains uPerson that Agrees with an argument. In (16) and (18), the uPerson of non-finite T is checked by the Person feature of the subject. Since the subject is internally-Merged with the non-finite T, the result is an {XP, YP} structure that is labeled with shared Person features. In (17), uPerson is checked by the Person feature of John. Since John is a Copy here, there is no visible {XP, YP} structure, and to labels by itself. Thus, non-finite T can either label by itself, or in an {XP, YP} structure with an argument that it shares a Person feature with.

The question then arises of whether or not there is evidence for agreement between a subject and non-finite T. Notably, agreement in infinitives can be found in a variety of languages. For example, standard Brazilian Portuguese has infinitives that can be inflected for person and number as in (19).

Inflected infinitives are also found in European Portuguese (Raposo, 1987), as well as other Romance languages such as Galician and Old Neapolitan (Groothuis, 2015; Scida, 2004). Hungarian also has inflected infinitivals, as shown in (20).

An anonymous reviewer points out that the distribution of subjects with agreeing infinitives and with non-agreeing infinitives is different. According to Pires (2006, p. 93), in Brazilian Portuguese a non-inflected infinitival requires a PRO subject (which has a local antecedent), and an inflected infinitival has a pro subject (which does not require a local antecedent). Furthermore, a non-agreeing infinitive requires a sloppy reading under ellipses, whereas an agreeing infinitive permits a strict or sloppy reading, and a non-agreeing infinitive does not permit split-antecedence but an agreeing infinitive does. Although an in-depth analysis is beyond the scope of this work, I suggest that these differences boil down to whether or not the non-finite T Agrees partially or fully with an argument. In some Portuguese infinitivals, there may be full agreement with an argument (the infinitival property is due to the lack of tense, not phi-features), and pro is permitted. In non-agreeing infinitivals, there can only be partial agreement, which is not sufficient to check Case on an argument and only PRO is permitted.

Even though there is no clear overt indication of agreement in modern English infinitives, it is possible that there is partial agreement, as found in languages such as Portuguese, Hungarian, etc. Thus, I assume that T can label either by itself or via shared Person features.

Furthermore, I assume that heads can generally label. Mizuguchi (2017, p. 331) suggests that “[h]eads can label only when they are without unvalued features.” If a head has unvalued/unchecked features, it is incomplete, and thus it is reasonable to assume that Labeling isn’t possible. I adopt this view in my model; heads can label by themselves as long as they lack unchecked features. A root, however, needs to be categorized. Thus, a root cannot label by itself. For example, the root walk can be labeled only after it combines with a categorizer N or v.

3.8 Box Theory

Chomsky (2024) proposes Box Theory, in which the traditional A-bar elements (wh-phrases, topicalized phrases) are essentially placed into a box structure that can be accessed by C (and possibly other functional heads associated with topic and focus). The derivation of (21) proceeds along the following lines. When v* is Merged, the lower v*P phase is complete. At this point, the elements within the v*P are no longer accessible. However, the wh-phrase what is in a Box, where the Box could be thought of as a structure that contains focused (A-bar) elements.¹⁹ When the matrix interrogative C_Q is Merged, it looks into the Box and finds what. The wh-phrase what ends up being pronounced at the position of C, but it still remains in the Box. Crucially, what in its base position must be converted into a Copy, and so FormCopy must apply.

Timing of insertion into the Box is an issue that arises. Chomsky (2024) suggests that boxing is contingent upon IM. Chomsky writes segregation of a boxed element is “established by IM [Internal Merge], which carries the derivation from the propositional to the clausal domain.” Chomsky further writes that “we can think of the element E that is IM-ed to the phase edge as being put in a box, separate from the ongoing derivation D.” Chomsky appears to be proposing that boxing results from IM of a particular SO to the phase edge. Note that it is simpler to put an SO into the Box, without boxing being contingent on IM, rather than to do IM of the SO followed by boxing. Furthermore, I also assume that IM of arguments is free (see Section 4). If IM to a phase edge results in boxing, then there could be overgeneration of boxed SOs.

Assuming that boxing happens without IM, it could be that as soon as a wh-phrase is externally Merged with an SO, it goes into the Box, or it could be that it goes into the Box at the phase-level. Also, when a phrase is accessed from the Box, its base position should be treated as a Copy. This means that FormCopy must apply. FormCopy could apply as soon as an SO goes into the Box, or it could apply as soon as the SO is accessed from the Box. Also, consider (22). In this case, C_Q and the subject who are within the same phase. So whether or not who has to go into the Box isn’t clear as C_Q should be able to access who without looking into the Box.

In my implementation, I had to make decisions about the timing of the Box operation. From an implementational perspective, it is easier to put a wh-phrase into the Box as soon as possible, rather than to wait until a phase is complete, since waiting requires checking an already formed structure for wh-phrases (or other phrases that need to go into the Box). Thus, my model places a wh-phrase into the Box as soon as possible. Since the Box is assumed to exist, this applies to wh-subjects too. The model places wh-phrases in the Box and C_Q can only see into the Box. FormCopy applies when C_Q accesses a wh-phrase from the Box.²⁰

Assuming the existence of the Box, successive-cyclic wh-movement is potentially an issue. If an argument is in the Box, there is no reason for it to undergo IM to an intermediary position. However, there is evidence for successive-cyclic wh-movement. Some well-known evidence is the existence of partial wh-movement (McDaniel, 1989). For example, in German and Albanian, a wh-phrase can appear in an intermediary position and a wh-phrasal scope marker (or some type of question element) can appear in the relevant scope position, as shown in (23) and (24). In Malay, as shown in (25), a wh-phrase can move to the edge of a clause in which it does not have scope, and be interpreted with scope in a clause in which there is no overt wh-marker.

The presence of a wh-phrase in an intermediary position has generally been taken to indicate that a wh-phrase undergoes movement through intermediary positions. Further well-known evidence for successive-cyclic movement is the existence of complementizers that agree with wh-phrases. For example, Irish has a particular complementizer that appears in a non-interrogative embedded clause when a particular wh-phrase undergoes long-distance movement (McCloskey, 1979, 2001).

Box Theory can deal with successive-cyclic wh-movement as follows. An intermediary complementizer can access the Box at externalization, but not in the syntax. This means that in some languages and constructions, there is access by an intermediary C, of an element in the Box, and this has an influence on Spell-Out. For example, an element in the Box can be pronounced at an intermediary position without actually being in that position in the syntax. Chomsky (2024) writes that in partial wh-movement constructions, presumably such as (23)–(25) above, there is a Labeling violation in the embedded clause in which a wh-phrase appears, as the wh-phrase does not share features with the non-interrogative C.²¹ To deal with this issue, Chomsky writes that “the boxed wh-XP is accessed” in the intermediary position “and under Externalization, spelled out, but with no labelling problem since the phrase does not appear in the derivation.” Although a variety of issues regarding when and how a wh-phrase is accessed for externalization require further examination, this analysis can account for partial wh-movement facts. If this approach is correct, then a wh-phrase does not actually undergo successive cyclic movement, but it can be accessed successive-cyclically, subject to language-internal and cross-linguistic variation, at the point of externalization.

The notion of the Box is beneficial in the following ways. First, it becomes possible to transfer a phase as soon as a phase head is externally-Merged. There is no need for an escape hatch at the v*P phase edge, which was previously assumed to exist to account for A-bar movement of a wh-phrase (e.g., see Chomsky, 2001). When v* is Merged, the v* head and its complement can be transferred. Furthermore, when C is Merged, transfer of the CP can occur. Elements at the edge of a CP presumably are accessed via the Box. Under the traditional Phase Theory view, the edge of a phase is transferred separately from the rest of the phase. But the traditional view complicates transfer of a matrix CP. Under the traditional view, when a matrix CP is formed, first the complement of C is transferred and then the CP edge is transferred, thus requiring two transfer operations to occur at the edge of a CP. This is no longer necessary. When C is Merged, the complete CP can be transferred.

While issues remain regarding the exact definition and nature of the proposed box structure, from an implementational perspective, it has some advantages.

3.9 Arguments as NPs

Although the status of determiners is peripheral to this work, it is necessary to explain how they are implemented in this model. I assume that arguments are NPs and not DPs (contra the typical DP hypothesis of Abney (1987), and much following work). The view that arguments are NPs, and not necessarily DPs, has been suggested by Chomsky (2007), as well as by Van Eynde (2006), Bruening (2009), Oishi (2015), Bruening, Dinh, and Kim (2018), and Bruening (2020), among others.

If an argument is really a DP, then problems arise regarding phi-features. As Bruening (2009, p. 28) points out, transitive verbs select nominal arguments, whereas they do not select for determiners. The typical approach in the Minimalist Program incorporates Agree relations between functional heads and arguments. Assume that unchecked phi-features on a probe on T or v* form an Agree relation with phi-features on an argument, in which uPhi on T Agree with phi-features on the argument. If the phi-features are on a nominal head N that is contained within a DP, then there is a potential problem because Agree(T, DP) would require T to see inside of the DP to the NP (or Agree would require features of N to percolate up to the D head). It is simpler for T to simply Agree with the head of the NP.

Although determiners can show phi-feature agreement in many languages (e.g., in Romance languages, etc.), gender, person, and number are properties of nominals, not determiners. Number shows up on nouns (e.g., cat vs. cats). Person and gender show up on pronouns (e.g., I vs. you, he vs. she). Agreement between determiners and nominals can occur in languages, such as in (26)a–b from Spanish, but as Bruening (2009, p. 30) points out, “every element in the nominal phrase must agree with the head noun in gender and number,” suggesting that the core element in these phrases is the N, rather than a determiner, quantifier, or adjective.²²

In my model, I indicate an argument as an NP as shown in Figure 4.²³ When there is a D, it is pair-Merged to the NP. Pair-Merge is indicated with a dotted arc. Given two SOs X and Y, when X is pair-Merged with Y, forming <X, Y>, X is less prominent than Y and generally not accessible to syntactic operations (see Chomsky, 2000, 2004). When the NP is Merged with another SO, the pair-Merged D is invisible to Agree relations.

Click to enlarge

Figure 4

NPs Pair-Merged With D