1
Doubt the Labeling Theory!* Takashi Munakata
Yokohama National University Interface_condition@leaf.ocn.ne.jp
0. Introduction
I pick up a number of problematic issues of the Labeling Theory advocated by Chomsky (2013, 2015) and give some critical comments on these issues to share the basic points of these theoretical problems as the starting point of more exciting discussion.
Basically, I raise controversial issues of the Labeling Theory, including the properties and status of Labeling and the Labeling Algorithm, pair-merge and the status of v*P. Then, I inspect these issues with scrutiny and attempt to identify and discuss the theoretically problematic points.1
Note that I don’t intend to spoil the Labeling Theory. Rather, I hope that the whole discussion of this talk drives the theory to proceed in a good and fruitful direction. For this purpose, I sometimes make alternative proposals.
1. The semantic characteristics and status of Labeling
In this section, I deal with the semantic status and properties of Labeling (especially, <φ, φ>), focusing on the interface requirements (Full Interpretation and Argument Structure), the treatment of categorial features, and the weak versus strong distinction of heads.
1.1 (Dis)connecting the interface requirements and the status and properties of Labeling
(1) Properties of Labeling a. Requirements by Interfaces
* I would like to express my gratitude to Shoichi Takahahasi to offer me to the chance of the presentation. Also, I thank Tadao Nomura for inspiring me. Without him, I won’t challenge this quite tough work. Finally, I am grateful to Roger Martin for his suggestion.
1 In this presentation, I intensively focus on theoretical arguments without almost positing concrete examples, because I limit the scope of this presentation to concentrate on the architectural aspects of the Labeling Theory.
2
Chomsky (2015) suggests that Labeling offer special service to the semantic interface, while the phonological interface may use Labeling, rather ignored by narrow syntax.
b. No categorial feature should contribute to Labeling
Abandoning endocentric headed structures (cf. Narita 2014) c. Ability to label or not: weak vs. strong distinction
d. Unseen labels made existent
Sometimes, non-topmost syntactic elements are made invisible (a non-head element of a chain), making no contribution to Labeling.
e. Shared features bear label
This seems to correspond to the intersection of a set.
(2) The requirements by interfaces
i. Semantic interface(s) [C-I/ C & I (Uriagereka 2008 and Munakata 2005, 2009)]2
a. Full Interpretation
b. Argument Structure/ Selectional Properties C-Selection or s-selection may be included
c. Intentional & discourse-oriented properties
ii. Phonological interface (Takita 2016, Dobashi 2017,Bošković to appear) a. Full Interpretation
b. Linearization
c. Formation of the phonological domain and phonological units
(3) Problems for Full Interpretation
Though Epstein, Kitahara and Seely (2014) and Sorida (2016) suggest no Labeling violates Full Interpretation, don’t the following posit the violation of it?
a. v* made invisible, and thus, uninfluential by pair-merge (Chomsky 2015) b. Disappearance of C in case of that-trace effect
(4) Useless but necessary function of v*
Is non-substantive v* necessitated by the interface? In other words, does it
2 These authors claim that Conceptual and Intentional Domain which is related to the interface at phases should be divided to C-Domain and I-Domain and narrow syntax maps syntactic structures from C-Domain to I-Domain.
3
violate Full Interpretation?
According to Chomsky, phase status and inflectional/functional properties including verbal status is transmitted to R. However, given that only R cannot label a node and Labeling by R becomes possible only when a functional head v* appears with it after pair-merge, eligible Labeling cannot be achieved unless the Labeling algorithm knows v* is adjoined to (pair-merged with) R. That is, v* is visible, at least, in the form of <v*, R> to the Labeling algorithm. As a result, this complex can express a verbal property.
Also, the interface needs to know R goes with v* to assign the appropriate theta-roles (argument structure) on argument NPs (syntactic structure). If Labeling doesn’t provide this information, it is mysterious how theta-marking is achieved.
Thus, v* needs to be interpreted at the interface. Then, the Labeling algorithm should retain the relevant information of v*. If not recovered, it may violate Full Interpretation.
(5) [ who do you v* think Ø [α who ]]
i. IM of who due to strengthen the label of T ii. Inheritance
Between ii and iii, some version of Agree (a greed version (cf. Bošković 2007) or a probe-goal where who is a probe for case) is necessary to assure iii.3
iii. Labeling of α as <φ, φ>
iv. C → Ø, “so that who within the embedded clause can remain in situ and still be accessible to IM in the next phase.”
v. Transfer of [note that the information on the label of is also transferred to the interface, letting who undergo a further IM]
(6) Labeling via the valuation of shared feature
[ I v* wondered Ø [α whoQ TQ ]]: the value of Q = interrogative
[ I v* understood Ø [α whatQ TQ happened]]: the value of Q = relative α = <Q, Q>
3 This Agree issue is unclear under Chomsky (2015), though he assumes Agree is a pre-requisite for the Labeling via feature sharing.
4
(7) Disappearance of C (that-trace effect)
Chomsky (2015) claims that at feature inheritance, phasehood as well as all other inflectional and functional properties of C (φ-features, tense and Q) is inherited by T, and this status is activated on T when an empty C with no substantive property is deleted.
However, C plays a role in expressing the clause status and sentential force. This information should be achieved with the label of <φ, φ>, which always appear when φ-features are shared by NP within TP-Spec and T or between NP within RP-Spec and R. This seems mysterious why such Labeling can indicate clause status and sentential force. Note that even though T inherits the clause status and sentential force from C, Labeling doesn’t express this information, unlike the label <Q, Q> in (6), where the value of Q can express relative, interrogative or exclamative (Chomsky 2015, footnote 16). Thus, Full Interpretation may care about the existence of C unless Labeling properly picks up the clause status and sentential force C indicates.
(8) Argument Structure and the Labeling via Sharing of φ-features
Traditional TP and VP become <φ, φ> due to the Labeling algorithm. However, it is necessary to give evidence for the interpretative contribution (or role) of <φ, φ> at the C-I/I, which satisfies Full Interpretation.4
a. The interpretability of <φ, φ>
Remember that φ features are uninterpretable for T and V/v* (cf. Oka 2016, Narita 2014). The immediate question arises as to the uninterpretability of φ; that is, what does semantic contributions this apparently uninterpretable <φ, φ> make? (cf. Narita 2014)
b. <φ, φ> instead of T
Unlike T, which has distinctive semantic properties, φ does not have any clear semantic property contributing to sentential structures. Worse, it is uninterpretable on T, which may mean that it is harmful for the interpretation of sentences, though its bundle has a semantic role in the interpretation of NP.5
The same problem is applied to the c-selection of embedded clauses by
4 See Tonoike 2014 for the relevant discussion.
5 Rizzi (2015b), following Shlonsky (2013), only person feature needs to be shared between T and subject NP. In that case, the interpretability of <φ, φ> is assured if it indicates the viewpoint of events, though it remains mysterious how the person of <φ, φ> behaves within the verbal domain.
5
verbs and argument structure in case of (5) and (7).6 Assuming label plays a major role in selection based on argument grid, it cannot be said that <φ, φ> can occupy the grid of argument structure instead of C. Also, it is unreasonable that the semantic interface admits verbs to take <φ, φ> not CP as complement. Put differently, this means these verbs can take CP which is actually <φ, φ>, which is also shared by R and NP, contrary to the fact. One may argue against this by saying that it is no problem because T inherits the functional properties of C. However, then, it becomes mysterious why Labeling is necessary.7
c. <φ, φ> instead of RP (VP)8
Remember that R is too weak to label, though R might make contributions at interfaces by itself as a lexical element. Clearly, it is difficult for the semantic interface to see how the relation between R and its complement is established, because R is indirectly connected with its complement via feature-sharing <φ, φ> after IM of the complement NP, not H-XP relation detected by minimal search. It is not reasonable to think this feature sharing expresses the relationship discussed here.
Note that Chomsky argues the entire relation v*-R-complement is involved in theta-marking. However, it seems questionable for the interface to see which relationship is established in nominal structures [<n, R> [φ NP R (NP)], spec-head or head-complement.9 This happens because the lowest lexical domain cannot be determined due to the lack of specific lexical heads, such as V and N, under Chomsky (2015).
(9) a. [v* [α DP [ R DP]]] → b. [v*-R <v*, R> [φ DP
(10) i. form R- by EM ii. IM of DP in α (EPP)
6 See Obata (2016) for the relevant discussion.
7 This also seems to posit a question on the status of C, because C may be dispensed with if the interpretive process of clauses goes well with C deleted.
8 Under Chomsky’s (2015) analysis, verbal status is not expressed by a functional head v*, not a lexical head V, because only R is a proper lexical element. Thus, <φ, φ> appears by replacing RP in the traditional sense.
9 As Chomsky (2015) notes in footnote 16, agreement is necessary for Labeling in this case. Then, this problem disappears, though it remains necessary to distinguish spec-head from head-complement in the nominal domain.
6
iii. Merge v*, reaching the phase-level
Between ii and iii, some version of Agree (a greed version (cf. Bošković 2007) or a probe-goal where who is a probe for case) is necessary to assure iii.
iv. Inheritance
v. Labeling; α is Labeled <φ, φ>
vi. R raises to v*, forming R with v* affixed, hence invisible, so phasehood is activated on the copy of R, and DP (which can be a wh-phrase) remains in situ, at the edge.
vii. Transfer of
(11) Argument Structure and Labeling10
Theta-roles are given to arguments according to the theta-grid/argument structure, by applying it to syntactic structures with Labeled units.
(12) Argument Structure and v*-R a. Internal Argument
Chomsky (2015) notes that theta-marking is undergone with the entire relation v*-R-complement. As noted, v*-R-complement relation is buried under <φ, φ>. Also, as shown in (9), there is no head-comp relation between v*-R and DP. Now, let’s take this Chomsky’s suggestion seriously. He might assume that the head-comp relation is reestablished by changing [v*-R <v*, R> [φ DP [R R into [v*-R <v*, R>, DP] (cf. Narita (2014) and Takita, Goto and Shibata (2015), Cecchetto and Donati (2015)), after pair-merge of R and v* and transfer of the complement of R triggered by the transmission of phase-status from v* to R. However, this faces at least two problems: that is, the deletion of label of <φ, φ> as in (i), which is fixed once and all at Transfer/the phase-level <v* R>, and the destruction of structural relation, as shown in (ii). Let’s inspect each derivation:
10 Nobu Goto told me that Chomsky suggested at the lecture in Arizona University last spring that theta criterion is not valid under the current theory, which Saito also claimed at the seminar in Osaka University last summer. But, it is important to note that argument structure should still be necessitated to make an appropriate interpretation of predicate and argument relations. Thanks to Goto Nobu for providing this information.
7
i. [v*-R <v*, R> [φ DP [R R
The removal of Transferred SO
→ [v*-R <v*, R> [φ DP [R R …]]
The lower copy of R is ignored, because it may be invisible to the Labeling algorithm.
→ [v*-R <v*, R> [φ DP […]]
→ [v*-R <v*, R> [φ DP […]]
Destroying the node encoded by φ, which is kept via phase memory, as shown in (ii), it may be the case that no barriers are put between <v*, R> and DP.
First of all, this violates Full Interpretation, because the lower copy of R cannot survive, even if it is a phase-head.11 Also, as suggested in note 12, it is unclear whether the node encoded by φ is recovered, which should violates Full Interpretation, too, given that Labeling should be used later at the interface.
ii. We may think of four possible derivations with set notation, all of which violates the No Tampering Condition requiring that neither X nor Y is modified by Merge:
1. {<v*, R> {α DP { R, DP}}} Step 1
The removal of Transferred SO, that is the lower copy DP.
→ {<v*, R> {α DP { R }}} Step 2
The removal of the lower copy of R, which is invisible to the Labeling algorithm
→ {<v*, R> {α DP { }}} Step 3
→ {<v*, R> {α DP { Ø}}} Step 4 Recreating the structure
→ { <v*, R>, DP} Step 5
A new merge occurrence and restructure is necessary at the step 4, where the syntactic computational mechanism picks up DP from the set α and merge <v*, R> and DP, to form . This structural change necessitates the syntactic computational mechanism to newly merge
<v* R> (X) and DP (dummy Y) again, destroying all structure of a true Y (the node with the label <φ, φ>) and deleting its interior structure;
11 Yuji Sugimoto points out the possibility in a different context that R could be visible because the lower copy is different in the form from the upper copy involved in <v*, R>. If so, this deletion of the lower copy of R poses a vital problem on this kind of the approach.
8
that is, this is the restructure in true sense which clearly violates NTC.12
2. {<v*, R> {α DP { Ø}}} Step 4 Delete of because it is null.
→ {<v*, R> {α DP { Ø}}} Step 5.1
→ {<v*, R> {α DP}} Step 6.1
The reduction of singleton set α: {DP} → DP. Without further merge, the head complement relationship is established between <v*, R> and DP.
→ {<v*, R> {α DP}} Step 7.1
→ { <v*, R>, DP} Step 8.1
This has two serious problems. First, narrow syntax heavily changes the interior structure of α by further deleting the hierarchical information at Step 5.1 and Step 7.1. However, under the set notation assumed, merge creates hierarchy (cf. Hornstein 2009) by connecting a new head X with an existent set SO. Thus, deleting the hierarchy encoded by sets mean remerge (i.e. unmerge) in that the relationship created by merge vanishes. Second, narrow syntax destroys the interior of α clearly by deleting the structure of , removing it from the structural relationship and derivational history.
These don’t violate the weak version of the No Tampering Condition but the strong version of the No Tampering Condition that states that efficient computation will leave X and Y unchanged. More concretely, this deletes the derivational history of derived syntactic structures, which syntactic computation should keep to offer original structural information to the interfaces. Also, this operation can be said to be an equivalent to reverse merge, against the spirit of Bare Phase Structure. 3. {<v*, R> {α DP { Ø}}} Step 4
By reducing null set {Ø} → Ø.
→ {<v*, R> {α DP Ø}} Step 5.2
→ { <v*, R>, DP} Step 6.2
The reduction of singleton set α: {DP} → DP. Without further merge,
12 Note that this restructure vanishes all derivational history of the node encoded by <φ φ>, which is not transferred to the interfaces. Even if phase-memory retains this derivational history, the disappearance of the node encoded by <φ φ> confuses the mechanism of the interfaces because the transferred syntactic structures are totally different from and absolutely incompatible with one that phase-memory provides.
9
the head complement relationship is established between <v*, R> and DP. However, this destroys the hierarchy twice, violating the No Tampering Condition, as mentioned.
The methodologies of 2 and 3 require the interfaces to interpret SO consisting of sets established by merge differently. That is, even though the interfaces usually deal with syntactic structures established by set-merge which set-formation plays a role in establishing hierarchies, these interfaces need to process syntactic structures where this hierarchical information can be deleted by reduction – newly creating operation contradicting merge. Thus, they are should be abandoned.13
4. {<v*, R> {α DP { R, DP}}} Step 1 Set-merge of the new copy of R with the higher copy of DP
{<v*, R> R {α DP { R, DP}}} Step 2
The other possibility is that, making use of the copy of R, R reestablishes head-comp relation with the upper copy of DP, given that the lower copy of DP cannot be used due to Transfer. Clearly, it violates both the weak version of and the strong version of the No Tampering Condition, in addition to back-tracking. In either case, the head-comp relation cannot be built between R-v* and DP.
b. External Argument
According to (10), this argument is merged with the complex SO made up of the head <v* R> and the SO α, making {XP, YP} structure [ DP [<v* R> [ =<φ, φ> …]]. The Labeling algorithm faces two problems here. First, the external argument makes a hole within for Labeling of XP-YP structure with <v* R>. However, the lower copy of DP is invisible to the Labeling algorism, failing to maintain a structural relationship with <v* R>. This means that it cannot match the positon of an external argument at the theta grid:14
13 Thanks to Shoichi Takahashi for suggestion. Also, I am grateful to Seiya Negami for relevant discussion.
14 Note that the interface not the Labeling algorithm should be able to notice the existence of the lower copy (cf. Narita 2014), though the structural relationship once established has gone, which should be read off via Labeling at the interface.
10
i. { T { DP { <v*-R> <φ, φ> ...}}} External merge of DP
→ {α DP { T { DP { <v*-R> <φ, φ>…}}}}
The lower copy of DP becomes In v is ib le due to IM of DP The list of every relevant case of set-merge
1. { DP {<v*-R>, <φ, φ> ...}} → { DP { <v*-R> <φ, φ> ...}}
Nullify the set due to the invisibility the lower copy of DP. Also, the relationship the lower copy of DP has, making it disconnected from structures.
2. { T { DP { <v*-R> <φ, φ> ...}}} → { T { DP { <v*-R> <φ, φ>…}}}
→ { T { Ø { <v*-R> <φ, φ>…}}}
In a viewpoint of derivational history, it could be said that this structure includes the occurrence redundant null merge of a dummy null element with the set , for which it seems that the indirect relationship between and disappears and is cut off from because of the existence of a nullified set .
As shown schematically above, the invisibility of the copy brings about disfavored consequences for the derivational history, which implies that it is good to recover the structural relationship the lower copy is involved even if the Labeling algorithm doesn’t take care of lower copies.
The second problem is related to the proposal that v* disappears, and, hence, made invisible for Labeling. Given that external theta role is assigned according to the theta-grid, the verbal properties of v* are necessary to be recognized at the semantic interface. However, only the label R is visible and this head contributes to the interpretation of argument structure, lacking the verbal head v*. It is mysterious how the interface recovers the verbal status of v* from only R, given that it is unclear how the interface discerns the verbal properties of v* transmitted to R through inheritance and pair-merge, which seems need to inspect the interior of the label R. This seems to weaken the role of label.
(13) Intentional & discourse-oriented properties a. Sentential force – Expressed by C
b. Syntactic structure including c-command relations responsible for
11
scopal properties and binding relationships
c. Basis for Discourse Structure/ Information Structure
Information on topic, focus, discourse-formation, evidentiality, other discourse-oriented properties, and so on (namely, the main concern of the cartographic approach)
(14) Semantic types and Labeling (Takita 2016 and Cecchetto and Donati 2015) Labels don’t correspond to semantic type one to one. That is, it is the case that different labels represent the same semantic type. Also, Takita points out that it is unclear why <φ, φ> is parallel to the semantic type t.
(15) Short Summary on the contribution of Labeling to the interface:
It doesn’t offer special service to the interface, though it partially meets duties imposed by the semantic interface.
1.2 Are Categorial features survivable?
(16) No contribution of categorial feature to Labeling
Lexical heads whose status is mainly expressed categorial feature is replaced with Root R, leaving the categorial marking to functional heads. Also, Chomsky (2013, 2015) abandons projections with category and endocentricity instantiated by X’-Theory
(17) Three types of Labeling (Chomsky 1995, Citko 2011, Uriagereka 1998) a. Intersection – Labeled with feature sharing
University {AGU, YNU}
√ Disfavored because no intersection is usually observed. b. Union – pair-merge proposed by Citko
Label Algorithm Inquirer Group {Goto, Sugimoto}
√ Disfavored because sometimes contradictory syntactic elements undergo merge.
c. Headness – original Labeling inscribed in merge
{α {α, } proposed by Chomsky (1995, 2000, 2001, 2004) V {V, DP}
√ Disfavored because it cannot be derived from the set theory. We need some speculation to derive α from the set as it is.
12
(18) H as the atoms of the computation (Chomsky 2013)
The Labeling algorithm selects H in {H, XP} as the label and the usual procedures of interpretation at the interfaces can proceed.
(19) Revised Labeling Algorithm proposed by Chomsky (2008, 2013, 2015) Minimal search labels H in case of {H, XP} viewing from over the phase unit, which is clearly the reminiscent of (17c).
√ Headness is retained at the first merge of any head (excluding R), though endocentricity is sometimes taken out of the final stage, because the Labeling algorithm makes use of the methodology of (17a).
(20) Is the categorial status of syntactic elements represented by the categorial feature extinguished?
√ No, we don’t need to put it into the extinct or endangered notions. It survives and walks openly.
This selection-based approach is explicitly and implicitly adopted by Bošković (2016), Hornstein (2009), Cecchetto and Donati (2015), Narita (2014), Collins and Stabler (2016) and Collins (2002, 2014).15 Some authors claim this type of Labeling is necessarily implemented in overt syntax.
(21) Reason for Labeling
Necessary for the selectional requirements of head, which clearly demands the categorial status of syntactic objects.
√ Thus, heads need to be visible to narrow syntax and/or the semantic interface and SO must express its categorial status not to be left alone in a workspace.
→ The categorial feature must express the categorial status of SO. Thus, Labeling should take care of this information.
√ Note that Chomsky (2007, 2008), as Hornstein (2009) and Cecchetto and Donati (2015) claim, assumes Labeling drives syntactic computation, because of the selectional requirements of categorial heads. Under the current theory, it is preferred that narrow syntax produces well-designed
15 Chomsky obviously rejects this view in Chomsky, Fukui and Zushi (2015), though.
13
syntactic structures fitting well to the Labeling algorithm.
→ The burden (i.e. importance) on the categorial head seems decreased …
(22) Another traditional role of categorial features/heads
Projection Principle and the regulation of movement derived from it
Categorial heads chaired by categorial features are responsible for Projection Principle as well as Labeling according to X’-theory.
√ Projections cannot be well-grounded now, because merge and the Labeling algorithm (or Labeling methodologies in (17)) do not incorporate the notion of projections.
(23) Survival of the disguises of Projection Principle
The ban of overt movement defined with the Labeling theory (see Bošković 2015, 2016a,b, Rizzi 2015a,b, 2016, Cecchetto and Donati 2015. Also, see Narita 2014 for discussion).
a. Rizzi (2015b, 2016) Maximality →Critical Freezing Effect Only maximal objects with a given label can be moved.
No maximal projection cannot be moved. [ … V [Q [Q whichQ book] CQ IP …]
Labeling as soon as possible when the Labeling pattern is discerned. b. Bošković (2016a) Anti-locality
Moved elements must at least move over nodes Labeled with two distinct labels. Undefined labels are non-distinct from others (namely, these are not counted). Here, Who moves over only CP.
Who … V [? Who [CP that [? Who T …]]] Undefined labels fixed at phases
(24) Does the Labeling algorithm come up to these appeals?
The stable and unstable Labeling gives answer to the halting problems and successive movement, but the Labeling algorithm leaves the duties to stimulating researchers when it comes to the locality of movement based on projections not phases.
(25) Categorial heads and Labeling problems a. Rizzi (2016)
Maximal projections (XP) and minimal projections (X-head) cannot be
14
easily distinguished under bare phrase structure without projections. b. Bošković (2016a), Sorida (2016)
Labeling of {H, XP} is determined as soon as the configuration is created; that is, a head is merged with XP.
(26) ♪ Privilege of the first external merge of SO in Labeling.16
When H undergoes first external merge with another SO, the Labeling algorithm marks H as a label.
√ Labeling must reflect derivational history of a phase and the label of {H, XP} is immediately determined from the derivational history.17 † This proposal is compatible with Goto’s (2016) analysis, because the
merge of Gianni with the complex of √partir-v*-a is involved in derivational history and is accessible to Labeling, based on which theta-roles are checked with the argument structure. Thus, it is no problem that the label of is D, not <v-R>:
a. [φ√partir-v*-aφ [Tφ [D= Gianni [√v* [√partiraφ]]]]]
Rizzi (2016a) notes that a postverbal subject Gianni is always focal in (b) – which I think identificational focus, probably. Thus, it can be said that the information structure allows this kind of information grid.
b. Il direttore è Gianni. ‘The director is Gianni.’
I’ll return to the endocentricity in the pair-merge section. 1.3 Why weak versus strong, not lack versus existent?
(27) Weak/Strong distinction of Labeling ability
The lack of Labeling by T and R derives the EPP effect (object shift) and that-trace effect, which necessitates the Labeling via feature sharing of <φ, φ>
16 The symbol ♪ is used when I make my own proposal.
17 As clarified later, I propose that Labeling is necessary to keep derivational history.
15
(28) Asymmetry between T and R
a. No parametric distinction in R but parametric difference in T
English T is poor at Labeling, thus necessitating the help by the external argument, while Italian T is good at Labeling. On the other hand, R universally lacks the Labeling ability.
Note that inheritance transmits inflectional and functional (cf. T, Q, φ) features from C/v* (and, maybe, other heads) to T/R (and, possibly, others), which implies that T/R originally possess no (or few) feature. Now, where does the parametric difference of T come from? Apparently, C and inherited features are responsible for this difference (cf. Gallego 2016). This appears plausible at first glance, since it is supposed that C selects finite T and infinite T. However, English matrix T is not weaker than infinitival T which is assumed to lack any T and φ features. Also, since T labels <φ, φ> with the help of the external argument, C is equipped with uninterpretable φ features (or tense) lacking important values in English, which Italian C/T have, which seems difficult to be handled given SMT. It is a quite heavy task to find the source of weak T. That is, a strange parametric distinction sounds stranger, complicating C’s selection of T. In this sense, the Goto’s (2016) analysis is preferred.
b. No semantically and categorially identical status of R
As mentioned, Italian T can manage to label itself after inheritance, which is not so mysterious because T is the locus of the inflectional properties φ and tense. On the other hand, R is always too weak to label. This is rather plausible given that it lacks any identical status semantically and categorially. Thus, this necessitates the Labeling via feature sharing with internal arguments in the lower domain where no categorial head appears, whereas R is definitely obliged to undergo pair-merge with categorial functional heads in the upper domain. However, why is R weak not labelless? It lacks any categorial and semantically identical status, which should lack any label in the historical sense. Weakness of the Labeling ability is good when it explains the necessity of <φ, φ>, but strong vs. weak is measured relatively, not absolutely, unlike zero vs. non-zero, which conceptually spoils the theory. That is, how strong does a syntactic item enough
16
need to be for Labeling?18
(29) The Labeling algorithm is sometimes discriminative …
The Labeling algorithm mostly can see SO, but it overlooks specific SOs carelessly (or intentionally …). For example, in addition to pair-merged functional heads, which are discussed before and later, non-topmost syntactic elements (i.e. trace) of chains cannot contribute to Labeling, which necessitates so-called phase-memory.
(30) Lower copies shout, “Please look at me!”
The Labeling algorithm turns its face away and neglect lower copies perfectly, but does the semantic interface see it even without Labeling? Chomsky suggests that SO, not copies of SO, needs to be licensed so that they can be given interpretation at the interfaces. If so, it may be the case that copies of SO are also licensed by side-effect, when a topmost copy of SO is licensed.
(31) Distinctive and sensible observer necessary
Under the copy theory, copies of SO remain intact after IM (Takita, Goto and Shibata 2015), which should be discerned by the Labeling algorithm. Chomsky suggests that lower copies of SO is invisible because a part of discontinuous element and the Labeling algorithm can recognize it,19 because phase-memory keeps the instance of IM of SO (Chomsky 2015) at least until transfer and Labeling occurs at transfer.20
√ However, the observer of the Labeling algorithm should take care of domestic rules which says that it must ignore SO until it doesn’t undergo further IM, or satisfies the requirement that it must establish a chain-like object phase by phase to know whether a copy is part of a discontinuous element or a top of the whole chain or where it is not located topmost
18 In this sense, it is necessary to fix a definition as to how poor inflection satisfies a too weak property of Labeling, whereas it should be necessary to know how rich is T enough to label? I am grateful for this clarification to the discussion with Nobu Goto, Masayuki Komachi, Kazuo Nakazawa, Yuji Sugimoto, Tadao Nomura, Shoichi Takahashi, and Shigeo Tonoike.
19 According to Nomura (2015, 2016), Chomsky gives the following definition of invisibility in the 2014 spring lecture at MIT:
i. α is in the domain of D if and only if every copy of α is in D,
20 In Chomsky (2013), transfer and IM are enforced simultaneously at phase levels, for which the Labeling algorithm dispenses with this type of phase-memory.
17
among brother copies, entertaining a topmost copy as well as discarding lower copies, though it knows copies are copies.
→ Is chain formation needed? Or, does SO always contribute to Labeling regardless of that IM is applied to it?
1.4 Invisible but Interpretable Copies
(32) Intermediate Copies
Almost no function and syntactic role assigned, though some reconstruction effects (cf. Legate 2003) and complementizer agreement (cf. Haegemen and von Koppen 2012, Maki and Ó Baoill 2011) observed. Thus, copies are ignored throughout successive cyclic movement
√ No contribution to Labeling and semantic but the interface should recognize the residue of these copies in occupied positions to find reference relationship based on syntactic structures, which are recovered from Labeling.
(33) Original copies
Needed to be licensed with theta-roles based on argument structure and wh-elements need to become variables in the original positions. In addition, reconstruction effects are observed depending on the syntactic environment.
√ As argued above, the semantic interface should need information of the original copies of SO for theta assignment (and, perhaps, variable binding), which are determined structurally. Thus, these copies ought to be recognized via Labeling. Namely, the Labeling algorithm needs to make the relationships (derivational history) between original copies of SO and merged elements including important theta information at the semantic interface.
(34) ♪ The Labeling algorithm identifies the most prominent head/feature from {XP, YP}.
i. The prominent head needs to be prominent in its chain or among its copies.
√ The prominence of a head can be detected after the formation of its chain is formed, or it undergoes further IM.
18
ii. A head H can be prominent after no further IM of H occurs within the phase domain.
♪ Narrow syntax may reuse the syntactic properties and features of SO for IM (Hornstein 2009), which result in the creation of an identical copy of SO.21 That is, the occurrence of reuse gives a hint of IM to the Labeling algorithm.
√ This doesn’t increase any economical burden like phase-memory if the Labeling algorithm contribute to programming derivational history of a phase domain, as supposed later.
Abe (2015), instead, the transferred material becomes inactive in syntactic computation assuming Transfer involves copy.
iii. All copies of SO are licensed when SO receives Labeling finally, which makes the interfaces read off all derivational history of SO.
√ As I said, the structural relationship copies of SO has is now recovered from the derivational history.22 Due to reused materials of SO, the interfaces can easily recognize identical copies of SO, and a chain is simply formed by using derivational history and identical copies.
(35) Sharing of φ-features, Q-features and what?23
A variety of syntactic analyses, especially those along with the cartographic approach advocated recently (cf. Cinque 1999, Shlonsky 2014, Rizzi 2015a,b, 2016) provide distinctively diverse and different features and functional heads.
Does the Labeling algorithm allow them to make significant contribution to its feature recipe which list possible labels? Or, does the Labeling algorithm love minimalism like its inventor so much that its feature recipe is confined to favorites φ and Q? Or, is the Labeling algorithm flexible and efficient, so it picks up familiar ones?
√ The issue is related to the standard by which features and labels are incorporated into feature recipe of the Labeling algorithm.
21 This is originally pointed out to me by Hisatsugu Kitahara. I appreciate him for fruitful discussion.
22 There is a caveat. That is, intermediate copies have few contributions to the semantic interface, which means they are rather useless.
23 Chomsky (2015) also assumes CP can be Labeled as exclamative or relative other than interrogative via Agree on Q (see note 9).
19
(36) a. Only traditional features
Q/wh and φ, as originally suggested by Chomsky, while case feature is excluded because it is the reflection of φ-agreement.
√ This approach forces the Labeling algorithm to stop its ears and skip all semantic contents encoded by functional areas within the CP-domain and other lower syntactic domains, setting aside claims by the advocators. Clearly, narrow syntax should resort to a different method, owing to the helpless Labeling algorithm.
b. Anything ok! Even all functional heads of Cinque hierarchy welcome! [Speech act franklyspeech act SAspeech act [Habitual usuallyhabitual Asphabitual […]]]
(Cinque 1999)
√ Under this approach, it is necessary to posit as many features and/or feature values as the number of distinct semantic interpretations which might express subtle differences. The Labeling algorithm likes to read Feature-Pad for cooking these features …
c. Familiar feature values with restricted features and relevant heads [Speech act franklyspeech act Moodspeech act [ … ]]
[Foc GianniOp<Foc> Foc [ Gianni … ]] (cf. Takano 2017)
[Top The AGU campusDiscourse<topic> Top [TP Nobu really wanted to visit the AGU campus]]]
√ The number of feature values and relevant features can be reduced under this approach, though this should be compatible with Inclusiveness Condition.
(37) ♪ Non-Labeling approach to the functional area within the CP-domain24 a. Labelless SO is legitimate, when the semantic interface can resort to its
interpretative mechanism. Thus, the Labeling algorithm only takes care of relevant derivational history without giving labels (cf. Ott 2015, Takano 2016, Cecchetto and Donati 2015).25
b. The interface can make use of information structure [Topic Focus Background (= TP)] (cf. Vallduvi 1990) to deal with labelless SO.
c. In the information structure, the first syntactic element basically receives topic interpretation, whereas the other elements are given
24 SC = Syntactic Structure, LA = Label, and IFS = Information Structure.
25 Instead, C might undergo multiple default agree with these syntactic elements for Labeling. Another possibility is appealing to pair-merge, which makes SO invisible, turning the eyes of the Labeling algorithm away.
20
focus interpretation.
d. SC [Q wh-Q [C C-Q [ XP-Topic YP-Focus … [φP Subj-φ [TP T-φ]]]] LA Q C labelless labelless φ T
IFS Topic Focus
(38) a. An-wa sakkaa-no-siai-wa mita.
An-Top (genuine) soccer-Gen-game-Top (contrastive) watched
“As for An, she watched a soccer game (not a baseball game or a lacrosse game).” (The second topic marker bears contrastive focus.) b. Rakurosu-no-siai-sae (focus) An-wa mita.
Lacrosse-Gen-game-even An-Top (contrastive) watched
“An (not Ten or Sun) watched even a lacrosse game.” (The second topic marker bears contrastive focus, because it does not come first in order.)
c. Dare-ni An-wa gakkoo-de atta-no? who-Dat An-Top (genuine) school-at met-Q
Lit. “As for An, who did she meet at school?” (The topic marker indicates a genuine topic because it comes first other than wh-element.)
d. [Q wh (Q) [NP-wa [φ …] ] no (Q)] IFS Topic
Either approach seems to shows that the Labeling algorithm doesn’t get along with these domains …
(39) Labeled by sharing features disguises the intersection of a set
If real, the Labeling by feature sharing makes so-called Spec-Head Agree dedicated (fitting) to the Labeling algorithm (Feature-equilibrium or featural symmetry in the sense of Fukui (2012), also Narita and Fukui 2012).26 If derivational history includes Spec-Head Agree or if the Labeling algorithm earnestly discerns the occurrence of Agree resulting in feature sharing, the Labeling algorithm finds out the prominent shared feature between two SOs rather easily and doesn’t need to conduct search deeply.27
26 The definition of Feature-equilibrium is given below:
i. For a syntactic feature F, an SO {α, β} is in an F-equilibrium ≡def. α and β share a matching feature F that is equally prominent in α and β.
27 It is better to make use of Agree proposed by Bošković (2007) than Agree via probe-goal, because the latter doesn’t necessitate XP to undergo IM into YP-Spec to form {XP, YP}. Also, the view of agreement by Seely (2016) is helpful to advocate the direction implied in (38).
21
(40) Is minimal search economical?
Search is required to inspect all interior not only of SO but syntactic heads to specify shared prominent features and, worst, it is necessitated to make deeper search when two lexical heads which shares a feature are buried under complex SOs.
√ The adjective ‘minimal’ is excellent (sheep head), but minimal search is not so minimal in reality (dog meat), necessitating extra search (cf. Hornstein (2009), Fukui (2012), Goto (2017)).
(41) What do represent φ-features?
As discussed above, φ-features are unvalued on C/T and v*/R and that it is unnatural to think <φ, φ> represents the clause status, propositionality as well as sentential status, which is mysterious given that φ-features are uninterpretable on T and it is unnatural that feature bundles of gender, number and person exhibits these semantic properties. Also, it follows that <v*, R> takes <φ, φ> complements, which includes some information on argumentation. Putting aside the question whether <φ, φ> expresses this kind of information, this label deals with so diverse semantic considerations that it is quite difficult to process the information as a single label.
Given labels encodes relevant information to the interfaces, it should be the case that <φ, φ> makes some semantic contribution, which seems unreasonable. Thus, oppositely, feature sharing <φ, φ> bears a different role for Labeling, which indirectly contributes to the interface. Especially, making use of the observed asymmetry of this agreement relation, I give a speculative solution by proposing a new Agree system, excluding <φ, φ> from label.
(42) ♪ Uninterpretable/unvalued features play a role in Labeling28
a. The Labeling algorithm gives a label with Y to {XP, YP} when X and Y shares a prominent feature and Y gets the feature value from X under Agree.
28 As suggested later, case is not a kind of uninterpretable feature but the reflex of the sequence of φ-agreement (cf. Seely 2016).
22
b. [? [X … X[ ] ] [… Y[_] ] ] … Y has a unvalued feature Before Agree
→ [? [X … X[ ] ] [… Y[γ] ] ] … Y gets valued by X After Agree
→ [YP [[X … X[ ] ] [… Y[γ] ] ] … Labeling
Unvalued features are more prominent for the syntactic mechanism in the sense that it needs Agree. If only valued features, no Agree and feature sharing is forced. Thus, unvalued feature are more easily detected by the Labeling algorithm.
√ The Labeling algorithm, which is accessible to derivational history including the instance of Agree, can discern the feature sharing via Agree instantly, detecting more prominent unvalued features on Y. Thus, it takes Y as a prominent element, making use of it for Labeling. c. [?→ TP [[DP … D[ ] ] [… T[Ø → γ] ]] …
√ Contrary to the ordinary view, when X and Y undergoes a feature sharing which lends a help to Labeling, either of X or Y must contain an unvalued feature asymmetrically. This speculation at least excludes
<φ, φ> from the Labeling.29
(43) ♪ When there is no Agree between X and Y in the form of [XP, YP], which is equated to [α [X … X …], [Y … Y …], the Labeling algorithm determines a label of α as the same label as that of unmoved SO if the sister of SO undergoes IM. Otherwise, the label of α is undefined.
√ That is, if XP undergoes IM, the label of α is automatically determined as Y, succeeding the label from the remaining SO, whereas the label of α becomes X, when YP undergoes IM. This is natural because derivational history includes the instance of IM at phases and the Labeling algorithm is accessible to the history.
Now, the Labeling algorithm can derive all Labeling from syntactic operation automatically, without resorting to minimal search.
1.5 The Final Issue: Projection and Labeling
(44) Returning to Projection
Traditionally, projections made important contributions, because it
29 Sorida (2016) derives the necessity of unvalued features from what he calls Blocking Feature, which cancels the search of the Labeling algorithm or makes the Labeling possible under feature sharing.
23
reflects the syntactic and semantic selectional requirements of heads, in addition to categorial status. That is, projections are the honored children of headness. Really, they can project necessary features up to maximal projections (that is, percolation (cf. Hornstein 2009)). Whereas, Labeling is inferior in that it cannot reflect the selectional requirements and the categorial status of heads after a head merges with its complement. This may be one of the factors disappointing some syntacticians.
√ As implied before, derivational history of phases is responsible for a half of tasks projections fulfill, the reflection of selectional requirements of heads on syntactic structures, though it is not enough to be alternative candidate for the regulation of movement.
(45) ♪ What does the Labeling algorithm offer?
With Labeling plus syntactic structures, narrow syntax satisfies the requirement that it lets the semantic interface make use of derivational history which is vital to fostering interpretive process.30
(46) No Tampering Condition (Chomsky 2013) Neither X nor Y is modified by Merge.
√ The No Tampering Condition holds because any deletion of derivational history is prohibited.
2. Magical Pair-merge
(47) Pair-merge creating an ordered pair = special magician
a. Making v* unseen and useless, and, worse, depriving the phase status of it, while giving R a privileged status and entertaining it as a phase. b. Pair-merge elaborates order-pair to a magic rod with which narrow
syntax achieve special activities in (a), though an order set only gives the order of some mechanism (combination of comparative, argument structure, etc.).
c. Magical pair-merge necessitates ordered-pair, but it is difficult to give a well-grounded motivation for this.
30 I adopt the definition of the derivational history by Collins (2014).
24
(48) Adjunction of v* to R, not vice versa
i. R raises to v* via pair-merge, forming R with v* affixed and <v*, R>, hence invisible, so phasehood is activated on the copy of R, and DP (, which can be a wh-phrase, remains in situ, at the edge.
a. Chomsky assumes that, after inheritance and consequence of Labeling
<φ, φ>, v* is adjoined to R, because a verbal affix appears on a root R, not conversely, even though v* categorizes R. This seems a bit strange if a categorizer takes the natural responsibility as a head.
b. Chomsky attributes this to inheritance including phase-status and inflectional/functional properties (φ-features and others) of v* and pair-merge <v*, R>, which makes v* invisible and let the inherited phase-status of v* on the copy of R consequently. That is, the activation of phase-status on R presupposes pair-merge.
(49) Pair-merge should be magical! i. Two scenarios of inheritance
a. Functional properties don’t include verbal status.
Rising of R via pair-merge is necessary, because R lacks categorial status. However, if so, why does v* not function as a head, because it is a categorizer and should be regarded as prominent, giving rise to Labeling.
b. Functional properties include verbal status.
Inheritance does all the work, helping R get the Labeling itself as verbal, though the Labeling of <φ, φ> is once fixed and unchanged because of Agree between it and an internal argument. Then, why is pair-merge necessary for raising? Exploited v*all possessions of which are deprived is now empty (and should be eliminated given Full Interpretation) and lack of phase-status, does syntax take care of it? Apparently, this pair-merge is enforced because the virtual invisibility of v* is a pre-requisite for the activation of phasehood on the copy of R. The bequest is executed when the pretentious predecessor dies, made unseen in syntax. Needless to say, this pair-merge is redundant. Maybe, a magician loves to celebrate and decorate R as a new phase king.
√ If inheritance can bear the role of categorization instead of a categorizer, why can’t the inheritance from C to T give a Labeling ability to be labeled as T, since a full-fledged C selects a finite T. So, I
25
think it problematic that v* is not a categorizer.
ii. A magician loves the lower copy of R but hates the upper copy of R pair-merged with R.
Note that the phasehood is activated on the lower copy (invisible to the Labeling algorithm but visible in syntax) not the upper one of R, which is clearly counter-cyclic.31 Syntax should target the upper copy of R given minimal search is the economy principle. The reason is simple. The internal argument remains in-situ at the edge of <φ, φ>. If the phasehood is activated on the upper copy of R, it should be trapped, unable to escape from <φ, φ>, when it is a wh-element.
(50) Historical pair-merge32
a. Adjunction (multi-segment category) (Chomsky 1995) Merge (α, K) → L = {<H(K), H(K)>, {α, K}}
Adjunction creates segments of H<K>, which causes α to be invisible in syntax.
√ In <α, > = {{α} {α, }}, too roughly speaking, α is highlighted in the set, which might be equal to the situation that K, which doesn’t undergo adjunction, continues to remain as the object of syntactic computation, whereas adjoined elements are technically separated from syntax due to segments. In that sense, it may be said that K, which is a part of an adjunction structure and is adjoined to by α, is distinctively picked out from the set (α, K).
b. Asymmetric pair-merge (Adjunction) (Chomsky 2001) Merge (α, ) → <α, >
c. SIMPL <α, > → {α, } (Chomsky 2004)
According to Chomsky, adjuncts are necessary to yield predicate composition in the semantic component, which requires pair-merge. Pair-merge takes α into a separate plane, leaving on the “primary”
plane,” the simple structure with all its properties. SIMPL plays a role in combining these planes.
31 Nomura (2015) also points out this point. He relates the activation of the phasehood to the inheritance of unvalued φ-features, not necessitating the inheritance of phasehood That is, the location of φ-features determines the phasehood and the transferred domain, which also regulates the escape hatch and the place of object shift.
32 Relevant discussions are dealt with in Irurtzun and Gallego (2007), Richard (2009), and Oseki (2014).
26
√ Given that order pair can distinguish which one is which, it might be possible to posit a fuzzy function <α, > → <P2, P1>, where P2 is a separate plane distinct from a main syntactic computational space in syntactic workspace, while P1 is the primary plane where a syntactic computation occurs. Seemingly, it has two merits. One is that pair-merge can be unseen in the primary plane, a main syntactic computational space, due to which adjuncts don’t participate in syntax. Second, because it becomes unseen in syntax, the No Tampering Condition and the Extension Condition cannot be violated.
Now, what can we say about pair-merged set <v*, R>? Now, v* is made invisible due to pair-merge, which needs to activate the phase-status on R. Second, the No Tampering Condition is not severely violated because v* virtually disappears due to pair-merge. At last, R gets privileged as a phase.
However, this is a risky and mysterious assumption. Also, pair-merge doesn’t give any solution to the categorizer problem, though SIMPL puts v* back into the primary plane, letting it making semantic contributions to the interfaces. However, SIMPL is itself a mysterious operation.
(51) Is the function <α, > → <P2, P1> apparently or really funny?
As mentioned, only the formation of order pair is not enough for the invisibility of an adjoined SO under pair-merge. This function requires two syntactic workspaces, that is, a normal syntactic workspace and an adjunction syntactic workspace, the latter of which is invisible to primary syntactic computation and unified with the former via SIMPL. Obviously, the latter syntactic workspace has a ‘ninja’ property, because it is too secret and obscure for syntactic computation to recognize. Evidently, it needs verification since narrow syntax deals with two totally different independent syntactic workspaces at parallel. Yet, this ninja workspace is dependent on the primary syntactic workspace, as adjoined SOs should be moved with adjoining syntactic elements when the latter undergoes IM and needs to be properly unified via SIMPL, by which syntactic computation integrates these blindly. Thus, these syntactic workspaces are funny in that they are contradictorily inaccessible but intelligible and consist of the two sides
27
of the same coin. Possibly, narrow syntax doesn’t face the problematic situation even if it goes through the parallel computation of two diverse syntactic workspaces, but it requires a heap of burden for justification because we need to introduce a totally new and different function to syntax: a hidden function mapping SO to two unconnected but intelligible workspaces
(52) ♪ Pair-merge as union
As mentioned in (17b), this method of Labeling should be available. Especially, v* functions as a categorizer, marking R as verbal, which simply means that the categorial (verbal) property of v* is combined with R to produce v*-R.
a. When a categorizer H is pair-merged with R, the Labeling of {H, R} should be determined via union, combining a categorial feature of H with R and producing the R with the category of H, the result of which is the label of H-R.
b. Pair-merge can be applied to R in the following two ways.
i. Before R undergoes the first merge, H is pair-merged with R. That is, external pair-merge. (cf. Epstein, Seely and Kitahara 2016, Nomura 2015, 2016, Obata 2016, Otsuka 2016, Sugimoto 2016)
ii. H is pair-merged with the copy of R before it undergoes the first merge by reusing the copy of R in the syntactic space. That is, internal pair-merge.33
c. φ-features
i. All Rs have a slot for φ-features.
ii. Each H is different in the possession of and values of φ-features, if any. The categorizer n has valued φ-features, whereas the categorizer v has
unvalued φ-features. The categorizer a lacks φ-features.34 d. Argument
Each of R and a categorizer may introduce one argument.35
√ The categorizer H must undergo pair-merge before set merge, since this categorizer need to fulfill the role to categorize R pair-merged with it. Thus, a pair-merged {H-R} undergoes the first set merge with its
33 This operation may be similar to sideward movement in Takano (2017).
34 Possibly, parametric differences might be observed in the categorizer a.
35 In case of three place predicates, R may introduce two arguments, or another functional head may introduce the other.
28
complement.
(53) Instantiation of pair-merge
A. Transitive verbs (Internal pair-merge) – Mihn knows a fact. i. [? NPφ, R( ), NP]36
ii. [? NPφ, R[φ], NP] φ-Agree between NP and R iii. Internal pair-merge {v*-R}
Category Argument φ
v* verbal One unvalued
R undetermined Zero (NP) φ
Union verbal One φ
iv. [v*-R v*-R [R NPφ, R[φ], NP]] Labeling
The Labeling algorithm should look at the copy of R, because R and v*-R is different. Also, the label of ? gets R due to (41).37
v. NPα[v*-R v*-Rφ [R NPφ, R[φ], NP]] Merge of External Argument NPα vi. NPα[v*-R v*-Rφ ] Transfer
a. In (i), an internal argument NP undergoes IM after it set merge with R. b. In (ii), NP undergoes φ-Agree with R, providing φ-features for the
φ-slot of R.
c. In (iii), internal pair-merge between v* and R occurs as soon as the categorizer v* enters syntactic workspace. Pair-merge unifies of the categorial status, argument slot and φ-features of R and v*. Note that v*-R has valued φ-features via R from the internal argument.
d. The Labeling of ? gets R because R and NP undergoes Agree in φ-features and R lacks φ-features, which makes R marked as the label based on (42).
♪ e. The complement of v*-R is transferred after the external argument is set merged with v*-R. The interfaces detect the sequence of [v*-Rφ, NPφ, R[φ]] in the syntactic structure, regarding R[φ] as an accusative
36 Given the Labeling algorithm in (26), R should label the lowest unit [R, NP], which might imply that the label of this unit can be regarded as labelless. However, as I said, this should be ok, because R will be licensed when it gets categorial status by pair merging with a categorizer (here, v*), which makes every copy of R visible to the Labeling algorithm, letting it reflecting relevant derivational history.
37 The semantic interface mainly interprets {v*-R}, simply making use of the lower copy of R to see its relationship with complement other than to detect the information of it as an accusative case marker.
29
case because its occurrence of φ-features are the same as that of v*-R. Before transfer, the internal argument undergoes further IM out of RP,
if it is a wh-element.
B. Unaccusative verbs (External pair-merge) - Mihn sleeps. i. External pair-merge {v*-R}
Category Argument φ
v* verbal One unvalued
R undetermined One empty
Union verbal One empty
ii. [v*-R v*-R, NP] Labeling & No Transfer
Because there is no syntactic derivation, v*-R doesn’t cause any Transfer.
iii. [T NPφ, Tφ [v*-R v*-R, NPφ]
a. First, an external pair-merge of v* and R occurs to categorize R as verbal via unifying these SOs. Because both of v* and R lack φ-features, no φ-features are assigned to <v*-R>, resulting in the diminishment of the φ-features slot.
b. Because {v*-R} set merged with its complement first, the label of the set {v*-R, NP} is v*-R. Due to the fact that v*-R is merged with a single SO, where the derivational history contains only the head complement relationship between v*-R and NP, v*-R doesn’t cause any transfer.
C. Bridge Verbs (Internal pair-merge) - Feliina thinks that Mihn sleeps. i. R [C C, [T NPφ, Tφ] […]]]
ii. Internal pair-merge {v*-R}
Category Argument NP φ
v* verbal One unvalued
R undetermined Zero (CP) empty
Union verbal One empty
iii. [v*-R v*-R [α R [C C, ]] Labeling
iv. NP[v*-R v*-Rφ [α R [C C, ]]
Merge of External Argument
v. NP [v*-R v*-R ] Transfer of α
30
a. In (i), R is set-merged with CP due to its selectional requirement.
b. In (ii), an internal pair-merge of v* and R occurs to categorize R as verbal via unifying these SOs. Because both of v* and R lack φ-features, no φ-features are assigned to <v*-R>, resulting in the diminishment of the φ-features slot.
c. Because {v*-R} undergoes the first set merge with its complement in (iii), the label of the set {v*-R, NP} is v*-R. Also, after an external argument is merged with the SO with the label of v*-R in (iv), it causes the transfer of α.
D. Unergative verbs i. [ R ]
ii. Internal pair-merge {v*-R}
Category Argument NP φ
v* verbal One unvalued
R undetermined One empty
Union verbal One empty
iii. [v*-R v*-R, R]38 Labeling
iv. NP [v*-R v*-R, R] Merge of External Argument
a. First, R is introduced to the syntactic workspace because R may take its complement, as mentioned in footnote 38, in (i).
b. In (ii), an internal pair-merge of v* and R occurs to categorize R as verbal via the unification of these SOs. Because both of v* and R lack φ-features, no φ-features are assigned to <v*-R>, resulting in the diminishment of the φ-features slot.
c. The label of the set {v*-R, NP} is v*-R, because {v*-R} is first set merged with its complement in (iii). There is no transfer after an external argument is merged with the phrase headed by v*-R, as v*-R is merged with R, where the derivational history involves only the head complement relationship.39
38 Note that unergative verbs can take a complement, like John slept a good sleep. In this case, the derivation proceeds in the same way as in (53A). I speculate that when an internal argument doesn’t undergo IM and stays in-situ, letting the internal pair-merge between v* and R, this argument gets default case, whereas the lower copy of R becomes a kind of unergative affixes, as suggested in note 39.
39 One may ask a question how the lower copy of the sequence of [v*-R, R] is represented. I suggest
