Desktop version

Home arrow History arrow A Minimalist View on the Syntax–Semantics Relationship: Turning the Mind into a Snowflake

Chains meet features

It is to be noted that the ‘viral theory’ of uninterpretable features, which had long dominated the landscape of the minimalist theory of syntactic computation (see Uriagereka (1998) and, in the present context, Chomsky (2005: 17)), has gradually grown faint, if it has not effectively faded away. On the one hand, there may be several grounds for being suspicious about the presence of features whose presence may cause an otherwise successful derivation to crash, be it during the syntactic computation itself or, even worse, only after the structure has been delivered to interpretive components (see e. g. Preminger (2014) for an extensive discussion of the issue, arguing for an abandonment of the uninterpretable-features model and a resurrection of the obligatory rules approach—applying obligatorily if appropriate conditions are met, and remaining silent otherwise). It is telling that the very term ‘uninterpretable’ was intended to be subject to elimination, the notion of ‘unvalued’ features being deemed more appropriate—interpretability crucially invokes interface-related properties, while the valued-unvalued distinction may be seen as a purely formal one, although strictly correlated with the interpretable-uninterpretable one (with some proposals retaining both distinctions as independent, as in Pesetsky and Torrego (2007) and related work). On the other hand, it has become clear that the presence of unvalued features in structures delivered to interpretive components need not be as harmful for interpretive processes to occur as it was supposed to be; in particular, instead of inducing crash at the interface or within the C-I component, they may be assumed to be inactive there—invisible for the interpretive procedure as it were, as hypothesized in Epstein, Kitahara, and Seely (2010). Such behaviour requires clearly further elucidation (see section 1.3.3), but it might be immediately noted that it fits better the current model of derivational process than its predecessor: on the basically label-based picture, where merge applies freely and the sole syntactic filter (if it is to be understood as a filter, actually a matter of dispute) is the requirement that labeling occur—which, in turn, does not mean any tampering with syntactic structures, labels as such not being added to syntactic objects in any way—labeling reduces syntactically to creating an appropriate configuration for a relationship between syntactic objects in the structure to be read by interpretive components, and no further operations on features—

‘valuation’ and/or elimination—is conceptually motivated (a conclusion reached on different grounds and embedded in a quite another view of the derivation in Boeckx (2015a)). It should be noted that the Chomsky-Richards line of deduction of feature inheritance from the necessity of elimination of uninterpretable features and the timing of transfer does not stand in a framework in which unvalued features are invisible rather than causing a derivation to crash, as noted already in Epstein, Kitahara, and Seely (2010), and the same goes for the most recent formulation of the phase-based theory of syntactic derivation in Chomsky (2015b); their own proposal, involving a doubly-peaked structure along the lines of Epstein, Kitahara, and Seely (2012) does not fare well, either. The issue was, to recall, how to make unvalued features be valued so that they could be eliminated before reaching the C-I component, such features being indistinguishable from their valued counterparts after valuation. Residing on phase heads, they would remain in the derivational workspace to undergo transfer at the next phase, since only the complement of the phase head is subject to the operation, as indicated in (3).


The Chomsky-Richards solution to the problem was to make unvalued features undergo a lowering operation—feature inheritance—so that they are valued immediately before the structure containing them gets transferred. Epstein, Kitahara, and Seely (2012) proposal ties feature inheritance to the valuation of Case, and the transfer is necessitated by the emergence of an anomalous, doubly-peaked structure. Once feature inheritance gets connected with inheritance of phase-head behaviour in general, though, all such motivations do not suffice for a deduction of the mechanism—what undergoes transfer in the model of Chomsky (2015b) is the complement of the derived phase head.


The phase head by inheritance crucially remains in the workspace, bearing unvalued features which, in cases discussed in Chomsky (2015b), are in configuration with features of an object in the ‘specifier’ of the derived phase head, leading to labeling under feature sharing, as in (5).


To be sure, the ‘specifier’ is no longer an official notion of the theory, and the relationship is actually established between two complex objects based on featural composition of their heads (lexical items most prominent for labeling purposes). The mechanism of feature inheritance provides in this model the way to label a syntactic object in case its head is a lexical item devoid of relevant features which might be shared with a ‘specifier’ (with the proviso that they constitute the valued-unvalued pair: ‘mere matching of most prominent features does not suffice (...) What is required is not just matching but actual agreement, a stronger relation...’, as Chomsky (2013c: 45) notes; see also Chomsky (2015b: 13 n. 16)) and it is unable to label the syntactic object it heads by itself—as in the case of roots and T-heads in languages like English. The inheritance of features is thus a prerequisite for a syntactic configuration necessary for labeling to be established. Now, whether a ‘copy-like’ or a ‘donate-like’ mechanism of feature inheritance is adopted (the former as in Chomsky (2013c), the latter in Chomsky (2015b); see further Ouali (2008, 2010) and the discussion in section 1.3), the features inherited by the derived phase head are bound to remain in the derivation—a problem noted with regard to phase heads in the ‘copy-like’ system in Chomsky (2013c: 47 n. 47), without a satisfactory solution, though: deletion of features on the derived phase head is not an option in (5) (see also the discussion in Carstens, Hornstein, and Seely (2016: 83)). Recall that the label (F, F) is only a notational convenience, merely registering that a proper configuration for labeling under feature sharing has been obtained—no objects like (F, F) are added to syntactic structures during the operation of the labeling algorithm:

Crucially, LA does not yield a new category as has been assumed in PSG and its various descendants, including X' theory. Under LA, there is no structure [a X], where a is the label of X. LA simply determines a property of X for externalization and CI. (Chomsky 2015b: 6)

Given this fact—explaining the syntactic inactivity of labels, noticed and discussed already in C. Collins (2002) and Seely (2006) (see further Chomsky (2012a: 4), who notes that ‘labelingshould not be marked at all in generated structures but rather regarded as a property of Generative procedure],’ and the discussion in Epstein, Kitahara, and Seely (2015b: 106-108))—it is impossible to delete unvalued features of the derived phase head (or even an entire head, as it happens with C-type phase heads in Chomsky (2015b)) without destroying the configuration required for the object to be labeled when it gets transferred. This is not in itself damaging for an account of phases in terms combining principles of minimizing the computational space and unvalued features together with the inheritance mechanism; it merely undermines the connection between ‘valuation’ and the necessity of transfer understood as the necessity to eliminate recently valued features. Non-phasal heads in v*P and CP are assumed to lack features which would allow them to be seen by the labeling algorithm as labels. Feature inheritance does equip them with features, although it does not always suffice to make them eligible as labels, as discussed in section 1.3: adopting the donate-like mechanism of Chomsky (2015b), together with the assumption that feature inheritance involves bundles of features and does not choose among them and target them separately (see Chomsky (2013c: 47) and Chomsky (2015b: 11) with n. 12, with the hypothesis that it is only unvalued features that are ‘donated’—deleted after inheritance, as per Chomsky (2015b)—while the status of valued ones is left unclear; see Carstens, Hornstein, and Seely (2016: 81-82) for relevant discussion, couched in terms of ‘lowering’ as distinct from ‘inheritance,’ and see also section 2.4.3), enables the root to participate in labeling, but does not make it a label itself:


Crucially, in (6) у remains still unlabeled after feature inheritance, while the root requires head raising to the phase head to be categorized. The activation of the phasehood on the root maximizes the object which should be labeled in all respects—to the exclusion of the root, which still requires further operations enabling у to be labeled (compositionality does not allow it to escape labeling, and it is the head that must be responsible for the process), although it does participate in labeling of a, hence it has entered the web of (interpretive) relationships. As for T, the situation seems in some cases analogous (see section 1.3.6 for some tentative remarks); the general mechanism should perhaps best be understood with respect to its derivational consequences as minimizing the size of the structure remaining visible for further operations and maximizing the size of labeled and transferred part thereof so that the the search space might be as small as possible. Whenever the configurations fulfill labeling requirements, the structure passes the labeling gate and is subject to interpretive procedures; but the procedure is general and applied non-teleologically, hence leading also to crashes (see sections 1.3.3-1.3.4 for a discussion of possible interpretive import of featural configurations). Interestingly for the present discussion, features relevant for labeling a are bound to remain, as much as the label (strictly speaking, the configuration of objects and their properties) they give rise to remains, visible for narrow syntax, as assumed explicitly for structures in which there is movement from a configuration in which a label has been determined, as in (7).


The structure in (7) is present in A' raising from the spec-VP position in RTO structures, as well as in A' raising from the Spec-TP; in both cases, the phase head transfers its features downwards and a correct labeling configuration is established, with inactivation of the original phase head (via head raising in the former case, via deletion in the latter). Importantly, movement of XP in (7) does not interfere with labeling provided that it takes place within the phase-limited memory window, when the object (F, F) ({ф, ф), in these cases) is still present as a unit in the derivational workspace:

The basic principle is that memory is phase-level—as, e. g. in distinguishing copies from repetitions. Hence at the phase level (...), if some element (...) has already been labeled, the information is available and [it] stays labeled. For interpretation at CI, labels are computed at the phase level, with cyclic transfer. (Chomsky 2015b: 11)

< Prev   CONTENTS   Source   Next >

Related topics