On the Uninterpretability of Interpretable Features

download On the Uninterpretability of Interpretable Features

of 22

description

Zeijlstra 2014

Transcript of On the Uninterpretability of Interpretable Features

  • This is a contribution from Minimalism and Beyond. Radicalizing the interfaces. Edited by Peter Kosta, Steven L. Franks, Teodora Radeva-Bork and Lilia Schrcks. 2014. John Benjamins Publishing Company

    This electronic file may not be altered in any way.The author(s) of this article is/are permitted to use this PDF file to generate printed copies to be used by way of offprints, for their personal use only.Permission is granted by the publishers to post this file on a closed server which is accessible to members (students and staff) only of the authors/s institute, it is not permitted to post this PDF on the open internet.For any other use of this material prior written permission should be obtained from the publishers or through the Copyright Clearance Center (for USA: www.copyright.com). Please contact [email protected] or consult our website: www.benjamins.com

    Tables of Contents, abstracts and guidelines are available at www.benjamins.com

    John Benjamins Publishing Company

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features

    Hedde ZeijlstraGeorg-August-Universitt Gttingen

    In this paper I present several arguments that argue against the assumption in current generative syntactic theory that certain formal features are semantically active as well (so-called interpretable formal features). Instead, I propose that the set of formal features and the set of semantic features (to the extent that they are featural in the first place) are fully independent. An acquisitional and diachronic theory further constrains the possible combinations of syntactic and semantic features that can be lexically stored, which results in the apparent overlap in the distribution of particular syntactic and semantic features (which has originally been the cause of taking them on a par).

    1. Introduction

    Ever since the introduction of functional projections in syntactic structures hosted by formal features, questions have arisen as to: (i) what constitutes the set of formal features in a particular grammar, i.e. is this pool of formal features in a particular grammar given by UG or does it emerge in the process of language acquisition; and (ii) what are the syntactic and semantic properties of formal features? Lets start by having a closer look at both questions.

    1.1 A universal set of formal features?

    As for the first question, over the past two decades several proposals have been formulat-ed that aim at accounting for the presence of the set of formal features in natural language grammars. Initially, it was argued that UG provides this set of formal features and that every language has the same set of formal features at its disposal, a view much in line with the so-called cartographic approach, which in its most radical version assigns a universal syntactic structure to all natural languages with variation lying in the way that (parts of) this structure are phonologically realized (cf. Pollock 1989, Beghelli and Stowell 1997, Rizzi 1997, 2004, Cinque 1999, 2002, 2006, Starke 2001, 2004, Miyagawa 2010).

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    110 Hedde Zeijlstra

    More recently, an alternative view arose that states that the set of formal features is as minimal as possible in every language. Under this view, sometimes referred to as building block grammars or the WYSIWYG (What You See Is What You Get) approach, formal features and consequently functional projections should only be assumed to be present if there is overt evidence for it (cf. Iatridou 1990, Grimshaw 1997, Bobaljik and Thrainsson 1998, Koeneman 2000, Nilsen 2003, Zeijlstra 2008)

    The main difference between the building block grammar/WYSIWYG approach and the cartographic approach (in its most radical sense) is that the visible presence of a particular formal feature in a particular language (for instance if it overtly heads a functional projection) does not, on the first approach, imply its presence in all lan-guages, whereas this is the basic line of reasoning under the latter approach (cf. Cinque 1999, Kayne 2002, Starke 2004). This reduces the question as to what constitutes the set of formal features to a question about the nature of UG. Is UG a rich body of knowl-edge that contains the set of potential formal features that a language may be sensitive to, or is UG, as has been proposed in more recent minimalist views (cf. Chomsky 2005), much poorer in nature and are the relevant formal features to be acquired in the course of first language acquisition? Even though the latter view should be taken to be the default hypothesis (given that one should only postulate things in UG that other-wise cannot be accounted for), its correctness can only be evaluated against a concrete proposal of how these formal features can be acquired in the first place. Formulating such a proposal and evaluating its consequences is one of the goals of this article. In or-der to do so, in this paper, I further elaborate Zeijlstras (2008) proposal, which argues that syntactic doubling is the only available cue to determine the presence of formal features. However, I also argue that the implementation of this proposal drastically changes the way that the nature of formal features should be considered.

    1.2 Syntactic and semantic properties of formal features

    For Chomsky (1995), the set of formal features, i.e. the set of features that may partici-pate in syntactic operations, is a set that intersects with the set of semantic features (see (1)). Consequently, formal features come about in two kinds: interpretable and unin-terpretable formal features. Interpretable formal features ([iF]s) are features that are part of the intersection of the two sets and therefore may both participate in syntactic operations and receive an interpretation at LF. Uninterpretable features, by contrast, are features that are only formal, and not semantic in nature and therefore cannot receive an LF-interpretation.

    (1)

    [P]

    [uF]

    Formal featuresPhonological features Semantic features

    [iF]

    [S]

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 111

    Chomsky (1995, 2002) furthermore argues that every feature that reaches the inter-faces must be interpretable (following the Principle of Full Interpretation):

    (2) Full Interpretation (FI): Every element of an output representation should provide a meaningful input to the relevant other parts of the cognitive system.

    To satisfy Full Interpretation, all uninterpretable formal features must be deleted in the course of the derivation. For Chomsky, Agree is the only available operation that is capable of deleting uninterpretable formal features: if an interpretable and a match-ing uninterpretable formal feature stand in a particular (c-command) configuration, the uninterpretable features can be checked off against the interpretable one and as consequence be deleted. Once every uninterpretable feature has been deleted, the deri-vation can be fully interpreted at the interfaces; after the deletion of those features that are only formal in nature, all features left are either phonological or semantic features.

    In later work, (Chomsky 2002), argued that this view should be modified as it would otherwise face a look ahead problem: since it can only be determined at the level of LF whether a particular feature is interpretable, the (un)interpretability of a feature is not visible in the course of the derivation (which precedes transfer at LF). Hence, deletion of uninterpretable features as such cannot be a trigger for syntactic opera-tions. For this reason, Chomsky argues that deletion of uninterpretable features does not form the trigger of syntactic operations, but rather feature valuation does: every feature that has not been valued in the lexicon needs to be valued in the course of the derivation; valuation then takes place under Agree.

    For this, Chomsky postulates that all formal features that are interpretable are also lexically valued and formal features that are uninterpretable are also lexically unval-ued. Again, only the former type of features (lexically valued and interpretable fea-tures) are members of the set of semantic features. Furthermore, Chomsky argues that during syntax all lexically unvalued features that are valued during the derivation, get deleted prior to LF. As a result, all and only those formal features that are interpretable survive at LF.

    Chomskys (2002) proposal in a way enriches the feature taxonomy by including a second parameter, feature (un)valuedness, but given that for him interpretability and valuedness always go hand in hand, the number of different types of formal features remains identical. For Chomsky, there are still only two types of formal features: lexi-cally valued and interpretable features and lexically unvalued and uninterpretable ones (see (3),), where __ means unvalued and val means valued).

    (3)

    [P]

    [uF: __]

    Formal featuresPhonological features Semantic features

    [iF: val]

    [S]

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    112 Hedde Zeijlstra

    Note, however, that it is a pure stipulation that (un)valuedness and (un)interpretabil-ity always coincide. If that stipulation is given up, as has been proposed by Pesetsky & Torrego (2007), who argue that valuedness and interpretability should be disen-tangled, formal features come about in four kinds: (i) interpretable and unvalued fea-tures; (ii) interpretable and valued features; (iii) uninterpretable and unvalued features and (iv) uninterpretable and valued features. Both two types of interpretable features form a subset of the set of semantic features (although no clear examples are known of the second type), both types of uninterpretable features do not. Pesetsky & Torregos (2007) taxonomy thus looks as follows:

    (4)

    [P]

    [uF: val][uF: __]

    Formal featuresPhonological features Semantic features

    [iF: ___]([iF: val])

    [S]

    Hence, different views on the nature of formal features are available, with both (un-)interpretability and lexical (un)valuedness playing a role. However, at the same time it should be mentioned that the primary motivation to adopt lexical (un)valuedness as a second parameter is to solve the look ahead problem that pops up immediately if feature uninterpretability is a property of features that is visible at LF only and if at the same time its deletion would be the trigger for syntactic operations that take place prior to LF.

    However, as to the best of my knowledge has never been noted thus far, if fea-ture uninterpretability would not be an LF phenomenon but instead a purely formal property, this look ahead problem would disappear as well and consequently the need to adopt lexical (un)valuedness as a second parameter to characterize properties of formal features. In this paper, I argue for this alternative stand.

    1.3 Outline and scope

    The central aim of this article is to show that the learnability algorithm for formal features that I propose entails that the connection between features that may check off other features and LF-interpretability is not direct but rather indirect (and a result of language acquisition rather than a property of the formal system) and that interpre-table features are not semantic features as they lack LF-interpretability. Rather, the car-rier of an [iF] generally (but crucially not always) carries the semantics of F as well. The feature taxonomy that I propose is, then, a simpler and more symmetric one, where the set of formal features is autonomous and consists of two types of formal features: [iF]s and [uF]s, where [uF] encodes a need to stand in a proper Agree configuration with [iF], and where [iF] encodes the ability to satisfy [uF]s configurational needs. The taxonomy is depicted below:

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 113

    (5)

    [P]

    [uF]

    Formal featuresPhonological features Semantic features

    [iF] [S]

    The question that then arises is how the language learner can acquire which element carries [iF] and which one carries [uF] and how the alleged connection between [iF] features (the checkers) is captured by the proposed learnability algorithm.

    In short, it is argued that the acquisition of formal features is governed, after Zeijlstra (2008), by so-called doubling cases, i.e. cases where the locus of interpreta-tion of a particular semantic property (i.e. a semantic operator or feature) does not correspond 1:1 with its morphosyntactic manifestation(s). Hence, formal features, [iF] and [uF], can only be acquired against the background of semantic mismatches. As a result most, but crucially not all, elements carrying [iF] will therefore also carry the semantics of F. The correspondence between semantic content and the ability to check off a particular formal feature must, then, be a property of language acquisition, not of grammar itself.

    2. Acquiring formal features

    As argued for in the introduction, the only way to evaluate the claim that the set of formal features in a particular grammar is not part of UG but rather emerges in the process of language acquisition (and therefore on the basis of language input only), is by evaluating a particular procedure that may account for this emergence. In this section, I propose such a procedure. In short, I argue that formal features can only be acquired if a particular semantic feature is morphosyntactically doubled; this will form a cue for language learners to assume that this semantic feature is formalized, i.e. that it should be taken to be a formal feature as well. In 2.1, I outline the general format of this learning algorithm and in 2.2, I illustrate it by applying it to the acquisition of formal negative features.

    2.1 Proposal

    Following/modifying earlier work (Zeijlstra 2008), I argue that formal features should only be postulated by the language learner if the language input provides overt evi-dence for it. In terms of learnability, this entails that the null hypothesis must be that formal features are absent and that morphemes map phonological content directly to semantic content. Hence, the starting assumption should be that any element that seems responsible for the induction of a particular semantic context should also be taken to be the carrier of this semantic content:

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    114 Hedde Zeijlstra

    (6) Assume a 1:1 correspondence between morphemes and semantic content.

    Assuming (6), it follows that only if a 1:1 relation between some morpheme and its semantic contribution proves to be absent, other properties than semantic and pho-nological properties must be assigned. So, if a particular element, for instance a verbal agreement marker, manifests the presence of some semantic context (e.g. the semantic phi-features of the subject), then it must be taken to carry a corresponding uninterpre-table formal phi-feature. If a language lacks any kind of phi-morphology, it therefore lacks formal phi-features as well; only if phi-agreement is present does this provide evidence for the language learner to assume that the verb contains some uninterpreta-ble person, number and/or gender features. So, only semantically redundant elements must be assigned [uF]:

    (7) If some morpho-syntactic element manifests the presence of some semantic context F, but cannot be assumed to be the carrier of F itself, then assign a formal feature [uF] to .

    Informally, (7) means that carrying [uF] may only take place if the element carrying [uF] is able to mark the presence of some semantic property F without actively carry-ing the semantics of F itself. Now, since a learnability requirement for the acquisition of [uF]s is that they appear in a particular semantic context, in principle no [uF] fea-ture can be present without a corresponding semantic operator F. But, if it is a formal requirement that the element carrying this corresponding semantic property F must be present if an element carrying [uF] is present, this element should not only carry the semantics of F, but also a formal property that states that an element [uF] cannot survive in the sentence without it; this formal property is, by definition, a feature [iF]:

    (8) Assign [iF] to all morphosyntactic elements that introduce the semantic context that is manifested by [uF]. If no overt morphosyntactic element is responsible, assume some covert element to be present that carries the semantics of F and that therefore should be assigned [iF].

    So far, this is not new: as pointed out in Zeijlstra (2008), under Chomsky's (1995, 2002) analysis of formal features as in (1), if formal features must be acquired during the process of language acquisition, uninterpretable features must form the cue, since interpretable formal features are still part of the set of semantic features and there-fore semantically indistinguishable from them. The only distinguishing property of interpretable features in comparison to purely semantic features under this approach is their ability to check their uninterpretable counterparts. Consequently, (9) must be true as well. As we will see later on, this is an important and necessary step, although it may appear to be redundant now, given (8).

    (9) Assign [iF] to all those elements that are responsible for the rest of the grammatical occurrences of [uF].

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 115

    To summarize, a feature [uF] is learnable. If some element does not carry the seman-tics of F, but at the same time may only appear in a semantic context F, this element carries [uF]. If some [uF] appears in a grammatical sentence, some element must carry [iF]. Hence, [iF] features are learnable as well.

    2.2 Example: Negative Concord

    Now, let us apply this proposal to the acquisition of formal negative features. As is well-known, some languages do and other languages do not exhibit doubling phenomena with respect to negation. In Dutch, every morphosyntactically negative element also contains a semantic negation, but in Czech and Italian this is not the case, as is shown in (10)(12) below. According to the algorithm in 2.1, this entails Czech and Italian have formal negative features at their disposal, but Dutch does not.

    (10) a. Jan ziet niemand. Dutch Jan sees n-body Jan doesnt see anybody. b. Niemand zegt niets. n-body says n-thing Nobody says nothing.

    (11) a. Gianni *(non) ha telefonato a nessuno. Italian Gianni neg has called to n-body Gianni didnt call anybody. b. Ieri *(non) ha telefonato nessuno. yesterday neg has called n-body Yesterday nobody called. c. Ieri nessuno (*non) ha telefonato (a nessuno). yesterday n-body neg has called to n-body Yesterday nobody called (anybody).

    (12) a. Milan *(ne-)vid nikoho. Czech Milan neg.saw n-body Milan doesnt see anybody. b. Dnes *(ne-)vol nikdo. today neg.calls n-body Today nobody calls. c. Dnes nikdo *(ne-)vol. today n-body neg.calls Today nobody calls.

    In Dutch, every morphosyntactically negative element corresponds to a semantic negation. These negative elements are either the negative marker niet or a negative

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    116 Hedde Zeijlstra

    quantifier, as illustrated below in (13). Note that the locus of the negative operator at LF does not coincide with its relative position at surface structure, but this is due to quantifier raising in (13b) or Verb Second in (13c), which both are independent phenomena.

    (13) a. Niemand komt. x.[person(x) & come(x)] n-body comes Nobody is coming. b. Jan doet niets. x.[thing(x) & do(j, x)] Jan does n-thing Jan does nothing. c. Jan loopt niet. walk(j) Jan walks neg Jan isnt walking.

    Since there is no element that marks the presence of negation but is not semantically negative itself, condition (6) is always fulfilled with respect to negation, and conse-quently, there is no need to assign any [uNEG] feature, along the lines of (7). Since there are no [uNEG] features to be assigned in Dutch, there is no reason to assign [iNEG] features either (cf. (8)(9)). The only types of negative elements in Dutch are the negative marker and the negative indefinites, and these contain a semantic nega-tion and no formal negative feature:

    (14) Dutch negative elements: Niet Negative indefinites

    Things are different, however, in Negative Concord languages. Let us start by discuss-ing the Non-strict Negative Concord language Italian. Both n-words (the term for negative indefinites in Negative Concord languages, after Laka 1990) and the negative marker may render a sentence negative:

    (15) a. Gianni non ha telefonato. call(g, x)] Gianni neg has called Gianni didnt call. b. Nessuno ha telefonato. x.[person(x) & call(x)] n-body has called Nobody called.

    However, in Italian a combination of the negative marker with the n-word gives rise to a single semantic negation. In fact, In Italian, postverbal n-words obligatorily need to be accompanied by the negative marker non or a preverbal n-word. This means that a large part of negative sentences in the L1 input consists of sentences such as (16).

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 117

    (16) Gianni non ha visto nessuno x.[person(x) & see(g, x)] Gianni neg has seen n-body Gianni has seen nobody

    Since (16) contains more than one negative element, but only one negation in its semantics, only one of the negative elements can be semantically negative and the other one must be semantically non-negative (otherwise semantic compositionality would be violated). An n-word, like nessuno, which in isolation gives rise to a semantic negation, together with a negative marker gives rise to just one semantic negation. Following (7), an n-word must therefore carry an uninterpretable formal negative fea-ture [uNEG]. Since non can still be assumed to be responsible for semantic negation, it must be assigned [iNEG] along the lines of (8).

    The fact that non is the carrier of [iNEG] and n-words carry [uNEG] remains problematic in one respect, namely that Italian also allows sentences such as (17), which provide evidence against nessunos semantic negativity. Here non is absent (and may not even be included). Hence, all overt negative elements carry [uNEG].

    (17) Nessuno ha telefonato a nessuno xy[person(x) & person(y) n-body have.3sg called to n-body & call(x, y)] No-one has called anyone

    Now, following (8), some abstract negative operator must be assumed to be present and carry [iNEG], otherwise no element could be responsible for the checking of the n-words [uNEG] features:

    (18) Op nessuno ha telefonato a nessuno [iNEG] [uNEG] [uNEG]

    The inventory of Italian negative elements is thus as follows: the negative marker car-ries [iNEG] and contains a semantic negation; so does the covert negative operator Op. The n-words in Italian, however, are semantically non-negative and only carry an uninterpretable formal negative feature [uNEG].

    (19) Italian negative elements: Non [iNEG] (+ ) N-words [uNEG] Op [iNEG] (+ )

    The learning algorithm in (6)(9) enables the language learner to acquire the negative inventory in (19). However, one cautionary note must be made here. Since Italian has two semantically negative elements (non and Op), one of them being phonologically null, the question what prevents overgeneralization of Op, inclusion? Why wouldnt many more sentences that the ones like (15b) and (17) contain Op? The answer to this question should receive a more general answer, since this is a general question about the distribution of covert elements, and not necessarily about the distribution

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    118 Hedde Zeijlstra

    of covert negative elements. In short, referring the reader to Zeijlstra (2008, 2012) for more discussion, I argue that covert elements may only be postulated to be present in some sentence if the grammaticality of the sentence could otherwise not be accounted for. This is indeed the case in (15b) and (17), since no overt element can be said to be responsible for the checking of any [uNEG] feature, but the other examples (which contain non or are simply positive), are grammatical without postulating any covert negative operator. Consequently, Op cannot be present in such sentences.

    In Czech, the application of the learnability algorithm, again, leads to slightly dif-ferent results. First, since Czech is (strict) a (strict) Negative Concord language, nega-tion must be formalized and n-words are attributed a feature [uNEG]. However, the (default) assumption that the negative marker carries [iNEG] cannot yet be drawn on this basis. To see this, take (20).

    (20) Nikdo ne-vol n-body neg.calls Nobody calls

    If ne carried a feature [iNEG], the negative subject would appear outside its scope, which is in contrast with the fact that nikdo marks the presence of a negative context in which it appears (see Zeijlstra (2004, 2008) for a much more extensive discussion of these facts).

    (21) Op Nikdo ne-vol x.[person(x) & call(x)] [iNEG] [uNEG] [uNEG]

    As a final consequence, single occurrences of ne cannot be taken to be realizations of the negative operator, but must be seen as a marker of such an operator. In (22) the negative marker indicates the presence of Op, which in turn is responsible for the negative semantics of the sentence.

    (22) Milan Op ne-vol call(m) [iNEG] [uNEG]

    A reviewer raises the question of why ne should be obligatorily present, as removing it would not affect the Agree relation between the [iNEG]-bearing abstract operator and the overtly realized [uNEG]-bearing n-words. This question, however, relates not to a property of (Strict) Negative Concord specifically, but to the more general problem of morphosyntactic agreement: why is it that particular elements, not only negative, but also person, number or gender markers whether inflectional morphemes or independent elements may not be omitted, despite the fact that they simply realize uninterpretable features? Presumably, the morphological grid of the verbs requires all kinds of agreement morphemes to be spelled out, but I should be acknowledged that here we face a much more general problem which cannot be fully addressed within the confines of the present paper. This question is, however, independent from the fea-tural status of ne: regardless what causes its obligatory presence in negative sentences,

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 119

    it does not have an interpretable [iNEG] feature, but rather an uninterpretable [uNEG] feature.

    Czech, thus, has a different inventory of negative elements than Italian. In Italian, the negative marker is semantically negative and carries [iNEG]. In Czech, on the other hand, it is semantically non-negative and carries [uNEG]. The only semantically negative element carrying [iNEG] in Czech is Op:

    (23) Czech negative elements: Ne [uNEG] N-words [uNEG] Op [iNEG] (+ )

    The acquisitional procedure outlined in Section 2.1 predicts that Czech and Italian are languages that have a formal negative feature at their disposal, whereas Dutch has not. The idea that these languages differ in terms of the formal features that are part of their formal feature inventory comes along with a number of predictions. One of these predictions is that languages without Negative Concord may not exhibit an overt negative marker that occupies a head position in the clausal spine. This prediction is indeed born out (cf. Zeijlstra 2004, 2008). Other predictions have been tested and con-firmed as well and the reader is referred to the aforementioned papers for an overview. It should also be noted that even though Czech and Italian do not differ with respect to the presence of a formal negative feature, they do differ with respect to what ele-ment has what kind of formal features. In fact, the grammatical differences between Czech and Italian with respect to the expression of negation follow directly from the differences in terms of the kind of negative feature ([uNEG]/[iNEG]) that each nega-tive element carries, fully in line with the Borer-Chomsky conjecture that states that parametric differences between languages reduce to lexical differences (cf. Borer 1984, Chomsky 1995).

    For the remainder of the paper, the crucial assumption is not, however, the ex-act inventory of negative elements, but rather the fact that the learning algorithm in (6)(9) predicts that only Negative Concord languages have formal negative feature ([iNEG]/[uNEG]) and Non-Negative Concord languages do not have any formal neg-ative feature. Other proposals, such as Brown (1999) and Progovac (2005), who take Czech ne to carry [iNEG], or other views on Negative Concord as an instance of syn-tactic agreeement (cf. Zanuttini 1991; Haegeman and Zanuttini 1991, 1996, Haegeman 1995 and Haegeman & Lohndal 2010) are thus still compatible with the grammatical feature taxonomy that is proposed in Section 3.

    3. Interpreting interpretable features

    A question that arises now concerns the interpretational status of interpretable for-mal features, like [iNEG]. Does an interpretable formal feature such as [iNEG] have

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    120 Hedde Zeijlstra

    semantic content itself or not? Under the outlined proposal, two logical possibilities arise:

    (24) [iNEG] (and therefore all [iF]s) are interpreted as carriers of the semantics of negation (or F).

    (25) The element carrying [iNEG] (or [iF]) must be taken to carry a semantic negation (or the semantics of F) as part of its lexical semantics; this means that it is not the feature [iNEG]/[iF] itself that is being interpreted at LF.

    Solution (24) represents the current view on formal features and, at first sight, appears to be the preferred option. First, it immediately reduces the ability of interpretable formal features to check and delete matching uninterpretable formal features to their semantic properties. Moreover, the fact that only uninterpretable formal features have to be delet-ed (and thus to undergo Agree) whereas interpretable features do not do so also imme-diately follows: all other elements are semantically interpretable and do not violate Full Interpretation. However, the option in (24) comes at a particular price, as it faces several severe and some hitherto unnoted theoretical and empirical problems. I discuss these problems below and argue that these problems do not surface under assumption (25).

    3.1 Theoretical problems

    First, as already mentioned in 1.2, the assumption that uninterpretable features must be checked and deleted, because otherwise they would make the derivation crash at LF (due to Full Interpretation), introduces a major look ahead problem. For Chomsky (1995), uninterpretable features must be deleted at the level of LF and feature checking is a necessary condition for feature deletion. However, at the stage in the derivation in which feature checking takes place, it is not yet known that the feature, if remained unchecked, would cause the derivation to crash at a later stage, as has been pointed out by Epstein et al. (1998), Epstein and Seely (2000) and others. Chomsky (2002), for this reason, introduces the notion of feature valuation, and by stipulation, connects that to feature interpretability (see 1.2) and thus has to expand the feature taxonomy with all related problems.

    However, this look ahead problem arises only under the view on feature interpret-ability in (24). The view on feature interpretability in (25) does not face this problem, as here the difference between [iF]s and [uF]s is only formal in nature and thus visible in the course of the derivation; the only property of [iF] is that it is able to check the configurational needs of [uF].

    Second, although it is an advantage that feature checking can be motivated in terms of Full Interpretation, it can only do so by virtue of the stipulation that fea-ture checking leads to LF deletion of [uF]s. However, it is unclear why feature check-ing should lead to deletion. Nothing principled motivates it, so the conjecture that checked uninterpretable features are deleted is at best a stipulated one.

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 121

    In fact, one may even wonder why the appearance of [uF]s at LF should make the derivation crash. Take for instance the structure in (26):

    (26) E

    C

    A B

    D

    Now suppose that A is semantically empty, i.e. it contains only formal features at LF. In that case, the denotation of D is identical to the denotation of B. If no other gram-matical condition is violated and D can be a semantic complement of C (or vice versa), nothing renders (26) illegible at LF. Hence, the presence of uninterpretable features does not a priori violate Full Interpretation.

    In fact, the idea that the presence of an uninterpretable feature at LF would violate Full Interpretation is even contradictory. Arguing that the presence of some element blocks the interpretation of a structure that would otherwise receive a proper inter-pretation at LF presupposes that this element has interpretational effects and as such cannot be said to be fully uninterpretable.

    Again, under the view in (25), such problems do not necessarily arise. There is no need anymore to allude to a principle such as Full Interpretation that gives rise to a contradiction. The triggering of syntactic operations, simply takes place as a result of the need of certain learnable formal properties of lexical elements. As long as the outcome of the derivation is legible to the interfaces, no further constraints on the derivation have to be imposed.

    3.2 Empirical problems

    Apart from these theoretical considerations, (24) also makes some predictions that do not hold empirically. First, it is predicted that only elements with the semantic prop-erty F may check [uF]s; second, it is predicted that elements that have the semantics of F may check [uF]s. Both predictions are, however, too strong. Certain elements that lack the semantics of F may sometimes check [uF]s and certain elements that carry the semantics of F sometimes fail to check [uF]. Again, this would be ruled out under (24), but is predicted to be possible under (25). Let me illustrate this again by means of Negative Concord:

    3.2.1 Non-negative contexts checking [uNEG]N-words in complement clauses of verbs expressing doubt or fear, prepositions as without, or in comparatives, receive a non-negative interpretation, as the following Spanish examples taken from Herburger (2001) illustrate:

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    122 Hedde Zeijlstra

    (27) a. Pedro compr el terreno sin contarselo a nadie. Spanish Pedro bought the land without telling to n-body Peter bought the land without telling anybody b. Antec de hacer nada, debes lavarle las manos. before of do n-thing, must.2sg wash.cl the hands Before doing anything, you should wash your hands. c. Dudo que vayana encontar nada. doubt.1sg that will.3pl.subj find n-thing I doubt they will find anything. d. Prohibieron que saliera nadie. forbade.3pl that went_out.3sg.subj n-body They forbade anybody to go out. e. Es la ltima vez que te digo nada. is the ultimate time that you tell.1sg n-thing This is the last time I tell you anything. f. Juan ha llegado ms tarde que nunca. Juan has arrived more late than n-ever Juan has arrived later than ever.

    One can only maintain (24) (under the view that Negative Concord is an instance of syntactic agreement) if all these non-negative licensers of n-words contain some underlying negation (as has been proposed by Postal 2000 among others). Since all these licensers are in some sense felt to be negative (they are all (Strawson-)Downward Entailing), such underlying negations are not unnatural to expect. However, it should be noted that languages vary to quite a large extent with respect to whether these con-texts may license n-words, as shown below for Russian and Czech without.

    (28) a. Bez nikoho Czech without n-body Without anybody b. * Bez nikogo Russian without n-body Without n-body

    Consequently, languages should then cross-linguistically (and language-internally) differ with respect to whether some (Strawson-)Downward Entailing contexts must be lexically decomposed into some negative element or not. However, as there is no independent motivation for this assumption, this would be an instance of circular rea-soning.

    These facts, therefore, show that it is problematic to assume that every element that may license an element carrying [uNEG] contains a decomposable semantic ne-gation. However, if that is not the case, there are semantically non-negative elements that may check a feature [uNEG], which is a clear violation of (24).

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 123

    3.2.2 Negative contexts that cannot check [uNEG]Another argument against the idea that [iF]s are semantic features is provided by ele-ments that have some clear semantic property F, but are unable to check appearing [uF]s in their local domain. One such example is provided by French negation. Take the following data from Zeijlstra (2010):

    (29) a. Personne (ne) mange. French n-body neg eats Nobody eats. b. Jean (ne) mange rien. Jean neg eats n-thing Jean doesnt eat anything. c. Personne (ne) mange rien. nobody neg eats n-thing Nobody doesnt eat anything.

    (30) Marie (ne) mange pas. Marie neg eats neg Marie doesnt eat.

    (31) Personne (ne) mange pas (rien) n-body neg eats neg n-thing Nobody doesnt eat (anything)

    This leads to the following question: what are the properties of French n-words and French negative markers ne and pas, such that ne can combine with both n-words and pas, while still yielding a single semantic negation, whereas pas and n-words may not be combined in such a way?

    Note that ne may also appear in various kinds of other (Strawson-)Downward Entailing contexts, such as restrictive focus, comparatives, complement clauses of ex-pressions of fear, avoidance, denial or doubt, conditionals and temporal before clauses, as shown below (all examples have been taken from Rooryck (2008: 34) and Zeijlstra (2012)):

    (32) a. Jean (ne) voit que Marie. Jean neg sees comp Marie Jean only sees Marie. b. Jean est plusmalinque Pierre (ne) lest. Jean is smarter Pierre neg it is Jean is smarter than Pierre is. c. Il a barricad la porte de peur/crainte quon (n) entre chez lui. he has blocked the door of fear thatthey neg enter with him He blocked the door for fear that people might come in.

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    124 Hedde Zeijlstra

    d. Jean a vit que Lucienne (ne) tombe. Jean has avoided that Lucienne neg fall.subj Jean prevented Lucienne from falling. e. Nie/doute-t-il que je (ne) dise la vrit? denies/doubtshe that I neg tell.subj the truth? Does he doubt/deny that I am telling the truth? f. Je viendrai moins que Jean (ne) soit l. I will-come to less that Jean neg is.subj there I will come unless Jean is there. g. Il est parti avant que nous (n) ayons mange. he is left before that we neg have eaten He left before we ate.

    This makes it plausible to assume that ne is actually an NPI (cf. Zeijlstra 2010), which may freely occur in a wide subset of all downward entailing contexts. Furthermore assuming, along the lines of Kadmon & Landman (1993), Krifka (1995), Chierchia (2006, 2011) that the licensing requirement by NPIs lies in their pragma-semantic properties, the fact that pas is semantically negative already accounts for nes possible co-occurrence with pas. However, if that were the case, pas should be expected to check the n-words [uNEG] feature, contrary to fact. Pas is never able to establish a Negative Concord relation with an n-words.

    These facts are mysterious under analyses where [uNEG] needs to be checked by a semantically negative feature. However, once it is dropped that pas, being a semantic negation, must carry a feature [iNEG], the facts follow immediately. Pas is semanti-cally negative but lacks a feature [iNEG] and therefore cannot establish Agree relations with n-words. The only element carrying [iNEG] is the abstract negative operator Op; n-words carry [uNEG] and therefore must be checked by this abstract negator. Ne, finally, is an NPI and may appear under the scope of Op, pas and other (Strawson-)Downward Entailing contexts.

    (33) French negative elements: Pas Ne NPI N-words [uNEG] Op [iNEG] +

    Again, the connection between so-called uninterpretable formal features and their corresponding semantics seems weaker than previously assumed, something unex-pected under (24), but predicted under (25).

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 125

    3.3 Proposal

    On the basis of the arguments outlined above, I adopt (25) instead of (24), which takes [iF] not to be a semantic feature, but a purely formal feature whose only property is that it can fulfil some checking requirement by [uF]. The fact that [iF] and F often correspond (i.e. elements meaning F carry [iF] and vice versa) is not a property of grammar but the result of the acquisition mechanism of formal features. Take again, the acquisition algorithm presented in Section 2, repeated below.

    (34) Assume a 1:1 correspondence between morphemes and semantic content.

    (35) If some morpho-syntactic element manifests the presence of some semantic context F, but cannot be assumed to be the carrier of F itself, then assign a formal feature [uF] to .

    (36) Assign [iF] to all morphosyntactic elements that introduce the semantic context that is manifested by [uF]. If no overt morphosyntactic element is responsible, assume some covert element to be present that carries the semantics of F and that therefore should be assigned [iF].

    (37) Assign [iF] to all those elements that are responsible for the rest of the grammatical occurrences of [uF].

    Now, the relevance of (37) becomes clear. [iF] does not denote that some element has to have the semantic property of F. The only requirement that is needed is that it is learnable that some element is able to check some [uF]s feature. For covert checkers this means that they must always carry the semantics of F, given (36), but for overt ele-ments that is not necessarily the case. This learnability algorithm now, in accordance with (25), solves the problems addressed in 3.12.

    Theoretically, the checking requirement of [uF]s is no longer semantically mo-tivated. [uF] is nothing than a formal encoding that this feature needs to stand in a c-command relation with some element carrying [iF]. Checking [uF]s thus takes place in syntax and no look ahead problem appears in the first place. Consequently, valua-tion is no longer needed to make feature checking possible within syntax.

    Concerning Full Interpretation, the stipulated requirement is dropped that [uF]s must be deleted/erased at LF. Since [uF]s and [iF]s lack any semantic import, they do not block legibility of the derivation at LF. The structure, including semantically vacu-ous elements, is still fully interpretable.

    The facts in 3.2 follow as well, since semantic negation is not a prerequisite for e.g. verbs expressing doubt in Negative Concord languages to carry [iNEG]. However, at first sight this may lead to an overgeneralization. In principle, now, every verb could be assigned a feature [iNEG], but the facts suggest that only (Strawson)-Downward Entailing elements carry them. However, this can be readily explained in diachronic terms. N-words historically emerge from NPIs (see Roberts & Roussou 2003, Jaeger

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    126 Hedde Zeijlstra

    2010). So, what used to be an instance of NPI licensing must have been reanalyzed as an instance of syntactic Agree.

    (38) DE context NPI DE context[iNEG] N-word[uNEG]The Slavic minimal pair in (28), can be explained as well, as it can now be assumed that Czech bez carries [iNEG], but Russian bez does not. So, the fact that only (Strawson-)Downward Entailing elements carry [iNEG] also follows without assuming that this is a formal requirement.

    Moreover, the facts regarding French follow as well. In fact, it is even predicted that French pas never got analyzed as carrying [iNEG], since it did not start out as a negation at the moment that n-words were reanalyzed as n-words (in those days pas was still an NPI itself, reinforcing the by then negative marker ne).

    Finally, it must be noted that even though the proposed system allows for a certain amount of non-correspondence between [iF]s and the semantics of F, it still ensures that there is a strong (though not absolute) correlation between semantic and formal features. After all, without doubling properties with respect to some semantic property F, no [i/uF] can be acquired in the first place. So, only by virtue of the relation between [iF] and F in the acquisitional domain, some [iF] may be assigned to an element that lacks the semantics of F. Some instances of this have been observed in the study to negation, but other cases of this may be instances where some kind of feature is ob-ligatorily present on all members of a particular syntactic category, instead of only on relevant ones. Examples to keep in mind, but also left to further study, are inherent case, gender and declination features.

    4. Conclusions

    In this article I argue that:

    i. The set of formal features in each language is acquired in the process of language acquisition and is thus not part of UG;

    ii. The proper cue for acquiring formal features are so-called doubling cases, i.e. cases where the locus of interpretation of a particular semantic property (i.e. a semantic operator or feature) does not correspond 1:1 with its morphosyntactic manifestation(s);

    iii. The set of formal features does not intersect with the set of semantic features: all formal features are uninterpretable syntax-internal features in the sense of Svenonius (2006); the close correlation between some formal features and some semantic properties follows as a by-product of the acquisition mechanism for for-mal features;

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 127

    iv. The set of formal features consists of two types of features: [iF] and [uF] features, where [uF] features need to be c-commanded in their local domain by [iF] fea-tures;

    This leads to a taxonomy of grammatical features as depicted below:

    (39)

    [P]

    [uF]

    Formal featuresPhonological features Semantic features

    [iF] [S]

    In this sense, formal features are as independent and module-internal as phonologi-cal and semantic features, and thus adhere to a very strict modular view, also between syntax and semantics. The only reason why syntactic features may have some semantic effect is that, in the process of language acquisition, formal features only emerge under mismatches between sound and meaning.

    References

    Beghelli, Filippo and Stowell, Tim. 1997. Distributivity and negation. The syntax of each and every. In Anna Szabolcsi (ed.), Ways of Scope Taking, 71107. Dordrecht: Kluwer. DOI: 10.1007/978-94-011-5814-5_3

    Bobaljik, Jonathan David and Thrainsson, Hskuldur. 1998. Two heads arent always better than one. Syntax 1: 3771. DOI: 10.1111/1467-9612.00003

    Borer, Hagit. 1984. Parametric Syntax: Case Studies in Semitic and Romance Languages. Dordrecht: Foris Publications.

    Brown, Sue. 1999. The Syntax of negation in Russian: A Minimalist approach. Stanford: CSLI Publications.

    Chierchia, Gennaro. 2006. Broaden your views. Implicatures of domain widening and the Logicality of language. Linguistic Inquiry 37: 535590. DOI: 10.1162/ling.2006.37.4.535

    Chierchia, Gennaro. 2011. Meaning as an inferential system. Polarity and free-choice phenom-ena. Ms., Harvard University.

    Chomsky, Noam. 1995. The Minimalist Program. Cambridge, MA: MIT Press.Chomsky, Noam. 2002. Derivation by phase. In Ken Hale: A Life in Language, Michael J.

    Kenstovicz (ed.), 154. Cambridge, MA: MIT Press.Chomsky, Noam. 2005. Three factors in language design. Linguistic Inquiry 36: 122. DOI:

    10.1162/0024389052993655Cinque, Guglielmo. 1999. Adverbs and Functional Heads. Oxford: Oxford University Press.Cinque, Guglielmo (ed.). 2002. Functional Structure in DP and IP. The Cartography of Syntactic

    Structures, Volume 1. Oxford: Oxford University Press.Cinque, Guglielmo (ed.). 2006. Restructuring and Functional Heads. The Cartography of Syntactic

    Structures, Volume 4. Oxford: Oxford University Press.Epstein, Samuel David, Groat, Erich M., Kawashima, Ruriko and Kitahara, Hisatsugu. 1998. A

    Derivational Approach to Syntactic Relations. Oxford: Oxford University Press.

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    128 Hedde Zeijlstra

    Epstein, Samuel David and Seely, Daniel. 2002. Derivation and Explanation in the Minimalist Program. Oxford: Blackwell Publishers. DOI: 10.1002/9780470755662

    Grimshaw, Jane. 1997. Projections, heads and optimality. Linguistic Inquiry 28: 373442.Haegeman, Liliane and Zanuttini, Raffaella. 1991. Negative heads and the neg criterion. The

    Linguistic Review 8: 233251. DOI: 10.1515/tlir.1991.8.2-4.233Haegeman, Liliane and Zanuttini, Raffaella. 1996. Negative concord in West Flemish. In

    Parameters and Functional Heads. Essays in Comparative Syntax, Adriana Belletti and Luigi Rizzi (eds), 117179. Oxford: Oxford University Press.

    Haegeman, Liliane and Lohndal, Terje. 2010. Negative concord and multiple agree: A case study of West Flemish. Linguistic Inquiry 41: 181211. DOI: 10.1162/ling.2010.41.2.181

    Herburger, Elena. 2001. The negative concord puzzle revisited. Natural Language Semantics 9: 289333. DOI: 10.1023/A:1014205526722

    Iatridou, Sabine. 1990. About AgrP. Linguistic Inquiry 21: 421459.Jger, Agnes. 2010. Anything is nothing is something. On the diachrony of polarity types of

    indefinites. Natural Language and Linguistic Theory 28: 878822. DOI: 10.1007/s11049-010-9113-1

    Kadmon, Nirit and Landman, Fred. 1993. Any. Linguistics and Philosophy 16: 353422. DOI: 10.1007/BF00985272

    Kayne, Richard. 2000. Parameters and Universals, Oxford: Oxford University Press.Koeneman, Olaf. 2000. The Flexible Nature of Verb Movement. Ph.D. dissertation, Utrecht

    University.Krifka, Manfred. 1995. The semantics and pragmatics of polarity items in assertion. Linguistic

    Analysis 15: 209257.Laka, I. 1990. Negation in Syntax: On the Nature of Functional Categories and Projections.

    Ph.D. dissertation, MIT.Miyagawa, Shigeru. 2010. Why Agree? Why Move? Unifying Agreement-based and Discourse

    Configurational Languages. Cambridge, MA: MIT Press.Nilsen, ystein. 2003. Eliminating Positions: Syntax and Semantics of Sentential Modification.

    Ph.D. disseration, Universiteit Utrecht.Pesetsky, David and Torrego, Esther. 2007. The syntax of valuation and the interpretability of

    features. In Phrasal and Clausal Architecture: Syntactic derivation and interpretation, Simin Karimi, Vida Samiian and Wendy W. Wilkins (eds), 262294. Amsterdam: John Benjamins. DOI: 10.1075/la.101.14pes

    Pollock, Jean-Yves. 1989. Verb movement, universal grammar, and the structure of IP. Linguistic Inquiry 20: 365424.

    Postal, Paul. 2000. An introduction to the grammar of squat. Ms., NYU.Progovac, Ljiljiana. 2005. Negative and positive feature checking and the distribution of polar-

    ity items. In Negation in Slavic, S. Brown & A. Przepirkowski (eds), Slavica Publishers.Rizzi, Luigi. 1997. The fine structure of the left periphery. In Elements of grammar. Handbook

    in generative syntax, Liliane Haegeman (ed.), 281337. Dordrecht: Kluwer.Rizzi, Luigi. 2004. The structure of CP and IP. The Cartography of Syntactic Structures, Volume 2.

    Oxford: Oxford University Press.Roberts, Ian and Roussou, Anna. 2003. Syntactic Change. A Minimalist Approach to

    Grammaticalisation. Cambridge: Cambridge University Press. DOI: 10.1017/CBO9780511486326

    Starke, Michal. 2001. Move Dissolves Into Merge: A Theory of Locality. Ph.D. dissertation, University of Geneva.

  • 2014. John Benjamins Publishing CompanyAll rights reserved

    On the uninterpretability of interpretable features 129

    Starke, Michal. 2004. On the Inexistence of Specifiers and the Nature of Heads. In Structures and Beyond. The Cartography of Syntactic Structures, Volume 3, Adriana Belletti (ed.), 252268. Oxford: Oxford University Press.

    Svenonius, Peter. 2006. Interpreting uninterpretable features. Linguistic Analysis 33: 375413.Zanuttini, Raffaella. 1991. Syntactic Properties of Sentential Negation. Ph.D. dissertation,

    University of Pennsylvania.Zeijlstra, Hedzer Hugo. 2004. Sentential Negation and Negative Concord. Ph.D. dissertation,

    University of Amsterdam.Zeijlstra, Hedzer Hugo. 2008. On the syntactic flexibility of formal features. In Theresa

    Biberauer (ed.), The Limits of Syntactic Variation, 143174. Amsterdam: John Benjamins. DOI: 10.1075/la.132.06zei

    Zeijlstra, Hedzer Hugo. 2010. On French negation. In Proceedings of the 35th Annual Meeting of the Berkely Linguistics Society, Iksoo Kwon, Hannah. Pritchett & Justin Spence (eds). Berkely, CA: BLS.

    On the uninterpretability of interpretable features1. Introduction1.1 A universal set of formal features?1.2 Syntactic and semantic properties of formal features1.3 Outline and scope

    2. Acquiring formal features2.1 Proposal2.2 Example: Negative Concord

    3. Interpreting interpretable features3.1 Theoretical problems3.2 Empirical problems3.2.1 Non-negative contexts checking [uNEG]3.2.2 Negative contexts that cannot check [uNEG]

    3.3 Proposal

    4. ConclusionsReferences