Intervention effects in grammar and language acquisition

2.1 Hierarchical vs. linear intervention

The intervening element must intervene hierarchically, not just linearly. So, from a baseline sentence like (4)a, extraction of the temporal adjunct when is ill formed, as in (4)b, whereas from a baseline like (5)a, movement of when is possible, as in (5)b:

The difference here is that who intervenes between when and its trace hierarchically in (4)b, in that who c-command the trace but not when. On the contrary, in (5)b who intervenes only linearly, not hierarchically, as it is buried within the complex subject Determiner Phrase (DP). So, this kind of contrast supports a hierarchical definition of intervention, along the following lines:

Under (6), who is not visible as an intervener in (5)b, hence the structure is predicted to be well-formed.

Drawing from much work in the antisymmetry tradition (Kayne 1994 and subsequent work), Kayne (2018) has recently argued that syntactic representations involve linear as well as hierarchical organization (i.e. linear order results not only from a post-syntactic linearization operation, but directly from core syntactic operations like merge). If this is correct, and linear properties are accessible to grammatical principles, one may expect a residual deviance also in cases of purely linear intervention. Is this residual deviance empirically supported?, i.e. is (5)b perceived as (residually) more deviant that corresponding structures without linear intervention (e.g. something like When did the uncertainty about the result dissolve?)? This is an area in which controlled acceptability judgments with methods such as those illustrated in Sprouse (2007) and Sprouse, Schutze & Almeida (2013) may be of relevance (see Section 2.3). In this connection, it may be worth mentioning that Franck et al. (2006) and Franck et al. (2007) found both linear and hierarchical interference effects in elicited production of agreement, with significantly stronger effects with hierarchical intervention, but observable effects also with purely linear intervention.

In the lack of more refined evidence bearing on wh-extraction, I will continue to focus on purely hierarchical intervention, as resulting from (6), and will leave the possibility open of (weaker) linear intervention effects.

2.2 The class of interveners is featurally selective

The initial typology of positions for phrasal movement was the traditional distinction between A-positions and A’-positions (Rizzi 1990). This led to the expectation that every intervening position would induce a minimality effect on any form of A’-movement, but this characterization turned out to be too coarse. For instance, whereas quantificational adverbials like beaucoup (a lot) in French would determine an intervention effect on combien extraction from the direct object, manner adverbials like attentivement (carefully) would not (Obenauer 1983, 1994; Laenzlinger 1998):

Analogously in Italian and many other languages a (Clitic Left Dislocated) Topic can freely escape from an embedded clause introduced by a corrective focus, or from a wh-island:

Conversely, certain cases of A’-movement, e.g. in relative clauses, seem to be able to freely move across a topic:

All these cases presumably involve A’-specifiers. So, some A’-specifiers do not induce minimality effects on cases of A’-movement, which shows that more fine-grained distinctions than the simple A/A’-typology are needed.

In Rizzi (2004), this point was addressed (building on Starke 2001) through the proposal that, in the calculation of intervention, RM takes into account morphosyntactic features defining syntactic positions (the approach is then sometimes referred to as “featural Relativized Minimality”, or fRM). If we focus on specifier positions, the typology will be provided by the features determining external or internal merge of the specifiers. Moreover, morphosyntactic features are organized into feature classes, such that minimality effects are found within the same feature class, but not across classes. So, the wh-phrase and the manner adverbial belong to two distinct feature classes (in the original terminology, the operator class and the modifier class, respectively, the latter including features which license adverbial positions in an approach to adverb syntax like the one developed in Cinque 1999), so that no minimality effect is found in (8). On the other hand, the wh-element and the quantificational adverbial beaucoup share an operator feature, and this gives rise to a minimality effect in (7). Topics are licensed by the feature Top, which forms a feature class of its own, so that topics (at least, topics of the Italian kind) do not trigger minimality effects on operator chains nor are they affected by intervening operators, as shown by the well formedness of (9), (10) and (11).

2.3 Graded judgments

One important property of weak-island environments is that they give rise to graded judgments, depending on the nature of the extracted element and of the intervener. One classical observation is that all other things being equal, wh-phrases corresponding to arguments are more easily extractable than wh-phrases corresponding to adjuncts (Huang 1982):

Of course, these examples differ in several respects. Not only is it the case that which problem is an argument whereas how is an adjunct: another potentially crucial difference is that the first is complex, and contains a lexical restriction, the noun problem, whereas the second is a bare wh-operator. Here, I would like to focus on the relevance of the lexical restriction. A more minimal comparison between a bare and a lexically restricted wh-element shows a systematic, if slight, preference for the latter in extraction contexts:

This preference was quantitatively assessed in Villata, Rizzi & Franck (2016) in French: linguistically naïve speakers assigned on average 2.7 points to the equivalent of (13)a and 3.6 points to the equivalent of (13)b in a 7-point Likert scale. The preference is slight, but systematic and reproduced in distinct experiments with different scales. The paper also shows that the amelioration effect in (13)b is independent from D(iscourse)-linking (Pesetsky 1987, 2000), another factor that is often assumed to play a role in the acceptability of weak-island violations: in an experiment in which both lexically restricted and bare wh-elements were contextually D-linked, contrasts of the type (13)a–b were preserved. Acknowledging the fact that other factors have a role (the argument–adjunct distinction, with additional gradations between different kinds of adjunct, Rizzi 1990:91; D-linking; contrast, Lahousse et al. 2014, the finite or non-finite status of the wh-island; etc.), I will focus here on the improvement in acceptability determined by the lexical restriction.

2.4 Gradations of acceptability as a grammatical phenomenon

At this point, the analysis is confronted with an important conceptual issue. Grammatical principles are typically expressed in binary terms: either a structure respects a given principle, and is grammatically well-formed in that respect, or it violates the principle, and is ill-formed. If the grammatical structure is fundamentally binary in this sense, the empirical domain that grammatical hypotheses are built on is typically more complex and nuanced: judgments of acceptability, naturalness, etc. are typically graded, and it is often possible to identify clear gradients of acceptability, such that a given structure A is reliably judged more acceptable than B, which in turn is more acceptable than C, where A, B and C are all judged deviant to some extent, compared to a fully acceptable baseline case D. How can one express such gradations?

A classical approach is to appeal to the competence-performance distinction. Perhaps structures violating a grammatical principle are all marked as deviant by the grammar on a binary (yes, no) basis, and observable differences in degrees of acceptability are a function of grammar-external performance factors. For instance, it could be that ungrammatical structures which are easier to understand, e.g. for the comprehension of theta role assignment (who does what to whom), are perceived as more acceptable than less semantically transparent ungrammatical structures (on this point see Villata 2017).

Understandability or other grammar-external factors may well have a role in modulating acceptability, but there is another possibility to explore. It could be that, in some domain, the grammar offers tools for directly capturing gradients of acceptability. The exploration of this line is not new. A variant of it was proposed as early as in Chomsky (1964), with the notion of grammar-based degrees of grammaticality. And an important part of the discussion on constraints on extraction in Government-Binding analyses (Chomsky 1981) was based on the observation that Subjacency violations are less severe, in terms of acceptability, than Empty Category Principle violations, which in turn implied that certain principles are more “inviolable” than others, a step suggesting again that the pure binary picture was oversimplified.

The next step is the observation that fRM offers particularly promising ingredients for a grammar-based approach to acceptability gradients. We can construe intervention locality as a system that penalizes configurations in which an element is “on the way” in the computation of a local relations, such as the formation of a syntactic chain. The penalty can be linked to the degree of similarity between target and intervener in the local relation: the more similar the intervener, the higher the penalty. How can one define “similarity” in a precise enough way? The featural conception offers a natural grammatical dimension to define the relevant dimensions of similarity. We have seen that the features that are taken into account are morphosyntactic features which define syntactic positions, the features which license external or internal merge; in particular, for specifier positions filled by internal merge (movement), the features which trigger the operation.

We can now look at the set-theoretic relations that may arise between such defining features in the target and in the intervener. Let us consider, in particular, the three set-theoretic relations of disjunction, inclusion and identity. They give rise to a natural ranking of distinctness, expressed by the arrow in the diagram in figure 1:

Given the intervention configuration … X … Z … Y …, let us call X “the target” of the local relation, and Z “the intervener”.

Disjunction in relevant features is the case of maximal distinctness: here target and intervener have nothing relevant in common, and no penalty arises. At the opposite, lower side of the scheme we have Identity: here the target and in the intervener have all the relevant features in common, hence they are non-distinct, and the maximal penalty arises. Inclusion is the intermediate case of distinctness: X has some features distinguishing it from Z, but also some features in common with Z. Here, an intermediate degree of penalty arises.

We may then express these properties through the following revised formulation of fRM:

Again, “Relevant Syntactic Features” are syntactic features which define syntactic positions, by licensing applications of external and internal merge.

Let us now illustrate the functioning of the system in connection with the graded acceptability in cases of extraction from an embedded domain:

Example (15)c involves a case of identity, as both target and intervener are characterized by the +Q feature: here target and intervener are minimally distinct, hence a strong level of disruption arises. In (15)b, under the assumption that the lexically restricted character of the phrase is expressed by +N (and that this specification is relevant in the calculation of locality: on this see Section 3), the target properly includes the featural specification of the intervener. This is the intermediate degree of distinctness between target and intervener, giving rise to a partial disruption. Finally, in cases of extraction from a declarative like (15)a, we have potential interveners, the main and embedded subject positions, which are disjoint from the target (in terms of the features taken into account here). No locality-based disruption arises here, and in fact the structure is fully acceptable.

Other set-theoretic relations may be relevant here. In Villata et al. (2016) also cases like the following are considered, dubbed “inverse inclusion”, in which it is the intervener’s specification which properly includes the specification of the target, rather than the other way around, as in (15)b:

In the controlled acceptability experiment, inverse inclusion turns out to be as degraded as identity (15)c. This is expected, given the formulation of the principle in (14) because X is not distinct with respect to Z in inverse inclusion, much as it is not distinct in the case of identity. This case is important because it shows that what matters is not simply a non-directional distinctness of target and intervener (in which case inclusion and inverse inclusion would be expected to be alike), but the distinctness of the target X calculated with respect to the intervener Z: when the specification of X properly includes the specification of Z, as in (15)b, X is partially distinct from Z, hence an intermediate level of disruption arises. On the contrary, when the specification of Z properly includes the specification of X, as in (16), the target is non-distinct from the intervener (much as in the identity case), so that a stronger disruption arises. Evidently, the directionality in the calculation of distinctness is crucial here. ^[1]

3.1 Positional differences in northeastern Italian dialects

Striking evidence for a positional difference comes from Bellunese and other northeastern Italian dialects studies by Munaro (1999). In the core case, lexically restricted wh-phrases appear at the beginning of the clause, whereas bare wh-elements appear at the end; both cases involve inversion with the subject clitic:

Under Munaro’s analysis, both kinds of elements move to the left periphery and trigger subject clitic inversion (notice that (17)b differs from the corresponding French example, arguably involving a genuine “in situ” strategy, in that inversion is triggered). But the two kinds of elements target distinct positions in the map of the left periphery, a higher position (notated +Q+N) for the lexically restricted one, a lower position for the bare wh (notated +Q). Subsequent remnant movement of the Tense Phrase (TP) to the Spec of the intermediate position H determines the clause final position of the bare wh-element in the surface order:

3.2 Obligatoriness of inversion in Italian and Portuguese

In Italian, inversion is obligatory with bare wh-elements, whereas lack of inversion becomes possible with lexically restricted wh-phrases (Rizzi 1996, 1997 and Ambar 2003 observe this kind of contrast in European Portuguese): ^[2]

From this viewpoint, lexically restricted wh-phrases pattern with perché (why), which also does not require inversion:

In previous work, I proposed that perché occupies the Spec of Int, a functional head also hosting the embedded yes–no marker se (if) and occurring higher than Q in the map of the left periphery. In the original analysis, this head is inherently endowed with an interpretable interrogative feature, which does not require checking through T to C movement, whereas the lower +Q position does require checking. If a lexically restricted element targets a higher position +Q+N in the complementizer field, it may also benefit from the high interpretable interrogative feature and be exempted from checking (and inversion):

That the lexically restricted position is higher than the bare position is also, and more directly, shown by the fact that, whereas the bare position is strictly incompatible with left-peripheral focus, the lexically restricted position is (at least marginally) compatible with it, much as perché:

This pattern follows from a representation like (22) if the position +Q is in fact identical to the focal position, so that a left peripheral focus and a bare wh-element are strictly incompatible, whereas perché and lexically restricted elements occupy higher positions, compatible (at least marginally) with a focus position. On the (in)compatibility between these left peripheral positions, see Bocci, Rizzi & Saito (2018).

3.3 Co-occurrence with CP-internal that

Bayer and Brandner (2007) have observed that in Bavarian “phrasal wh” and why are natural with a post-wh occurrence of dass (that), whereas bare wh-elements are not (with some variation):

This pattern differs from the one of other varieties (e.g. various Germanic and Romance dialects) in which all wh-elements can (or must) co-occur with that (e.g. wie dat “who that” in Dutch varieties: Haegeman 1994). Apparently, that-like elements are versatile elements which can normally occur in the highest position of the C-system expressing declarative force, but also in a very low position, plausibly expressing finiteness, in varieties permitting who that sequences, and also in intermediate positions in cases like (24), where that is consistent with high wh-positions (lexically restricted and why) and inconsistent with low positions. So, the patterns can be captured by a map like (25):

Notice that a similar pattern is observed in standard Italian, not in questions but in exclamatives: che can co-occur with an exclamative phrase, but not with a bare exclamative element (on these constructions see Botteri 2018 and references quoted there):

3.4 Elicited production patterns in child English

Adult English does not seem to offer direct evidence for positional differences between bare and lexically restricted wh. Nevertheless, experimental work on language acquisition suggests that learners of English actively explore the option of a positional difference of the kind directly illustrated by other adult languages.

For instance, Thornton (1995) and Guasti, Thornton & Wexler (1995) conducted elicitation experiments with children acquiring English, trying to elicit negative wh-questions. Children clearly resist expressing negation on interrogative do. When such structures as (27) are elicited, children try out non-target-consistent structures, a typical case being a reduplication strategy, as in (28):

Just to fix ideas, I would like to speculate here that the child tries to avoid satisfaction of both the Q-criterion and the negative criterion (Haegeman 1995) through the same element, possibly in the attempt to minimize reconstruction: didn’t in (27) would carry both the Q and the Neg feature, and its higher occurrence would satisfy the Q-criterion, whereas its lower (reconstructed) occurrence would satisfy the Neg criterion, under Haegeman’s analysis:

(consider also a certain analogy here with the configuration barred by Wexler’s 1994 Unique Checking Constraint).

In the reduplicated structure (28) produced by the child, the two criteria are satisfied by distinct occurrences of do, with no need of reconstruction (and/or of multiple checking).

Thornton (1995) and Thornton, Guasti & Wexler (1995) show that reduplication is by and large the only systematic strategy followed by children with bare wh-elements, whereas it alternates with two other strategies with lexically restricted wh-phrases:

Why do lexically restricted wh-phrases have more options? Thornton (1995) argues that lexically restricted wh-phrases may target a higher position then bare wh-phrases, one which does not trigger inversion and may co-occur with the overt complementizer that, whereas bare wh-phrases are forced to occur in a lower position, requiring inversion and inconsistent with that.

From the vantage point of the phenomena discussed in the previous sections, it is remarkable that English learners hypothesize non-target-consistent structures which are in fact used in other languages: uninverted structures like (30)b are akin to the Italian–Portuguese structures discussed in 3.2, and the structures with post-wh that in (30)c are akin to the Bavarian structures discussed in 3.3 (as well as to the corresponding Italian exclamatives). What makes the parallel particularly striking is that the “special” options are restricted to complex wh-phrases in both child English and the adult varieties taken into account. The crosslinguistic evidence thus provides strong support for Thornton’s analysis. What all this suggests is that the learners of English, trying to avoid costly configurations like (28), revert to Universal Grammar (UG) options which may be selectively available only for certain elements. So, they limit the use of the higher +Q+N position to lexically restricted elements, much as many adult varieties do.

3.5 Languages allowing multiple wh-movement

Languages like Bulgarian and other Slavic languages, and also Rumanian, Hungarian, etc. permit (or require) movement of more than one wh-element to the left periphery in multiple questions. Moreover the movement must reproduce the order of elements within the TP structure, e.g. the subject must precede a complement. The phenomenon is illustrated through Rumanian data (Alboiu 2002; Soare 2009):

Many analyses have been proposed of this ordering preserving constraint ever since Rudin’s (1988) original observation (Richards 1997; Boskovic 2002; Krapova & Cinque 2008 a.o.). The relevant point here is that the order can be subverted if the complement wh-phrase is lexically restricted:

Whatever mechanism is responsible for the order preservation constraint (e.g. Richards’ “tucking in”, or Krapova & Cinque’s interpretation of RM, etc.), the special status of lexically restricted elements may be amenable to the fact that they target a special position in the left periphery, distinct from and higher than the positions of bare wh-elements, so that the mechanism imposing order preservation (whatever its ultimate nature) is not operative. The analysis here is only sketched out and many complications should be taken into account; nevertheless, I believe that this sketch suffices to support the view that there are distinct attractors for bare and lexically restricted elements.

In conclusion, different languages offer a varied array of arguments supporting the conclusion that lexically restricted wh-phrases (+Q+N) target a distinct and higher position than the position targeted by bare wh-elements (+Q) in a cartographic map of the left periphery. This systematic finding in turn supports the view that the specification +N (our notation for the presence of a lexical restriction in the phrase) actively participates in the identification of the landing site of wh-movement. Thus, it is not surprising that the +N feature, syntactically active in the triggering of movement, may be taken into account in the computation of locality, e.g. in modulating the acceptability of extractions from weak islands. In the last decade, the presence of +N has been exploited also in connection with intervention effects in language acquisition, which are reported in the next section.