1. Introduction

This chapter is, in some sense, an update on the ‘nature versus nurture’ debate in the realm of developmental semantics. We report the findings from experimental studies designed to adjudicate between a usage-based approach to language acquisition, and a nativist approach. The studies we review were inspired in large part by the model of the language apparatus proposed by Gennaro Chierchia (see especially Chierchia Reference Chierchia and Belletti2004). Several features of the model have been the focus of experimental investigations of child language, including children’s sensitivity to scalar implicatures and their knowledge of several linguistic phenomena that, on the surface, do not appear to be related, but which Chierchia’s model has sought to unify. More specifically, the same linguistic contexts that cancel scalar implicatures (downward-entailing contexts) are also contexts that license negative polarity items and ones that generate a “conjunctive” interpretation of disjunction. We have been investigating these properties in children’s emerging grammars in two typologically distinct languages, Mandarin Chinese and English. In this chapter, we summarize some of the main findings of these investigations.

2. Two interpretations of disjunction

This section reviews the basic tenets of the alternative approaches to language acquisition that will feature prominently in the remainder of the chapter: the usage-based (nurture) approach and the nativist approach. To frame the discussion, we will focus on the interpretation of disjunction as seen from these alternative perspectives. First, preliminary remarks about the semantics of disjunction are in order.

3. The usage-based approach

According to many scholars, knowledge of the meanings of linguistic expressions is acquired by language learners based on observations of the linguistic behavior of others; this is especially true of the meanings of logical expressions. We call this the usage-based account. Meaning must be manifested in use, on this perspective, because all there is to the meaning of a logical expression is the use that speakers make of it. Modern philosophy of language yields many endorsements of this approach to the acquisition of language, and especially the acquisition of the meanings of logical expressions. For example, the philosopher W. V. O. Quine (Reference Quine1992: 37–38) asserts:

Dummett (Reference Dummett1978: 216–217) explicitly adopts the usage-based approach to the meaning of logical expressions:

Advocates of the usage-based approach are also well represented in the literature on child language. In this domain, advocates of the approach contend that “children acquire word uses closely related to those used in natural language input, only later using a word to convey a broader range of meanings” (Morris, Reference Morris2008: 68; cf. Lievenet al. Reference Lieven, Pine and Baldwin1997; MacWhinneyReference MacWhinney, Givón and Malle2002; TomaselloReference Tomasello2003). The acquisition of the meanings of logical connectives, such as disjunction, is no exception. Citing the finding mentioned earlier – that exclusive-or dominates in the input to children – Morris (Reference Morris2008: 70) identifies an advantage to children in initially assigning the exclusive-or interpretation to disjunction words:

Lacking relevant input, children learn the “formal logic” uses of disjunction words much later, according to Morris (Reference Morris2008: 82–84):

As all of these quotations make clear, the idea that logic is innate does not go down easily with many scholars. It almost seems to be presupposed in much of the literature, both on child language and on reasoning, that children acquire logic by observing the outward behavior of adult language users. Yet, as we will see, there is a growing body of evidence that clearly demonstrates that children initially assign the inclusive-or meaning to disjunction words across languages, despite the absence of decisive evidence from the input. This evidence is problematic for the usage-based account (Crainet al. Reference Crain, Goro and Thornton2006; Crain et al. Reference Crain, Gualmini, Pietroski, Carruthers, Laurence and Stich2005; Crain and KhlentzosReference Crain and Khlentzos2008, Reference Crain and Khlentzos2010; Crain and ThorntonReference Crain, Thornton, Traxler and Gernsbacher2006). Before we turn to the child language laboratory, the next step is to explain the apparent paradox: how the majority of disjunctive statements (and hence the input to children) can have the exclusive-or meaning of disjunction and, at the same time, all human languages, including all child languages, assign the inclusive-or meaning to disjunction.

4. The Gricean model

Grice (Reference Grice, Cole and Morgan1975) proposed that the basic meaning of disjunction across human languages is indeed inclusive-or. To explain how the exclusive-or meaning is derived in “real-life” circumstances, the standard Gricean account takes the following line. Disjunction words and conjunction words form a scale < or, and >. The scale is ordered by information strength, from weakest to strongest, where a term α is “weaker” than another term β if β asymmetrically entails α (HornReference Horn1996a). Since the truth conditions assigned to statements with conjunction are a subset of those that verify the corresponding statements with disjunction, statements with “and” asymmetrically entail the corresponding statements with “or”. So, “or” is weaker than “and.” When a speaker uses the weaker term “or”, a scalar implicature of “exclusivity” is engaged. The trigger to the implicature is usually attributed to the Gricean conversational Maxim of Quantity, which entreats speakers to make their contributions as informative as possible. In adherence with this maxim, hearers generally assume that a speaker who uses “or” in describing a situation is not in a position to use the corresponding statement with the stronger term “and” (Grice Reference Grice, Cole and Morgan1975). Upon hearing a statement with “or”, hearers therefore remove the truth conditions associated with “and” from the meaning of “or”, yielding the exclusive-or reading of disjunctive statements (A or B, but not both). To see that sentences with “and” are true in a subset of the circumstances associated with “or,” consider sentences (3) and (4).

On the basic meaning of disjunction (inclusive-or), (3) is true in circumstances C1-C3. Notice that the corresponding statement with conjunction (4) is true in just one of these circumstances, C3. According to the Gricean view, the statement with the weaker term (“or”) induces a derived meaning that includes the negation of the stronger term (“not both”).

The Gricean view appeals to scalar implicatures to explain the findings from the reasoning literature that have been used to impugn the conclusion that disjunction words are inclusive-or in human languages. It is expected, on the Gricean approach, that when scalar implicatures are cancelled, the basic inclusive-or truth conditions will be assigned to sentences with disjunction. As Grice observed, scalar implicatures are cancelled, or at least diminished, in situations of uncertainty. When one is making a prediction, or making a bet, it is possible to use disjunction without invoking a scalar implicature.

Consider example (5). In a study by Gualminiet al. (Reference Gualmini, Crain, Meroni, Howell, Fish and Keith-Lucas2000), this sentence was uttered by a puppet, Kermit the Frog, who was invited to make wagers about the outcomes of a series of stories.

Following Kermit’s statement in (5), several alternative outcomes were presented to children. On one outcome associated with (5), Batman ended up taking both a cake and an apple. The use of disjunction in (5) does not guarantee that this will be the outcome, but neither is this possibility excluded, since the outcome was still in dispute when Kermit produced (5). The finding was that 3- to 5-year-old English-speaking children accepted statements like (5) 98% of the time in situations where both disjuncts turned out to be true, i.e., when Batman took both a cake and an apple. When sentences describing past events were produced (Batman took a cake or an apple), children continued to accept statements with “or,” whereas adults rejected them. In statements about past events, there is less uncertainty, so adults at least compute the scalar implicature when disjunction is used to describe past events. Where children are less sensitive to scalar implicatures than adults, children appear to be more logical. We will review some of the literature on children’s sensitivity to implicatures in Section 8.

The Gricean model explains why an implicature of exclusivity is engaged in ordinary assertions with disjunction like (3), repeated here as (6). The Gricean model was extended by Gazdar (Reference Gazdar1979) to explain why people assign the truth conditions associated with inclusive-or in negated disjunctions like (8). According to Gazdar, scalar implicatures are often suspended under negation because the relative strength of scalar terms is the reverse of what it would be without negation. The use of ‘or’ in (7) results in a stronger statement than the corresponding statement with ‘and’ (8).

As indicated by the contexts given below the examples, statement (8) Ted did not order pasta or sushi is true in one circumstance, when Ted ordered neither dish. Notice that the corresponding statement with conjunction (9) Ted did not order pasta and sushi, is true in that circumstance (Ted ordered neither dish), and is also true in other circumstances (Ted ordered just one of the dishes). Therefore, negated disjunctions are true in a narrower range of circumstances than negated conjunctions, hence negated disjunctions are more informative (stronger). In subsequent work, Horn (Reference Horn1989) proposed that scalar implicatures are not just suspended under negation but, more generally, they are suspended in downward entailing contexts.

5. The unification of linguistic phenomena

If linguists are struck by the diversity of human languages, they are also struck by the common themes, i.e., the patterns that occur in individual languages, and across languages. Here is where Chierchia enters the picture. First, Chierchia (Reference Chierchia and Belletti2004, Reference Chierchia2006) sought to verify the conjecture by Horn (Reference Horn1989) that scalar implicatures are suspended in downward-entailing contexts. The strategy he followed was to see if three apparently unrelated linguistic phenomena were tied together by the semantic property of downward entailment. Although this correlation is mentioned in passing in Chierchia et al. (Reference Chierchia, Guasti, Gualmini, Meroni, Crain, Foppolo, Noveck and Sperber2004) and in Guastiet al. (Reference Guasti, Cherchia, Crain, Foppolo, Gualmini and Meroni2005), a detailed exposition of the correlation is reported in Chierchia (Reference Chierchia and Belletti2004, Reference Chierchia2006).

As we saw, one phenomenon that was hypothesized to be governed by downward entailment was the suspension of scalar implicatures (SIs). This indeed proved to be the case. Chierchia showed that, time and again, SIs were suspended, or ‘recalibrated,’ in downward-entailing environments. But, just as importantly for the present discussion, Chierchia showed that the suspension of SIs was correlated with two other linguistic phenomena. One is the interpretation of disjunction. Because it is anticipated that SIs are suspended in downward-entailing linguistic contexts, it follows that if the exclusive interpretation of disjunction is derived by a SI, then the “basic” inclusive meaning of disjunction should arise in downward entailing linguistic contexts. The third linguistic phenomenon was already well known from the seminal work of Fauconnier (Reference Fauconnier1975) and Ladusaw (Reference Ladusaw1979). This phenomenon is the licensing of negative polarity items (NPIs), such as English any. As Ladusaw (Reference Ladusaw1979) had established, downward-entailing linguistic contexts license NPIs. Chierchia sought to show, then, that downward-entailing contexts tie these different linguistic phenomena together. In downward-entailing contexts, NPIs are licensed, and disjunction is assigned the inclusive-or interpretation, yielding conjunctive entailments.

All these facts are expected, as Chierchia pointed out, for the simple reason that downward-entailing linguistic contexts reverse entailment relations, as compared to ordinary (non-downward-entailing) linguistic contexts. To appreciate this insight, we first need to state what it means for an operator, or a linguistic environment, to be downward entailing (DE). The defining property of DE linguistic contexts is that they guarantee the validity of inferences from general statements to more specific statements. More specifically, these contexts license inferences from sentences with expressions that refer to sets of things (e.g., fruit, car) to sentences with corresponding expressions that refer to the subsets of those things (e.g., apples, Prius). Formally, a function f is DE iff f(A) entails f(B), whenever B ⊆ A.

The class of downward-entailing operators in human languages includes negation, negative adverbs, certain prepositional phrases (but not others), certain determiners (but not others), verbs expressing minimum conditions, comparatives, and many others. In addition, there are downward-entailing linguistic environments, such as the antecedent of conditionals (but not the consequent). These operators and environments form a natural class in human languages, despite appearing to be a fairly motley collection of expressions and linguistic contexts.

Now let us look at how the semantic property of downward entailment provides a unifying account of several linguistic phenomena which, on the surface, look unrelated. To illustrate, we will use the three linguistics structures in (9). Notice first, that each of these linguistic contexts satisfies the defining property of downward entailment – they validate inferences from sets (speaking a Romance language) to their subsets (e.g., speaking French, speaking Spanish, speaking Italian). Example (9a) shows that the phrase headed by the determiner every in subject position is DE. The antecedent of a conditional statement is shown to be downward entailing in (9b), and (9c) shows that the preposition before is DE.

We chose these linguistic contexts because, in each case, downward entailment is restricted to one structural position, or to one of a pair of lexical items. To see this, consider the “invalid” inferences from sets to their subsets in (10). The examples in (10) reveal that the set-referring expression, Romance language, cannot be replaced by the subset-referring expression, French, in the consequent clause of conditional statement, in the predicate phrase (nuclear scope) of the universal quantifier every, or in the complement of the preposition after. These “asymmetries” in DE linguistic contexts, as illustrated in examples (9) versus (10), will loom large in our later discussion of language learnability.

A second feature of downward entailment is the licensing of negative polarity items (NPIs), such as English any (also ever and at all). Examples in (11a)–(13a) illustrate that any is welcome in the phrase headed by the determiner every in subject position, and in the antecedent of conditional statements, and following the preposition before. Examples (11b)–(13a) show, by contrast, that any is not licensed in the predicate phrase (nuclear scope) of the universal quantifier every, or in the consequent clause of conditional statements, or following the preposition after. Such asymmetries are potentially problematic for language learners.

A third phenomenon associated with downward entailment pertains to the interpretation of disjunction. In the scope of a DE operator, or in DE linguistic environments, disjunction is assigned its basic inclusive-or meaning. We know this, because disjunction licenses a conjunctive entailment in these environments (Boster and CrainReference Boster and Crain1993; ChierchiaReference Chierchia and Belletti2004, Reference Chierchia2006; Crain et al. Reference Crain, Gualmini, Pietroski, Carruthers, Laurence and Stich2005; Crain and KhlentzosReference Crain and Khlentzos2008). Example (14) shows that or generates a conjunctive entailment in the phrase headed by the determiner every in subject position; (15) shows that disjunction generates a conjunctive entailment in the antecedent of a conditional; and (16) shows that disjunction generates a conjunctive entailment in the scope of the preposition before.

Downward-entailing (DE) linguistic contexts are a common feature of human languages, so the unification of the linguistic phenomena observed for English should extend to even historically unrelated languages, such as Mandarin Chinese. We will illustrate this, first, using ruguo conditionals in Mandarin. Ruguo-conditionals correspond most closely to English if. . ., then. . . statements. We saw that the antecedent of an English conditional statement is downward entailing, so the statement If a linguist bought a car, he got a rebate entails If a linguist bought a Prius, he got a rebate. We also saw that the consequent clause of an English conditional is not downward entailing, so the statement If a linguist got a rebate, he bought a car does not entail If a linguist got a rebate, he bought a Prius. If the English examples are translated into Mandarin Chinese, the same asymmetry appears. That is, the antecedent clause of a ruguo-conditional is downward entailing in Mandarin, as shown in (20), but the consequent clause is not, as indicated by the “*” in (21).

Another linguistic phenomenon governed by downward-entailing linguistic contexts is the licensing of negative polarity items, such as English any. In Mandarin, the expression that corresponds most closely to the English NPI any is renhe. Mandarin renhe has the same pattern of distribution in ruguo-conditionals, as the English NPI any does in if . . ., then. . . conditionals. This is shown in (22).

A third linguistic phenomenon that is governed by downward entailment involves the interpretation of words for disjunction in human languages, including English or and Mandarin huozhe. In the antecedent clause of a conditional statement, disjunction generates a conjunctive entailment. By contrast, when disjunction appears in the consequent clause, it does not generate a conjunctive entailment; instead, disjunction is assigned “disjunctive” truth conditions in the consequent clause. Mandarin works in exactly the same way. In the antecedent clause of a conditional statement, disjunction generates a conjunctive entailment, as illustrated in (23). When disjunction appears in the consequent clause in Mandarin, the interpretation is “disjunctive,” as shown in (24).

As Chierchia (Reference Chierchia and Belletti2004) points out, it is instructive to ask whether children could have plausibly learned when disjunction words generate a conjunctive entailment, and when they do not. If these facts are learned, then both English-speaking children and Mandarin-speaking children must have access to evidence that words for disjunction (English or; Mandarin huozhe) have different meanings when they appear in the antecedent clause of a conditional, as compared to when these same words appear in the consequent clause. Because this is a fact about meaning, Chierchia remarks, it seems highly unlikely that children have abundant experience relevant to this distinction.

Consider the asymmetries in the Mandarin examples (23)–(24). In both sentences, the disjunction word (huozhe) is permitted. So the relevant distinction is not based on a distributional analysis of the occurrence or non-occurrence of a particular (kind of) word. The relevant distinction is in the interpretation of disjunction: in (23) disjunction makes a conjunctive entailment, whereas in (24) disjunction is assigned “disjunctive” truth conditions. The distinction that children must draw concerns the different interpretations that the same disjunction words receive when they appear in different linguistic environments. Nevertheless, 3- to 5-year-old children have been found to correctly distinguish the meaning of disjunction words in these two positions (cf. Gualminiet al. Reference Gualmini, Meroni, Crain and Weisgerber2003; Meroniet al. Reference Meroni, Gualmini, Crain and van Geenhoven2006; see SuReference Su2011, for results from studies with Mandarin-speaking children).

Based on these joint facts that (a) young children know when disjunction does and does not generate a conjunctive entailment, and (b) children lack evidence in the input for this distinction, Chierchia (Reference Chierchia and Belletti2004) concludes that acquiring the meaning of disjunction “yields a particularly strong version of the poverty of the stimulus argument.” The findings from cross-linguistic research on child language, therefore, reinforce the nativist account of language acquisition, according to which children are innately endowed with ‘deep-seated’ linguistic principles that tie together a variety of linguistic phenomena which appear to be unrelated on the surface. Advocates of the theory of Universal Grammar refer to deep-seated regularities as “core” linguistic properties, as compared to ‘peripheral’ properties of languages. We discuss the core/periphery distinction in the next section.

6. Core versus periphery

Advocates of the theory of Universal Grammar have argued that the core/periphery distinction is significant. In contrast, advocates of a usage-based model of language acquisition have argued that the core/periphery distinction has little merit. For example, Goldberg (Reference Goldberg2006) makes the following remark: “[t]he impossibility of making a clear distinction between the core and the periphery of linguistic structure is a genuine scientific discovery, and it has far-reaching theoretical implications.” According to usage-based theorists like Goldberg (Reference Goldberg2006) and Tomasello (Reference Tomasello2003), if a learning mechanism suffices for learners to acquire linguistic phenomena that lie at the periphery of human languages, then the same mechanism will surely also suffice for learners to acquire core linguistic phenomena.

This argument reveals how the notion “core” is understood by advocates of the usage-based approach. This view of the core/periphery distinction is revealed in another statement by Goldberg (Reference Goldberg2003: 14):

Tomasello (Reference Tomasello2003: 104–105) concurs:

As these quotations indicate, according to the usage-based model of language acquisition, core properties are ones with “maximal generality”. Their effects are expected to appear with greater regularity in a language, as compared to peripheral phenomena. This invites the inference that children should have an even easier time learning core properties, as compared to peripheral properties, of the local language.

Proponents of Universal Grammar also contend that there are core properties shared by human languages, and that these properties express regularities. But, the kinds of regularities are not the same as those discussed by usage-based theorists. Here is an instructive quote by Chomsky (Reference Chomsky1965):

There are two noteworthy characteristics of the kinds of “deep-seated” regularities that are stated in the theory of Universal Grammar. First, these regularities are expressed universally and, second, they tie together a number of linguistic phenomena which, on the surface, appear to be unrelated.

7. Putting scalar implicatures into the semantics

Up to this point, all the research findings are compatible with the standard Gricean view of scalar implicatures. We have seen that scalar implicatures (SIs) are responsible for the exclusive-or interpretation of disjunction and SIs tend not to arise in downward-entailing (DE) contexts. Chierchia followed up some of the implicit predictions of the (neo-)Gricean approach, by showing that DE contexts have broader empirical coverage than had previously been documented. Not only do DE contexts suspend SIs, they also govern another two apparently unrelated linguistic phenomena; disjunction generates a conjunctive entailment in DE contexts, and NPIs are licensed in DE contexts. Chierchia introduced an additional claim, however, which represented a challenge to the standard Gricean view.

According to the standard Gricean view, there is a division of labour between the semantics and the pragmatics. SIs are a pragmatic phenomenon, and are computed globally, after the truth-logical (truth-conditional) meaning of the sentence is composed. On Chierchia’s model, by contrast, there is no division of labour; SIs are computed within the semantic component, alongside the logical meaning, as part of the on-line incremental composition of sentence meaning. Although we will leave it to others to discuss this particular feature of Chierchia’s proposal in detail, it will be useful to present the basic idea of the “localist” computation of SIs. This is useful in order to appreciate one of the conclusions that followed from this perspective on sentence meaning. It should suffice, for our purposes, to present one simple argument in favor of the localist view, and against the globalist view of the computation of scalar implicatures. This will give you the flavor of the difference that holds between these alternative views of the architecture of the language apparatus.

Consider sentence (25), which contains the existential indefinite some. This quantificational expression forms a scale with other quantificational expressions many, most, and every. The expression some is the weakest term on the scale <some, many, most, every>. This follows from the fact that a statement with some is asymmetrically entailed by the corresponding statements in which some has been replaced by many, most, or every. According to the standard Gricean account, therefore, (25) implies (but does not entail) that (26) is false.

On the Gricean account, scalar implicatures are a root phenomenon. They are applied to the output of the computational system, wherein the compositional semantic meaning of sentences are constructed. So, we might say that the semantics is responsible for the “logical meaning”, and the pragmatics is responsible for implicatures. In any event, example (27) represents the Gricean “globalist” view of the application of the scalar implicature that is operative in (25).

On the “localist” view, by contrast, the logical meaning and implicatures are both computed within the semantic component of the language apparatus. This is represented in (28).

Now we can appreciate the disjunction problem. Consider the disjunctive statement (29), in which some resides in the second disjunct.

On any account of SIs, (29) implies that Gennaro is not seeing every student. The accounts differ, however, on how this implicature comes about. According to the Gricean account, SIs are a pragmatic phenomenon, and are applied only after the compositional meaning of a sentence is constructed. On this approach, the implied denial of the stronger statement (that Gennaro is seeing every student) is a root phenomenon, as represented in (30).

Because negation is applied to the output of the compositional semantics, this leads to the unwanted consequence that both disjuncts are subsumed under negation. In English, however, negated disjunctions generate a “conjunctive” entailment, as in one of de Morgan’s laws of propositional logic: ~(A ∨ B) ⇒ (~A & ~B). Applying this law to the negated disjunction in (30) yields the invalid inference in (31), which is clearly not part of the intended meaning of the sentence Gennaro is either at the pub or seeing some students.

Now suppose, instead, that implicatures are factored into semantic representations locally. Then, the semantic representation in (32) provides a snapshot of the output of the SI, as it applies to the example under consideration.

There are far-reaching consequences of the localist perspective for language learnability. According to this view, scalar implicatures are computed alongside the truth-conditional meaning of a sentence. That is, SIs apply within the computational system of Universal Grammar. All things being equal, then, children should be expected to compute both aspects of the compositional meaning of sentences as early as they can be tested. Here is how Guastiet al. (Reference Guasti, Cherchia, Crain, Foppolo, Gualmini and Meroni2005: 670) put it:

In view of the expectations of the localist account of scalar implicatures, it is worth taking the time to consider what we have found out about children’s abilities to compute scalar implicatures.

8. Children’s knowledge of scalar implicatures

In a series of studies, children have been found to compute scalar implicatures relatively late in the course of language acquisition. Even children as old as 11 have been found to accept weak scalar terms (some, or) to a far greater extent than adults do, in contexts in which the corresponding statements with stronger terms (all, and) are (also) true. For example, Smith (Reference Smith1980) found that 4- to 7-year olds interpreted some as meaning some, and possibly all in responding to questions like Do some giraffes have long necks? and Braine and Rumain (Reference Braine and Rumain1981) found that 7- to 9-year olds interpreted or inclusively, assigning the meaning A or B, and possibly both to sentences of the form “A or B.” In an important study, Noveck (Reference Noveck2001) established that even 11-year-old children did not derive scalar implicatures to the same extent as adults do, though adults, too, failed to consistently compute scalar implicatures in some experimental conditions. Studies by Papafragou and Musolino (Reference Papafragou and Musolino2003) and by Guastiet al. (Reference Guasti, Cherchia, Crain, Foppolo, Gualmini and Meroni2005) replicated the findings by Noveck (Reference Noveck2001). In addition, they included experiments in which children were explicitly instructed to make decisions about the “informativeness” of different descriptions of the same object, rather than a judgment about truth or falsity. For example, children were shown a picture of a grape, and were asked whether it was better to use “grape” or “fruit” as a description. In this task, children’s performance improved following training. However, when the same children were tested one week later, without additional training, it was found that the effect of the previous training did not persist, for most of the children. A number of other experimental studies have also shown that children’s performance is task dependent (cf., Foppoloet al. Reference Foppolo, Guasti and Chierchia2012; Katsos and BishopReference Katsos and Bishop2011; Pouscoulouset al. Reference Pouscoulous, Noveck, Politzer and Bastide2007).

Children have been found to successfully compute scalar implicatures, without training, in three experimental contexts. One was a new task reported in Chierchiaet al. (Reference Chierchia, Crain, Guasti, Gualmini, Meroni, Do, Domínguez and Johansen2001), called the Felicity Judgment task. In this task, two puppets each offer a different description of a story, and children reward the puppet that “said it better.” On the test trials, one of the puppets used or, and the other one used and. For example, one trial was about a group of farmers who were cleaning their animals. At the completion of the story, all the farmers had decided to clean a horse and a rabbit, among their other animals. The puppets then produced the statements in (33).

Fifteen children participated in the study, with a mean age of 4;8 (range 3;2–6;0). Children correctly rewarded the puppet that produced the statement in (33b), with and, on 93% of the trials. Children clearly favored the puppet that had made the more informative statement, demonstrating that children as young as 3-years-old have one piece of pragmatic knowledge, Grice’s Maxim of Quantity.

In addition to the Felicity Judgment task, Meroni and Gualmini (Reference Meroni and Gualmini2012) found that children successfully computed the scalar implicature associated with the indefinite NP some (i.e., some, but not all) in another task, developed by Zondervan (Reference Zondervan2006). In the task, the puppet made an assertion in response to a specific question, called the Question Under Discussion (QUD). In the assertions corresponding to some, children computed a scalar implicature if the QUD contained all, as in the question/answer pair in (34).

In the context corresponding to (34), all the hot-dogs had been delivered. A group of fifteen English-speaking children (average age = 4;8) rejected the puppet’s statement in (34b) 87% of the time.

Taken together, the findings show that children have no trouble computing scalar implicatures when they are exposed to both terms on the relevant scale: <or, and>, <some, . . . all>. Clearly, the observation made in previous studies – that children over-accept sentences with scalar terms, including disjunction and the indefinite some – was not the result of a lack of pragmatic knowledge, since exposure to both scalar terms sufficed to enable children to compute the corresponding scalar implicatures. However, children are not consistently exposed to both the weaker and the stronger term in ordinary conversational contexts. In many circumstances, children are exposed only to the weaker term, e.g., or or some, and they must compute the “derived” meaning representation, with the stronger term, without the aid of a puppet who produces a statement with the stronger term (as in the Felicity Judgment task), or an experimenter who poses a question that contains the stronger term (as in the QUD task).

According to the localist approach advocated by Chierchia, the computation of SIs involves two components. One component is the recursive interpretation of a sentence, including both its truth conditions, and the set of alternatives. The second component involves building and maintaining different representations of the sentence in working memory. Based on the findings of the Felicity Judgment task, and the task involving a QUD, it appears that children experience difficulty when they are required to compute the recursive interpretation of a sentence and calculate a scalar implicature, at the same time. In such circumstances, children’s responses appear to be driven by considerations of truth or falsity, whereas adults are able to focus directly on the felicity of what is said, in preference to truth or falsity. This suggests that children’s failure to compute SIs is due to processing limitations, such as verbal working memory, and not a deficit in semantic/pragmatic knowledge.

If this line of reasoning is correct, then children should perform more adult-like in tasks that dramatically reduce processing demands, and in tasks that focus on the felicity of a statement rather than on its truth or falsity. A direct attempt to verify this prediction was conducted by Guastiet al. (Reference Guasti, Cherchia, Crain, Foppolo, Gualmini and Meroni2005), who sought to make the conversational exchanges in which children are engaged as close as to those that we all experience in everyday life. In ordinary conversational exchanges, speakers and hearers share a common conversational background, which they modify on the basis of the events that transpire in the context. We evaluate the information strength of the statements made by others by comparing them to a given context, to see if they conform to standard conversational norms. Based on these observations, Guasti et al. tested a group of 7-year-olds (i.e., the age of the youngest children in the Noveck study) using a Truth Value Judgment task (Crain and ThorntonReference Crain and Thornton1998). This task permits the experimenter to control the situational context and, “thereby, to establish the conditions that are prerequisite for computing scalar implicatures” (p. 685). More specifically, the experimental context in a Truth Value Judgment task can be partitioned into a Possible Outcome and the Actual Outcome. In the study under consideration, the Possible Outcome was part of the story that corresponded to the test sentence itself, and the Actual Outcome was a subsequent event that transpired which made the test sentence infelicitous. In short, the Actual Outcome served as the basis for rejecting the test sentence on the grounds that it was a possible outcome, but one that did not eventuate.

An example story featured five soldiers. The soldiers were instructed to travel abroad to collect treasure. Each of the soldiers could either go by motorbike or ride a horse. There was some discussion among the soldiers. Some soldiers said that they would like to go by motorbike, since motorbikes are fast. But, some other soldiers disagreed. These soldiers argued that gasoline is expensive and that it would be more fun to ride a horse. After more discussion, one by one, the soldiers decided to ride a horse. So at the conclusion of the story, all the soldiers rode a horse. Following the conclusion of the story, the puppet produced the test sentence (35).

Now we can partition the story into the Possible Outcome and the Actual Outcome. The Possible Outcome corresponds to the point in the story at which some of the soldiers said that they preferred to ride a horse, since motorbikes consumed too much fuel. That point in the story (the Possible Outcome) corresponds to the test sentence: Some soldiers are riding a horse. The Actual Outcome is “what really happened” in the story; in the end, all of the soldiers rode a horse. This represents the reason that the test sentence is infelicitous.

Unlike many previous experiments, the rejection rate for the child subjects (75%) in the Truth Value Judgment task was not significantly different than the rejection rate for an adult control group (83%). The conclusion reached by Guasti et al. (Reference Guasti, Cherchia, Crain, Foppolo, Gualmini and Meroni2005: 692) is that the “crucial factors for enhancing the computation of implicatures are the availability of the relevant evidence and naturalness of the situation.” Both of these factors were absent in the tasks in which children failed to compute implicatures. Although the findings are encouraging, the fact that the child subjects were 7-years-old tempers the conclusions that can be reached. Children younger than 7 typically fail to compute scalar implicatures in tasks that do not involve making the alternative scalar terms explicit but, to date, it has not been demonstrated that children younger than 7 can successfully compute scalar implicature in the absence of explicit instruction, or tasks in which the relevant alternative scalar terms are not made overtly available for comparison.

9. Cross-linguistic variation

In the remainder of the chapter our attention will be directed to studies that were designed to assess children’s knowledge of the semantic property of downward entailment. We will concentrate on negation, because this proved to be the most interesting, and the most challenging. We saw that negative statements in English, such as Ted didnotorder pastaorsushi generated a conjunctive entailment: (i) Ted did not order pasta, and (ii) Ted did not order sushi. We argued that negated disjunctions yield a conjunctive entailment in English because English “or” is inclusive-or, just as in de Morgan’s laws of propositional logic, where it is valid to infer the negation of two propositions (~A & ~B) from one negated disjunction ~(A ∨ B). The model Chierchia proposed anticipates that disjunction words in all human languages should be interpreted as inclusive-or, and that appearances to the contrary will be due, largely, to the influence of pragmatic implicatures. We would not expect disjunction to be interpreted as inclusive-or in some languages, but as exclusive-or in others. This would just make languages harder for children to learn.

We soon began to test this prediction in Japanese and in Mandarin Chinese, and we made a quite unexpected discovery (but cf. SzabolcsiReference Szabolcsi, Kenesei and Siptar2002). When English sentences with negation and disjunction are translated into Japanese or into Mandarin Chinese, adult speakers of these languages did not judge these sentences to generate a conjunctive entailment. Consider the Mandarin example in (36).

It turned out that Mandarin contrasts with English in the way that the disjunction word, huozhe, is interpreted in sentences with negation, meiyou, as in (36). Adult speakers of Mandarin judge (36) to be true in three circumstances: (i) when Ted didn’t order pasta, (ii) when Ted didn’t order sushi, and (iii) when Ted didn’t order either pasta or sushi. This is clearly not the conjunctive entailment that is generated by English speakers. How can this cross-linguistic variation be reconciled within the theory of Universal Grammar? Our proposal is that the variation involves a parameter. There are other possibilities to consider, however.

Let us quickly put to rest the possibility that the Mandarin disjunction word, huozhe, is exclusive-or. If this were the case, adult speakers of Mandarin would not interpret sentences like (36) in the way that they do. To see this, consider the logical formula (A ⊕ B), where the symbol ‘⊕’ represents exclusive-or. This formula is true if exactly one of {A, B} is true. It follows that the negated disjunction, ~(A ⊕ B), is false if exactly one of {A, B} is true. But as we saw, adult speakers of Mandarin accepted the negated disjunction in (36) when exactly one of the disjuncts is true, i.e., when Ted ordered pasta, but not sushi, or when Ted ordered sushi, but not pasta. These are the very circumstances in which negated disjunctions would be false if huozhe were exclusive-or. If huozhe were exclusive-or, sentence (36) would be true in just two circumstances: (i) when Ted ordered neither pasta nor sushi, and (ii) when Ted ordered both pasta and sushi. It is highly implausible, in our view, that negated disjunctions could be judged to be true, in any human language, in situations in which both disjuncts are true (see Crain and KhlentzosReference Crain and Khlentzos2008 for further arguments that disjunction is not exclusive-or in human languages).

The observed differences in the interpretation of negated disjunctions across languages could be reconciled with the theory of Universal Grammar in (at least) two ways. First, negated disjunctions could have a different syntactic structure in some languages as compared to others. On this scenario, languages like Mandarin and Japanese might analyze negated disjunctions (e.g., (36)) as having two negated verb phrases, where the second verb phrase has undergone “disjunction reduction”. Example (37) shows the underlying syntactic structure of the sentence Ted did not order sushi or pasta, according to this analysis. If the semantic interpretation is computed before the second VP is elided, this could explain the judgments made by speakers of Mandarin and Japanese in response to negated disjunctions

On this account, languages like English do not undergo disjunction reduction. Because the semantic interpretation of negated disjunctions does not involve a second instance of negation in languages like English, disjunction directly licenses a conjunctive entailment.

A second account of the observed cross-linguistic variation was initially suggested by Szabolcsi (Reference Szabolcsi, Kenesei and Siptar2002), and proposed more formally in work by Goro (Reference Goro2004, Reference Goro2007). (For subsequent research adopting this proposal, see CrainReference Crain2008; Crain et al. Reference Crain, Goro, Minai, Schalley and Khlentzos2007; Crain, et al. Reference Crain, Goro and Thornton2006; Crain and KhlentzosReference Crain and Khlentzos2008.) On this account, the cross-linguistic variation is due to a parameter, called the Disjunction Parameter. The Disjunction Parameter partitions languages into two classes. In one class of languages, disjunction is a positive polarity item. By definition, a positive polarity item (PPI) takes scope over (local) negation. In the other class of languages, according to the Disjunction Parameter, disjunction is not a positive polarity item, so the logical form mirrors the surface syntax in negated disjunctions, with negation taking scope over disjunction. So, the two values of the disjunction parameter distinguish languages according to the possible scope relations between negation and disjunction. On one value of the parameter, disjunction and negation are related by “inverse scope,” because disjunction is a PPI. On the other value of the parameter, disjunction is not a PPI, so the scope relations are dictated by the surface syntax.

Let us indicate the parameter value on which disjunction takes scope over negation as OR = +PPI, and let OR = −PPI indicate the scope relations in languages where negation takes scope over disjunction. English takes the OR = −PPI value of the parameter. This is why Ted didn’t order pasta or sushi generates a conjunctive entailment, as in classical logic. Notice that the symbol for negation takes scope over the disjunction operator “∨” in the logical formula ~ (A ∨ B). Following de Morgan’s laws, a negated disjunction generates a conjunctive entailment: (~A & ~B). Languages in which negation takes scope over disjunction include English, German, French, Greek, Romanian, Bulgarian, and Korean (Szabolcsi Reference Szabolcsi, Kenesei and Siptar2002). In Mandarin, disjunction has the other parameter value, OR = +PPI. Therefore, disjunction fails to generate a conjunctive entailment in negative sentences like (36). The logical formula corresponding to (36) is (~A ∨ ~B). Languages in which disjunction takes scope over negation include Mandarin, Japanese, Hungarian, Russian, Serbo-Croatian, Slovak, and Polish (Szabolcsi Reference Szabolcsi, Kenesei and Siptar2002; Goro and AkibaReference Goro, Akiba, Chand, Kelleher, Rodríguez and Schmeiser2004a,b).

10. Independent evidence for a scope analysis

We have used tests of polarity sensitivity to assess the viability of the Disjunction Parameter. More specifically, disjunction fails to “scope out” over negation in linguistic contexts that cancel the polarity sensitivity of PPIs. We will illustrate one such context using the English PPI some. First, let us establish that some is indeed a PPI. To this end, we would point to the difference in interpretation between sentence (38), with some, and sentence (39), with any. Assuming that both the PPI some, and the NPI any are instances of existential quantification, “∃”, the difference between the meaning of (38) and the meaning of (39) is indicated alongside the examples: some is interpreted as taking scope over negation, ∃ > NOT, whereas any appears inside the scope of negation, NOT > ∃. So example (38) can be paraphrased as “There are kangaroos that Julia didn’t chase,” and example (39) can be paraphrased as “There do not exist kangaroos that Julia chased.”

11. Avoiding subset problems in language acquisition

Introducing the Disjunction Parameter raises new questions. One question is whether children are expected to adopt an initial, default setting of the Disjunction Parameter. Alternatively, children could be free to select either value. Following the line of reasoning first advanced in Berwick (Reference Berwick1985), children are expected to initially adopt the same value for a certain class of parameters. The learning mechanism that determines the initial, default value for this class is called The Subset Principle. The Subset Principle orders parameter values according to the number of sentences and their corresponding meanings that can be assigned to linguistic expressions.

The Subset Principle is operative when the class of languages that adopts one setting of a parameter, P (call these P1 languages), generates fewer sentence/meaning pairs for a given type of expression than the class of languages that adopts the alternative setting of P (call these P2 languages). In such cases, the Subset Principle compels children learning P1 languages and children learning P2 languages to initially hypothesize that the local language is a P1 language, rather than a P2 language. Children learning P2 languages are therefore expected to speak a fragment of a ‘foreign’ language for a while. If it can be established that children differ from adults in this way, this would be strong presumptive evidence against the usage-based account of language development. Because the usage-based account is “input matching”, these kinds of differences between child and adult languages are not anticipated.

The subset principle that is required to explain children’s use of the Disjunction Parameter is a semantic version of the familiar subset principle described by Berwick (Reference Berwick1985) and by Pinker (Reference Pinker1984). Both of these researchers observed that a learnability problem could arise for children when one language generates a subset of the sentences generated by another language. In the absence of negative evidence, children are compelled to initially adopt the “subset” language. Since the early 1990s, it has been claimed, albeit controversially, that when children are presented with a semantic ambiguity, they are guided by a learnability constraint that compels them to initially adopt the subset interpretation in order to guarantee that the superset reading can be learned from positive evidence, if the superset interpretation is assigned by adult speakers of the local language (Crainet al. Reference Crain, Ni, Conway, Clifton, Frazier and Rayner1994). This constraint on semantic interpretation is sometimes called the Semantic Subset Principle or the Semantic Subset Maxim, to distinguish it from the (syntactic) Subset Principle proposed originally by Berwick and by Pinker. The Semantic Subset Principle has been critically discussed by Musolino (Reference Musolino2006) and by Gualmini and Schwarz (Reference Gualmini and Schwarz2009). Responses to the criticism and new evidence in favor of this principle are provided in Notley, Thornton, and Crain (Reference Crain2012) and Notley et al. (Reference Notley, Zhou, Jensen and Crain2011). We offer additional evidence in Section 12.

The Semantic Subset Principle is expected to be operative in the case of the Disjunction Parameter. It can easily be verified that the circumstances in which sentences are true on the OR = +PPI value comprise a superset of those circumstances that make sentences true on the OR = −PPI value. In other words, the binary values are in a subset/superset relation. To see this, note that the OR = −PPI value yields a neither reading, whereas the OR = +PPI value yields a not both reading. A statement to the effect that neither A nor B is true in a subset of the circumstances corresponding to the statement endorsing not both A and B.

Based on learnability considerations (i.e., in the absence of negative evidence), Goro (Reference Goro2004, Reference Goro2007) predicted that Mandarin- and Japanese-speaking children should adopt the OR = −PPI value of the Disjunction Parameter. If so, they would be expected to interpret sentences with negation and disjunction in the same way as English speakers, OR = −PPI, but not in the same way as adult speakers of Mandarin or Japanese. If children start with the subset language, OR = −PPI, this guarantees that there will be positive evidence to compel children to override their initial preferences for the scope relations between negation and disjunction, if these preferences are not exhibited by adult speakers of the local language. For disjunction, the positive evidence for Mandarin-speaking children will be sentence/meaning pairs exhibiting the preference by Mandarin-speaking adults for disjunction to take scope over negation. Because disjunction takes scope over negation, these sentences will typically engage a scalar implicature. This guarantees that negated disjunctions will be used by Mandarin and Japanese speakers in circumstances in which only one disjunct is false, not both. In Mandarin, then, the sentence corresponding to Ted didn’t order sushi or pasta means that it is either sushi or pasta (but not both) that Ted didn’t order. In contrast to adults, the same sentence is expected to be true for Mandarin and Japanese-speaking children only if both disjuncts are false, NOT > OR. So, the truth conditions assigned to negated disjunctions by adults are inconsistent with the truth conditions that correspond to children’s initial interpretation. Assuming that scope preferences take time to reverse, a number of negated disjunctions must be encountered by children before they jettison their initial preference in favor of the adult parameter setting.

12. Experimental studies of the Disjunction Parameter

The predictions were confirmed in a series of experimental studies. The initial studies were conducted with Japanese-speaking children, by Goro and Akiba (Reference Goro, Akiba, Chand, Kelleher, Rodríguez and Schmeiser2004a,b). The Goro and Akiba studies examined children’s interpretation of Japanese negated disjunctions like (50). For adult speakers of Japanese, such negative sentences lack the conjunctive entailments associated with de Morgan’s laws. Adult Japanese speakers interpret (50) to mean that the pig didn’t eat the carrot or didn’t eat the pepper. Despite the appearance of ka within the scope of sentential negation in surface syntax, ka is interpreted by adults as having scope over negation.

Thirty 3- to 6-year-old children’s understanding of sentences like (50) was assessed using a version of the Truth Value Judgment task. On a typical trial, subjects were asked to judge whether or not (50) was an accurate description of a situation in which the pig had eaten the carrot but not the green pepper. The findings were as predicted by the Semantic Subset Principle. A control group of Japanese-speaking adults consistently accepted the target sentences, whereas children rejected them 75% of the time. The findings are even more compelling once the data from four children, who responded like adults, were set aside. The remaining 26 children rejected the target sentences 87% of the time.

Based on the findings of Goro and Akiba (Reference Goro, Akiba, Chand, Kelleher, Rodríguez and Schmeiser2004a,b), and the observation that the Mandarin Chinese disjunction word huozhe is a positive polarity item, further evidence of a disjunction parameter was pursued. According to the Semantic Subset Principle, children acquiring Mandarin should initially interpret the disjunction operator huozhe in conformity with de Morgan’s laws in simple negated sentences, despite the absence of this interpretation for Mandarin-speaking adults.

Here are the main findings. In response to negated disjunctions, twenty Mandarin-speaking children (mean age 4;5) rejected the target “not . . . or” statements 97% of the time, as did English-speaking children and adults; by contrast Mandarin-speaking adults accepted them 95% of the time. To justify their rejections, Mandarin-speaking children pointed out the animals in question had only eaten one vegetable. This indicates that children had assigned the “neither” (OR = −PPI) reading, as in English (also see Jinget al. Reference Jing, Crain, Hsu and Otsu2005).

It is worth pointing out that Chinese and Japanese differ in word order. In Japanese, the word for disjunction ka comes before negation nakat, as seen in example (50). In Chinese, the word for negation meiyou comes before the word for disjunction huozhe, as seen in example (51). We have seen also that adult speakers of both Chinese and Japanese interpret disjunction as having scope over negation (OR > NOT). Since the ordering of disjunction and negation is reversed in these languages, it follows that the interpretation assigned by adult speakers of both languages, according to which disjunction takes scope over negation, cannot be based on linear order. We also saw that children acquiring both Chinese and Japanese initially adopted the opposite scope relations, with negation taking scope over disjunction. It follows that children were not assigning this interpretation based on linear order, since the ordering of negation and disjunction differs across these languages. In short, no principle based on linear order could explain either the interpretation assigned by adults, or by children, across languages.

A parallel between Chinese and Japanese is the way in which adults generate the “neither” reading. In both Japanese and Chinese, this reading is assigned to simple negative sentences with conjunction. For example, the Mandarin Chinese conjunction word he appears in the simple negated sentence in (52), which conveys the ‘neither’ interpretation. Notice that this interpretation of conjunction under negation is not consistent with De Morgan’s laws, since “not (A and B)” is logically equivalent to ‘not A or not B’, where (52) is interpreted as requiring both A and B to be false.

Disjunction is interpreted in the same way in human languages as it is in classical logic. But it would appear that this is done at the expense of conjunction, which appears to be non-Boolean in (52). Yet, again, appearances can be deceiving. It is also possible that conjunction is a PPI in languages like Mandarin. As we saw, this possibility can easily be verified by examining the scope relations between conjunction and negation when these operators appear in different clauses (see CrainReference Crain2012).

If scope parameters are based on subset/superset relations in logical entailments, as we have suggested, then we expect that both children’s productions and their interpretations may be different from adults, depending on the default parameter setting, and the setting that is operative for adults. In further support of this view, it has been found that many English-speaking children produce the existential indefinite some in contexts where it is not licensed in the adult grammar. For adult English-speakers, some is a positive polarity item (PPI) and must take scope over local negation. For English-speaking children, however, it appears that some is not a PPI, and can be interpreted in the scope of negation. In fact, some can even be produced in the scope of local negation. Children have been found to use some or something in contexts where adults use any or anything to express the same message, e.g., None of the people had some presents (Musolinoet al. Reference Musolino, Crain and Thornton2000).

Negative polarity items, by contrast, are rarely, if ever, misunderstood or produced in the wrong linguistic environments by children at any age. Children adhere to the syntactic and semantic constraints on the use of negative polarity items from the earliest stages of language acquisition. Large-scale reviews of the spontaneous production data of both English-speaking children (aged 0;11–5;2) and Dutch-speaking children (aged 1;5,-3;10) have revealed children almost never produce negative polarity items without a downward-entailing licensor of some sort (TieuReference Tieu2010; Van der WalReference Wal1996). In elicited production tasks, it has also been found that children do not produce negative polarity items in non-downward-entailing environments, while they do produce them in downward-entailing environments (Crain and ThorntonReference Crain, Thornton, Traxler and Gernsbacher2006; O’LearyReference O’Leary1994; Van der Wal Reference Wal1996). The fact that children produce negative polarity items in just the right contexts shows that they are sensitive to the difference between downward-entailing environments and non-downward-entailing environments.¹

13. Conclusions

The model of the language apparatus proposed by Chierchia is innovative in several respects. Although the model is heavily influenced by Gricean principles, it clearly differs from the traditional Gricean account. Chierchia’s model locates the computation of scalar implicatures within the semantic component of the language apparatus, rather than in the pragmatics, as on the Gricean approach. This reassessment of the semantics/pragmatics interface leads to expectation that young children, across the globe, should manifest adult-like abilities to produce and comprehend sentences with logical operators and quantificational expressions, including the computation and cancellation of scalar implicatures. We briefly reviewed progress that has been made on that front.

The model proposed by Chierchia has inspired cross-linguistic studies of child language, including investigations of children’s knowledge of the licensing conditions of negative polarity items, and their interpretation of disjunction in different linguistic contexts. The findings of studies of children acquiring typologically distinct languages, including English and Mandarin Chinese, have verified that downward entailment is a unifying property in human languages, including child language. In conducting this research, we have uncovered several interesting and unexpected differences between child and adult languages. The observed discrepancies between child and adult language are difficult to explain on the usage-based account of children’s emerging linguistic competence. The observed differences could be explained, however, by the theory of Universal Grammar, according to which child language is expected to differ from the local adult language along the natural seams of human languages, i.e., parameters. We discussed one “scope” parameter, the Disjunction Parameter. We saw that, across languages, children appear to initially favor the value of the Disjunction Parameter that generates scope relations that make sentences true in the narrowest range of circumstances. This ensures that children will have access to positive evidence if the local language favors the alternative scope possibilities, those that make sentences true in a broader range of circumstances. This brings us back to the nature versus nurture debate. The research findings we have reviewed in this chapter are difficult to reconcile on a usage-based approach to language acquisition because the usage-based approach lacks the kinds of core principles, such as downward entailment, which underpin disparate linguistic phenomena that appear systematically in individual languages, and across human languages. The findings provide support for a theory of Universal Grammar that includes a computational system that expresses deep seated semantic regularities of the kinds anticipated on Chierchia’s model, and which help explain children’s rapid and uniform mastery of human languages

One last point. Nearly everyone admits that there are linguistic universals in some sense. At issue is whether the universals of human languages are specific to language, or whether cross-linguistic generalizations simply owe to the fact that humans are born with the “same basic conceptual apparatus” (GoldbergReference Goldberg2003: 16). In fact, there are compelling reasons for thinking that the interpretations generated by logical expressions in human languages are governed by principles that are specific to language, and are not simply a system of inferences that any rational system would make. Much previous research in child and adult language, and in the literature on reasoning, has emphasized differences between the meanings of expressions in classical logic and in human languages. Although there are some clear differences, we have offered some reasons to suppose that logic and human languages share some of the same basic meanings, including the meaning of disjunction. We began this chapter by citing some differences of opinion about the meaning of disjunction in human languages, as well as some observations about the nature of the input to children. Many scholars have reached the conclusion that disjunction is exclusive-or in human languages; the input to children is certainly consistent with this conclusion. However, the finding is that children, across languages, prefer to interpret disjunction as inclusive-or even in contexts in which adults prefer to interpret it as exclusive-or. This finding invites the inference that the inclusive-or meaning of disjunction is a contingent a priori fact about the minds/brains of human beings, and not a fact that is acquired by observing how others behave, or the meaning that any rational creature would adopt (for further discussion, see CrainReference Crain2012).

Acknowledgments

The research reported in this chapter was funded by an Australian Research Council (ARC) Discovery Grant (DP0879842) and further supported by the Australian Research Council Centre of Excellence for Cognition and its Disorders (CE110001021). We would also like to acknowledge three anonymous reviewers whose thorough and very helpful comments made this a much better paper.

1. Introduction

In this chapter, we review some works on the acquisition of adjectives and present novel experimental data on their interpretation. We interpret these data as suggesting that younger children start by interpreting relative gradable adjectives (GAs henceforth) like tall in a categorical way, i.e., as referring to sets of objects, and only at a later stage they switch to the comparative-like interpretation. We propose that this evolutionary trend can be easily explained within a semantic framework that assumes that relative GAs denote a partial function from individuals to truth-values. We further suggest a parallelism with the phenomenon of scalar implicature computation in children, a phenomenon that has been studied extensively by Gennaro Chierchia, whose valuable contribution extends from semantic theory (Chierchia Reference Chierchia2006; Chierchia et al. Reference Chierchia, Fox, Spector, Maienborn, von Heusinger and Portner2012) to experimental investigation (Chierchia et al. Reference Chierchia1998; Chierchia et al. Reference Chierchia, Crain, Guasti, Gualmini, Meroni, Do, Domínguez and Johansen2001; Chierchia et al. Reference Chierchia, Guasti, Gualmini, Meroni, Crain, Foppolo, Noveck and Sperber2004; Foppoloet al. Reference Foppolo, Guasti and Chierchia2012; Panizzaet al. Reference Panizza, Chierchia and Clifton2009). This chapter is organized as follows: after introducing the well-known distinction between intersective and relative adjectives (Section 1.1), we review the major findings in the literature on the acquisition of adjectives (Section 1.2); we then summarize the two major theoretical approaches that have been put forth to analyze the meaning of relative GAs, i.e., the degree-based analysis and the partial function account (Section 1.3). In Section 2, we present the experimental data of two studies and, in the final section, we discuss these results within the partial function approach.

2. Testing GAs: non adult-like children and children-like adults

In a first experiment we investigated the interpretation of relative and absolute GAs in children and adults with the aim of attesting a difference within the class of GAs and, eventually, to identify a developmental trend between younger and older children in the comprehension of these adjectives (cf. also Foppolo and Panzeri, in press). Relative and absolute GAs differ in the way the standard of comparison is set: relative GAs necessitate the retrieval of a contextual standard, whereas the interpretation of absolute GAs does not vary from situation to situation, since the standard is identified with the absolute GA’s upper or lower boundary (i.e., its maximum or minimum standard). In our first study, we aimed at detecting whether children are aware of this difference in the first place,⁶ and how they react when they are asked to evaluate a description such as “This is relative GA” when the standard of comparison is not retrievable from the context. In a second study, we employed similar material to test adults when considerations about informativeness were suspended.

2.1 Experiment 1: a developmental study on the interpretation of GAs

2.1.1 Participants, procedure, and materials of Experiment 1

We tested 3 and 5-year-old Italian speaking children (20 3-year-olds (3;0–3;11, MA=3;6) and 18 5-year-olds (5;0–6;1, MA=5;4)) and 23 adults that volunteered to participate. Participants were randomly assigned to one of two lists.

In a single experimental session, we administered a Truth Value Judgment Task followed by a Scalar Judgment Task. The participants were tested by two experimenters, one that manipulated a puppet and the other that showed the experimental objects and asked questions. As for the Truth Value Judgment Task, the procedure was as follows: the experimenter put a single object on the table in front of the participant. The puppet described that object by using an adjective (e.g., “This is tall”), and the participant had to evaluate the puppet’s description by using one of three possible responses: “yes, correct,” “no, incorrect,” or “I can’t tell, it depends.” In order to familiarize the participants with this procedure, they were first given a training session in which different toys were presented and described by the puppet. In particular, a toy boy with a big hat was shown in the training session and described by the puppet as “This is blond.” Given that the big hat covered the whole head of the boy, it was impossible to see what color his hair was. Thus, the response “I can’t tell” was prompted.⁷

The second part of our experiment consisted of a Scalar Judgment Task like the one used by Syrett (Reference Syrett2007); see also Syrett et al. (Reference Syrett, Kennedy and Lidz2010). In this task, the participants were presented with a series of seven objects and, for each of these items, they were asked “Is this adjective?” An option between “yes” and “no” was given in this case.

The experimental material for the Truth Value Judgment Task consisted of 6 pairs of positive/negative relative GA antonyms (wide/narrow, heavy/light, fast/slow, big/small, tall/short(height), long/short(length)⁸ and 6 pairs of absolute GA antonyms, classified as maximum standard (clean, straight, dry, smooth, closed, full, empty) and minimum standard adjectives (dirty, bent, wet, rough, open). Critical trials were interspersed with controls for a total of 24 items for each list. Besides the controls that prompted a genuine “yes” or “no” response (for example, participants were shown a blue object described as “This is yellow”), we inserted two trials that should prompt participants to answer “I don’t know.” In particular, a toy zebra was described as “This is obedient” and a toy man was described as “This is married.”

In the Truth Value Judgment Task, we tested adjectives with respect to single objects that were as abstract as possible, i.e., not directly recognizable as having a specific function, so that they should not evoke any comparison class.⁹ Each object was extracted from an ideal scale of seven objects decreasing along a relevant dimension. For example, for the scale of height we used seven different wooden rods ranging from the tallest (20 cm tall) to the shortest one (5 cm tall), with intervals of 2.5 cm between two consecutive elements in the series. We presented one object in isolation in the Truth Value Judgment Task and the whole series in the Scalar Judgment Task. Here is a sample of the scales used in the Scalar Judgment Task for tall/short (Fig.10.1), straight/bent (Fig. 10.2), full/empty (Figs. 10.3 and 10.4), with the indication of which element(s) were used in isolation in the first part of the experiment.

Figures 10.1–10.4 Sample of the material used to test relative and absolute GAs in the Truth Value Judgment Task and in the Scalar Judgment Task.

Figure 10.2 (cont.)

Figure 10.3 (cont.)

Figure 10.4 (cont.)

We tested antonyms for relative GAs (e.g., tall/short) between lists with respect to the same type of object (e.g., a vertical wooden rod as in Fig. 10.1), but using different ends of the scale: in particular, we used the 2nd element from the left in the series of seven to test the positive antonym and the 6th element of the same series to test the negative antonym. We tested antonyms for the absolute GAs in which an abstract object was used between lists by using the same object in the series of seven: for example, the 2nd element from the left in Fig. 10.2 was used in List 2 to test straight, which is a maximum standard adjective, and in List 1 to test bent, which is its minimum standard counterpart. Finally, we tested full/empty and open/closed within lists but using different types of objects. In particular, two different series of objects were created for each pair of antonyms: seven bottles and seven paint-tubes ranging from being completely full to completely empty (Figs. 10.3–10.4), and seven purses with a zipper and a box with a lid ranging from being completely open to completely closed. As for the Scalar Judgment Task, a total of 9 scales were tested. Of these, 3 were series of “abstract” objects that were used to test couples of relative GAs antonyms (big/small, long/short, wide/narrow, tall/short), 2 were used to test absolute GAs with abstract objects (clean/dirty, straight/bent), and 3 were used to test full/empty and open/closed. Again, antonyms were tested between lists, while the order of presentation was balanced across participants and lists: half of the participants in each list were shown a relative GA first, the other half encountered full as first item.¹⁰ In testing antonyms across lists, we always started from the most representative element in the series, which was crucially reversed across conditions: e.g., we started from the straightest rope when asking “Is this straight?” (1st element from the left in Fig. 10.2) and from the most bent rope when asking “Is this bent?” (7th element from the left in Fig. 10.2).

2.1.2 Results of Experiment 1

As far as the controls are concerned, participants in each age group behaved almost at ceiling when asked to evaluate descriptions that were genuinely true or false with respect to the item shown. In case of the controls for which an “I don’t know” response was expected, adults behaved as expected and almost always used this option (96%). Interestingly, though, the percentage of children that did so decreased as a function of age: while 40% of the 5-year-olds answered “I don’t know” in these trials, only one of the 3-year-old children selected this option. Instead, the group of the younger children split between a “yes” (50%) and a “no” answer (48%).

Turning to the experimental trials, the distribution of the three types of answers (yes; no; I don’t know) used by the three age groups (children aged 3 and 5; adults) is plotted in Fig. 10.5, differentiated by type of GA: relative GAs¹¹ (Rel); minimum standard absolute GAs (Abs Min); maximum standard absolute GAs (Abs Max):¹²

Figure 10.5 Answers’ distribution per Age of participants and Type of GA.

We can evince some very broad generalizations from the graph. First of all, adults’ performance was at ceiling in case of absolute GAs: adults always accepted “This is minimum standard absolute GA” (e.g., “This is bent”) when the target object possessed the GA property to a degree that exceeded the minimum standard degree (e.g., the minimally bent rope in Fig. 10.2) and they nearly always rejected the description “This is maximum standard absolute GA” (e.g., “This is straight”) when the target object did not have the GA property to its maximum degree (e.g., the same minimally bent rope in Fig. 10.2). This behavior was predicted on the basis on Kennedy and McNally (2005). In case of relative GAs, adults answered “I don’t know” 60% of the time, and they accepted the description only 30% of the time. A high percentage of “I don’t know” answers was predicted on the basis of the fact that we did not provide a contextual standard of comparison, and thus the description “This is relative GA” was not interpretable. On the basis of adults’ performance, we can conclude that our experimental design can serve as a basis for distinguishing, experimentally, between these two classes of adjectives.

Turning to children, a difference emerged between 3 and 5-year-olds with respect to relative and maximum standard absolute GAs especially. First of all, in the case of absolute GAs, the performance of the 5-year-olds parallels that of adults: children aged 5 consistently rejected the description “This is maximum standard absolute GA” referred to an object that did not possess the property denoted by the adjective at a maximum degree (e.g. straight referred to the minimally bent rope), and consistently accepted the description “This is minimum standard absolute GA” even when referred to an object that possessed the property denoted by the adjective to a minimum degree (e.g., bent referred to the minimally bent rope). However, while the 3-year-old children did not differ from the older children and adults in this latter case (e.g., they accepted bent referred to a minimally bent rope), they also accepted the description “This is maximum standard absolute GA” referred to an object that did not possess the property denoted by the adjective at a maximum degree (e.g., they accepted straight referred to the minimally bent rope) more than the adults and the 5-year-olds. In the case of relative GAs, a developmental trend seems to emerge: while children at age 3 showed a prevalence of “yes” answers and adults a prevalence of “I don’t know” answers, the distribution of the 5-year-old children’s answers seems to lay in between.

Statistical comparisons by means of Fisher Exact Tests for count data revealed a significant general effect of Age in case of relative GAs (p<0.001). A significant difference was attested in paired comparisons between the 3 and the 5-year-olds, between the 3-year-olds and the adults and between the 5-year-olds and the adults (all p<0.001). However, this effect might be partly due to the fact that children, in general, used the option “I don’t know” much less than adults. Considering the incidence of “yes” answers over the total in case of relative GAs, the difference between the 3 and the 5-year-olds disappears (p=0.29), while it does not comparing the 5-year-olds and the adults (p<0.01). In case of minimum standard absolute GAs, a significant difference was observed between children aged 3 and adults (p<0.05), but not between children of age 3 and 5 (p=0.14), nor between children aged 5 and adults (p=0.53). In case of maximum standard absolute GAs, a significant general effect of Age was revealed instead (p<0.001). This difference was probably due to the fact that, differently from the older children and the adults, a lot of the 3-year-olds accepted the description “This is maximum standard absolute GA” referred to an object that did not possess the property denoted by the adjective at a maximum degree (all ps<0.001). No statistical difference is revealed for maximum standard absolute GAs between the 5-year-olds and the adults instead (p=0.11). Further statistical comparisons revealed that, like the adults, children at age 3 and 5 distinguished between relative and absolute GAs (minimum and maximum standard considered separately, all ps<0.001), while children at age 3 distinguished relative GAs from maximum standard absolute GAs (p<0.001) but not from minimum standard absolute GAs (p=0.29). This was probably due to the fact that, while a “yes” answer was expected in case of minimum standard absolute GAs, the proportion of “yes” answers in the 3-year-old children was higher than expected in the case of relative GAs (more than 85%).

To better grasp children’s behavior, we will first focus on their responses to relative GAs and then we will compare children’s and adults’ performance in the Scalar Judgment Task. As for children’s behavior with relative GAs, we plotted in Fig. 10.6 the distribution of participants’ “yes” answers (differentiated by Age) for each of the relative adjectives tested.

Figure 10.6 Relative-GAs: proportion of “yes” answers per Age and Adjective.

A very clear picture emerges: first of all, the incidence of “yes” responses in general is very high in children, in the younger group especially, where the rate of “yes” answers is always over 70%, reaching peaks of 90% and 100% in some cases. In the second place, except for the first two pairs of antonyms (fast/slow, heavy/light), the 5-year-old children seem to pattern more like the 3-year-olds than the adults, providing more than 60% of “yes” answers to the adjectives that presumably they are most familiar with (big/small, tall/short, long/short), and, most importantly, not differentiating between antonyms. We will get back to this result in the discussion sessions of this paper.

Finally, we summarized the results of the Scalar Judgment Task in the graph in Fig. 10.7, in which the percentage of “yes” answers obtained for each of the seven items in the series are plotted for each age group separately, differentiating between antonyms. We report the percentage of “yes” answers by adults (solid line), 5-year-old children (dashed line), and 3-year-old children (dotted line) in the Scalar Judgment Task (SJT). Maximum standard absolute GAs, indicated as ABS (max. std.), and positive relative GAs, indicated as REL (pos.), are in the top panels. Minimum standard absolute GAs, indicated as ABS (min. std.), and negative relative GAs, indicated as REL (neg.), are in the bottom panels.

Figure 10.7 Percentage of “yes” by adults (solid line), 5-year-old children (dashed line) and 3-year-old children (dotted line) in the SJT, differentiated between max-std abs-GAs, positive rel-GAs (top panels), min-std abs- GAs and negative rel-GAs (bottom panels).

As it is evident from the graphs, the pattern of response differs between the two conditions (relative vs. absolute) for all age groups: in general, participants were much more categorical in the absolute GAs condition (left panels)¹³, independently of age. In particular, both adults and children drastically dropped their “yes” answers on item #2 in case of maximum standard absolute GAs but not in case of relative GAs, as expected and as already found by Syrett (2007): in case of relative GAs (right panels), adults’ “yes” responses constantly decreased and dropped below 50% acceptance only on the 4th item and the same pattern is observable in children, even if the curve is smoother for the 5-year-old children, and even smoother for the 3-year-olds (cf. Foppolo and Panzeri in press for more details about the statistical comparisons for the Scalar Judgment Task).

2.4 Experiment 2: turning adults into children

In a second study, we aim at prompting adults to accept a description like “This is relative GA” applied to an abstract object in isolation (i.e., when the standard of comparison was not retrievable from the context) by manipulating the experimental setting so as to instruct participants that informativeness was not at stake.

2.4.1 Participants, procedure and materials of Experiment 2

A total of 73 Italian adults were tested, and randomly assigned to one of two lists. These were undergraduate students at the University of Milano-Bicocca that volunteered to participate.

The materials and procedure used was similar to the one used in the Truth Value Judgment Task in the first experiment, except for some crucial manipulations that we will describe in detail. Differently from the former study that involved children, in which the objects were physically presented to participants, items in this experiment were recorded in a sequence with a digital video camera and then were shown on a video. As before, two experimenters were involved: one introduced the study and provided instructions, and the other manipulated a puppet. As in the other study, the puppet provided a description for each of the items presented by using an adjective (e.g., “This is tall”) and the participants were asked to express their judgment on a pre-compiled answer sheet. In this case, though, adults were given only two options of response: “yes” and “no.” This solution was dictated by the fact that children in the first study did not use the option “I don’t know.”

First, we manipulated the procedure so that a training session was added, in which adults were asked to be as tolerant as they could with respect to the description provided by the puppet. To make sure that the participants understood what we meant by “being tolerant,” we provided them with the following scenario. The puppet was introduced as being an alien on Earth. He was sent from his planet to test some hypotheses about the functioning of the languages spoken on Earth that his population had developed on the basis of fragments of conversations that they have been receiving from our planet. The participants were told that their aim was that of helping this alien in refining his hypotheses about Italian. In particular, they were instructed to be cautious in rejecting descriptions that, though not being optimal, could nevertheless apply to the situations presented. To this end, the participants were provided with some examples, to help them understand what they needed to do to solve this delicate task. In the first example, the puppet was asked to describe a situation in which a toy boy has four coins in his wallet. The puppet made a first guess by saying “The boy has seven coins.” The experimenter warned the puppet that this description was not correct, and instructed the participants to straightforwardly reject incorrect descriptions of this sort. Then the puppet made a second guess by saying “The boy has three coins.” At this point, the experimenter acknowledged the fact that the alien’s description was not optimal, but at the same time she warned the participants that they might confuse the alien and prevent him from converging on the target language if they told him that “The boy has three coins” cannot be used to describe a situation in which the boy has four coins. To this aim, the experimenter suggested that the participants imagine a situation in which the toy boy wanted to buy an ice cream, and that the ice cream cost three coins. If the alien was told that it is not correct to say that the boy has three coins when he has four, then he would conclude that the boy could not buy an ice cream, which in fact is not true. By using a similar strategy, the experimenter instructed the participants to be cautious in rejecting the description “The boy pulled some of the flowers” in a situation in which the boy pulled all the flowers. By suggesting the scenario in which the boy was told “If you pull some of the flowers, you’ll be punished,” the experimenter aimed to convince the participants that the alien could be brought to an erroneous conclusion in this case too, if they were too strict in their evaluation. As before, if he were told that the description with some cannot be used in a situation in which all is at stake he will, erroneously, conclude that the boy wouldn’t merit a punishment in case he pulled all the flowers.

As before, different classes of adjective were tested within subjects while antonyms of the same pair were presented between lists. Each list comprised a total of 29 items. Of these, 15 were controls: 6 were clearly true, 2 were clearly false and the remaining 7 were cases in which the adjective used was not applicable to the object, that should prompt a “no” response by the participants. For example, we tested adults with the description “This is vegetarian” referred to a toy plane and expected them to reject it on the basis that the property predicated by the adjective (i.e., “being vegetarian”) cannot apply, under any circumstance, to a plane.

Of the critical items, 4 were pairs of absolute GA antonyms (clean/dirty, straight/bent, open/closed, and full/empty) and 4 were pairs of relative GAs (fast/slow, big/small, tall/short, long/short). These were tested using the same objects used in the former study, but with one crucial difference: in this study we did not pick objects from the different ends of the series to test antonyms. In particular, we always used the 6th element in the series, the one that was closer to the negative pole of the scale: for example, we used the sixth rod in Fig. 10.1 to test both tall and short.¹⁴ This manipulation was dictated by the aim of this study: we wanted to prompt a “yes” answer from adults whenever the adjective could apply to a certain object. At the same time, though, we believed that it was too “easy” to prompt a yes answer to positive antonyms by using the element in the series that was the closest to the positive pole. In order to give more strength to our claim, we thus tried to make the task harder for the participants, who had to judge tall with respect to a pretty short item.

2.4.2 Predictions

We predicted, in the first place, an increase of “yes” answers in case of relative GAs and maximum standard absolute GAs in this study compared to our previous study. For example, we expected more adults to accept the description “This is tall” referred to the short rod (item #6 in Fig. 10.1) on the hypothesis that the property predicated by the adjective (i.e., tallness) applies to any object that has some vertical dimension. Similarly, we expected more adults to accept the description “This is straight” referred to a minimally bent rope (item #2 in Fig. 10.2) on the hypothesis that the property predicated by the adjective (i.e., straightness) applies to any object that has some degree of straightness, even if this doesn’t reach the maximum standard on the scale of straightness. At the same time, though, we expected adults to be tolerant with critical cases only and to reject the descriptions in case it was obviously false or when the adjective was not applicable to the object.

2.4.3 Results of experiment 2

Our results confirmed our predictions. These are summarized in Fig. 10.8:

First of all, adults’ performance with controls was at ceiling. In particular, they did not hesitate in rejecting a description when the adjective used was not applicable to the object (indicated as Contr-NA in Fig. 8) or when the description was false (indicated as Contr-False in Fig. 10.8). As we said, these controls were crucial for our experiment. As far as GAs are concerned, these gained a high percentage of “yes” answers from adults in all the categories considered: not only did the participants accept the description “This is minimum standard absolute GA” when the target object possessed the GA property to a degree that exceeded the minimum standard; half of the time, they also accepted the description “This is maximum standard absolute GA” referred to the same object. In case of relative GAs, adults accepted the description more than 70% of the time despite the fact that, as in the first experiment, no contextual or perceptual standard of comparison was provided.

Figure 10.8 Distribution of adult’s responses in Experiment 2.

Plotting the proportion of adults’ “yes” responses in the two studies¹⁵, the picture that emerged is shown in Fig. 10.9

Figure 10.9 Distribution of adult’s responses in Experiment 1 and 2.

For what concerns relative and maximum standard absolute GAs, the increase of “yes” answers in this second experiment is significant, even taking into consideration the fact that in the second study the probability to answer “yes” was 0.50, while it was 0.33 in the first experiment, where three options of response were given (χ²(2)=85.35, p<0.0001 for relative GAs; χ²(2)=14.19, p<0.001 for maximum standard absolute GAs). In fact, for both types of adjective, the proportion of “yes” answers in the second study was different from chance (χ²(1)=46.21, p<0.0001 for relative GAs; χ²(1)=7.27, p<0.01 for maximum standard absolute GAs). This result is all the more surprising if we consider the fact that, differently from the previous study, here the “positive” relative adjective was referred to an object that had the property predicated by the adjective at a minimal degree (e.g., tall in this study was predicated of item #6 from the left in Fig. 10.1, while it was predicated of item #2 in the previous study) and the maximum standard absolute GAs full was referred to an object that had the property predicated by the adjective at a minimal degree (e.g., it was predicated of the almost empty bottle or paint tube in this study, while it was predicated of the almost full object in the previous study). Plotting adults’ performance in the second study with children’s performance in the first one, the picture that we get¹⁶ is shown in Fig. 10.10.

Figure 10.10 Proportion of “yes” answers by children and adults across studies, differentiated by Adjective type: relative (left), max-std abs-GAs (centre), min-std abs-GAs (right).

As it is evident from the graph, adults’ performance in the case of relative and maximum standard absolute GAs did not differ from the younger children. Statistical comparisons by means of Fisher exact test confirmed that: (i) in the case of relative GAs, adults did not differ from 3-year-old and 5-year-old children (χ²(1)=2.27, p=0.13; χ²(1)=1.80, p=0.18 respectively); (ii) in the case of maximum standard absolute GAs, adults did not differ from the 3-year-olds (χ²(1)=0.83, p=0.36) but, interestingly, they differed from the 5-year-olds (χ²(1)=7.38, p<0.01).

2.4.4 Discussion of Experiment 2

In this experiment, we aimed at turning adults into (young) children by banning informativeness considerations from the judgments of relative GA descriptions. We did so by training adults to be tolerant with respect to underinformative statements. For example, we asked them not to straightforwardly reject a description like “The boy pulled some of the flowers” in a situation in which the boy pulled all the flowers in consideration of the fact that, though not fully felicitous, that description was not logically wrong. This manipulation seemed to work, in the end: adults’ acceptance rate of relative GA descriptions rose above 70% in Experiment 2. We interpreted this result as compatible with our hypothesis that relative GAs can be assigned a minimal, nominal-like interpretation when a comparison class or standard is not available in the context.

Within a degree-based semantic framework, one could explain adults’ behavior in this second study by assuming that adults are in fact quantifying over possible contexts, and eventually, over possible standards of comparison. Under this assumption, adults would interpret “This is relative GA” not as meaning “This object can fall under the extension of GA” (as we maintain), but as “There could be a context c that sets a standard of comparison d_c such that this object is more GA than d_c.” We believe that, if this were the case, this would imply that, in principle, it would be possible to interpret relative GAs as if they were equivalent to minimum standard absolute GAs, i.e., as if an object is relative GA if it has at least a minimum degree of relative GA. We believe that this assumption would be less parsimonious because it would require two distinct interpretations for GAs: one in terms of degrees as abstract entities ordered along a scale (and, as KennedyReference Kennedy2007 extensively argued, in this interpretation the scale of degrees evoked by relative GAs does not have a lower boundary that could constitute the minimum standard); the other interpretation would allow for the identification of a minimum standard for being relative GA, either by identifying an abstract degree that could serve as a lower boundary of the scale, or by assuming that relative GAs also evoke scales of individuals that are ordered along the relative GA dimension.

3. Conclusions

Our first study showed that 3-year-olds had a very high acceptance rate for relative GAs when the context did not supply a standard of comparison, and had a higher acceptance rate of maximum standard absolute GAs compared to adults’ and 5-year-olds’. To explain this result, we hypothesized that, when they first encounter adjectives, children start by interpreting them as if they were common nouns, i.e., as denoting sets of individuals that share a property. Under this hypothesis, they would start by interpreting “S is long” as “S has a horizontal dimension” and “S is tall” as “S has a vertical dimension.” This assumption is in line with the conceptual biases that have been suggested to explain infants’ first acquisition of words (see MarkmanReference Markman1990) and with a series of studies that have further shown that even 4-year-olds tend to interpret a novel word that is distributionally and morphologically presented as an adjective as if it were a common noun, i.e., as referring to stable properties of objects (see Hallet al. Reference Hall, Waxman and Hurwitz1993; Klibanoff and WaxmanReference Klibanoff and Waxman2000; Mintz and GleitmanReference Mintz and Gleitman2002; Waxman and Guasti, Reference Waxman and Guasti2009). Moreover, we think that this hypothesis would also account for children’s systematic errors reported in the literature, substitution errors in particular, and for the fact that younger children did not seem to treat relative antonyms differently in our Experiment 1 (i.e., they accepted “This is long” as much as “This is short”).

We further maintained that 3-year-old children are still in the process of switching from the nominal to the comparative interpretation. This would explain why they perform well in selection tasks in which they are asked to choose (or label) the unique (more) relative GA of a group on the one hand, and why they exhibit extreme labeling on the other, i.e., they attribute the relative GA property only to the extreme of a series. We proposed that this behavior ensues from the fact that children haven’t reached the adult-like stage in the ability to retrieve the standard of comparison yet and thus stick to the nominal-like meaning first, when such a standard is not directly available. Building on the assumption that this ability undergoes development, we believe that a parallelism with the phenomenon of scalar implicature is worth exploring at this point. As it has been attested in different studies (cf. Foppoloet al. Reference Foppolo, Guasti and Chierchia2012 for a recent overview), children tend to accept underinformative sentences like “Some Ss are P” in a situation in which “All Ss are P” would be more appropriate. Nonetheless, children have been shown to know that the description with “all” would be pragmatically more appropriate, and indeed choose that in a Felicity Judgment task (a.o., cf. Experiment 5 in Foppolo et al. Reference Foppolo, Guasti and Chierchia2012). Also, they have shown sensitivity to pragmatic infelicity despite the fact that they were found to be more tolerant than adults in case of underinformative statements (Katsos and BishopReference Katsos and Bishop2011). Analogously, we can view 3-year-old children’s behavior with relative GAs as a hint that they are assigning them a minimal, nominal-like interpretation in the absence of a perceptual or a normative standard to exploit.

In Experiment 2 we tested whether this minimal, nominal-like interpretation was also detectable in adults. To this end, we tried to turn adults into children by manipulating the experimental setting so as to instruct them to be tolerant and accept underinformative descriptions, as long as they were true. In fact, the experimental manipulation was effective in prompting children-like responses in adults: adults accepted descriptions of the form “This is relative GA” in situations in which informativeness was left aside. We argue that, in such a situation, adults behaved as children resorting to the nominal interpretation of relative GAs.

At this point, we would like to speculate about how the semantic analysis of GAs could account for these facts. The main difference between the degree-based approach and the partial function perspective lies in the fact that the former assumes the existence of degrees as primitive entities, and associates scales of these abstract entities to GAs. According to this view, the GA tall evokes a scale of degrees of height. In order to understand a sentence containing a relative GA in its bare form, one needs first to evoke the scale of degrees associated to the relative GA dimension, then to identify the intended standard of comparison (by making reference to the normative class or perceptual cues for instance) and, finally, to compare the degree to which an individual possesses the relative GA to that standard. As we have seen, younger children accept “This is relative GA” even in the absence of a contextual standard of comparison.

The degree-based approach should assume that, in the absence of contextual cues, younger children posit a minimum standard of comparison for relative GAs (“S is tall” would mean “S has at least a minimum degree of height”), and thus accept the description. Even if this hypothesis could explain the results of the Truth Value Judgment Task, it would clash with the reports in the acquisition literature, and from the Scalar Judgment Task in Experiment 1: when presented with a series of objects, younger children tend to label only the extremes of a series, and this suggests that they posit the standard of comparison to the upper (and not lower) boundary of the series. Recall that in our first study, the 3-year-olds considered the same object to be big or tall in the Truth Value Judgment Task with a percentage of 85%; that dropped to 60% in the Scalar Judgment Task when the same object was shown as part of a series of seven.¹⁷

In the partial function analysis, on the other hand, the core meaning attributed to relative GAs is the same as that of common nouns: a function from individuals to truth-values, that is, a way to identify sets of individuals that share a property. The source of context-dependency of relative GAs lies in the identification of the intended domain of application of the function, and their vagueness is located in the partiality of the function, i.e., in the fact that for some individuals one could be unsure whether they belong to the extension of the relative GA predicate. The hypothesis that children start by interpreting GAs nominally can easily meet a partial function approach: they would first guess that GAs behave as common nouns, denoting total functions that identify sets of individuals, those that can in principle fall under the extension of the GA. The comparative interpretation, on the other hand, requires them to understand that different choices of the domain of application can lead to different outputs of the function, and this, in turn, would make the comparative interpretation more informative than the nominal one.

We believe that the 3-year-old children tested in Experiment 1 were in a phase in which they were switching to the comparative interpretation: when the intended normative class was not provided, as in our Truth Value Judgment Task in which the standard of comparison was not retrievable from the context, children would resort to the minimal, nominal-like interpretation (and behaved differently from the adults); when the normative class was explicitly presented to them, e.g. when a series of seven objects was given, as in the Scalar Judgment Task, they could switch to the comparative interpretation instead. Even in this case, though, they still had problems in partitioning the restricted domains, thus exhibiting extreme labeling, as attested also in the literature on the acquisition of adjectives summarized in the introduction. Observing the developmental trend in children in Experiment 1, we suggested a parallelism with the phenomenon of scalar implicature computation in children. In case of scalar implicatures, younger children are shown to be aware of the different informativeness status of scalar alternatives like “Some Ss are P” and “All Ss are P.” Nonetheless, they have difficulties in retrieving the strong alternative with all (or rejecting the weak one with some) when the underinformative statements are given in isolation, as happens in a classical Sentence Evaluation task (see Foppoloet al. Reference Foppolo, Guasti and Chierchia2012 for a recent overview). Conversely, when the scalar alternatives are provided, as in the Felicity Judgment task, children have no difficulty in selecting the more appropriate description. Also, a recent study shows that children switch to the strengthened meaning of underinformative statements when they are made aware of the ambiguity associated with some (cf. Foppolo et al. Reference Foppolo, Guasti and Chierchia2012, Experiment 6). This seems analogous to what happens in case of relative GAs: children access the comparative reading of relative GAs when the comparison class is explicit (as in the studies in which a perceptual or a normative class is presented, or in the Scalar Judgment Task), while they accept the weaker, nominal reading of the relative GAs whenever the class of comparison, whether contextual, perceptual, or normative, is not explicitly provided.

Acknowledgments

We are very honored and pleased to contribute to this volume that celebrates Gennaro Chierchia. We would like to offer Gennaro our paper as a warm hug for his birthday, in return for the many he has given us over the years. We also thank two anonymous reviewers for their valuable feedback and Ivano Caponigro and Carlo Cecchetto for their editorial help. For academic purposes, Francesca Panzeri is responsible for Sections 1.1, 1.3, 2.3, and 3; Francesca Foppolo for Sections 2.1, 2.2, and 2.4; and Maria Teresa Guasti for Section 1.2.

1. Introduction. The issue

It is a notorious fact that A′-dependencies with object gaps are mastered later in development than subject dependencies. In previous work in collaboration with Naama Friedmann (Friedmann et al. 2009) we have traced back this property to a particular implementation of the syntactic locality principle known as Relativized Minimality (RizziReference Rizzi1990, Reference Rizzi and Belletti2004): Object A′-dependencies inevitably cross the subject position, giving rise to an intervention configuration which is hard to compute for the child’s system. This approach is supported by a rich array of experimental evidence showing that the difficulty that children experience with object A′-dependencies is modulated by the featural specification of the moved object and the intervening subject. The quoted chapter briefly hints at the possibility that the same explanatory scheme could be extended to account for the fact, familiar from the adult psycholinguistic experimentation, that object A′-dependencies are harder to process than subject dependencies also for adults. In the present chapter we would like to develop this insight further by reinterpreting some substantive results of psycholinguistic research (Gordonet al. Reference Gordon, Hendrick and Johnson2001, Reference Gordon, Hendrick and Johnson2004; Warren and GibsonReference Warren and Gibson2002, Reference Warren and Gibson2005) in terms of our syntactic Relativized Minimality approach. This will lead us to propose a featural typology of nominal expressions inspired in part by Chierchi’s influential work on nominals at the syntax–semantics interface (Chierchia Reference Chierchia1998). We intend to show that a combination of the Relativized Minimality approach and a properly defined featural typology of nominal expressions provides a fine-grained system capable of predicting subtle gradations in intervention effects, such as those uncovered in the experimental work both with children and adults. The system thus provides a basis for a unified account of formal syntactic generalizations (e.g., the selectivity of weak islands), developmental facts, and results of real-time psycholinguistic experimentation in the adult.

2. Background

3. On some relevant psycholinguistic evidence

Object relatives such as (7a) are more difficult to process than subject relatives like (7b) for adult speakers:

This type of contrast has been highlighted through very different psycholinguistic techniques, from phoneme monitoring, to response accuracy to probe questions, to self-paced reading times (Frauenfelderet al. Reference Frauenfelder, Mehler and Segui1980; King and JustReference King and Just1991; Waterset al. Reference Waters, Caplan, Hildebrandt and Coltheart1987, etc.). A central idea in classical linguistically inspired psycholinguistic approaches is that the extra-complexity observable in object dependencies is linked to a structural cause (such as De Vincenzi’sReference De Vincenzi1991 Minimal Chain Principle). The analysis we intend to explore closely follows this tradition.

We concentrate here on two major recently developed approaches, which we will refer to as:

1. A. The similarity-based approach (Gordon et al. Reference Gordon, Hendrick and Johnson2001, Reference Gordon, Hendrick and Johnson2004)

   A dependency is hard to compute when an element intervenes which is similar to the target of the dependency.

2. B. The increased referential processing approach (Warren and GibsonReference Warren and Gibson2002, Reference Warren and Gibson2005)

   A dependency is hard to integrate when, in the course of the computation of the dependency, elements occur which require substantial referential processing.

The point of departure of the first approach is the observation due to Bever (Reference Bever and Cohen1974) according to which double center embedded relative clauses which are completely unparsable when all noun phrases correspond to a lexical description (e.g., the reporter), become much easier to understand if the noun phrases are sufficiently different, and do not all correspond to lexical descriptions. The contrast presented in the references quoted is given in (8) vs. (9):

Bever’s original idea was that somehow a crucial role in the amelioration should be attributed to the difference among the three noun phrases, which should be submitted to different types of “rules” in the assignment of their grammatical roles. In other words, the embryo of a similarity-based type of account. The specific version that Gordon et al. (Reference Gordon, Hendrick and Johnson2001) assume is in terms of storage in working memory during processing: in object relatives, two noun phrases (the relative head and the subject) must be stored in working memory before the A′-dependency is resolved and the appropriate thematic roles are assigned, and this temporary storage is facilitated if the two noun phrases are sufficiently different along dimensions to be made precise. Gordon et al. (Reference Gordon, Hendrick and Johnson2001) contrast an approach inspired by Bever’s idea with Gibson’s (Reference Gibson1998) original account of the processing of complex sentences, the source of Warren and Gibson (Reference Warren and Gibson2002, Reference Warren and Gibson2005, to be discussed below), according to which the difficulty involved in dependency resolution is a function of the processing cost for the assignment of discourse referents to nominals occurring while the dependency is computed.

Let us discuss the two approaches in turn. A fundamental insight of Gordon et al. (Reference Gordon, Hendrick and Johnson2001) has been to manipulate the intervening subject in object relatives and test how this could affect processing. In a series of reading time and comprehension experiments devoted to the comparison between the processing of subject vs. object relative clauses, the authors have shown that in object relative clauses whose head is a lexically restricted noun phrase in our terminology, manipulating the nature of the subject of the relative clauses as in (10) has clear effects in reducing the reading time and improving the accuracy in the comprehension of the structure. In two distinct experiments the authors have compared the case of lexically restricted subjects with proper names and pronominal subjects, as in the following pairs:

What emerged from these experiments was that the intervention of the lexically restricted subject determined a difficulty both in terms of reading times at the critical words (the verb of the relative clause and the main clause verb) and in terms of accuracy of comprehension, in comparison to the case of an intervening (1st or 2nd) pronominal or proper name subject. As Gordon et al. (Reference Gordon, Hendrick and Johnson2001) pointed out, the amelioration effect brought about by a pronominal or a proper name subject is expected under a similarity-based approach of the kind inspired by Bever’s (Reference Bever and Cohen1974) original insight. The authors have also tested the processing of subject relatives under comparable experimental conditions (with a lexically restricted relative clause head and the object argument either a lexically restricted noun phrase or a pronominal or a proper name) and have found no comparable effect in reading time at the critical words or in comprehension accuracy. In line with the authors’ account, we can then draw the conclusion that similarity-based effects only arise in the case of object dependencies.

Gordon et al. (Reference Gordon, Hendrick and Johnson2001) have also tested object extracted clefts, comparing cases in which the extracted object was a lexically restricted noun phrase with cases in which it was a proper name, with the subject of the cleft sentence also varying in the same way. The relevant material in the object extracted cleft experiment is constituted by sentences like (11):

The reason for using clefts to test a further case of object extracted construction is that with clefts the comparison between the two relevant noun phrases, i.e., the object extracted phrase and the subject of the cleft sentence, can be directly performed in all conditions relevant for the similarity account, a possibility which is precluded in relative clauses since a proper name cannot be a (restrictive) relative head (except in special cases such as The Paris I love). The result of the cleft experiment is particularly revealing: the cases in which the object extracted noun phrase and the subject of the cleft sentences were both of the same type, i.e., both a proper name or both a lexically restricted noun phrase, were the hardest cases (e.g., It was John that Bill saw in the parking lot); when there was mismatch, lexically restricted/name or name/lexically restricted (e.g., It was the barber that Bill saw in the parking lot), the object extracted clefts were both read faster at the critical words and the sentences were more accurately comprehended (the difference was particularly strong in the comprehension test).

The general conclusion that Gordon et al. (Reference Gordon, Hendrick and Johnson2001) drew from their studies is that a similarity-based account which takes into consideration the nature of both relevant noun phrases – the head and the subject of the relative clause in the case of object relatives, the extracted object and the subject of the cleft sentence in the case of clefts – and their (dis-)similarity can explain their results better than competing theories. In particular, these authors argue that the similarity-based account fares better than a discourse integration account à laGibson, in which the crucial role in affecting performance is solely played by the nature of the subject of the object relative clause, or of the subject of the cleft sentence in object extracted clefts.⁴ This is so because, in a discourse integration account, the more or less complex computation of the referential properties of the subject of the relative/cleft is the only relevant factor, while the (dis-)similarity between the two noun phrases is not expected to affect the computation of an A′-dependency with an object gap.

Gordonet al.’s (Reference Gordon, Hendrick and Johnson2001) similarity approach is essentially cast in the same mould as our syntactic and developmental analysis based on Relativized Minimality, reviewed in Sections 2.1, 2.2. The crucial questions are: how is similarity determined? And, is there any added value if we can derive the (dis-)similarity effect from a formal syntactic principle, rather than solely ascribing it to the functioning of the processing system?

Before addressing these points, let us turn to the second approach, the increased referential processing approach, repeated here for convenience:

1. B. The increased referential processing approach (Warren and Gibson Reference Warren and Gibson2002, Reference Warren and Gibson2005)

   A dependency is hard to integrate when, in the course of the computation of the dependency, elements occur which require substantial referential processing.

According to this approach, an object relative like (7a) is harder to process than a subject relative like (7b) because the lexically restricted noun phrase in subject position, the senator (a kind of expression requiring substantial processing to establish its referent in the discourse context) occurs while the A′-dependency is being resolved between the head of the relative, the reporter, and the gap in the object position of attack. In (7b), nothing involving significant referential processing occurs in the dependency between the relative head and the subject gap, which is therefore easier to process.

This theory makes more refined predictions if coupled with an independent ranking of nominal expressions along a hierarchy of referential accessibility. Consider such a theory stating that first and second person pronouns are ranked as the most accessible, since they have their referents immediately given by the speech act (speaker and hearer); proper names are intermediate, and definite descriptions are the hardest in referent accessibility (the authors quote ArielReference Ariel1990; Gundelet al. Reference Gundel, Hedberg and Zacharski1993 in this connection). The approach then predicts an increasing processing difficulty to resolve the dependency in the following cases of object relatives, as Warren and Gibson (Reference Warren and Gibson2005) discuss in detail:

This prediction is borne out: Warren and Gibson (Reference Warren and Gibson2005) report a series of experiments in which sentences like (12a) are on average 30 ms faster to read at the critical word (the embedded verb attacked, at which the A′-dependency must be resolved) than sentences like (12b), which are in turn on average 28 ms faster to read at the critical word than sentences like (12c). The gradation of the reading times of attacked in these examples is, in this approach, a symptom of the increasing referential processing load required by the subject noun phrase intervening in the object dependency, hence a consequence of the increased referential processing approach, in combination with a reasonable hierarchy of referential accessibility, or givenness.

Notice that also the similarity-based approach is consistent with Warren and Gibson’s observation that (12c) involves the slowest reading time at the critical word attacked. On the other hand, as Warren and Gibson observe, the similarity-based approach does not immediately predict that reading time at the critical word is faster with the intervening pronoun (12a) than with the intervening proper name (12b). It should be noticed in this connection that the similarity-based approach can be made consistent with this observation if it can be supplemented by a precise theory of structural similarity from which it can be derived that a proper name is more similar to a definite description than a pronoun is (on this, see below). In that case, the similarity-based approach could also predict the experimental finding reported above from Warren and Gibson (Reference Warren and Gibson2005).

More experimental work has been brought to bear on adjudicating between the two hypotheses. Recall that Gordonet al. (Reference Gordon, Hendrick and Johnson2001) compared subject and object clefts with either proper names or definite descriptions in clefted and subject position; here we only report the relevant object clefts in better detail (definite description = DefD, i.e., a lexically restricted noun phrase in our terms, proper name = PropN):

The similarity-based approach predicts that (13a) and (13d), involving the intervention of a structurally identical subject in the object dependency, should be harder to process than (13b–c), involving structurally distinct elements. The prediction is borne out, as Gordon et al. (Reference Gordon, Hendrick and Johnson2001) have shown. This effect is not predicted under the increased referential processing approach, according to which the only relevant factor should be the referential accessibility of the intervening subject. So, the increased referential processing approach correctly predicts that (13a) should be harder to process than (13b), but incorrectly expects that (13c) should be harder than (13d) because the referential processing involved by the definite description the lawyer in (13c) should be higher than the referential processing cost involved by the proper name John in (13d), while the empirical evidence shows the opposite. So, this appears to be critical evidence in favor of the similarity-based approach, which makes the right prediction on (13).

Warren and Gibson (Reference Warren and Gibson2005) responded to this critique by introducing into the picture also the case of pronominals, both in clefted and in subject position. They thus considered the following sentence type, involving 9 possible combinations:

Warren and Gibson (Reference Warren and Gibson2005) studied the processing of such sentences through different online and offline measures, but the most relevant result for the comparison between the two hypotheses is that the pronoun – pronoun condition (15a) showed a very fast reading time at the critical verb avoided, while both the PropN – PropN (15b) and the DefD – DefD (15c) conditions were much slower, with reading at the critical verb involving about 50 extra ms:

The authors have pointed out that the fast reading time at the critical verb in (15a) is unexpected under a similarity-based approach, as the pronominal nature of both the clefted element and of the intervening subject should cause about as much difficulty as the two other cases with match (PropN – PropN, DefD – DefD), contrary to fact.⁵

So, the very fast reading at the critical verb in the Pron – Pron condition (15a) is a problem for the similarity-based approach. However, it seems to us that this very same set of data related to object extraction in clefts also raises a question for the increased referential processing approach. As we reported above, Gordon et al. (Reference Gordon, Hendrick and Johnson2001) observed a similarity effect in cases like (11)/(13): object extraction across a subject matching its status of definite description or proper name gives rise to a detectable difficulty at the critical verb, clearly a similarity-based effect. Figure 3 of Warren and Gibson (Reference Warren and Gibson2005) seems to confirm such an effect. Relevant examples are the following, in which the intervening subject is always a PropN:

If we compare reading times of the three cases of (16) (see Warren and GibsonReference Warren and Gibson2005, Figure 3), a substantially higher reading time is found at the critical word for the match condition (16b) compared to the two mismatch conditions (16a) and (16c). This is unexpected under the increased referential processing approach, according to which the crucial factor is the referential accessibility of the intervening subject (which is held constant in (16a–c)); on the other hand, the slower reading time at the critical word in (16b) is what is predicted by the similarity-based account, as in that approach what counts is the similarity of the two noun phrases, the target and the intervener, hence a slowing effect is expected in the match condition (16b), but not in the two mismatch conditions, as observed by Gordon et al. (Reference Gordon, Hendrick and Johnson2001) in connection with (11)/(13).

In conclusion, this array of experimental evidence raises at least one empirical problem for each approach:

4. Back to syntactic locality: selective intervention effects in psycholinguistics

In this Section we would like to show that the approach developed in Section 2.2 to the difficulties children experience with object relatives and based on syntactic locality as presented in Section 2.1, can offer some insight on how to address the similarity issue and the results on adults from the psycholinguistic literature just reviewed, specifically the questions raised in (17).

Clearly, our approach has much in common with Gordon et al. (Reference Gordon, Hendrick and Johnson2001): both approaches attribute a crucial role to the similarity between, in our terms, the target and the intervener of the local dependency. The two approaches differ in that for Gordon et al. similarity solely affects the processing system by conditioning the operation of working memory, to the effect that the temporary storage required for dependency resolution is facilitated if the target and the intervener are sufficiently dissimilar, whereas our approach directly appeals to a formal grammatical principle.

A question that is central to both accounts is: how is similarity determined? Various refinements in the Relativized Minimality tradition, in particular featural Relativized Minimality, represent an attempt to make the relevant dimension of similarity precise. For us, the expression of the operative distinctions is in terms of the specifications of morphosyntactic features triggering syntactic movement, while other imaginable dimensions of similarity – i.e., purely semantic, purely phonological, etc. – are not relevant.

Gordon et al. (Reference Gordon, Hendrick and Johnson2004), a follow up of the original study of Gordon et al. (Reference Gordon, Hendrick and Johnson2001), provide strong empirical evidence for the conclusion that only morphosyntactic features define the relevant dimension of similarity causing the effect in adult reading time. Thus there is strong converging evidence with the conclusion we reached on the basis of the syntactic and developmental facts.

Specifically, Gordon et al. (Reference Gordon, Hendrick and Johnson2004) showed that, given lexically headed object relatives (e.g., the salesman), manipulation of the definite, indefinite, or generic nature of the subject of the relative clause (e.g., The salesman that the accountant vs. an accountant vs. accountants contacted spoke very quickly) does not ameliorate the reading time and accuracy in comprehension, nor is there any amelioration if the subject of the relative clause is impoverished in its lexical descriptive content (e.g., The salesman that the accountant vs. the person contacted spoke very quickly).⁶ In contrast, the authors found a strong amelioration in cases in which the subject of the relative clause was a quantified expression like everyone. The relevant contrast is given in (18)–(19):

The reading time difference at the critical word, the verb contacted, was quite robust (736 vs. 487 ms, respectively). Also the accuracy in answering the comprehension question reached ceiling in cases like (19), while it was lower in (18). The amelioration with everyone is coherent with the amelioration found in the previous study from Gordon et al. (Reference Gordon, Hendrick and Johnson2001), brought about by a (1st and 2nd person) pronominal subject in object relatives as in the relevant case of (10). In our terms the amelioration effect in (19) relates to the lack of a lexical restriction in the quantified expression everyone.⁷

It is natural to explore the hypothesis that the developmental effects illustrated in Section 2.2 and the processing effects with adults addressed in this Section have a common explanation. Since for us the developmental effects are a consequence of the operation of the formal syntactic principle Relativized Minimality, the natural next step is to make the proposal that the processing asymmetries between subject and object A′-dependencies in adults can be ascribed to the same theoretical construct. In fact, the generalization arrived at empirically by Gordon et al. (Reference Gordon, Hendrick and Johnson2004) is precisely what is expected under our assumption that a formal syntactic principle is involved in the processing asymmetries: the fact that only morphosyntactic features are relevant in the computation of similarity follows from the fact that the causal operative principle is syntactic in nature.⁸

We see this convergence as a welcome outcome, one that goes in the direction of looking at processing and grammar as tightly integrated systems which operate with the same computational ingredients. We take up this point again in the conclusions.

Let us now go back to the problems in (17i) and (17ii), respectively left open by the competing processing accounts considered. In order to deal with them, we have to sharpen our assumptions on the featural system involved in nominal expressions.

Starting from (17ii), a major finding of the Gordonet al. (Reference Gordon, Hendrick and Johnson2001) paper is that a similarity effect is found in (13), repeated in (20) for convenience, in which (20a) and (20d), involving the crossing of the same kind of nominal expression, show a significant slowing down in the reading time of the critical verb:

How can this effect be captured by our syntactic approach? We assume that proper names always involve an N component which must be somehow specified in the lexicon as expressing singularity (ChierchiaReference Chierchia1998), hence they form a special class of nouns, stored as such in the lexicon. Let us refer to this class as N_proper as opposed to N_common. This featural distinction is clearly morphosyntactic in nature as it triggers different kinds of syntactic behaviors; e.g., in various Romance languages an N_proper requires obligatory movement to D, a movement motivated by the need of acquiring argument status (LongobardiReference Longobardi1994, Chierchia Reference Chierchia1998). Since N_proper and N_common have distinct morphosyntactic status, we expect that they will give rise to selective intervention effects, with N_proper being a stronger intervener in an N_proper dependency, and similarly for N_common in an N_common dependency. Remember that cases like (20a) and (20d) are well-formed under featural Relativized Minimality because the target has a featural specification richer than the intervener, as it also contains the feature attracting the noun phrase to the left periphery, presumably +Focus in the case of a cleft (BellettiReference Belletti2008, Reference Belletti2009). So, for instance the noun phrase Bill in (20d) is specified [+Foc, +NP, +N_proper ], a specification that properly includes the one of the intervening subject, which is just [+NP, +N_proper ], a permissible configuration under Relativized Minimality.

All the examples of (20) are thus well-formed and accessible to the adult grammatical system. Nevertheless, the complexity of the computation of the inclusion relation is detectable through the slowing effect in reading time and the decreased accuracy in comprehension.⁹ So, the difficulty with the computation of featural inclusion, insurmountable for young children, persists in adults in a subtler form, which is highlighted through the sophisticated techniques of experimental psycholinguistics.

Consider now problem (17i), illustrated by the fact that a pronoun crossing another pronoun does not give rise to a similarity-based effect, i.e., (21a) is characterized by a fast reading time at the critical verb, even though two nominal expressions in the crossing configuration are present which belong to the same syntactic category. Why is this case different from the cases in (21b) and (21c)?

Here we want to capitalize on the fact that pronominals lack a lexical restriction. We suggest that this can be expressed in featural terms by assuming that they lack the feature +NP altogether. So, the A′-dependencies in (21b) and (21c) involve a target and an intervener both specified with the feature +NP, whereas (21a) does not involve any such feature. Of course the two pronominals will share the category label, presumably D; however, in this system shared categorial features do not necessarily imply membership to the same class for intervention: only features triggering movement play a role (see also note 8). What are such features? On the one hand we have criterial features such as Q(uestion), R(elative), Foc(us), Top(ic), etc. If Rizzi and Shlonsky (Reference Rizzi, Shlonsky, Sauerland and Gärtner2007) are on the right track, +N also is the attracting feature involved in the Subject Criterion, the principle capturing the EPP and accounting for the attraction of a nominal expression to the canonical subject position. Some ϕ-features (e.g., Case, Person, Number. . .) also presumably function as attractors, as in the standard minimalist analysis of A-movement, or in cliticization. Other features may have a morphological realization and may enter into the labeling of expressions; however, they may be inert for triggering movement. We assume that this is precisely the case for the categorial feature D. Again this discussion reiterates the point that what is relevant for triggering minimality effects is not a generic notion of similarity: not only purely semantic similarity, but also categorial identity may be uninfluential if the relevant features are not movement related.¹⁰ Under this set of assumptions, it is expected that a pronominal crossing another pronominal, as in (21a), will not give rise to an increase of complexity analogous to the cases in which a proper name or a definite description cross a nominal of the same kind, as in (21b–c).

An alternative that comes to mind to account for the ease of processing characterizing examples like (21a) would be to appeal to the difference in person between the two pronouns. This could also be the source of the observation reported above that 1st and 2nd person pronouns never give rise to interfering effects with 3rd person definite descriptions. This alternative would predict that 3rd person pronouns should give rise to interference among themselves and with definite descriptions, which are inherently 3rd person. While we do not know of adult psycholinguistic evidence bearing on this prediction, we would like to go back to acquisition and report on some elicited production data with children, showing that children from very young ages are able to produce the difficult headed object relatives when a third person (null) pronominal subject is crossed in Italian, in examples like (22) (from Belletti and Contemori to appear):

Already the youngest group tested (3;4–3;11) was able to produce target sentences of the type in (22) in 94% of the cases. If the person feature was (solely) responsible for the facilitating effect in (21a) one would not expect a 3rd person subject pronoun to determine any facilitating effect for children in an object relative with a 3rd person relative head. The acquisition data suggest that the facilitation with pronominals holds independently of the person mismatch; hence, as we have suggested above, there is something inherent in the pronoun inducing the facilitation effect: in our terms, the lack of a lexical restriction (see also Belletti and Contemori to appear for relevant discussion).

5. Concluding remarks

Formal linguistic research has elaborated a detailed theory of locality based on the notion of intervention. In previous work we have proposed that linguistic locality can play an explanatory role in accounting for developmental effects such as the delay in acquisition of certain object A′-dependencies. In this chapter we have proposed that the same theoretical apparatus can successfully address the important empirical results arrived at in the experimental studies of adult processing that have been reviewed here. Psycholinguistic research has observed that there are robust asymmetries in reading times and accuracy to probe questions between subject relatives and subject extracted clefts on the one hand and object relatives and object extracted clefts on the other. Moreover, such effects are modulated by the similarity of the target with the intervening subject in the case of object A′-dependencies, as explicitly claimed by Gordonet al. (Reference Gordon, Hendrick and Johnson2001). We have explored the natural hypothesis that such asymmetries stem from the same source as intervention effects analyzed in syntactic theory and in the study of development. We have proposed a unified account of the empirically distinct manifestations of the intervention effects through featural Relativized Minimality. The unification is supported, among other considerations, by Gordon et al.’s (Reference Gordon, Hendrick and Johnson2004) results that among the imaginable dimensions of similarity (purely semantic, purely phonological, morphosyntactic, etc.) only certain morphosyntactic featural distinctions are relevant. The highly selective nature of the effect strongly suggests that the locus of interference is not “general working memory,” which would lead us to expect unselective similarity effects, but rather the computational devices specifically dedicated to syntactic computations. This is precisely what is expected under the assumption that a formal syntactic principle like Relativized Minimality plays a causal role here: only featural specifications involved in the triggering of movement operations are taken into account in the computation of the locality principle in (movement-related) syntactic dependencies.

We may think of the relevant morphosyntactic features as instructions for movement, and of Relativized Minimality as a principle requiring distinctness in potentially ambiguous situations arising from intervention: when an intervener is not distinct from the target, the system blocks; when the intervener is disjoint with respect to the target, the system works smoothly; when the intervener is only partially distinct from the target (inclusion in the relevant feature specification; on intersection see note 9) the system has to cope with a more complex computation which is inaccessible in early stages of development, and which gives rise to the persistent complexity effects uncovered in the adult psycholinguistic research that we have reviewed.

We take the conclusion that syntactic locality enters in a unified manner into the explanation of development and of adult parsing as a welcome result, as it goes in the direction of a substantial unification of grammar and processing, taking them as systems operating with largely overlapping ingredients and in tightly integrated manners.

Acknowledgments

We are indebted to Daniele Panizza for insightful comments on an earlier draft of this article. We take the opportunity to also thank our dear friend Gennaro for his long lasting friendship, scientific and otherwise. We look forward to his illuminating contributions to come.