14 Binding theory: Syntactic constraints on the interpretation of noun phrases

Old version (Fall 2006) - current version here

This chapter is devoted to binding theory, the part of syntactic theory that is concerned with how the interpretation of noun phrases is constrained by syntactic considerations. For the purposes of binding theory, it is useful to distinguish several types of noun phrases: full noun phrases (the question, the student that asked the question, and so on), ordinary pronouns (I, you, they, and so on), reflexive pronouns (myself, yourself, themselves, and so on) and the reciprocal pronoun each other.

We discuss two alternative approaches to binding theory. The first approach, due to Hellan 1988, was proposed on the basis of Norwegian, a language with a unusually rich set of pronouns. We first discuss the Norwegian facts and then extend Hellan's analysis to English. The second approach, due to Chomsky 1981, was proposed on the basis of English and does not cover the full range of Norwegian facts. However, it includes an important condition on the distribution of ordinary noun phrases that is missing from Hellan's binding theory.

Coreference and coindexing

The primary function of a noun phrase like Bill Clinton, my two cats, the king of France, Santa Claus, or colorless green ideas is to refer - that is, to stand for a particular discourse entity, its referent. As the examples show, referents can be entities in the actual world, entities in some possible world, or even entities that could not possibly exist in principle. One of the characteristic features of human language is the absence, in general, of a one-to-one relation between noun phrases and referents. On the one hand, it is possible to use different noun phrases to refer to the same referent. The classic example is the fact that the expressions the morning star and the evening star both have the same referent - the planet Venus. On the other hand, the same noun phrase can have different referents. For instance, my apartment, used either by the same person at different points in time or by different persons at the same point in time, can refer to lodgings of vastly different size and attractiveness in completely different locations. Similarly, Lois's checking account balance can refer to widely differing dollar amounts.

In general, then, determining the intended referent of an expression requires recourse to a particular discourse context (who is speaking when, to whom, and so on). But there are certain expressions whose interpretation is particularly context-dependent - namely, pronouns. For instance, it is perfectly felicitous to introduce a new topic in a conversation with a friend using (1a) (provided that the speaker and the friend have in common a single acquaintance by the name of Vanessa). But replacing the proper noun Vanessa with a pronoun, as in (1b), in the same context (that is, as a new topic of conversation) is decidedly odd.

(1) a.   I ran into Vanessa the other day.
b. I ran into her the other day.

On the other hand, if Vanessa has already been mentioned in the discourse, then (1b) is perfectly felicitous, as illustrated in the mini-discourse in (2).

(2)     A: Have you seen Vanessa recently?
B: Yup. I ran into her the other day.

Pronouns, then, in contrast to ordinary noun phrases, are referentially dependent on some antecedent in the discourse. The term 'antecedent' is potentially misleading. Since it is derived etymologically from Latin ante-cedens 'one who walks before', it suggests that antecedents are required to precede a referentially dependent expression. However, precedence is neither a necessary nor a sufficient condition for antecedenthood. For instance, a sentence like Zelda's brother likes herself is ungrammatical even though Zelda precedes herself.. A less misleading term might be 'referential anchor,' but we continue to use the 'antecedent' because it is the standard term in the literature.

It is possible for a sentence to contain more than one possible antecedent for a pronoun. For instance, in (3), he can refer to either Tim or Tom.

(3)     Tim told Tom that he needed some time off.

In any particular discourse context, of course, one interpretation may well be favored over the other, given what is known about Tim, Tom, their respective work loads, and so on. In case the antecedent for he must be explicitly specified in an unambiguous way, this can be achieved as in (4).

(4) a.   Tim told Tom that he, Tim, needed some time off.
b.   Tim told Tom that he, Tom, needed some time off.

When (3) has the interpretation in (4a), he and Tim are said to corefer. On the alternative interpretation in (4b), it is he and Tom that corefer.

It is convenient to introduce a notation to represent coreference relations. Let us begin by asssociating any noun phrase with a referential index. In the syntactic literature, it is standard to use the letters of the alphabet, beginning with i (for 'index'), both as referential indices and to indicate movement. For clarity, however, we will use the natural numbers as referential indices. We now introduce the following convention. In order to indicate an interpretation where two expressions refer to the same discourse entity (in other words, where they corefer), we assign the same index to both expressions, or coindex them. On the other hand, in order to indicate an interpretation in which two expressions refer to distinct discourse entities (that is, where they do not corefer), we assign distinct indices to each of the two expressions. Such expressions are said to be contraindexed. In neither case are the specific indices important - only whether the indices are the same or not. For instance, both indexings in (5) represent the interpretation in (4a), and both indexings in (6) represent the interpretation in (4b).

(5) a.   Tim1 told Tom2 that he1 needed some time off.
b. Tim1097 told Tom18 that he1097 needed some time off.
(6) a.   Tim1 told Tom2 that he2 needed some time off.
b. Tim380 told Tom7 that he7 needed some time off.

(7) gives a further grammatical indexing for the string in (3). Of course, in any particular discourse, this indexing is pragmatically felicitous only if a discourse entity with the index 3 (say, Tim's brother Mike) has already been mentioned.

(7)     Tim1 told Tom2 that he3 needed some time off.

Reference and indexing must be carefully distinguished. Reference relations are actual linguistic relations that we have intuitions about. For instance, we have the intuition that Tim and him can corefer in (8a), but not in (8b).

(8) a.   Tim thinks that everyone admires him.
b. Tim admires him.

By contrast, indices are a notational device intended to represent arbitrary reference relations; the indices themselves have no independent linguistic or psychological status. It is perfectly possible to assign referential indices to noun phrases so as to represent interpretations of a sentence that are possible in principle, but ungrammatical in fact. Two such ungrammatical indexings are illustrated in (9).

(9) a. * Tim1 admires him1.
b. * He1 admires Tim1.

The proposition (= state of affairs) that both (9a) and (9b) are trying to express is not inherently semantically anomalous. It can be expressed perfectly well by the grammatical sentence in (10).

(10)     Tim1 admires himself1.

Notice furthermore that what makes the sequences in (9) ungrammatical is the index assignment. The same sequences of words as in (9) are grammatical sentences when associated with the indices in (11).

(11) a. ok Tim1 admires him2.
b. ok He1 admires Tim2.

In other words, the grammaticality of a sentence is always determined with reference to a particular intended interpretation.1

Representations like those in (9) and (11) are generally abbreviated as in (12). The descending numerical order of the indices on the object noun phrases is intended to unambiguously indicate the scope of the asterisk.

(12) a. ok Tim1 admires him2,*1.
b. ok He1 admires Tim2,*1.

Why can't we use sentences like (9a) or (9b) to express the proposition that is expressed grammatically in (10)? This is exactly the type of question that is addressed in the rest of this chapter.

Hellan 1988

The co-argument condition

English makes a distinction between reflexive pronouns ending in -self (myself, yourself, and so on) and the corresponding ordinary pronouns (I, you, and so on). Norwegian, too, distinguishes between pronouns with forms containing selv 'self' and ones without. As (13) and (14) show, selv is in complementary distribution with zero (that is, where a form with selv can occur, a form without selv cannot, and vice versa).

In what follows, we restrict our attention to reflexive pronouns with local or long-distance antecedents (that is, with an antecedent in the same clause as the reflexive pronoun or in a higher clause, respectively). Many languages also have inherently reflexive pronouns, which are special in that they don't function as an argument of their predicate, even though they occupy a syntactic argument position. Inherently reflexive pronouns are rare in English (to behave oneself, to perjure oneself), but much more common in other languages.

(13) a.
Jon1 beundrer seg selv1.
     admires SEG self
b. *
Jon1 beundrer seg1.
'Jon1 admires himself1.'
(14) a. *
Jon1 bad   oss beundre seg selv1.
     asked us  admire  SEG self
Jon1 bad oss beundre seg1.
'Jon1 asked us to admire him1.'

In an effort to understand what determines the complementary distribution, let us consider the derivation of the sentences. Both sentences require the elementary tree in (15) for the verb beundre(r) 'admire(s)'.


Observe now that in the derivation of (13a), seg selv and Jon both substitute into the elementary tree for beundrer. In the derivation of (14a), on the other hand, the subject position of the elementary tree for beundre is filled by PRO, which is referentially dependent on the matrix object oss 'us', whereas Jon substitutes into the elementary tree for the matrix predicate bad 'asked'. The structures resulting from substitution in the two cases are shown in (16a) and (16b).

(16) a.    
b. i.       ii.  

Given this derivational difference, we can define a notion of co-argument as in (17a) and propose the condition governing the distribution of selv in (17b).

(17) a.   Definition of co-argument (to be revised):
A and B are co-arguments iff they substitute into the argument positions (specifier or complement) of the same elementary tree.
b. Co-argument condition (to be revised):
Selv and its antecedent (more precisely, the minimal DP nodes dominating them) must be co-arguments.

One type of syntax/semantics mismatch. While adequate for (13) and (14), the definition and condition in (17) does not extend to examples like (18).

(18) a.  
Jon1 snakket om    seg selv1.
     talked  about SEG self
b. *
Jon1 snakket om seg1.
'Jon1 talked about himself1.'
(19) a. *
Jon1 bad   oss snakke om    seg selv1.
     asked us  talk   about SEG self
Jon1 bad oss snakke om seg.
'Jon1 asked us  talk about himself1.'

As is evident, the distribution of selv in (18) and (19) is parallel to that in (13) and (14). However, the grammatical (18a) violates the condition in (17b) because the DPs dominating seg selv and its antecedent do not substitute into the same elementary tree. Rather, seg selv substitutes into the elementary tree for om 'about', whereas its antecedent Jon substitutes into that for snakket 'talked'.

One way of addressing this problem is to introduce a semantically-based notion of argument, based on the idea that situations can be conceptually decomposed into events (activities, states, qualities, and so on), on the one hand, and different types of participants in such events (agents, experiencers, beneficiaries, goals, etc.), on the other. In general, the participants in an event denoted by a lexical item correspond to the substitution nodes in the elementary tree associated with that lexical item. Indeed, this correspondence is so general that substitution nodes are often equated with participants in a situation, with both being referred to as arguments. However, locutions like snakker om DP 'talk about DP' provide evidence that syntactic arguments (constituents of the elementary tree projected by a lexical item) must be distinguished from semantic arguments (participants in the event denoted by the lexical item). In such cases, there is a mismatch between the syntax and the semantics. Specifically, the DP, being a complement of the preposition, stands in no local syntactic relation with the verb. Nevertheless, it is a semantic participant in the 'talking about' activity. For the purposes of binding, the prepositional complement's status as a semantic argument is sufficient to license selv. Based on the facts presented so far, it is therefore possible to revise (17b) to read as in (20b).

(20) a.   Semantic co-argument:
A and B are semantic co-arguments iff they denote co-participants in the same situation.
b. Co-argument condition (to be revised):
Selv and its antecedent must be semantic co-arguments.

It should be noted that the status of a particular phrase as a semantic argument is not always easy to determine, and that there is some individual variation regarding the relevant judgments. This is reminiscent of the situation that we encountered in connection with the distinction between syntactic arguments and adjuncts discussed in Chapter 4.

Another type of syntax/semantics mismatch. As we have just seen, a noun phrase can be a semantic argument of a head denoting an event without being one of its syntactic arguments. Also possible is the converse state of affairs, in which a noun phrase counts as a syntactic argument of a head without being its semantic argument. We illustrate this second type of syntax/semantics mismatch with reference to perception verb complements. Consider (21).

Jon1 så seg selv1 sitte i stolen.
    saw SEG self  sit   in chair.def
'Jon1 saw himself1 sit in the chair.'

As the tree for (21) in (22) shows, the matrix verb 'saw' takes as its complement a VP small clause, so Jon and seg selv are not co-arguments in the sense of either (17a) or (20a).2


In fact, seg selv is not a semantic argument of at all, since the complement subject position can be occupied by expletive det, the Norwegian counterpart of expletive there.

Vi så  det   sitte ett spøkelse i  stolen.
we saw there sit   a   ghost    in chair.def
'We saw a ghost sit in the chair.'

The facts discussed so far motivate extending the concept of syntactic co-argument as in (24).

(24)     Syntactic argument:
A and B are syntactic co-arguments iff they are locally head-licensed by the same head C.

The notion of local head-licensing was introduced earlier (though not by that name) in connection with case licensing. There, we distinguished three configurations in which case was licensed: spec-head, head-spec, and head-complement. These structural notions are also relevant for binding theory. In (16), Jon and seg selv are co-arguments because they stand in the spec-head and head-complement relation, respectively, with beundrer 'admires'. In (23), (the trace of) Jon and seg selv are co-arguments because they stand in the spec-head and head-spec relation, respectively, with så.

Our final revision of the condition on selv is given in (25).

(25) Co-argument condition (final version):
Selv and its antecedent must be semantic or syntactic co-arguments.

Note that the statement of the condition is not yet ideal. The fact that (25) contains a disjunction suggests that we do not have a complete understanding of the notion of co-argument.

The predication condition

In addition to the distinction between elements with and without selv, Norwegian distinguishes between the two third person singular forms seg and ham. Like seg, ham occurs on its own as well as before selv. As (26) and (27) show, seg and ham are in complementary distribution.

Note that the noun phrases in (26) and (27) are all semantic arguments. Therefore, since the co-argument condition in (25) is satisfied, the contrast in (26) and (27) must be due to independent conditions governing the distribution of seg and ham.

Jon1 snakket om    { seg / *ham } selv1.
     talked  about   SEG    HAM   self
'Jon1 talked about himself1.'
(27) a.  
Vi snakket med  Jon1 om    { ham / *seg } selv1.
we talked  with      about   HAM    SEG   self
'We talked with Jon1 about himself1.'
Vi fortalte Jon1 om   { ham / *seg } selv1.
we told          about  HAM    SEG   self
'We told Jon1 about himself1.'

The difference between the two cases is that the (potential) antecedent is a subject in (26), but not in (27). Further evidence that the antecedent of seg must be a subject comes from (28), which shows that what is crucial is the antecedent's status as a subject of predication.

Vi gjorde Jon1 glad i  { seg / *ham } selv1.
we made        fond of   SEG    HAM   self
'We made Jon1 fond of himself1.'

The facts in (26)-(28) are accounted for by the condition in (29).

(29)     Predication condition:
The antecedent of seg must be a subject of predication, whereas the antecedent of ham must not be.

The co-argument condition in (25) and the predication condition in (29) are independent of one another. This has the result that the domain in which the referentially dependent elements under discussion can appear can be partitioned without overlap as in (30).3

Antecedent co-argument?
Antecedent subject of predication? Yes No
Yes seg selv seg
No ham selv ham

The tensed IP condition

The contrast in (31) shows that seg is subject to a final condition, given in (32).

(31) a.  
Jon1 bad   oss forsøke å  få  deg til     å  snakke pent   om    seg1.
     asked us  try     to get you towards to talk   nicely about SEG
'Jon1 asked us to try to get you to talk nicely about him1.'
b. *
Jon1 var ikke klar  over at   vi hadde snakket om    seg1.
     was not  aware over that we had   talked  about SEG
'Jon1 was not aware that we had talked about him1.'
(32)     The minimal tensed IP dominating seg must also dominate its antecedent.

Strict vs. non-strict co-arguments

As we saw earlier, selv is subject to the co-argument condition in (25), repeated in (33).

(33)     Co-argument condition:
Selv and its antecedent must be semantic or syntactic co-arguments.

The contrast in (34) might at first glance be taken to indicate that the reciprocal pronoun hverandre 'each other' is not subject to such a condition.

(34) a. *
Jon1 traff noen venner  av seg selv1.
     met   some friends of SEG self
'Jon1 met some friends of his1.'
[ Jon og  Marit ]1 traff noen venner  av hverandre1.
      and          met   some friends of each other
'[ Jon and Marit ]1 met some friends of each other's1.'

Nevertheless, as (35) shows, the distribution of hverandre is not completely unconstrained.

(35)   ?*
[ Jon og Marit ]1 leste mine bøker om    hverandre1.
      and         read  my   books about each other
Intended meaning: 'Jon read my book about Marit, and Marit read my book about Jon.'

It is possible to account for the contrast between (34b) and (35) by introducing a distinction between strict and non-strict co-arguments. Strict arguments are defined as in (17) or (20), but non-strict co-arguments are more loosely related. Consider the two configurations in (36).

(36) a.       b.  

In (36a), we will say that the absence of a specifier in BP allows B to extend its argument-taking domain up to the first maximal projection that is 'closed off' by a specifier (in this case, AP). In such a domain, which we will call a projection chain of B, the arguments of the heads in the projection chain count as co-arguments in a non-strict sense. In (36b), on the other hand, the presence of Spec(BP) prevents the formation of such a projection chain.

The similarities and differences in the distribution of selv and hverandre can now be succinctly stated as in (37).

(37) a.   Selv and its antecedent must be strict co-arguments.
b. Hverandre and its antecedent must be co-arguments (possibly non-strict ones).

We illustrate the application of the condition in (37b) to (34b) and (35) with reference to the structures in (38).

(38) a.     b.

(38a) is a (recursive) instantiation of the projection chain configuration in (36a). None of the maximal projections in the object DP is closed off by a specifier, and the projection chain of the preposition that governs hverandre therefore extends to VP. As a result, (the trace of) Jon og Marit is a non-strict co-argument of hverandre and can serve as its antecedent. (38b) turns out to be an instantiation of the projection chain configuration as well. But here, the projection chain of the P containing hverandre extends only to the object DP, which is closed off by mine. The ungrammaticality of the sentence can then be derived from the feature mismatch (number and person) between hverandre and its potential antecedent mine.

Extending Hellan's binding theory to English

As discussed earlier, Norwegian distinguishes between two third person pronouns: seg and ham, which can both stand alone or be combined with the morpheme selv, yielding four referentially dependent forms (seg, seg selv, ham, and ham selv). English lacks the distinction between seg and ham and has only two referentially dependent forms (exemplified in what follows by him and himself). The question that guides the discussion in this section is how these forms divide up the syntactic territory covered by the four forms of Norwegian.

The co-argument condition

Since English -self is cognate with Norwegian selv, it is reasonable to assume that English reflexive pronouns are subject to a condition analogous to the co-argument condition in (25). Such a condition is given in (39a). The companion condition on ordinary pronouns in (39b) makes explicit the complementary distribution between the two types of pronouns.

(39) a.   Reflexive pronouns and their antecedents must be (semantic or syntactic) co-arguments.
b.   Ordinary pronouns and their antecedents must not be co-arguments.

The conditions in (39) are borne out by the data in (40) and (41).

(40) a.   John1 admires { himself1 / *him1 }.
b. John1 talked about { himself1 / *him1 }.
c. We made John1 fond of { himself1 / *him1 }.
(41) a. John1 asked us to admire { him1 / *himself1 }.
b. John1 asked us to talk about { him1 / *himself1 }.
c. John1 made us fond of { him1 / *himself1 }.

In (37a), we proposed that the antecedents of Norwegian selv must be strict co-arguments, and so we are led to wonder whether this is true of English reflexive pronouns as well. Evidence bearing on this question comes from examples like (42), which has the structure in (43).

(42)     John1 bought some books about himself1.

Himself, a semantic argument of books, has no strict co-arguments (the PP dominating it has no specifier). But for precisely this reason, (43) is an instance of a projection chain. In particular, given that the entire object DP contains no specifiers, the projection chain of about, the head governing himself, extends to VP, making himself and (the trace of) John non-strict co-arguments. Since (42) is grammatical, we conclude that the antecedent of himself need only be a non-strict co-argument. In this respect, then, himself differs from Norweian seg selv and resembles hverandre 'each other.'

As expected, if the projection chain of about is closed off by a specifier of the object DP, then it is that DP that must contain an antecedent that agrees with himself in the relevant grammatical features (number and gender). This is shown by the contrast in (44).

(44) a.   Mary bought John1's book about himself1.
b. * John1 bought Mary's book about himself1.

Finally, the contrast between (42) and (45) on the one hand and (44) and (46) on the other confirms that reflexive and ordinary pronouns are in complementary distribution.

(45)   * John1 bought some books about him1.
(46) a. * Mary bought John1's book about him1.
b. John1 bought Mary's book about him1.

The predication condition

A further question that arises in comparing the English pronouns with their Norwegian counterparts is whether himself covers the syntactic territory of seg selv alone or of both of the selv forms (seg selv and ham selv). (41) and (46) show that the antecedent of him, unlike that of ham, can be a subject of predication, and (44a) suggests that the antecedent of himself needn't be (we say 'suggests' rather than 'shows' because the status of prenominal genitives with respect to predication is a bit murky). This evidence is consistent with the view that himself corresponds to both of the selv forms, and that English reflexive and ordinary pronouns are not subject to a predication condition analogous to the one in (29). This view is clearly confirmed by the facts in (47), where the antecedent of himself is not a subject of predication.

(47) a. We talked with John1 about { himself1 / *him1 }.
b. We told John1 about { himself1 / *him1 }.

The tensed IP condition

Since the tensed IP condition governs the distribution of seg versus ham and English does not distinguish between these two elements, we would expect the tensed IP condition to be irrelevant in English. This conclusion is confirmed by the parallelism between (48a) and (48b).

(48) a.   John1 asked us to try to get you to talk nicely about { him1 / *himself1 }.
b. John1 was not aware that we had talked about { him1 / *himself1 }.

In summary, then, English himself corresponds closely to the two selv forms of Norwegian, seg selv and ham selv. The correspondence is not perfect, however, because the antecedent of English himself, unlike that of the selv forms, needn't be a strict co-argument.

Chomsky 1981

We have presented an approach to binding theory that was developed in order to account for the particularly rich data from Norwegian. In this section, we compare this approach with the standard binding theory of Chomsky 1981, which was developed on the basis of morphologically simpler data of English. Chomsky's binding theory contains three conditions (or principles, as they are more commonly referred to), which govern the distribution of reflexive and reciprocal pronouns, ordinary pronouns, and full noun phrases, respectively. We present each of these principles in turn.

Principle A

Consider the English binding facts in (49).

(49) a. [ Zelda ]1 helped [ herself ]1.
b. [ Zelda's sister ]1 helped [ herself ]1.
c. * [ Zelda's ]1 sister helped [ herself ]1.

In Hellan's binding theory, (49a) is grammatical because herself and its antecedent Zelda are syntactic co-arguments. The same is true of herself and Zelda's sister in (49b), and so it is grammatical as well. In (49c), on the other hand, the intended antecedent Zelda is not a co-argument of herself, and so the sentence is ungrammatical under the intended interpretation.

Chomsky 1981 derives the grammaticality pattern in (49) on the basis not of co-argumenthood, but on the basis of the structural relation of c-command, defined as in (50).

(50)     A c-commands B iff (= if and only if)
a. neither A nor B dominates the other, and
b. the first branching node that dominates A also dominates B.

The structures for the sentences in (49) are shown in (51). In all three of the structures, the intended antecedent of the anaphor herself and the anaphor itself are boxed in red. The first branching node dominating the intended antecedent is boxed in black. Notice now that the black-boxed nodes dominates that anaphor in (51a,b), but not in (51c). In other words, the anaphor is c-commanded by the intended antecedent in (51a,b), but not in (51c).

(51)   a.     b.       c.  
Intended antecedent c-commands anaphor? yes yes no
These configurational relations suggest the condition in (52).

(52)     Principle A (to be revised):
An anaphor must be c-commanded by a coindexed antecedent.

The term anaphor in Chomsky's usage refers to reflexive and reciprocal pronouns. This usage is potentially confusing because in general linguistic usage, anaphora refers to referential dependence regardless of morphological form. In other words, ordinary pronouns and even full noun phrases can count as anaphors in this wide sense. Here, we will use the restricted sense of the term, in keeping with Chomsky's usage.

(52) is generally expressed more succinctly as in (53a), where the notion of 'binding' is defined as in (53b)

(53) a.   An anaphor must be bound.
b.   A binds B iff
i.   A c-commands B, and
ii. A and B are coindexed.

As it turns out, (53a) is a necessary but not sufficient condition on anaphors in English, since it fails to account for data involving complex clauses like (54).

(54)     John1 thinks that Mary2 will help { herself2 / *himself1 }.

The grammaticality of the variant with herself is unproblematic (herself is bound by Mary ). But given (53a), the ungrammaticality of the variant with himself is surprising, because himself is bound by John.

In Hellan's version of the binding theory, John is ruled out as an antecedent of himself by the co-argument condition in (39a). In Chomsky's version, a similar effect is achieved by introducing the notion of governing category, a locality domain within which an anaphor must be bound. Since the definition of governing category is fairly complex, we will work our way up to it in several steps, motivating each complication of the definition in turn.

(55)     Principle A (final version):
An anaphor must be bound within its governing category.
(56)     Governing category (to be revised):
The governing category for an expression is the lowest IP that contains that expression.

(55) correctly describes the contrast in (54), but is unable to account for contrasts as in (44), repeated as (57).

(57) a.   Mary bought John1's book about himself1.
b. * John1 bought Mary's book about himself1.

In both sentences, the lowest IP that contains himself (the only IP in the sentence) also contains John, in accordance with the definition of governing category in (56). The ungrammaticality of (57b) is therefore unexpected.

A first step in accounting for such contrasts lies in making reference to DPs in addition to IPs in the definition of governing category, as in (58).

(58)     Governing category (to be revised):
The governing category for an expression is the lowest IP or DP that contains that expression.

But now (58) overshoots the mark by incorrectly ruling out sentences like (42), repeated as (59).

(59)     John1 bought some books about himself1.

A further refinement of the definition of governing category yields (60).

(60)     Governing category (to be revised):
The governing category for an expression is the lowest IP or DP that contains that expression, and a specifier.

The reference to a specifier in (60) is the formal counterpart to the distinction between strict and non-strict co-arguments in Hellan's approach to the binding theory. There is a difference, however, in what the two approaches regard as the unmarked domain within which an anaphor must be bound. In Hellan's approach, the availability of non-strict co-argument antecedents extends a binding domain that would otherwise be smaller. By contrast, in Chomsky's approach, the addition of clause (ii) in (60) restricts a binding domain that would otherwise be larger.

There remains one final revision to make to the definition of governing category. The revision is motivated by sentences like (61a), which have the structure in (61b).

(61) a.   Joan1 believes herself1 to be indispensable.

Let's consider in detail why (61) poses a difficulty for the definition in (60). As is evident from (61b), the lowest IP or DP that contains the anaphor and a specifier is the complement IP (indicated by the box). (The anaphor and the specifier turn out to be the same node, but nothing in (60) rules this out.) But since the complement IP contains no antecedent for the anaphor, (61a) is incorrectly predicted to be ungrammatical.

In order to accommodate sentences like (61a), the governing category must be extended in just the right way to allow matrix subjects to act as antecedents in sentences like (61a), but not in ones like (54). This is achieved by the formulation in (62).

(62)     Governing category (final version):
The governing category for an expression is the lowest IP or DP that contains (i) that expression, (ii) a specifier, and (iii) the expression's case-licensing head.

Note that the addition of clause (iii) in (62) has the same effect as Hellan's extension of the notion of syntactic argument in (24) (recall the discussion of the second type of syntax/semantics mismatch discussed earlier in connection with the distribution of Norwegian selv).

Principle B

We turn now to the distribution of pronouns.

As noted earlier, reflexive and reciprocal pronouns are combined in the Chomskyan tradition under the rubric of anaphors. This leaves the unqualified term 'pronoun' to refer to ordinary personal pronouns (I, you, he, and so on). Unless otherwise noted, we follow this usage in what follows.

Replacing the anaphors in (49) by pronouns yields the sentences in (63).

(63) a. * [ Zelda ]1 helped [ her ]1.
b. * [ Zelda's sister ]1 helped [ her ]1.
c. [ Zelda's ]1 sister helped [ her ]1.

The resulting grammaticality pattern, which is the converse of that in (49), suggests the condition in (64).

(64)     Principle B (to be revised):
A pronoun must be free (= not bound by an antecedent).

As in the case of our first formulation of Principle A, the formulation of Principle B in (64) is not ye quite inadequate. This time, however, the condition errs on the side of caution, incorrectly ruling out grammatical sentences like (65).

(65)     John1 thinks Mary will help him1.

Since John c-commands him (along with everything else in the sentence), (64) incorrectly leads us to expect (65) to be ungrammatical.

The fact that the antecedent and the pronoun are in different clauses in (65) suggests reformulating (64) to incorporate the concept of governing category, as in (66).

(66)     Principle B (final version):
A pronoun must be free in its governing category.

As is evident, both Hellan's and Chomsky's approaches to the binding theory agree that anaphors and pronouns are in complementary distribution.

Principle C

Consider the sentences in (67).

(67) a. * She1 treats Mary1 well.
b. * She1 claims that they treat Mary1 well.
c. * She1 claims that we know that they treat Mary1 well.

In Hellan's approach to the binding theory, (67a) is ruled out by the condition in (39b), according to which an ordinary pronoun and its antecedent cannot be co-arguments. It is not, however, ruled out by Principle B, the counterpart to (39b) in Chomsky's binding theory, since the pronoun is free. In order to rule out sentences like (67a), Chomsky's binding theory therefore contains a third principle that governs the distribution of full noun phrases (referred to in Chomsky's usage as r(eferential)-expressions).

(68)     Principle C:
R-expressions must be free.

Principle C is reminiscent of Principle B, but differs from it in that the anti-binding requirement is not relativized to a binding domain. The absolute character of the anti-binding requirement in Principle C is borne out by the ungrammaticality of (67b,c).

The ungrammaticality of the sentences in (67b,c) means that an anti-binding condition on full noun phrases also needs to be incorporated into Hellan's approach to binding theory. This is because the co-argument condition in (39b), which ruled out (67a), fails to apply to (67b,c). They must therefore be ruled out by separate means. The necessity for an independent anti-binding condition on full noun phrases is underscored by sentences like (69).

(69)   * We made herself1 treat Mary1 well.

In Chomsky's binding theory, (69) is ruled out by both Principles A and C. But in Hellan's version of the binding theory, (69) satisfies the co-argument condition in (39a), the equivalent of Principle A, because Mary and herself are syntactic co-arguments (treat locally head-licenses Mary). It is therefore only with reference to an anti-binding condition on full noun phrases that (69) can be ruled out as required.

In traditional grammar, the examples in (67) and (69) would all be classified as instances of cataphora, instances of referential dependence in which the antecedent follows the referentially dependent element. It is worth pointing out explicitly that not all instances of cataphora are ungrammatical. Examples like (70), where the referentially dependent element precedes, but does not c-command the antecedent, are grammatical, as expected under Principle C.

(70)     If he1 calls, tell John1 I'll be back in an hour.

The term 'anaphora' is generally used to subsume both cataphora and what might be called strict anaphora, where the antecedent precedes the referentially dependent element. According to the metaphor underlying the terms, discourse flows along like a stream, with temporally earlier elements located upstream from later ones. The antecedent is then upstream of the referentially dependent element in strict anaphora (< Greek ana 'up, against the current'), but downstream in cataphora (< Greek kata 'down, with the current').

A final point needs to be made about Principle C. We pointed out earlier that Principle C is an absolute requirement in the sense that it makes no reference to binding domains. It is absolute in a further sense as well: it makes no reference to whether the binder is a referentially dependent element. Thus, Principle C rules out (71) on a par with (67a).

(71)   Mary1 treats Mary1 well.

It has been argued, however, that under certain discourse conditions, sentences structurally parallel to (71) are in fact acceptable, as illustrated in (72a). Two further acceptable violations of Principle C from unedited usage are given in (72b,c).

(72) a.   Nobody likes Oscar. Even Oscar1 doesn't like Oscar1.
b.   Phil1 said that if he2 came by, Phil1 would show him2 around. (manager@babel.ling.upenn.edu to beatrice@babel.ling.upenn.edu)
c.   Luke1 thinks that everyone has as much integrity as Luke1 has. (overheard at the White Dog Cafe, Philadelphia, PA, 17 Feb 01)

However, equivalent sentences in which the full noun phrase is bound by a referentially dependent element are degraded.

(73) a. * Nobody likes Oscar. Even he1 doesn't like Oscar1.
b. * He1 said that if { he2 / Sean2 } came by, Phil1 would show him2 around.
c. * He1 thinks that everyone has as much integrity as Luke1 has.

Although judgments regarding contrasts of the type illustrated by (72) and (73) can be delicate, it seems reasonable to weaken Principle C to include reference to the status of the binder. This is done in (74).

(74)     Weakened version of Principle C:
Full noun phrases must not be bound by a referentially dependent element.


1. The same point arises in other connections as well. For instance, the question in (i) is grammatical on the interpretation in (ii.a), but not on that in (ii.b).

(i)   How have they forgotten which problem they should solve?
(ii) a.   Howi have they forgotten ti which problem they should solve? (By succumbing to Alzheimer's.)
b. * Howi have they forgotten which problem they should solve ti? (By using Fourier analysis.)

2. The Trees program doesn't support the special character å, so we substitute the conventional Scandinavian orthographic variant aa in the trees.

3. Certain difficult cases are set aside; for further discussion, see Hellan 1988, Chapter 3.