Selection for a Surface Anaphor*

Mark Baltin
New York University


Janet Dean Fodor
City University of New York

1 Defining surface anaphora

Hankamer and Sag (1976) showed that anaphora comes in two varieties, which they termed deep anaphora and surface anaphora. A deep anaphor does not require an overt linguistic antecedent (i.e., it can be pragmatically controlled), while a surface anaphor does need an overt antecedent, either in the same sentence or earlier in the discourse. Null Complement Anaphora (NCA) is an example of deep anaphora, as is do it anaphora, while VP Ellipsis (VPE) is surface anaphora, as illustrated in the following examples.1

(1)(based on Hankamer and Sag's (47)):
 [Indulgent father feeds baby a chocolate bar for dinner]
 Mother: I don't approve. (NCA)
   I forbid you to do it. ("do it" anaphora)
  *I forbid you to. (VPE)

Hankamer and Sag correlate this first property of anaphors, optional pragmatic control versus obligatory linguistic control, with a second property: evidence of internal structure within the anaphor. For example, an elided VP can contain the underlying position of an extracted wh-element, while do it VPs cannot:

(2)(Hankamer and Sag's (27)):
 He knows how to fly but I don't know how to.
(3)(modelled on (58) of Johnson (2000)):
 I know which book Jose read for class, and which one Lulumae did (*it).

Hankamer and Sag propose, therefore, that deep anaphors are present as anaphors (internally unstructured elements) at the initial level, while surface anaphors arise by syntactic processes such as deletion in the course of the derivation. Hence, the deep/surface anaphor classification goes beyond mere symptoms and reflects the etiology of the anaphor.

As Hankamer and Sag note, however, it would also be possible to account for the distinction among anaphors in a way that makes the presence or absence of internal structure more fundamental than the contrast between deep and surface structure. As Wasow (1972) suggested, one might posit underlying anaphors that have internal structure (the 'surface' anaphors) and underlying anaphors that do not (the 'deep' anaphors), and require that the former have their lexical content filled in later by copying of an overt antecedent. This would explain why the antecedent cannot be merely situationally implied but must have linguistic structure. See Williams (1977) for an implementation of Wasow's idea. Hankamer and Sag say (in their note 15): "We will not discuss the details of making this work; syntactically, it is at best a notational variant of the deletion theory."

Nearly a quarter of a century later it is still not clear whether these two basic approaches to ellipsis are indeed trivial variants—that any facts that one can handle, the other can too. In this squib we draw attention to a kind of ellipsis in English that has not previously been noted, and we observe that it creates discomfort for both the deletion approach and the copying approach. Each is forced to make a difficult choice between possibly unwelcome ancillary assumptions.

Our novel observation is that a predicate can obligatorily select a null surface anaphor as its complement. This stems from examples such as (4), which contains the transitive verb please with experiencer subject:

(4)You may read whichever book you please.

We will first set out the properties of this rather odd construction, and then we will consider what demands it makes on a theory of ellipsis.

2 Non-finite complement ellipsis

The fulcrum for our argument is the strange behavior of the transitive verb please in (4). One puzzle about (4) is: What is the object of please? (4) doesn't imply that you will please a book, so although this is some sort of relative clause construction, it is not merely a case of movement of a wh operator (or of the head nominal; Kayne, 1994, and references there). The meaning of (4) suggests VP ellipsis following wh-movement: You may read whichever book you please to read. But though it means this, for many speakers this expansion of (4) is not grammatical; while (4) is colloquial, (5) is bad or at best archaic. However, for other predicates, as in (6), the comparable expansion is acceptable. For yet other verbs, as in (7), the elided form is impossible though the full form is fine.

(5)?* You may read whichever book you please to read.
(6)  You may read whichever book you wish / would like (to read).
(7)  You may read whichever book you care / plan / ask / are hoping *(to read).

In (8) we show examples similar to (6) but with a gerundive complement when it is not elided. In (9) the examples are ambiguous since these predicates permit either the ellipsis construction or simple wh-movement of an object.

(8) Our students just read whatever books they fancy / are capable of / feel like (reading).
(9)a. Our students read any books they choose.
 b.Throw away anything you like
   (cf. Throw away anything you don't like.)

In (9a) choose can mean select without elision, or elect to read with elision; (9b) is silly on the non-elision reading but sensible with elision.

The experiencer-subject please of (4) may be a relic from earlier days when it was more productive. It occurs in fixed forms such as if you please, as you please. Whatever its history, it is presently less common than please with experiencer-object and cause-subject (e.g., The book pleased the students). Though the experiencer-subject form exists in modern American English its distribution is very limited; in particular, its complement must apparently be null. This contrasts with the verbs in (6) which can take either null complements or fully overt complements. Both they and the verbs in (7) (but not please) also permit standard VPE: a null infinitival VP following to:

(10)a.  You may read whichever book you wish / would like / care / plan / ask / are hoping to.
 b.* You may read whichever book you please to.

If (10) represents 'normal' VPE, the ellipsis without to in (4), (6), (8) and (9) must be something else. In fact, the elided constituent is demonstrably not VP but IP or perhaps CP. This is clear from the comparison between (11) and (12):

(11)For the seminar the students will read any books Professor Jones wishes to.
(12)For the seminar the students will read any books Professor Jones wishes.

Sentence (11) has VPE with overt to, and refers to books that Professor Jones wishes to read herself. Sentence (12) without to cannot mean this but refers to books that Professor Jones wishes the students to read. Thus, ellipsis of the whole of to VP is in fact ellipsis of the whole clause including its subject. When the subject is PRO, as in (6), the difference is not evident, but it emerges clearly in ECM constructions with unlike subjects such as (12).2

We will call the phenomenon in (4), (6) and (12) non-finite clause ellipsis (NFCE), a label we intend to be neutral between the deletion and copying analyses. NFCE is acceptable only in certain contexts. In (4) the elided clause is inside a relative clause, and the antecedent is the matrix clause, so it is an instance of antecedent-contained deletion (ACD; May, 1985; Baltin, 1987). NFCE occurs not only in free relatives (with wh in the head DP) like (4), (6) and (13), but also in normal headed relatives like (14). But it does not occur in non-embedded contexts such as (15); this contrasts sharply with VPE as in (16) (using the absence of to to distinguish NFCE from VPE).

(13)  You may read however many books you please / decide.
(14)  You may read any of my books (that) you please / wish.
(15)* The students were required to read Moby Dick but they didn't wish.
(16)  The students were required to read Moby Dick but they didn't wish to.

Though the judgments are somewhat delicate, it seems that the head that is modified by the relative clause must be quantified rather than specific, or more generally that it should have the semantics of a 'free choice' construction. The quantifier emphasizes the broad extension of the DP, while the relative clause, containing some sort of desiderative or 'choice' predicate, expresses a criterion which limits that extension. Examples like (17) which don't meet this semantic condition are not fully acceptable; and there is a striking contrast in the acceptability of (18) depending on whether any is present.

(17) ?* He is reading some books that he pleases / wishes.
(18)a.* You can read three books you please.
 b.  You can read any three books you please.

It is noteworthy that NFCE does not occur in wh-questions or in topicalized declaratives.

(19)I know you plan to read some books.
 a.*Which ones do you please?
 b.  Which ones do you wish *(to)?
(20)I plan to read some books;
 a.*but these, I don't please.
 b.  but these, I don't wish *(to).

We have found NFCE in two other contexts. It occurs in comparatives, as in (21), and in if-clauses as in (22) though not in other adverbial clauses as in (23):

(21)  Susan can work as hard as she pleases / wants.
   Thomas never eats more than he feels like.
(22)  Sam can read this book if he pleases / chooses / sees fit.
(23)* Sam can read this book because he pleases / chooses.

Note that the comparative clauses have a variable within the elided material, as the relative clauses in (4) and (14) do, but this is not so for the if-clauses. Thus, like VPE (see (2) above), NFCE can elide a variable or not do so, depending on whether or not there is an operator in the context which needs a variable to bind.

It seems clear that there is some sort of semantic commonality that unites the possible contexts for NFCE and distinguishes them from the impossible ones, but despite some helpful advice from Anna Szabolsci, we are not brave enough to try to spell out here what the semantic distinction is. We focus on two points. One is that there can be a bound variable inside the null complement, indicating that it must have internal structure at some derivational stage, at least at LF. The other is that the predicates which permit NFCE are restricted, and in a partly arbitrary way; though they have some semantic content in common, it does not seem possible to predict exactly which predicates do select NFCE and which do not. For example, intend does but mean in the sense of intend does not; want and like do but prefer does not. Furthermore, the various structural contexts for NFCE (if-clauses, ACD relative clauses, etc.) tolerate slightly different subsets of predicates, some broader (e.g., comparatives: Thomas ate more than he planned) and some narrower (e.g., if-clauses: I'm sure Thomas will turn up if he planned *(to)).

On these grounds we draw the conclusion (only a little tentative)that lexical selection for NFCE is not semantic selection but requires specific subcategorization of individual lexical items. Certainly, in the case of experiencer-subject please there appears to be no way other than lexical stipulation to ensure that it takes only NFCE. (To treat experiencer-subject please in these constructions as a pure idiom, with completely unpredictable properties and hence no consequences for the nature of the grammar, would be implausible given that its only peculiarity is to do obligatorily what other predicates do optionally.)

Since NFCE requires an internally structured null complement, Hankamer and Sag's classification predicts that it also requires a linguistic antecedent. This is confirmed. If Ivan urges Jorge to jump from the plane, Jorge can't say (24):

(24)*But I don't feel like.

So NFCE qualifies as surface anaphora. Like other surface anaphors, its antecedent need not be in the same sentence as long as it is explicit in the discourse.

(25) Martha must read a biology book. Whichever one she pleases / decides is OK with her instructor.

To summarize the facts presented here: Experiencer-subject please selects NFCE obligatorily, and other predicates select it as one among other alternatives. NFCE involves an elided IP or perhaps CP complement that is subject to some unusual and stringent contextual restrictions, probably semantic in origin. (There are also some odd locality restrictions on NFCE in relative clauses, which we do not have space to discuss here.) An important question is: what precisely does the lexical entry for experiencer-subject please stipulate, which forces its complement to be null? And of more general theoretical concern: Is there a satisfactory answer to this question if ellipsis is derived by deletion of lexical material, and/or if it is derived by copying into an underlyingly null constituent?

3 Deletion versus copying

If surface anaphors arise by deletion, then the fact that a predicate can obligatorily select a surface anaphor will entail obligatory deletion of the complement. Earlier theories of generative grammar sanctioned rule features, which were used to mark heads for whether transformations must apply, could apply, or could not apply in their domains (Lakoff, 1971). For NFCE, we might suppose that the selecting predicate could carry a rule feature which forces the deletion rule to apply. Baker (1979) argued against rule features on learnability grounds. We cannot discuss the learnability issues here, but we note that Baker showed how subcategorization can do much the same work as rule features. However, in the present case the subcategorization would have to be for surface constituents—the outcome of the deletion process. The deletion theory without rule features assumes that underlyingly there is no difference between complements that will later be deleted and those that will remain. The only way for please to insist on NFCE would therefore be by surface selection for a null complement.

It should be noted that under the Minimalist Program (Chomsky, 1995), a rule feature would be a very odd duck, since there are no specific rules in a grammar but only general processes such as Attract which drives movement for purposes of feature-checking. The features that are checked as a result of movement operations represent independently motivated properties of constituents such as syntactic category, Case, Tense, or person, number, and gender; they do not refer to derivational processes per se. Rather, derivational processes are held to be responses by the grammar to the need to ensure that only appropriate features are present at each derivational level. In the case of deletion, it is unclear what the relevant properties would be or how they would control the derivation. Possibly the mechanisms outlined by Lasnik (2000) could be adapted to this purpose, but otherwise there is nothing in the theory at present that looks as if it could the basis for resuscitating the concept of rule features for deletion.

In short, a deletion analysis of surface anaphora could accommodate obligatory selection for a null surface anaphor but only, it seems, by introducing selection for surface constituents, or introducing some kind of 'strong' deletion features analogous to the strong features checked by movement. In the latter case, it would be necessary to articulate a mechanism such that a structure could be saved only by deleting a constituent that carries an illicit feature. Ideally, also, these features should have some independent motivation, as the features that drive movement do; but it is uncertain whether this condition can be met.

For a copying analysis of surface anaphors, obligatory selection such as NFCE selection by please may require some adjustments also. Consider first a simpler copying theory than Wasow's, in which a surface anaphor (like a deep anaphor) is generated as such, and (unlike a deep anaphor) its internal structure is created at LF by copying the complete structure of the antecedent phrase at the anaphor position. Assuming there are no objections to structure-building operations at LF, this would be satisfactory—except for the problem raised by Ross (1967) and taken seriously by Wasow: that there would then be no source for the wh-phrase in VPE examples like (2) and (3), and in the NFCE examples in relative clause contexts such as (4), (6), and (8).

For those who can embrace generation of a wh-phrase in its 'moved' position (e.g., Pollard and Sag, 1994; Brody, 1995) this is not a problem, as long as the trace/variable that the wh-operator must bind can be created in the course of copying the antecedent (as in the derivation of ACD constructions). On the other hand, if wh-fronted constructions must be created by movement, then Wasow's version of the copying hypothesis is necessary. The surface anaphor must be internally structured at the outset, with null lexical items except where overt material such as a wh-phrase (or internal argument of a passive, etc.) is needed because it will have to move out of the anaphor during the derivation. Since the grammar cannot look ahead, many derivations will crash, but that is not necessarily an impediment. What will need attention, on this approach, is to make sure that non-null elements that are inside the anaphor underlyingly do move out, or else block the derivation when copying applies—perhaps by some sort of full interpretation constraint. Otherwise, a surface filter would be needed to ensure that the anaphor is completely null at PF though it may be only partly null at earlier stages.

What is it that please would select, on each of these accounts, in order to ensure that it has a null complement? For deletion controlled by a feature that must be eliminated from the structure, please would select a complement clause, otherwise normal but bearing the relevant feature. For deletion which applies freely but whose output is filtered by surface selection, please would select a complement with normal syntactic and semantic structure but marked as obligatorily null at PF. On the simple copying approach with a non-movement treatment of wh-dependencies, please would select an IP node realized as the empty element (and hence not further expandable). This surface anaphor would be distinguished from a deep anaphor by a feature stipulating that interpretation of the anaphor is possible only after semantically relevant structure has been created by copying. On the Wasovian copying approach, as we have construed it here, please would select a complement clause that will be expanded in the normal way except that some or all of its terminal lexical items will be phonologically null. The latter property might be controlled by some sort of [optional-null] feature that could trickle down from the top IP node to the terminals. But this may be unnecessary if grammars permit null lexical items to be merged freely in any context, and filtered out later by the usual general principles requiring null elements to be licensed.

This squib represents our belief that, despite Hankamer and Sag's passing remark about notational equivalence, the issue of deletion versus copying as the derivational mechanism for surface anaphors is a substantive issue of theoretical importance. It is also still a live issue, very far from being resolved (see for example Tancredi, 1992, for a strong defense of the deletion approach to ellipsis). Obligatory NFCE for please adds to the empirical pressure on an account of ellipsis, and helps to expose theoretical commitments and needed innovations.


Baker, C. L. (1979) "Syntactic Theory and The Projection Problem", Linguistic Inquiry, Vol. 10, No. 4, pp. 533–581.

Baltin, M. (1987) "Do Antecedent-Contained Deletions Exist?", Linguistic Inquiry, Vol. 18, No. 4, pp. 279–295.

Brody, M. (1995) Lexico-Logical Form, MIT Press, Cambridge, Mass.

Chomsky, N. (1981) Lectures on Government and Binding, Foris, Dordrecht.

Chomsky, N. (1995) The Minimalist Program, MIT Press, Cambridge, Mass.

Hankamer, J., and Sag, I. A. (1976) "Deep and Surface Anaphora", Linguistic Inquiry, Vol. 7, No. 3, pp. 391–428.

Johnson, K. (2000) "What VP Ellipsis Can Do, And What It Can't, But Not Why", in M. Baltin and C. Collins, eds., The Handbook of Contemporary Syntactic Theory, Blackwell Publishers, Oxford, U.K., pp. 439–479.

Kayne, R. S. (1994) The Anti-Symmetry of Syntax, MIT Press, Cambridge, Mass.

Lakoff, G. (1971) Irregularity in Syntax, Holt, Rinehart and Winston, New York.

Lasnik, H. (2000) "Derivation and representation in modern transformational syntax", in M. Baltin and C. Collins (eds.) The Handbook of Contemporary Syntactic Theory, Blackwell Publishers, Oxford, UK. pp. 62–87.

May, R. (1985) Logical Form: Its Structure and Derivation, MIT Press, Cambridge, Mass.

Pollard, C., and Sag, I. A. (1994) Head-Driven Phrase Structure Grammar, University of Chicago Press, Chicago, Illinois.

Ross, J. R. (1967) Constraints on Variables in Syntax, Unpublished Doctoral Dissertation, MIT, Cambridge, Mass.

Tancredi, C. (1992) Deletion, De-Accenting, and Presupposition, Unpublished Doctoral Dissertation, MIT, Cambridge, Mass.

Wasow, T.(1972) Anaphoric Relations in English, Unpublished Doctoral Dissertation, MIT, Cambridge, Mass.

Williams, E. (1977) "Discourse and Logical Form", Linguistic Inquiry, Vol. 8, No. 1, pp. 101–139.


* This squib was written as a 60th birthday greeting to our friend Jorge Hankamer. In our haste—since Jorge's birthday wouldn't wait—we fear we may have overlooked some pertinent discussion in the literature. If so we apologize for this, as for all other defects, and hope to remedy some of them in a later and longer draft of this work. [Back]

1 Hankamer and Sag use the term anaphor for expressions (overt or null) that pick up their reference from other elements (verbal or nonverbal) in the discourse. This differs from anaphors in the sense of the binding theory of Chomsky (1991, 1995). [Back]

2 The tense and modality of the antecedent clause are not inherited by the null anaphor. For example: The students must read every book the professor wishes does not mean every book that the professor wishes that the students must read; it means every book she wishes them to read, i.e., wishes that they will read. Typically, the antecedent is a finite clause, while the elided complement is non-finite, with no overt tense or modal. The fact that the elided complement includes the subject but not the Infl of the antecedent clause suggests that the match is checked at a stage at which the subject is still internal to VP. If so, the contrast with what is standardly known as VP ellipsis needs to be re-defined, but we cannot undertake that here. [Back]