Research on Language and Computation (2005) 3:411–428 DOI 10.1007/s11168-006-0005-9 © Springer 2006 n-ary Quantifiers and the Expressive Power of DP-Compositions SEUNGHO NAM Department of Linguistics, Seoul National University, Sillim-dong, Kwanak-gu, Seoul 151-742, Republic of Korea (E-mail: [email protected]) Abstract. This paper, extending Keenan’s (1987, 1988) Case-extension of generalized quantifiers, proposes a natural algebraic semantics of DP-coordination and DP-composition. DPs in subject and non-subject positions are uniformly identified as a case-extension of a usual generalized quantifier, and DPs with different semantic cases combine with each other to yield polyadic quantifiers. The paper proves that each set of case-extensions forms a complete atomic Boolean algebra consisting of a full set of unary quantifiers, and it further shows that natural language requires the full power of binary quantification, i.e., the full set of (Fregean) reducible and unreducible binary quantifiers. This result, Type < 2 > Effability, is derived from the fact that the set of all binary quantifiers can be constructed by taking the meet/join closure of the set of (composed) reducible type < 2 > quantifiers. This fact is illustrated with non-constituent coordination constructions (e.g., gapping in English and Korean), whose interpretation requires arbitrary meets and joins of reducible binary quantifiers. Key words: expressive power, generalized quantifier, non-constituent coordination, polyadic quantifier, reducibility. 0. Introduction This paper addresses the following two related issues: (i) How to give a unified semantics of non-subject DPs (determiner phrases) as well as subject DPs in the spirit of Generalized Quantifier Theory; and (ii) Can we determine the expressive range of n-ary quantifiers (functions from n-ary relations into truth-values) in natural language? Regarding (i), this paper, based on type-polymorphism, generalizes the case-extension approach of Keenan (1987, 1988). This forces us to a somewhat more general than usual notion of n-ary relation – one motivated on grounds of linguistic analysis, specifically, the need to represent all possible scope ambiguities without inducing argument ambiguities. On the other hand, our polymorphic DP-type makes it possible to directly compose two DPs without invoking Type-lifting operations, as in Dowty (1988) 412 S. NAM and Steedman (1990). Equally, the case-extension approach leads us to certain observations concerning the uniform interpretation of DPs in English (cf. (5)) as well as a natural way of representing quantifier scope ambiguities (cf. Proposition 1). The second issue (ii) has been discussed by van Benthem (1989), Keenan and Moss (1985), Keenan (1992) and Keenan and Westerståhl (1997). Illustrating various types of unreducible binary quantifiers, they conclude that the expressive power of English goes well beyond (Fregean) reducible binary quantifiers. But they do not assess how far beyond. Here we argue that the possibility of forming non-constituent coordinations requires, in principle, all binary quantifiers (type 2 functions) as possible denotations of natural language expressions over a finite universe. The sort of non-constituent coordinations of concern here are those illustrated by (1) from English and (2) from Korean, where we get a conjunction (or disjunction) of binary quantifiers. In (1), (EVERY-BOY◦ A-NOVEL) and (EVERY-GIRL◦A-PLAY) are conjoined (or disjoined), and in (2), (JOHN◦APPLE) and (MARY◦PEAR) are conjoined. (1) Every boy read a novel, and/or every girl a play. (2) John-i sakwa-lul, (kuliko) Mary-ka pay-lul Bill-hantey cwuessta J.-Nom apple-Acc (and) M.-Nom pear-Acc B.-Dat gave ‘John gave an apple to Bill, and Mary gave a pear to Bill.’ We prove that such boolean compounds of n-ary quantifiers are not (Fregean) reducible to iterated applications of unary generalized quantifiers. Specifically, limiting ourselves to the functions from binary relations to truth values [R → 2], we show that conjunction and disjunction of binary quantifiers provide English with the full expressive power on [R → 2] over finite universe. (This paper uses the notation [A → B] to refer to the whole set of functions with domain A and range included in B.) 1. Case-extensions of Generalized Quantifiers Generalized quantifiers (GQs) are functions from properties into truth values [P → 2], i.e., sets of properties. In a simple intransitive sentence, a subject DP takes a one place predicate as an argument to give a truth value. The theory of GQs can be extended to non-subject DPs in a sentence. For example, in a transitive sentence, the object DP can be seen as a function taking a two place predicate (P2 ) to give a one place predicate (P1 ), then the subject takes the derived P1 to give a truth value (zero place predicate, P0 ). Fully generalizing this view, we can treat DPs polymorphically as denoting functions from (n + 1)-ary predicates into n-ary predicates, that is, we extend GQs to take (n + 1)-ary relations to give n-ary relations for all non-negative integers n.1 EXPRESSIVE POWER OF DP-COMPOSITIONS 413 Adopting the notation of Keenan and Timberlake (1988), we might sketch the interpretation of everyone loves someone as in (3). Here, we assign DPs the “polymorphic” type Pn /Pn+1 , which reveals the idea that a DP saturates one argument of a predicate, reducing its arity by one. An expression of this type will denote a single function whose domain is the set of all (n + 1)-ary relations for n ≥ 0.2 (3) everyone loves someone Pn/Pn+1 P2 Pn/Pn+1 P1: SOMEONE(LOVE) P0: EVERYONE(SOMEONE(LOVE)) This representation, however, forces the first generalized quantifier that applies to an n-place predicate to saturate its n-th argument. Thus SOMEONE in (3) saturates the second argument of LOVE and takes narrow scope relative to EVERYONE. We need, obviously, a notation for allowing scope assignment of DPs and the arguments they saturate to be independent. And while such representation is relatively unproblematic for binary relations, the linguistic issues loom larger for n ≥ 3-ary relations as in (4), containing a ternary relation: (4) Every teacher sents a letter to exactly three students. Traditionally, n-ary relations are treated as sets of ordered n-tuples of elements of the universe E. Interpreting send in (4) as a set of triples α, β, γ we see that there are three ways to derive a binary relation from it according to which argument is saturated first: (i) saturating the Recipient argument first we derive {α, β| α sent β to exactly three students}. (ii) Saturating the Theme argument first we derive {α, γ | α sent a letter to γ }, and (iii) saturating the Agent argument first we derive {β, γ | every teacher sent β to γ . The traditional treatment of relations which interprets any binary relation as a set of ordered pairs, however, does not distinguish such different types of binary relations derived from ternary ones. Now, let us suppose we interpret (4) in such a way that EVERY TEACHER has narrowest scope (cf. (iii)). Then, A LETTER is looking at a binary relation, but which argument is it to saturate? Intuitively, the first member of the set of pairs. But that is not the argument it would saturate if EXACTLY THREE STUDENTS had been interpreted with narrowest scope – then, the remaining argument for A LETTER to saturate would be the second of the remaining pairs (cf. (i)). Thus, intuitively what we want to be able to do here, as linguists, is to remember which arguments of an n-ary predicate remain when some have been saturated. This is what motivates our enriched notion of n-ary relations to be proposed shortly. 414 S. NAM Note that (i)–(iii) reflect different scope relations among the three argument DPs of send, i.e., (i) exactly three students gets interpreted as having narrow scope, and in (ii) a letter does, and in (iii) every teacher does. This suggests that if we allow the three DP functions to apply to SEND in any possible order, we can interpret scope ambiguities in a straightforward way. But we must be able to do this without altering the argument a DP saturates. This problem arises less dramatically if attention is restricted to just one and two-place predicates, since saturation of one argument leaves little choice for the remainder. So to capture our linguistic intuitions concerning argument choice and scope ambiguities, we shall in the first instance represent an n-ary sequence of elements of a universe E as a function s from {1, 2, . . . , n} into E. Each such s corresponds to a unique n-tuple s(1), s(2), . . . , s(n) and the map sending each such s to that n-tuple is easily seen to be a bijection. So n [{1, 2, . . . , n} → E ] is isomorphic to E . And we shall represent an n-ary n relation over E not directly as a subset of E but rather as a subset of [{1, 2, . . . , n} → E ], a set of functions with domain {1, 2, . . . , n} and range included in E. This approach to n-ary relations is just a notational variant of the standard representation. But it generalizes slightly in a way that is useful to us. Namely, for each finite subset K of positive integers we define a K-ary relation to be a set of functions from K into E. We write RK for the set of K-ary relations over E. Formally, DEFINITION 1. For all non-empty universes E, all finite sets K of positive integers, R ∈ RK iff R ⊆ [K → E]. If |K| = n we call R an n-ary relation over E. We note that when |K| = |K | then RK and RK are isomorphic. Our interest in the (slight) generalization is that we will treat R{2,5} , for example, as the set of binary relations derived from relations of arity 5 or greater by saturating all but the second and fifth arguments. We will indulge in one traditional notational simplification however. A unary relation (property) p over E should be a set of functions from e.g. {1} into E. So for example the set {a, b, c} on this representation is {1, a, 1, b, 1, c}. But this representation is uniquely recoverable from the set {a, b, c} by the function that sends any subset X of E to {1, x|x ∈ X}. So we shall continue to note subsets of E as usual, and P, the set of unary relations, is just the set of subsets of E, as usual. We think of determiner phrases initially as denoting functions from properties to truth values, and we extend them in various ways to take (n + 1)-ary relations to yield n-ary ones, according to the arguments they saturate. Following Keenan (1987, 1988), we call this way of extension EXPRESSIVE POWER OF DP-COMPOSITIONS 415 “case-extension,” since the Case of a DP (e.g., Nominative or Accusative, etc.) determines the argument it affects semantically. DEFINITION 2. For all f ∈ [P → 2], fi (i ≥ 1), or i-th case-extension of f , is defined as follows: For all finite sets K of positive integers, ∀R ∈ RK , R, if i ∈ / K and fi (R) = {h ∈ [K − {i} → E]|f {α|h ∪ {i, α} ∈ R} = True}, otherwise. DEFINITION 3. The set CXi of i-th case-extensions of generalized quantifiers is defined by CXi =df . {fi | f ∈ [P → 2]}. As stated in Definition 2, given a generalized quantifier f and an (n + 1)-ary relation R ∈ RK , if i ∈ K, fi (the i-th case-extension of f ) saturates the i-indexed argument of R, to yield an n-ary relation which is a set of functions from K −{i} into E, and thus a relation where the i-indexed argument has been saturated. Thus suppose SEND is a set of functions from K = {1, 2, 3} into E, and let F, G, H be DP denotations. Then, applying G2 to SEND, we get a binary relation G2 (SEND) = {f ∈ [{1, 3} → E]|f (1) sent G to f (3)}.3 In this way, we can express all the six possible scope readings without altering the argument saturated by F1 , G2 , or H3 : F1 (G2 (H3 (SEND))), F1 (H3 (G2 (SEND))), G2 (F1 (H3 (SEND))), G2 (H3 (F1 (SEND))), H3 (F1 (G2 (SEND))), H3 (G2 (F1 (SEND))). The set CXi of i-th case-extensions has a straightforward (if tediously shown) but important property: it is a complete, atomic boolean algebra (see the Case-extension Theorem in (Theorem 1)).4 Linguistically, in what follows, we need this because we want to interpret conjunctions, disjunctions, and negations of DPs as boolean meets, joins and complements, as in John saw [every teacher and exactly two students], and the intended denotation sets must be rich enough to guarantee this. We will see later that the simple composition classes we use to interpret non-constituent coordinations in (1) and (2) fail to satisfy all the boolean conditions, and it is the additional requirement of boolean closure that forces FULL expressive power. THEOREM 1. Case-extension For all i, (CXi , ≤) is a complete atomic boolean algebra, where for all f, g ∈ CXi , f ≤ g iff f (R) ⊆ g(R). We include the proof of this theorem in the appendix, since while basically straightforward, it is important to realize that the enriched notion of n-ary 416 S. NAM relation we are using does not induce any booleanly unexpected properties (like being undefined at certain points, etc.). The proof of Theorem 1 essentially shows that the case-extension algebra CXi is closed under pointwise meets, joins, and complements. For example, saw [every boy and some girl ] gets the same interpretation as saw every boy and saw some girl, and love [not [every teacher]] gets interpreted as not [love every teacher]. Further, we have the following: COROLLARY 1. For all i = j, CXi and CXj are disjoint, i.e., CXi ∩ CXj = φ, and CXi is isomorphic to CXj . So, each CXi is isomorphic to [P → 2], the set of generalized quantifiers. This shows the fact that the i-th extensions of the set of GQs just extends them to a larger domain, but does not really introduce any “new” functions. The algebraic characterization of case-extensions, as shown in Theorem 1, accounts for one basic universal: (5) Universal: In natural language, two DPs interpreted by different Caseextensions cannot be conjoined. Since every boolean algebra is closed under meets and joins, we can interpret a conjunction of DPs with the same case as denoting their greatest lower bound in their case-extension algebra. However, we cannot interpret conjunctions of DPs which are interpreted in different case-extensions CXi and CXj (i = j ), since boolean meets and joins of two elements lying in different algebras do not make sense. This accounts for the semantic unnaturalness of coordinations like he and him, and I and her. Note that this semantic unnaturalness is not to be confused with syntactic ill-formedness. A variety of languages – Nandi (Nilo-Saharan), Malagasy (Malayo-Polynesian), Irish (Celtic) – present surface coordinations of DPs in different morphological cases. Essentially the first DP conjunct goes in the case required by the syntactic construction, and the remaining conjuncts fall into a default case (Accusative in Irish and Nandi, Nominative in Malagasy). (6) below from Irish (McCloskey 1986:265) illustrates a case of default-case assignment in coordination. (6) Chuaigh se-isean agus Went he(Nom)-contr and ‘He and he went home.’ e-isean ‘na bhaile him(Acc)-contr home Thus whether he and him sounds odd in a language depends in part on the mechanism assigning default morphological case. Our claim in (5) above, however, is a claim about how expressions may be semantically interpreted, regardless of their surface form. One is tempted to generalize (5) as follows: 417 EXPRESSIVE POWER OF DP-COMPOSITIONS (7) Speculation: For all type-preserving binary operators natural language, [DP∗ DP ]i = DP∗i DPi .5 ∗ on DPs in That is, case-extension preserves the binary operation ∗ , so the speculation does hold when ∗ is a Boolean operator: [DP and DP ]i = DPi and DPi . 2. Functional Composition and n-ary Quantifiers We now consider the increase in expressive power forced by non-constituent coordinations (e.g., gapping and right/left-node raising) illustrated by English (8) and Korean (9) (or equivalently (1) and (2) mentioned earlier). It is basically this increase which forces full finite type 2 expressive power. (8) a. The teacher showed [every girl two plays], and [every boy three novels] b. John loves Mary, and Harry Sue. (9) a. John-i Mary-lul, (kuliko) Harry-ka Sue-lul salanghanta J-Nom M-Acc (and) H-Nom S-Acc love ‘John loves Mary, and Harry Sue.’ b. John-i Mary-lul, (kuliko) Harry-ka Sue-lul Bill-hantey sokayhayssta J-Nom M-Acc (and) H-Nom S-Acc B-Dat introduced ‘John introduced Mary to Bill, and Harry introduced Sue to Bill.’ Having interpreted DPs as functions with all (n + 1)-ary relations in their domain, pairs of DPs like [every-girl two plays] or [John-i Mary-lul ] are naturally interpreted as the composition of the elements of the pair, yielding a function from (n + 2)-ary relations into the n-ary ones. Formally, for example, JOHN1 ◦ MARY2 will be an element of the composition class CX1 ◦ CX2 , defined in general by DEFINITION 4. For all case-extension algebras CXi and CXj , CXi ◦CXj =df . {f◦g|f ∈ CXi , g ∈ CXj and f◦g(R) = f (g(R))}. Elements of CXi ◦ CXj are called basic composite functions of an i-th caseextension and a j -th case-extension. One advantage of our polymorphic DP-type (Pn /Pn+1 ) is that we can directly compose two DPs without Type-changing operations like Lifting which is necessary for Dowty (1988) and Steedman’s (1990) account of DP-compositions. For example, they lift two DPs in different ways to allow them to be composed: thus, DP can be lifted to either (S/DP)/((S/DP)/DP) or ((S/DP)/DP)/(((S/DP)/DP)/DP) or higher types (directionality of slash ignored). We also note here that our DP-composition approach based on case-extension is more linguistically motivated than Hendriks’ (1988) type 418 S. NAM raising mechanism. Hendriks applies type changing operations (i.e., Argument Raising or Lowering) to “verbs” so as to derive different scope relations among DPs. For example, ditransitive verbs with three arguments like send can be assigned six different functional types, each of which derives a unique scope reading according to its order of the semantic saturation of the argument DPs. However, verbs in natural language do not indicate its possible saturation orders or the scope-relations among arguments. Instead, DPs usually carry a mark to indicate their CASE, focal/topical features, etc. Further, the semantic characteristics of DPs – e.g., monotonicity, (in)definiteness, etc. – constrain their scope interaction with other scopal elements. Let us also briefly compare our semantics with Oehrle’s (1987). In order to interpret gapping structures in general, Oehrle first takes a transitive verb like eat as a function f mapping the Cartesian product NP×NP to truth values, i.e., eat denotes a function EAT such that EAT(np1 , np2 ) → EAT(np2 )(np1 ). Then he extends this function f to f ∗ , which acts on conjoined/disjoined/negated pairs of arguments, so the domain of f ∗ is the boolean closure of [NP×NP], which he labels L[NP×NP]. Oehrle, without proof, takes the domain of f ∗ as a distributive lattice, and claims that each f ∗ is a unique homomorphism preserving meets (∧), joins (∨), and complements (–) in L[NP × NP]. That is, f ∗ (x ∧ y) = f ∗ (x) ∧ f ∗ (y), and f ∗ (x ∨ y) = f ∗ (x) ∨ f ∗ (y), and f ∗ (−x) = −(f ∗ (x)). Oehrle (1987) only deals with the gapping structures which contain individual-denoting NPs, and he interprets the conjoined/disjoined NP-pairs as an argument of the verb. Our semantics proposed here, however, deals with full range of boolean compounds of pairs of generalized quantifiers, and each pair of generalized quantifiers denotes a composed function, i.e., a binary quantifier. Now we note two basic properties of the CXi ◦ CXj defined under Definition 4: namely, Propositions 1 and 2. PROPOSITION 1. Scope Dependency For all i = j , a. CXi ◦CXj = CXj ◦CXi and b. F ∈ CXi ◦ CXj ∩ CXj ◦CXi if ∃f, g such that F = fi ◦gj = gj ◦fi . It is this which accounts in general for the possibility of generalized quantifier scope ambiguities. The proof of Proposition 1a given in Appendix shows that the different composition orders of the universal and existential quantifiers determine different type 2 functions (binary quantifiers): (EVERY-E)i ◦(SOME-E)j = (SOME-E)j ◦(EVERY-E)i , thus scope dependency is just failure of commutativity. However, (EVERY-E)i ◦(SOME-E)j ∈ CXi ◦CXj − CXj ◦CXi , whence the two composition classes are distinct. EXPRESSIVE POWER OF DP-COMPOSITIONS 419 As stated in Proposition 2b, if F = fi ◦gj = gj ◦fi , there would not arise a scope ambiguity between fi and gj , and F is in both CXi ◦ CXj and CXj ◦ CXi . For instance, as proved by Zimmerman (1987), proper names are scopeless, whence [JOHN1 ◦NO-STUDENT2 ] is the same function as [NO-STUDENT2 ◦ JOHN1 ], and so shared by CX1 ◦ CX2 and CX2 ◦ CX1 . PROPOSITION 2. a. CXi ◦ CXj is closed under complements (since −(fi ◦ gj ) = −fi ◦ −gj ), and b. CXi ◦ CXj is not closed under pointwise meets or joins. This last property is of our special interest, since the form of the sentences in (8) and (9) suggests precisely that we must take meets and joins of basic composite functions. Keenan (1992) points this out in terms of Fregean reducibility of Type 2 functions. In what follows, using this method of testing reducibility (Theorem 2), we prove that such boolean compounds of DP-compositions are not interpretable by basic composite functions in CXi ◦CXj (see also van Benthem’s (1989:446) theorem). Further, we argue that non-constituent coordinations in natural language force FULL finite type 2 expressive power, i.e., Type 2 Effability. Type 2 functions are ones from binary relations into truth-values, and type 1 functions are ones from properties into truth values. We will use these terms in broader sense to refer to functions from (n + 2)-ary relations to n-ary ones, and functions from (n + 1)-ary relations to n-ary ones, respectively. Following Keenan (1992), we define, DEFINITION 5. A type 2 function F is Fregean (= reducible) iff there are type 1 functions f and g such that F = f ◦g. In effect, with the generalized notion of ‘type 2’ functions, the set of (Fregean) reducible functions from (n + 2)-ary relations into n-ary relations is ∪i=j CXi ◦ CXj . Keenan (1992) uses the following theorem as a basis for testing (Fregean) reducibility of type 2 functions. THEOREM 2. Reducibility Equivalence (RE ) Let F, G be reducible functions of type 2. Then, F = G iff for all subsets P , Q of E, F (P ×Q) = G(P ×Q). This theorem says that the values of a type 2 reducible function on the cross product relations (P ×Q) determine its values at any binary relation.6 Thus, to show a function F to be unreducible, we only have to find a reducible function which gives the same values to all the cross product 420 S. NAM relations but which is not identical to F . We refer Keenan (1992) for a variety of linguistic contexts where unreducible type 2 functions derive. Specifically, he shows that the compositions of a universal quantifier and an anaphoric DP-function (e.g., (EVERY-STUDENT)◦HIMSELF) are not (Fregean) reducible. As illustrated earlier in (8) and (9), non-constituent coordinations involve conjunctions or disjunctions of DP-compositions, but such conjunction or disjunction is not guaranteed to be interpreted as a reducible type 2 function, i.e., a function in CXi ◦ CXj . For example, each of the composite functions in (8b) and (9), (JOHNnom ◦MARYacc ) and (HARRYnom ◦SUEacc ), is reducible, i.e., they are in CXnom ◦CXacc , but we prove that their conjunction F = (JOHNnom ◦MARYacc )∧(HARRYnom ◦SUEacc ) is not (Fregean) reducible: To prove this, we find a reducible function G = F such that for all product relations R, G(R) = F (R). Let G = (JOHNnom ∧ HARRYnom )◦(MARYacc ∧ SUEacc ) ∈ CXnom ◦CXacc . Then, G takes the same values at all cross product relations R as F does. We see that F = G, since, for all binary relations R, F (R) = True iff John and Harry bear the relation to Mary and Sue, respectively, but G(R) = True iff each of John and Harry bears the relation to each of Mary and Sue. Therefore, by Theorem 2, F = (JOHNnom ◦ MARYacc ) ∧ (HARRYnom ◦ SUEacc ) is not reducible, i.e., F ∈ / CXnom ◦ CXacc . Equally, for suitable non-trivial choice of GIRL and PLAY etc., the function [EVERY-GIRLdat ◦ TWO-PLAYSacc ] ∧ [EVERY-BOYdat ◦ THREE-NOVELSacc ] needed to interpret (8a) is unreducible. Essentially, this proves Proposition (2), that is, CXi ◦ CXj is not closed under meets and thus under joins since meets can be defined in terms of joins and complements, and so forces us to extend the set of possible denotations of DP-compositions. In other words, the nonconstituent coordinations force expressive power beyond the (Fregean) reducible functions in CXi ◦ CXj . The following gapping examples reveal that boolean compounds of composite functions can be obtained in various ways: (10) a. Neither does John love Mary, nor Fred Susan. = John doesn’t love Mary, or Fred Susan. b. Either John loves Mary, or Fred Susan. For example, (10a) can be paraphrased as ‘John doesn’t love Mary, and Fred doesn’t love Susan,’ and so we get a type 2 function, -(JOHN1 ◦ MARY2 ) ∧ -(FRED1 ◦ SUSAN2 ) which takes a binary relation to give a truth value. To prove this function to be unreducible, we observe first from Keenan (1992): EXPRESSIVE POWER OF DP-COMPOSITIONS 421 (11) If F of type 2 is unreducible then so is its (boolean) complement -F , its post-complement F -, and its dual -F -, where these notions are defined as follows:7 For all functions F of type n, n ≥ 1, and R ⊆ E n , i.e., an n-ary relation, a. -(F )(R) = -(F (R)) b. (F -)(R) = F (-R) c. -F - = -(F -) = (-F )Observe also: (i) the post-complement operation distributes over meets: ((F ∧G)-)(R) = (F ∧G)(-R) = F (-R) ∧ G(-R) = (F -)(R) ∧ (G-)(R) = (F - ∧ G-)(R), thus (F ∧ G)- = (F - ∧ G-). Dually (F ∨ G)- = (F -) ∨ (G-). And (ii) individual denoting functions like JOHN1 and MARY2 are self-dual, i.e., john1 = -JOHN1 -. That is, John laughed if and only if it is not the case that John didn’t laugh. Then the type 2 function derived from (10a) can be reduced as the post-complement of (JOHN1 ◦ MARY2 ) ∧ (FRED1 ◦ SUSAN2 ) as shown below:8 (12) -(JOHN1 ◦ MARY2 ) ∧ -(FRED1 ◦ SUSAN2 ) = -(-JOHN1 - ◦ -MARY2 -) ∧ -(-FRED1 - ◦ -SUSAN2 -) = (JOHN1 ◦ MARY2 )- ∧ (FRED1 ◦ SUSAN2 )= [(JOHN1 ◦ MARY2 ) ∧ (FRED1 ◦ SUSAN2 )]- We have already shown that (JOHN1 ◦ MARY2 ) ∧ (FRED1 ◦ SUSAN2 ) is unreducible. Thus, due to (11), its post-complement we get from (10a) is also unreducible, i.e., the function in (12) is not in CXi ◦ CXj . (10b) above gives a disjunction of two composite functions, that is, (JOHN1 ◦ MARY2 ) ∨ (FRED1 ◦ SUSAN2 ). This is type 2 function and (13) shows it to be the dual of (JOHN1 ◦ MARY2 ) ∧ (FRED1 ◦ SUSAN2 ), and thus unreducible. (13) (JOHN1 ◦ MARY2 ) ∨ (FRED1 ◦ SUSAN2 ) = -[-(JOHN1 ◦ MARY2 ) ∧ -(FRED1 ◦ SUSAN2 )] De Morgan Law = -[(JOHN1 ◦ MARY2 ) ∧ (FRED1 ◦ SUSAN2 )] from (11) above This completes the proof of Proposition 2: CXi ◦CXj is not closed under meets or joins. The following examples further illustrate that we need a much extended set of type 2 functions to give proper interpretations to English expressions. (14) a. Either John interviewed Mary, and Fred Susan, or else John Susan, and Fred Mary. b. John interviewed Mary, and Fred Susan, but neither John Susan, nor Fred Mary. Let α = (JOHN1 ◦ MARY2 ), β = (FRED1 ◦ SUSAN2 ), γ = (JOHN1 ◦ SUSAN2 ), and δ = (FRED1 ◦ MARY2 ), then from (14a and b) above, we get the following 422 S. NAM boolean compounds of type 2 functions: (a) (α ∧ β) ∨ (γ ∧ δ), (b) (α ∨ β) ∧ (γ ∨ δ), (c) (α ∧ β) ∧ (-γ ∧ -δ). These complex type 2 functions suggest that we need to extend the possible denotations of DP-compositions to the closure of basic composite functions CXi ◦CXj under pointwise meets and joins, noted CF(CXi ◦CXj ) in Definition 6. Further, from Proposition 3, we see that CF(CXi ◦CXj ) is closed under complements, too. DEFINITION 6. For all CXi and CXj , CF(CXi ◦CXj ), or the set of composite functions of CXi and CXj , is the closure of CXi ◦CXj under arbitrary meets and joins pointwise. PROPOSITION 3. CF(CXi ◦ CXj ) is closed under complements. It follows then that for all F, G ∈ CF(CXi ◦ CXj ), (F ∧ G)(R) = F (R) ∩ G(R), and (F ∨ G)(R) = F (R) ∪ G(R), and (-F )(R) = -(F (R)). And one shows: THEOREM 3. Composite Algebra (CF(CXi ◦ CXj ), ≤) is a complete atomic boolean algebra, where for all F, G ∈ CF(CXi ◦ CXj ), F ≤ G iff for all R ∈ RK (|K| ≥ 1), F (R) ⊆ G(R). For example, the zero and the unit elements of the algebra are given as follows: ∀R ∈ RK , R, if i ∈ / K or j ∈ / K and (a) 0CF (R) = Ø, otherwise. R, if i ∈ / K or j ∈ / K and (b) 1CF (R) = [K−{i, j } → E], otherwise. In effect, 0CF = 0CXi ◦ 0CXj , and 1CF = 1CXi ◦ 1CXj . Thus they are in CF(CXi ◦CXj ), and ∀F ∈ CF(CXi ◦ CXj ), 0CF (R) ⊆ F (R), and F (R) ⊆ 1CF (R). From Definition 6, we have that, for all F, G ∈ CF(CXi ◦ CXj ), the greatest lower bound and the least upper bound of {F, G} are (F ∧ G) and (F ∨ G), respectively. And, ∀F ∈ CF(CXi ◦ CXj ), hF , the complement of F , is given by: for all R ∈ RK , R, if i ∈ / K or j ∈ / Kand hF (R) = [K−{i, j }→E] − F (R), otherwise. Then (hF ∧)(R) = 0CF (R), and (hF ∨ F )(R) = 1CF (R). EXPRESSIVE POWER OF DP-COMPOSITIONS 423 To establish the major result of this section – Type 2 effability (Theorem 4), first, we restrict the domain of type 2 functions F ∈ CF(CX1 ◦ CX2 ) to the set of binary relations, R{1,2} , and characterize the atoms of CF(CX1 ◦ CX2 ) with the restricted domain. As we saw above, CF(CX1 ◦CX2 ) is a boolean algebra, and the atoms of the algebra can be given as follows: define, for all R ∈ R{1,2} , FR = df . (∧f ∈R (If (1) ) ◦ (If (2) )) ∧ -(∨g∈R / (Ig(1) ) ◦ (Ig(2) )) = (∧f ∈R (If (1) ) ◦ (If (2) )) ∧ (∧g∈R / -(Ig(1) ) ◦ (Ig(2) )), where for all b ∈ E, Ib is an individual (ultrafilter) in (℘ (E), ≤) generated by the property {b} ∈ ℘ (E). Thus, for all S ∈ R{1,2} , FR (S) = True iff S = R. One sees easily that FR is a minimal non-zero element, that is, an atom: It is not zero since FR (R) = True; and since it holds of just one relation, no other functions but zero can be strictly less than it. And if G is a non-zero function in CF(CX1 ◦ CX2 ) – so G holds of some R, then FR ≤ G. Therefore, every such G dominates an atom, and so the algebra is atomic. Finally, we have: THEOREM 4. Type 2 Effability Let R{1,2} be the set of binary relations, i.e., ℘[{1, 2}→E]. Then, F ∈ [R{1,2} →2] iff F ∈ CF(CX1 ◦ CX2 ) with the restricted domain R{1,2} . Proof (i) Trivially, from the definition of CF(CX1 ◦CX2 ), for all F ∈ CF(CX1 ◦CX2 ) with the domain restricted to R{1,2} , F ∈ [R{1,2} →2]. (ii) Now, for R∈R{1,2} , define FR as follows: FR (S) = True iff S = R. Thus, FR = (∧f ∈R (If (1) )◦(If (2) )) ∧ -(∨g∈R / (Ig(1) )◦(Ig(2) ) = (∧f ∈R (If (1) )◦(If (2) )) ∧ (∧g∈R / -(Ig(1) ) ◦ (Ig(2) )). (The same definition of FR is given on the previous page.) Since f and g above are functions in [{1, 2} → E], both (If (1) ) ◦ (If (2) ) and -(Ig(1) ) ◦ (Ig(2) ) are in CX1 ◦CX2 . Thus, each R ∈ R{1,2} , FR is in CF(CX1 ◦CX2 ), since it is a boolean function of basic composite functions of the form (Ia ◦Ib ). Now, for all F ∈ [R{1,2} →2], F = ∨R,F (R)=1 FR , thus any type 2 function F can be represented as joins of meets of the reducible functions in CX1 ◦ CX2 . Therefore, for all type 2 functions F ∈ [R{1,2} →2], F ∈CF(CX1 ◦CX2 ). This proves that CF(CX1 ◦CX2 ), the closure of (CX1 ◦CX2 ) under meets and joins, contains all the functions from binary relations into truth values. In other words, conjunction and disjunction in gapping sentences in English provide the FULL expressive power on (un)reducible type 2 functions from binary relations into truth values over finite universe E, since joins of atoms can be denoted by finite disjunctions. 424 S. NAM 3. Concluding Remarks The first half of this paper argues that case-extensions of GQs are motivated by linguistic intuitions concerning grammatical relations and scope ambiguities between quantifiers. Further, proving that each set of case-extensions form a complete atomic boolean algebra, we saw that the algebraic semantics proposed here accounts for the universal (5): Two DPs interpreted by different Case-extensions cannot be conjoined. Type-theoretically, our polymorphic DP-type Pn /Pn+1 allows direct composition of DPs without invoking Type-lifting operations. In section 2, we built up Composition algebras (Definition 6, Proposition 3, and Theorem 3) from the Case-extension algebras (Definition 2 and 3, Theorem 1) by taking the meet/join closure of the set of reducible type 2 functions. This is shown to include all type 2 functions and is needed as the denotation set for the non-constituent coordinations exhibited in the text. Thus such boolean compounds tell us that all type 2 functions over a finite universe are needed as possible denotations of natural language expressions. Acknowledgements I would like to thank Edward Keenan for comments on earlier versions of this article. I also thank anonymous reviewers for their critiques. All errors are my own. Appendix Proof of Theorem 1. Case-extension Theorem. For all i, (CXi , ≤) is a complete atomic algebra, where for all f, g ∈ CXi , f ≤ g iff f (R) ⊆ g(R). Proof. (i) Since the relation (≤) is defined in terms of subset relation (⊆), it is a partial order, i.e., reflexive, antisymmetric, and transitive. (ii) The zero (0CXi ) and the unit (1CXi ) elements: Define, for all K, and ∀R ∈RK , Ø, if i ∈ K and (a) 0CXi (R) = R, otherwise. [K−{i}→E], if i ∈ K and (b) 1CXi (R) = R, otherwise. We show the functions, 0 and 1 given as the above, are the zero and the unit element of CXi : ∀f ∈ CXi , ∀R ∈ RK , if i ∈ / K, f (R) = R, thus f (R) = EXPRESSIVE POWER OF DP-COMPOSITIONS 425 0(R) = R; and if i ∈ K, 0(R) = Ø ⊆ f (R). Therefore, ∀f ∈ CXi , 0 ≤ f . Dually, ∀f ∈ CXi , f ≤ 1. (iii) We show for all F, G ∈ CXi , {F, G} has a greatest lower bound and a least upper bound: Let F = fi , and G = gi for f, g ∈ [P→2]. Define hf,g as follows: for all R ∈ RK , fi (R) ∩ gi (R), if i ∈ K and hf,g (R) = R, otherwise. Then, we show hf,g is the greatest lower bound of {fi , gi }: First, we show hf,g ∈ CXi . If i ∈ / K, hf,g (R) = R = (f ∧ g)i (R); and if i ∈ K, (f ∧ g)i (R) = {h ∈ [K−{i}→E]|(f ∧ g){α|h ∪ {i, α} ∈ R} = 1} (by Definition 2) = {h ∈ [K−{i}→E]|f {α|h ∪ {i, α} ∈ R} ∧ g{α|h ∪ {i, α} ∈ R} = 1} = {h ∈ [K−{i}→E]|f {α|h ∪ {i, α} ∈ R} = 1} ∩ {h ∈ [K−{i}→E]|g{α|h ∪ {i, α} ∈ R} = 1} = fi (R) ∩ gi (R). Thus, hf,g = (f ∧ g)i ∈ CXi . Obviously, for all R ∈ RK , if i ∈ K, hf,g (R) ⊆ fi (R) and hf,g (R) ⊆ gi (R); and if i ∈ / K, hf,g (R) = fi (R) = R. Thus hf,g is a lower bound of {fi , gi }. Now, for all lower bounds w for {fi , gi }, w(R) ⊆ fi (R), and w(R) ⊆ gi (R), thus w(R) ⊆ fi (R) ∩ gi (R) = hf,g (R). Therefore hf,g is the greatest lower bound of {fi , gi }. Dually, we prove that (f ∨ g)i is the least upper bound of {fi , gi }. (iv) Now we show CXi is complemented: For all fi ∈ CXi , and R ∈ RK , define hf i as follows: R, if i ∈ / K and hf i (R) = [K−{i}→E] − fi (R), otherwise. To show hf i is the complement of fi , we show hf i ∈ CXi , and (hf i ∧ fi ) = 0CXi , and (hf i ∨ fi ) = 1CXi . To show hf i ∈ CXi , we show hf i = (-f )i : From the definition of Case-extension in Definition 2, for all R ∈ RK , if i ∈ / K, (-f )i (R) = R = hf i (R); and if i ∈ K, (-f )i (R) = {h ∈ [K−{i}→E]|(-f ){α|h ∪ {i, α} ∈ R} = True} = {h ∈ [K−{i}→E]|f {α|h ∪ {i, α} ∈ R} = F alse} = [K−{i}→E] − fi (R) = hf i (R). Thus, hf i = (-f )i ∈ CXi . Now, for all R ∈ RK , if i ∈ K, (hf i ∧ fi )(R) = hf i (R) ∩ fi (R) = ([K−{i}→E] − fi (R)) ∩ fi (R) = φ = 0CXi (R); and if i ∈ / K, (hf i ∧ fi )(R) = hf i (R) ∩ fi (R) = R ∩ R = R = 0CXi (R). 426 S. NAM Thus, (hf i ∧ fi ) = 0CXi . Dually, (hf i ∨ fi ) = 1CXi . Therefore, CXi is complemented. From (i) to (iv) above, CXi is a boolean algebra. To show CXi is atomic and complete, first we prove that the Case-extension map f ⇒ fi defined in Definition 2 is an isomorphism from [P→2] to CXi : we have already shown in (iii) and (iv) above that Case-extension preserves meets, joins, and complements, i.e., ∀f, g ∈ [P→2], (f ∧ g)i = fi ∧ gi , (f ∨ g)i = fi ∨ gi , and (-f )i = -(fi ), thus Caseextension is a homomorphism from [P→2] to CXi . And from the definition of Case-extension, we have that it is a surjection from [P→2] to CXi : that is, for all g ∈ CXi , there is a function f ∈ [P→2] such that fi = g. Now we show that Case-extension is an injection from [P→2] into CXi : For all f = g ∈ [P→2], ∃p ∈ P, such that, without loss of generality, f (p) = 1 and g(p) = 0. Set ψ ∈ RK where K = {i}, i.e., ψ ⊆ [{i}→K], such that ψ = {i, α|α ∈ p}. Then, from (Definition 2), fi (ψ) = {h ∈ [Ø→E]|f {α|h ∪ {i, α} ∈ ψ} = True} = {Ø}, but gi (ψ) = {h ∈ [Ø→E]| g{α|h ∪ {i, α} ∈ ψ} = True} = Ø. Thus, for all f = g ∈ [P→2], fi = gi ∈ CXi , and so Case-extension is an injection. Now by the following Lemma, we see that CXi is a complete and atomic boolean algebra. LEMMA. Let B = B, ≤B be isomorphic to D = D, ≤D , for B,D boolean algebras. Then, (a) if B is atomic then so is D, and (b) if B is complete then so is D. Here, the atoms of CXi are i-th case-extensions of the atoms in [P→2], thus, for all atoms Fp in [P→2], the isomorphic images (Fp )i in CXi are given by Fp = (∧a∈p Ia ) ∧ -(∨b∈p / Ib ); and (Fp )i = (∧a∈p (Ia )i ) ∧ -(∨b∈p / (Ib )i ). Where for all b ∈ E, Ib is an individual (ultrafilter) in (℘ (E), ≤) generated by the property {b} ∈ ℘ (E). Further, for all F ⊆ CXi . the greatest lower bound (∧F ) and the least upper bound (∨F ) of F are given by: ∧F (R) = ∩{f (R)|f ∈ F }; and ∨F (R) = ∪{f (R)|f ∈ F }. Proof of Proposition (1a): Scope Dependency For all i = j, CXi ◦ CXj = CXj ◦ CXi . Proof. To prove this, we find a function F such that F ∈ CXi ◦ CXj but F∈ / CXj ◦ CXi . Let F = (EVERY-E)i ◦ (SOME-E)j , where the universe E has at EXPRESSIVE POWER OF DP-COMPOSITIONS 427 least two elements. We represent this type 2 function as the following: F = (∧a∈E Ia )i ◦ (∨b∈E Ib )j ∈ CXi ◦ CXj , where for all b ∈ E, Ib is an individual (ultrafilter) in (℘ (E), ≤) generated by the property {b} ∈ ℘ (E). Then, F = (∧a∈E ∨b∈E (Ia )i ◦ (Ib )j ) = (∧a∈E ∨b∈E (Ib )j ◦ (Ia )i ). To show F ∈ / CXj ◦ CXi , that is, F is not a reducible type 2 function in CXj ◦ CXi , from the Reducibility equivalence theorem (Theorem 2), we find a distinct reducible type 2 function ϕ ∈ CXj ◦ CXi such that for all product relations R ∈ RK over E (where |K| = 2), F (R) = ϕ(R). Observe that R ∈ R{i,j } is a product relation iff ∃P , Q ⊆ E such that R = {f ∈ [{i, j }→P ∪ Q]|f (i) ∈ P and f (j ) ∈ Q}. Set ϕ = (SOME-E)j ◦ (EVERY-E)i = (∨b∈E Ib )j ◦ (∧a∈E Ia )i ∈ CXj ◦CXi . Then for all product relations R = P ×Q ∈ RK , (i) if K = {i, j }, F (R) = (EVERY-E)i ◦(SOME-E)j (R) = True iff P = E and Q = Ø; and ϕ(R) = (SOME-E)j ◦(EVERY-E)i (R) = T rueiff P = E and Q = Ø. And if K = {i, j }, by the definition of case-extension in Definition 2, we have: (ii) if i, j ∈ / K, F (R) = (EVERY-E)i ((SOME-E)j (R)) = (EVERY-E)i (R) = R, ϕ(R) = (SOME-E)j ((EVERY-E)i (R)) = (SOME-E)j (R) = R; (iii) if i ∈ K, and j ∈ / K, F (R) = ϕ(R) = (EVERY-E)i (R); and (iv) if i ∈ / K, and j ∈ K, F (R) = ϕ(R) = (SOME-E)j (R). Thus, from (i)–(iv), for all product relations R ∈ RK , |K| = 2, F (R) = ϕ(R). Now we see F = ϕ from the following: Suppose R is the identity relation, i.e., R = {f ∈ [{i, j }→E]|f (i) = f (j )}, then F (R) = (EVERY-E)i ◦ (SOME-E)j (R) =True, but ϕ(R) = (SOME-E)j ◦ (EVERY-E)i (R) = False. Therefore, by the Reducibility Equivalence Theorem, F is not reducible to a basic composite function in CXj ◦ CXi , i.e., F ∈ / CXj ◦ CXi . Notes 1 Composing two DP-functions, we would get a function from (n + 2)-ary relations to n-ary relations, which saturates two arguments. In section 2, the category name Pn /Pn+2 is used for the compositions of two DPs. 2 In this paper, the polymorphic types are used exactly in the same spirit as Keenan and Timberlake’s (1988) use of “n-tuple categories.” Moortgat (1988) takes polymorphic types as ambiguous ones brought in either by Type-changing operations (e.g., lifting) or by assigning incompletely instantiated categories to lexical items like and and or. 3 Sometimes we write fnom(inative) for f1 , and facc(usative) for f2 , and so forth. 4 What we have actually defined in Definition 2 and 3 is a boolean lattice, but as they are interdefinable with boolean algebras we shall use this latter term for its greater familiarity. See Keenan and Faltz (1985) for the definitions. 5 This equality fails if * is replaced by ◦ (composition), but the latter is not typepreserving. The domain of two composed DP functions is a proper subset of their common domain. 428 S. NAM R ∈ R{i,j } is a product relation iff ∃P , Q ⊆ E such that R = {f ∈ [{i, j } → P ∪ Q]|f (i) ∈ P and f (j ) ∈ Q} and we write standardly R = P × Q. 7 This proposition says that the set of unreducible type 2 functions is closed under postcomplement and dual as well as complement. Further, it follows that the set of reducible type 2 functions, CXi ◦CXj , is also closed under these operations. 8 We use the following facts: for all functions f and g, (i) (f - ◦ -g)(R) = (f -)(-g(R)) = (f -)(-(g(R))) = f (g(R)) = f ◦ g(R), and (ii) -(-f ◦ g)(R) = -(-f (g(R))) = -(-(f (g(R)))) = f (g(R)) = f ◦ g(R), and (iii) (f ◦ g-)(R) = f ((g-)(R)) = f (g(-R)) = (f ◦ g)(-R) = ((f ◦ g)-)(R). 6 References Dowty D. (1988) Type Raising, Functional Composition, and Non-Constituent Conjunction. In: Oehrle R. T., Bach E., Wheeler D. (eds.), Categorial Grammar and Natural Language Structures. D. Reidel, Dordrecht. Hendriks H. (1988) Type Change in Semantics: The Scope of Quantification and Coordination. In: Klein E. and van Benthem J. (eds.), Categories, Polymorphism and Unification. Center for Cognitive Science, University of Edinburgh, and Institute for Language, Logic and Information, Amsterdam. Keenan E. L. (1987) Semantic Case Theory. In: Groenendijk J. and Stokhof M. (eds.), Proceedings of the Sixth Amsterdam Colloquium. Foris Publications, Dordrecht. Keenan E. L. (1988) On Semantics and Binding Theory. In: Hawkins (ed.), Explaining Language Universals. Basil Blackwell, New York. Keenan E. L. (1992) Beyond the Frege Boundary. Linguistics and Philosophy 15, pp. 199–221. Keenan E. L. and Faltz L. M. (1985) Boolean Semantics for Natural Language. D. Reidel, Dordrecht. Keenan E. L. and Moss L. (1985) Generalized Quantifiers and the Expressive Power of Natural Language. In: van Benthem J. and ter Meulen A. (eds.), Generalized Quantifiers in Natural Languages. GRASS Vol. 4. Keenan E. L. and Timberlake A. (1988) Natural Language Motivation for Extending Categorial Grammar. In: Oehrle, R. T. et al. (eds.), Categorial Grammar and Natural Language Structures. D. Reidel, Dordrecht. Keenan E. L. and D. Westerståhl (1997) Generalized Quantifiers in Linguistics and Logic. In: van Benthem J. and ter Meulen A. (eds.), Handbook of Logic and Language. Elsevier, Amsterdam. McCloskey J. (1986) Inflection and Conjunction in Modern Irish. Natural Language and Linguistic Theory 4(2), pp. 245–281. Moortgat M. (1988) Categorial Investigation: Logical and Linguistic Aspects of the Lambek Calculus. ICG Printing, Dordrecht. Oehrle R. T. (1987) Boolean Properties in the Analysis of Gapping. In: Huck, G. J. and Ojeda A. E. (eds.), Syntax and Semantics, Vol. 20. Academic Press, New York. Sag I. (1977) Deletion and Logical Form. Doctoral dissertation, MIT. Steedman M. J. (1990) Gapping as Constituent Coordination. Linguistics and Philosophy 13, pp. 207–263. Van Benthem J. (1989) Polyadic Quantifiers? Linguistics and Philosophy 12, pp. 437–464. Zimmermann E. (1987) Transparent Adverbs and Scopeless Quantifiers. In: Groenendijk J., de Jongh D. and Stokhof M. (eds.), Foundations of Pragmatics and Lexical Semantics. Foris Publications, Dordrecht.
© Copyright 2026 Paperzz