Relational adjectives as properties of kinds

ing and generalizing, the schema for the type they propose for relational adjectives is that represented in (25), where the adjective is of type <,> – in this sense, a predicate modifier – but effectively intersective in the sense that it introduces a first-order property which is predicated of one of the modified noun’s arguments. (25) λPλx/y [P(x, ..., y) ∧ N(x/y)], where N is the noun from which the adjective is derived. F&K;’s analysis, though technically a predicate modifier analysis, is intersective in spirit, and is similar to ours in that the first order property introduced by the relational adjective does not (generally) modify the referent of the modified nominal. However, the similarities end there: F&K;’s central concern is to account for the apparent argument-saturating effect of the relational adjective, while, as will become clear below, this is not our first priority. We will leave additional comments on the differences between our analysis and F&K;’s until the final section of the paper. 4.2 A Larsonian intersective semantics As noted above, our proposal is inspired in Larson’s analysis of event-related adjectives (Larson 1998). Larson proposed that certain adjectives (in fact, many) denote properties of events rather than, or in 7 The most extensive analysis of relational adjectives (though not under that name) in the generative linguistics tradition is probably Levi’s (Levi 1978). Levi’s transformational syntactic account treated relational adjectives as nouns at Deep Structure, which were converted to adjectives in the course of deriving Surface Structure. Although she does not offer any explicit semantic type assignment for these adjectives, her analysis of predicative uses of them, sketched briefly in footnote 3, strongly suggests that her conception of relational adjectives was very close to a predicate modifier analysis. See also Bolinger 1967 for extensive informal discussion of these adjectives, which suggests an analysis similar to that proposed by Levi. 8 F&K; allow for the possibility that the adjective might modify any of the noun’s arguments, hence the “x/y” notation; however, this aspect of their analysis is not crucial for our purposes and we will not comment on it further here. 188 L. McNally & G. Boleda addition to, denoting properties of ordinary individuals; thus, an adjective like bona, ‘good’, could be translated as in (26a). In addition, he posited that nouns quite generally have an event argument in addition to their other, more familiar arguments, as illustrated with violinista in (26b). On the event related reading of the adjective, the adjective modifies the event argument of the noun, as in (26c): (26) a. T(bona) = λe.good(e) b. T(violinista) = λxλe.violinist(x,e) c. T(bona violinista) = λxλe.good(e) ∧ violinist(x,e) In this representation, bona denotes a first order property and restricts the denotation of violinista, but does so without being ascribed to the individual argument of violinista. The fact that bona denotes in type accounts for its acceptability in predicative positions; the fact that it modifies the noun’s event argument and not its individual argument accounts for its apparent nonintersectivity. The analysis we propose is analogous to Larson’s analysis for event-related adjectives, with the difference that we make use of kinds rather than events. First, we posit that all common nouns have an implicit kind argument, which is related to the individual-sort argument typically associated with nouns via the Carlsonian realization relation R (Carlson 1977). We represent the general translation for nouns, closely following Krifka, et al. 1995, as in (27), where the subscript k indicates a kind-level entity and the subscript o, an object-level entity. Put informally, this analysis states that objects realize the kinds of things that nouns describe: (27) For all common nouns N, T(N) = λxk λyo[R(yo,xk) ∧ N(xk)] Thus, a noun such as arquitecte, ‘architect’, would receive the translation in (28): (28) T(arquitecte) = λxk λyo[R(yo,xk) ∧ architect(xk)] Second, we posit that those adjectives traditionally described as relational denote properties of kinds. That is, they fall into the same sortal class as adjectives such as widespread or extinct in English. Thus, an adjective such as tècnic will have the translation in (29); it can be truthfully applied to any number of kinds – the kind architect, solution, translation, etc.: (29) T(tècnic) = λxk[technical(xk)] As under this analysis tècnic denotes a property of a kind, and not of an individual, we need a special noun-adjective (or more precisely, noun-adjective phrase) composition rule to combine the adjective with the noun, as in (30): 9 Of course, treated this way, this adjective would have to have other translations as well, corresponding to properties of the other sorts of individuals it can describe. No doubt a better analysis would assign a single translation to the adjective, on which it denoted a property of the most general sort of entity, a sort encompassing both entities and events. However, for the sake of illustrating Larson’s analysis, we will use more specific translations like that found in the text. 10 This assumption neither excludes nor presupposes the presence of an event argument; however, we will leave any possible event arguments out of the representations that follow to keep things simple. 11 The intuition is also expressed in Bosque and Picallo 1996, and more indirectly in Bolinger 1967, though neither of these works develop it into a specific semantic proposal. Relational adjectives as properties of kinds 189 (30) If noun N translates as λxk λyo[R(yo, xk) ∧ N(xk)] and adjective phrase AP translates as λxk[A(xk)], then [N AP] translates as λxk λyo[R(yo,xk) ∧ N(xk) ∧ A(xk)] The effect of this rule is to restrict the kind described by the modified noun to one of its subkinds. After the adjective phrase and the noun have composed, the resulting phrase still needs to be saturated with two arguments – one corresponding to a kind, and the other corresponding to the object-level individual described by the noun. We propose that the kind argument gets saturated by a contextually-determined kind. This seems plausible because in most or perhaps all cases, this kind will be uniquely identifiable in the context (indeed, this is the assumption behind Carlson’s claim that kind terms are like proper names). Thus, the noun phrase arquitecte tècnic translates as in (31), where we use an indexed free variable (analogous to a free pronoun) to saturate the kind argument: (31) λxk λyo[R(yo,xk) ∧ architect(xk) ∧ technical(xk)](kj) = λyo[R(yo,kj) ∧ architect(kj) ∧ technical(kj)] This property of individuals can then be applied to an argument such as Martí: (32) a. El Martí és arquitecte tècnic. ‘Martí is a technical architect.’ b. λyo[R(yo,kj) ∧ architect(kj) ∧ technical(kj)](m) = [R(m, kj) ∧ architect(kj) ∧ technical(kj)] This analysis has the advantage that it does not directly ascribe “technicalness” to Martí, while still entailing that a technical architect is an architect. It also predicts the unacceptability of #El Martí és tècnic mentioned above; that is, it predicts the apparent nonintersective behavior of the adjective. The key here is that if the argument of the adjective does not denote a kind, the adjective cannot be predicatively used: the sort of the adjective and its argument will conflict, and this sortal mismatch will make the predication infelicitous. In contrast, if the subject of a copular sentence containing a relational adjective does plausibly denote a kind, the predication will be acceptable, as in (33): (33) La tuberculosi pot ser pulmonar. ‘Tuberculosis can be pulmonary.’ Thus, the analysis both predicts that relational adjectives can be used predicatively and accounts for the conditions under which this use is possible. 4.3 A further argument for the analysis Our analysis makes yet another correct prediction, which amounts to an additional argument in its favor. This prediction involves adjective order. It has been noted (e.g. by Demonte 1999, Picallo 2002) that relational adjectives always appear closer to the head noun than do other intersective adjectives, illustrated in the following contrast: (34) a. inflamació pulmonar greu b. #inflamació greu pulmonar ‘serious pulmonary inflammation’ 190 L. McNally & G. Boleda Our analysis predicts precisely this pattern of word order possibilities. We first show how it predicts the unacceptability of (34b). We assume the following translations for inflamació, pulmonar, and greu: (35) a. T(inflamació) = λxk λyo[R(yo,xk) ∧ inflammation(xk)] b. T(pulmonar) = λxk[pulmonary(xk)] c. T(greu) = λxo[serious(xo)] Let us assume that adjective ordering reflects order of composition, and that if greu appears closest to the head noun, it must combine with it first. In order for this combination to take place, we must first saturate the noun’s kind argument; only then will it denote a property of individuals that can be intersected with the denotation of greu. That is, the translation of inflamació greu will be as follows: (36) T(inflamació greu) = λyo[R(yo,kj) ∧ inflammation(kj) ∧ serious(yo)] But this resulting translation is not of the right sort to combine with pulmonar: the latter can only be combined with something whose translation contains a lambda-bound kind argument. Thus, the phrase inflamació greu pulmonar is ruled out. This problem does not arise if we combine inflamació with pulmonar first and then with greu: (37) a. T(inflamació pulmonar) = λyo[R(yo,kj) ∧ inflammation(kj) ∧ pulmonar(kj)] b. T(inflamació pulmonar greu) = λyo[R(yo,kj) ∧ inflammation(kj) ∧ pulmonary(kj) ∧ serious(yo)] After the relational adjective combines with the noun, we can saturate the kind argument and the result will denote a property of the same sort as that denoted by greu. Note that this prediction is not contradictory with examples such as those in (23) above (producció mundial pesquera vs. producció pesquera mundial), as in these latter cases both adjectives are relational. As discussed in Section 3, our analysis correctly predicts that both orders should be possible and lead to no difference in denotation. 4.4 Some complications in the data 4.4.1 Relational adjectives predicated of nonkinds As noted in Section 4.2, our analysis predicts that relational adjectives should only take as arguments in a predicative construction NPs that denote kinds (as opposed to ordinary individuals), and this prediction appears to be largely borne out by facts such as (33) and (1c), repeated below for convenience: (38) a. La tuberculosi pot ser pulmonar. (=33) b. #El Martí és tècnic. (=1c) However, we have also found ostensible counterexamples to this prediction, in which a relational adjective is predicated of a NP that arguably does not denote a kind in the context. (39) presents an example. 12 Our analysis does not make any specific predictions concerning the ordering of relational adjectives with respect to each other. In principle it permits variation in the ordering of relational adjectives (see the discussion of (23) above), but of course other factors independent of semantic type per se may limit the ordering possibilities. However, we must leave further exploration of this issue for future research. Relational adjectives as properties of kinds 191 (39) Infection with tuberculosis spreads in two ways, by the respiratory route directly from another infected person or by the gastrointestinal route by drinking milk infected with the tubercle bacillus....In infections with M. tuberculosis, the tubercle bacilli commonly affect the lungs, in which case the disease is known as pulmonary tuberculosis. By contrast, infections with M. bovis often affect the bones and joints. About 90 percent of all clinically recognized tuberculosis in humans is pulmonary. (the Britannica Guide to the Nobel Prizes, http://www.britannica.com/nobel/micro/606_50.html) The sentences in (10) above, repeated here in (40), constitute additional examples: (40) a. El domini del Tortosa va ser només territorial. b. Aquest congrés és internacional. c. El conflicte és polític. In all of these examples, the property denoted by the adjective is used to classify individual instances of a kind that could typically be described using the adjective. For example, (40b) asserts that a particular conference belongs to the (sub)kind of international conferences. A thorough study of such examples (including their frequency and distribution in different types of corpora) will have to await future research, but we would like to make a few preliminary observations. Perhaps the most salient fact about such examples is that we have only found them attested with common noun subjects, and the contrast between (38b) and the examples in (39)-(40) is sharp. One way to explain this contrast is to hypothesize that relational adjectives are susceptible over time to extending the domain over which they denote. Perhaps they originate as properties of kinds and then, as those properties become useful for subclassifying instances of these kinds directly, their extension is expanded to include such instances themselves. Such an explanation would predict that, statistically speaking, it will sound more felicitous to predicate a relational adjective of an individual that is described using a noun denoting a kind for which that adjective is a well-established modifier than it will be to predicate such an adjective of an individual that is described by an expression that does not denote such a kind. While we must evaluate this prediction carefully in future research, the following case study bears it out. As (40b) sounded very natural to us, we did a simple Google search for the expressions “international conference” and “conference is international”. The first search returned about 3,720,000 hits, and the second, 251 hits. While these lists of hits contain irrelevant examples, certainly they returned many, many relevant ones. We then did a search for “international bakery” and “bakery is international”. This time, the former returned 1,910 hits (again, not all of which are relevant), and the latter, none. This dramatic difference in hits correlates with our intuition that, even though (41a) is perfectly acceptable (and was in fact attested), (41b) sounds very odd. 13 Moreover, this account might well also lead to explanation for the widely noted fact that many relational adjectives also have a nonrelational meaning, as in (49): (49) a. Aquests pantalons són molt econòmics. ‘These trousers are very economical (i.e., cheap)’ If our analysis is correct, cases in which a relational adjective is predicated of an individual for the purposes of subclassification could be a first step in the development of such derived meanings. 192 L. McNally & G. Boleda (41) a. Droubi's Bakery is an international bakery that is currently located only in Houston. http://www.droubisbakery.com b. ??That bakery is international. We suggest that while international might be plausibly used to describe and subclassify any number of kinds of things (including bakeries), its use as a classifier of bakeries has not become sufficiently established for the adjective to serve as a direct property of individual bakeries. It could be thought that this contrast is merely due to frequency: more frequent adjective-noun pairs (be they relational adjectives or not) would lead to predicative examples, whereas less frequent ones would not. However, this appears not to be the case: Google searches for “nice mouse” and “pink table” returned a number of hits comparable to the “international bakery” case (2,680 and 1,660, respectively), and their predicative counterparts (“mouse is nice” and “table is pink”) returned 183 and 65. This suggests that both the analysis and the explanation for cases like (40) and (41) are on the right track, for it reveals two related facts. First, for some adjective-noun pairs involving relational adjectives it is not possible to use the adjective predicatively at all ((41b)), or it is only possible under very constrained conditions. Second, even for adjective-noun pairs where we find the adjective predicatively applied to an NP headed by the noun in question, we find that predicative uses involving a given relational adjective and a given noun are proportionally much less frequent than predicative uses of a given nonrelational adjective in combination with a given noun. For example, while the attributive uses of nice in NPs headed by mouse are approximately 15 times more frequent than the predicative uses of nice with NPs headed by mouse (based on the figures mentioned above), the attributive uses of international in NPs headed by conference are 15,000 times more frequent than the predicative uses of the same adjective with NPs headed by conference. This may explain why many people have the strong intuition that relational adjectives cannot be used predicatively, even though this is clearly not the case. However, a thorough statistical analysis should be performed in order to test whether these differences in distribution are robust through the different classes of adjectives. Given this explanation, we would predict relational adjectives to sound anomalous when predicated of proper names because proper names are not classificatory expressions, and the set of individuals described by a proper name, generally being a singleton, will not permit further subclassification by a property such as one described by a relational adjective. We would also not be surprised to find that the use of relational adjectives predicatively to subclassify ordinary individuals is most frequent in specialized discourses, where the adjectives used for subclassification of a given kind of entity are well known and the interest in such subclassification is obvious. Obviously, this explanation for the facts in (40) runs the risk of weakening our analysis: If we stand by it, we must admit that at least some relational adjectives can denote properties not only of kinds but also of individuals. Nonetheless, we think this weakening is more apparent than real. First, our analysis clearly accounts for the classic subsective behavior of relational adjectives. Second, it forms the basis for a promising explanation of the complex distribution of relational adjectives described in this section. Relational adjectives as properties of kinds 193 4.4.2 The use of more familiar kind-level predicates within NP Because we propose that relational adjectives denote properties of kinds and because nothing in our analysis prevents any kind-level predicate from modifying a noun within an NP, we also expect that we should find examples of the more familiar kind-level predicates such as extinct in NPs which are predicated of ordinary individuals. However, as pointed out to us (Satoshi Tomioka and Olivier Bonami, p.c.), sentences such as the following sound extremely odd: (42) a. ??Dino is an extinct dinosaur. b. ??Tweety is a widespread bird. We suspect that the oddness/nonexistence of examples such as (42a), involving extinct, is that such sentences can never be true. If Dino is or was a dinosaur, it is entailed that that species of dinosaur exists or existed (whether in reality or fiction) at the relevant time of evaluation, and if the species is or was entailed to exist, it cannot simultaneously be or have been extinct, which is what the semantic rule for combining adjectives and nouns requires. Thus, (42a) may well be odd, and similar examples inexistent, because of their contradictory nature. A similar explanation can be provided for (42b): it is pragmatically odd to assign the property of being a widespread bird to a single individual. However, if this is true, we should find other, pragmatically plausible instances of kind-level predicates modifying a noun within an NP predicated of ordinary individuals. In order to test this prediction, we performed a series of Google searches for occurences of the adjectives extinct, widespread, scarce, abundant, common and rare in this construction. The searches were of course only approximations, as no linguistic constraints can be set on current web search engines: we searched for exact matches for “is a(n) A”, where A was one of the six adjectives just listed. For four of the adjectives (rare, scarce, common, and, perhaps surprisingly, widespread), we found relevant examples in the first 20 to 40 matches. These results clearly confirm our prediction. Some of the examples, together with the original URLs, are the following: (43) a. There are a number of reasons such a clamorous stir has developed with collectors over this find: (1) It Is a truly a vintage piece from the early 1980's. (...) (5) It was not printed in the United States, but is a scarce overseas piece. http://www.findarticles.com/cf_0/m0FCM/4_32/112904360/p1/article.jhtml b. This is a scarce figure of a railway engineer in fair to good all original condition.(...)This is a scarce figure in good condition. (...) . This is a scarce item in good all original condition. (...)This is a scarce Britains nurse in fair to good all original condition. http://www.collectorsworld.net/lead.htm 14 These are the kind-level adjectives listed in Krifka, et al. 1995, one of the standard references on genericity and kinds. We chose to search English examples because this class of adjectives is even smaller in Catalan than it is in English, and the number of web pages in Catalan, much smaller as well. 15 Google returns an approximate total number of matches for the searches, which we report here: extinct (7,370), widespread (109,000), abundant (22,700), rare (709,000), scarce (22,700), common (1,670,000). 16 It seems that this use of scarce is mostly found in collectors’ vocabulary. 194 L. McNally & G. Boleda (44) a. The Ageing Labour Force is a Common Challenge for Europe [Title, hence capital letters] http://presidency.finland.fi/netcomm/news/showarticle279.html b. "Sweet potato" is a common nickname for what small musical instrument? http://www.themusicstand.com/info/trivia/questions/0,1936,t,00.html (45) Charlie Kaufman is a Rare Scribe [Title, hence capital letters] http://www.scre.com/cgibin/news.cgi?v=news&c;=Screenwriting_Coverage&id;=031820048187 (46) SHIN SPLINTS is a widespread term for a variety of generalized symptoms for pain in the lower legs. http://www.doctorsexercise.com/journal/sum01.htm Note that the example in (46) is parallel to that in (42b), which we suggested was unacceptable for pragmatic reasons. What makes (46) different is that, while its subject does not denote a kind, it does denote an entity which can have distinct realizations at distinct points in time, making it easier to satisfy the truth conditions of the predicate: An individual term such as shin splints can qualify as widespread because it is used on many occasions. To sum up, it seems that kind-level predicates modifying nouns within NPs predicated of individuals are in fact attested; however, it is also clear that they are relatively rare. Our hypothesis is that this is for pragmatic, rather than semantic, reasons: Not many individual-denoting subjects fulfill the restrictions that a kind-level adjective imposes on the predicate. We are currently undertaking a statistical analysis that should shed more light on the facts discussed both in this subsection and in the previous one.