Word-class-changing inflection and morphological theory

One of the most common claims made about the difference between inflection and derivation in the morphological literature is that derivational affixes change the word-class of their base, while inflectional affixes do not change the word-class. In this paper I argue that this view is wrong, and that important insights about the nature of inflection and derivation are lost if word-class-changing inflection is not recognized. In §2, I present a number of examples of word-class-changing inflection, and in §3–5 I discuss several potential objections to my analysis. I show that the cases in question can be regarded neither as word-class-changing derivation (§3–4) nor as non-word-class-changing inflection (§5), and that a description in terms of feature neutralization is not a general solution (§6). In §7 I argue that to account for the syntactic properties of words, two types of word-class have to be distinguished: lexeme word-class and word-form word-class. §8 discusses some problems that arise in the formal representation of this proposal in constituent-structure trees and observes that Tesniere’s dependency grammar provides an interesting perspective. Finally, §9 discusses the universal correlation between inflection and preservation of internal syntax, and derivation and the non-preservation of internal syntax.


INTRODUCTION
One of the most common claims made about the difference between inflection and derivation in the morphological literature is that derivational affixes change the word-class of their base, while inflectional affixes do not change the word-class. In this paper I argue that this view is wrong, and that important insights about the nature of inflection and derivation are lost if word-class-changing inflection is not recognized. In §2, I present a number of examples of word-class-changing inflection, and in §3-5 I discuss several potential objections to my analysis. I show that the cases in question can be regarded neither as word-class-changing derivation ( §3-4) nor as non-word-class-changing inflection ( §5), and that a description in terms of feature neutralization is not a general solution ( §6). In §7 I argue that to account for the syntactic properties of words, two types of word-class have to be distinguished: lexeme word-class and word-form word-class. §8 discusses some problems that arise in the formal representation of this proposal in constituent-structure trees and observes that Tesniere's dependency grammar provides an interesting perspective. Finally, §9 discusses the universal correlation between inflection and preservation of internal syntax, and derivation and the non-preservation of internal syntax.

DESTROYING A MYTH: WORD-CLASS-CHANGING INFLECTION EXISTS
As is well known, derivational affixes often change the word-class! of their base, i.e. they are transpositional. 2  The PARTICIPLE,illustrated in (2a), is a familiar phenomenon, and the view that participles are verbal adjectives is very 01d. 3 The examples in (2b-d) are less familiar from European languages, but similar cases are found again and again in the languages of the world. Example (2b) shows an inflectional VERBAL NOUN from Lezgian. Verbal nouns (also called action nominals) are generally derivational in European languages, so I prefer to use the term masdar (originating in the Arabic linguistic tradition) for such inflectional verbal nouns. In the Lezgian example (2b), the masdar is the head of the equivalent of a complement clause. In (2c) we see the Turkish suffix -ki, which turns an adverb into an adjective. I call this an ATIRIBUT-IV1ZER. Finally, (2d) shows an equational sentence from Blackfoot, expressed by turning the predicate noun into a verb by means of an inflectional affix (called PREDICATlVIZER here). These are my fust counterexamples to the claim that word-class-changing affixes always have derivational status. One should think that linguists would have become aware of the existence of these cases, but nevertheless the myth that transpositional morphology is always derivational is found again and again in the literature, as the quotations in (3) show. The claim is repeated especially in introductory works like Nida (1946), Hockett (1958), Langacker (1972), Bauer (1988), but also in purely theoretical contributions like Selkirk (1982), Scalise (1988), Anderson (1992). Sometimes the claim is only indirect and implicit, as in Hockett (1958) and Langacker (1972), or is even presupposed, as is clear from the quotation of Jensen (1990).
(3) a. Nida (1946:99): derivational formations: may exhibit changes in major distribution class membership; inflectional formations: exhibit no changes in major distribution class membership b. Hockett (1958): p. 240: "When all inflectional affixes are stripped from the words of a language, what is left is a stock of stems ... "; p. 221: "The part of speech of a word is that of its stem." c. Langacker (1972:75): " ... derivational affixes have the potential to change the grammatical class of the elements to which they are attached. For example, the addition of the derivational suffixjUl to the noun care results in an adjective ... By contrast, an adjective inflected to agree in gender and number with the noun it modifies remains an adjective." d. Selkirk (1982:77): "One nice result of [the generalization that inflectional affixes are never heads] would be the apparently universally attested fact that inflectional affixes are not "category-changing". If the inflectional affix is not the head, then its sister category is, and hence always shares category features with the mother node." e. Scalise (1988:562): "Derivation rules change the syntactic category of their base, while inflection rules do not." f. Bauer (1988:12): "If an affix changes the part of speech of the base, it is derivational." g. Anderson (1992:78): .. It is also often noted that inflection and derivation, as usually understood, may differ with respect to the syncategorematicity of the relation marked by a particular rule. Inflectional rules do not change word class (that is, a verb marked [+Past] is still a verb), while derivational ones can." h. Jensen (1990:133): " ... adjective formation is clearly derivational, since it results in a change of category."

Martin Haspelmath
Thus, the position that I argue against in this paper is not a straw man, but seems to be widely accepted in the literature by linguists of various theoretical persuasions. Now there are two obvious routes of argumentation that are open to those who wish to defend the view that transpositional morphology is always derivational. First, one could maintain that examples such as those in (2a-d) do not show inflectional morphology, but derivational morphology after all. Second, one could argue that the affixes in (2a-d), while being inflectional, do not change the word-class of their bases. In the following three sections I will examine the possible counterarguments and argue against them.

INFLECTION AND DERIVATION
In this section I will discuss and argue against the claim that my cases of word-classchanging morphology are in fact instances of derivational morphology. In order for a meaningful debate to be possible, we must of course agree on what we mean by inflection and derivation. I will argue that on an understanding of the terms that is in accord with the traditional intuition behind them, the examples in (2a-d) cannot be interpreted as derivation.
It is of course possible to arbitrarily define the terms inflection and derivation in such a way that transpositional morphology automatically falls under derivation. This is done, for instance, by Drijkoningen (1992). His classification of affixes is shown in (4) below. According to Drijkoningen, an affix is "derivational" if it is the categorial head of the construction, i.e. if it determines the resulting word class; otherwise it is inflectional. Another distinction that cross-cuts the first distinction is that between "syntactic" and "lexical" affixes.
(4) four-way classification of affixes according to Drijkoningen (1992)  What I would call word-class-changing inflection is called "syntactic derivational" morphology by Drijkoningen. My objection against Drijkoningen's classification is purely terminological. An important general principle of good terminological usage is that old, well-established terms should not be used in a completely novel sense, but that the intuitions underlying the old terms should be reconstructed. If that is not possible in a coherent way, the term should be dropped. Completely new senses should be expressed by completely new terms. It is clear that Drijkoningen's distinc-tion "syntacticllexical" is much closer to the traditional inflection/derivation distinction than his own "inflection/derivation" distinction.
So what is the intuition underlying the inflection/derivation distinction? I think that the most basic property of inflectional forms is that they are described exclusively in grammatical paradigms, whereas derivational formations are described by listing them individually in a dictionary (there may also be a section on derivation in a grammar, but it is never organized in paradigms and it never gives all the relevant information). A brief look into any reference grammar or dictionary will confIrm that this characterization of inflection and derivation by their modes of description is universally valid. Since paradigms and dictionaries are among the most elementary components of any language description, the characterization is also highly theoryindependent, and I fInd it intuitively satisfying.
But we have to ask further: Why do linguists universally describe some morphological formations by paradigms and others by listing them in a dictionary? The answer is of course, because they have quite different properties. The following is my defInition of inflection and derivation that follows from the general characterization that I have just proposed:

(5) DefInition of inflection and derivation
Formations are inflectional to the extent that they are regular, general and productive; formations are derivational to the extent that they are irregular, defective and unproductive.
This defInition follows from my initial characterization because words can be described by means of abstract paradigms only if their formation is productive (i.e. if new words can be formed according to the rule), regular (i.e. if the words do not have any additional idiosyncratic properties), and general (i.e. if all the bases to which a rule could apply do in fact allow the formation of the word). By contrast, if a rule is unproductive, irregular and defective, an abstract paradigm is not sufficient for the description, and each form must be listed individually in the dictionary. By this defInition, there is no doubt that the formations in. (2a-d) are indeed inflectional forms. For instance, the German participle can be formed from any verb, i.e. it is maximally general, and its forms and meanings show almost no idiosyncrasies, i.e. it is very regular, and of course it is also productive. The same is true for the Lezgian masdar, which serves as the citation form of verbs in Lezgian dictionaries, i.e. it is the paradigmatic word form par excellence. The Turkish attrubutivizer and the Blackfoot predicativizer have analogous properties.
At this point I should mention that the inflection/derivation distinction is not absolute but allows for gradience and fuzzy boundaries. As Stephany (1982), Bybee (1985), Corbett (1987) and especially Plank (1994) have demonstrated, we are dealing with a continuum from clear inflection to clear derivation, with ambiguous cases in between. Note that this view of inflection and derivation follows from my defInition in (5): Productivity, regularity and generality are gradient properties, not binary, all-or-nothing features. Some linguists (e.g. Dressler 1989, Luraghi 1994) have discussed inflection and derivation in terms of prototypicality, pointing out that some types of affixes are prototypically inflectional (e.g. case affixes and person/number agreement affixes), whereas others are prototypically derivational (e.g. instrument nouns). It is possible and even probable that word-class-changing inflection is not prototypical inflection, and that it is not on the extreme inflectional end of the inflection-derivation continuum, but I insist that the examples that I discuss in this paper are much closer to prototypical inflection than to prototypical derivation. Thus, there may be some languages is which participles are more derivational than inflectional, but German is a good example of a language where participles are truly inflectional; in many languages masdars (verbal nouns) are more or less derivational, but in Lezgian they are clearly inflectional, and so on.
Before concluding this section, I would like to mention another definition of inflection that has become quite popular recently. According to Anderson (1992), inflection is morphology that is relevant (or visible) to the syntax. 4 According to this definition, too, the examples in (2a-d) constitute inflection because all of them are highly relevant for the syntax. Their relevance lies precisely in the fact that they change the word class: Thus, the participle in (2a) allows the verb (or the clause) to be used as a modifier, the masdar in (2b) allows it to be used as an argument, and so on. Thus, even if my definition in (5) is not accepted, there are very good reasons to accept my claim that (2a-d) are inflectional.

INFLECTIONAL (SUPER-)CATEGORIES
Another potential objection to the inflectional status of participles, masdars etc. is that inflectional forms normally belong to an inflectional super-category,5 e.g. accusative is an instance of the super-category of CASE, past is an instance of TENSE, plural is an instance of NUMBER, and so on. What super-category would participles, masdars, attributivizers and predicativizers belong to? In a sense, the answer to this question is trivial, because new super-categories can easily be created ad hoc for any inflectional form, e.g. PARTICIPIALITY, which has only the two sub-categories participle and non-participle. Another way to counter the objection would be to point out that super-categories are artifacts of a certain kind of (structuralist) linguistic description that have no reality for the speakers (cf. Bybee et al. 1994: 3).
However, it may be interesting to point out that such a super-category has already been postulated in the structuralist literature, although it did not become widely known. In several publications, the Moscow linguist Aleksandr Smirnickij has proposed a morphological super-category REPRESENTATION, whose members are the major word-classes (e.g. Smirnickij 1959: 247). At least in Moscow linguistic circles Smirnickij's idea has been adopted widely. In this parlance, one says that the English participle is the adjectival representation of the verb, and the English gerund is the substantival representation. Analogously, we could say that the Turkish formflimdiki (example 2c) is the adjectival representation of the adverbflimdi, and that the Blackfoot form nitaakiiyihpinnaan in (2d) is the verbal representation of the noun aakii. Thus, for instance Boguslavskaja 1989 says that the suffix -se in (6) marks the attributive, i.e. adjectival representation of the noun 'cow'.

MORPHOLOGICAL AND SYNTACTIC GENERALIZATIONS
After 1 showed in §3 that transpositional inflection is indeed inflection, skeptics might still doubt that these forms are indeed transpositional. One could say that the participle singende is a verb and not an adjective, the masdar qaraguni is a verb and not a noun, and so on. This view seems to be widespread at least for participles (e.g. Scalise 1988:566). However, there are good reasons for rejecting this view because it does not allow us to capture the morphological and syntactic generalizations that are expressed by the word-class-changing hypothesis. Morphologically, German participles behave just like adjectives, sharing the same two types of inflection patterns. The Lezgian masdar behaves just like other nouns, forming all sixteen cases that other nouns have;6 and so on for the other instances. Thus, a significant morphological generalization is lost if one does not recognize the transpositional nature of these forms. Similarly, there are syntactic properties in which participles and masdars behave just like adjectives and nouns. For instance, German participles are always preposed, like other adjectives but unlike most genitive modifiers and relative clauses, and the Lezgian masdar can appear in any argument position, like other nouns.
Let us now look at a few further cases of transpositional inflection. 'the teacher's table' The English DEADJECTIV AL ADVERB in -ly is sometimes classified as a derivational affix, but this is rarely justified. In the grammar by Quirk et al. (1985) it is treated in the section on derivation, but the authors say: "-ly can be very generally added to an adjective in a grammatical environment requiring an adverb, so that it could almost be regarded as inflexional". Example (7b) shows a SUBSTANTIVIZED ADJECTIVE, (7 c) shows a CONVERB (often called "adverbial participle", "gerund" or "conjunctive participle" in descriptions of individual languages), and (7d) shows a POSSESSIVE ADJECTIVE, quoted from Corbett's (1987) comprehensive study of Slavic possessive adjectives.
The examples in (7a-d) again illustrate the crucial properties that we discussed earlier: They are formed by highly productive, regular and general rules, i.e. they are inflectional, they are relevant to the syntax, and they behave like the derived wordclass morphologically and syntactically, so that the conclusion is in~scapable that they are cases of word-class-changing inflection.
Thus, the myth that word-class-changing inflection does not exist is not more than a myth and should be quickly forgotten. 7 But it is nevertheless interesting to ask how it arose and how it was possible that it could survive for such a long time. I suspect that it has to do with the fact that grammatical theory has been dominated by thinking about English grammatical structure for the past decades and continues to be dominated by Anglophone linguists. English has very little morphology, and although it has several cases of transpositional inflection (participle, gerund, adjectival adverb), these were not sufficient to direct grammatical theory in the right direction. I have not investigated the history of this question in any depth, but I have not found a single example of the view exemplified in (3a-h) in the literature of the 19th century or the first half of the 20th century, when linguistics was not yet dominated by English-speaking linguists. For Charles Bally and Lucien Tesniere, who developed their theories of grammar in the 1920s and 1930s, it was clear that word-class-changing inflection is possible (on Tesniere, see below §8). Thus, it may be that the myth discussed above is another example of an error introduced into the mainstream of grammatical theory due to insufficient consideration of linguistic diversity. In addition to participles, only three other types of "intermediate" word-classes can be described in this way. There is no way of describing, for example, masdars (cf. 2b), because nouns and verbs are not a natural class in Chomsky's system and do not share a feature value. 8 My second objection is that feature neutralization does not specify the precise relation between the two word-classes. Thus, the feature combination [+N, ~V] could describe either substantivized adjectives (Adj ~ N) or possessive adjectives (N ~ Adj).
I conclude that feature neutralization is neither general enough nor precise enough to account for transpositional inflection in general. Equally problematic is the account of Quechua nominalizations proposed by Lefebvre and Muysken (1988), which is similar in spirit to the neutralization view. These two authors argue that Quechua nominalizations (Le. participles/masdars) have the feature values [+N, +V] and thus constitute "mixed categories" (in the Chomskyan system, these feature values would characterize adjectives, but Lefebvre and Muysken claim that adjectives are a subclass of nouns in Quechua). This approach is even less general than neutralization, and it is equally incapable of distinguishing between nominalized verbs and verbalized nouns.

EXTERNAL AND INTERNAL SYNTAX
Above in §5 I said that words derived by word-class-changing inflection behave syntactically like their derived word-class. This is only half of the truth: These words generally have only part of the syntactic properties of their derived word-class, namely those that concern the combination with the head that governs them or that they modify -I call these properties EXTERNAL SYNTAX here because they concern an element outside their phrase. By contrast, their INTERNAL ,SYNTAX, i.e. their combination with dependents inside their phrase, is mostly identical to'that of their base word-class. 9 For example, in the German expression der im Wald laut singende Wanderer (example 2a), the participle singende behaves like an adjective only with respect to the head noun Wanderer that it modifies. With respect to its dependents, im Wald and laut, the participle singende behaves like a verb and unlike an adjective (adjectives cannot be modified by manner or locative adverbials, cf. *die im Wald laut schOne Blume 'the flower beautiful loud in the forest'). The same can be illustrated by the other examples (2b-d, 7a-d). For instance, the Kannada converb hee.lade in (7c) behaves like an adverb only with respect to the verb it modifies, but it behaves like a verb with respect to its dependent, the dative argument yaariguu 'anyone'. A particularly nice example comes from Sorbian, where the possessive adjective behaves like an adjective only with respect to the noun it modifies. When it comes to its own dependents, e.g. genitive modifiers, it looks like a noun, as illustrated in (9). Thus, when describing the complete set of syntactic properties of a word derived by transpositional inflection, two sets of word-class properties have to be taken into account: The word's WORD-FORM WORD-CLASS, which determines its external syntax, and its LEXEME WORD-CLASS, which determines its internal syntax. To make this quite clear, the lexeme word-classes and word-form word-classes of the examples that we have seen so far are listed in (10).

qaragv-un N flimdi Adv -ki Adj nit-aakiiN-yiv-hpinnaan beautijulAdj-Iy Adv c'exiAd{di N hee.lv-ade Adv muzN-owa Adj
Thus, my conclusion is not as radical as it may have seemed at the beginning: I do not say that participles are adjectives pure and simple, but they are adjectives with respect to their word-form word-class. In terms of lexeme word-class, they are verbs. However, this leads to radical consequences for widely accepted methods of formal representation.

PROBLEMS OF FORMAL REPRESENTATION
First of all, we note that Lieber's percolation mechanism does not work. If we assume that word-class-changing inflectional affixes are heads, we get representations like (11a) for Sorbian mui-owa (see example (9». In (1Ia), the whole word turns out to be an adjective, so this representation fails to capture the fact that the internal syntax is noun-like. If the feature of the non-head percolates, or if the stem is considered the head, as in (11 b), then the whole word turns out to be a noun, again not the correct result. What we need are two independent layers of word-class features that both percolate up. The lexeme word-class becomes the inner feature, the word-form word-class becomes the outer feature of the whole word. This can be represented as in (12).
It seems that there is no way to capture the facts without an innovative notational device such as the layered word-class features in (12). Another difficulty is that expressions with transpositional inflection are not easily shown in trees that represent immediate constituent structure, but simultaneously have only morphological words as their terminal nodes. Let us take, for example, the tree representation in (13), which shows the relevant aspects of the structure of the Lezgian sentence (2b). Here we see a VP that represents the subordinated subject clause: wun fad ijarai-'you getting up early'. Since this clause is in an argument position, and the nominalizing masdar suffix -un is present, it seems reasonable to assume that the VP is dominated by an NP which consists of the VP plus the masdar suffix.
The problem with the representation in (13) is that it violates the lexicalist principle, which says that morphological units play no role in syntactic structure (cf. Wunderlich 1987). And what is more, the word ijarai-un-i is not a constituent in (13), although it is clearly a word by all morphological and phonological criteria. That is, we seem to be dealing with a mismatch between morphology and syntax. Such mismatches have sometimes been discussed in the literature, e.g. by Sadock (1991), Drijkoningen (1992). Drijkoningen assumes a structure very much like (13) for English gerund clauses such as Joan writing the novel, without however discussing the fact that his proposal means a fairly radical departure from standard assumptions. Similar structures are proposed by Weber (1983) for data from Huallaga Quechua, e.g. (14) 'He had hit you.' Weber argues explicitly against the lexicalist hypothesis, against the separation of syntax and morphology, and against the view that the immediate constituents of syntax are words. 1o It seems to me that Weber is throwing out the baby with the bathwater. It is true that cases of morphology-syntax mismatches can be observed time and again in languages. However, these cases are not the normal state of affairs. I propose that the lexicalist principle should be seen as a preference principle, i.e. a "soft" constraint that can be overruled in certain circumstances, but that is observed in the large majority of cases (cf. Dressler 1987). Some cases of real morpQology-syntax mismatches in which the lexicalist principle is violated are shown in (15).
(15) a. English The woman I saw yesterday's hat. b. Somali (Berchem 1991:51) Macallim-iin-ta dugsi-gu waxay tegayaan qolal-ka. In (15a) the genitive suffix is found on the last word of the relative clause; in (15b) the nominative suffix is found on the possessor dependent; and in (15c) the directobject marker nii is found on the word preceding the direct object. Cases such as these truly violate the lexicalist principle, and I would predict that they are unstable and tend to be eliminated or restructured. The cases of word-class-changing inflection are of a very different nature. They are not rare or unusual across languages, and they show no signs of diachronic instability. 11 Thus, it would not be correct to treat them in the same way as true mismatches. How can we represent constructions with transpositional inflection in such a way that they do not involve a mismatch? One possibility is to project the word structure that was proposed in (12) into phrase structure, i.e. to admit two layers of word-class features not just at the word level, but also at the phrase level. The representation ofthe Lezgian sentence (2b) would then be as in (16) Thus, the subordinate clause wun fad ijarai-un-i is an NP for external purposes, but a VP for internal purposes. This kind of representation departs fairly radically from standard assumptions about word-classes and phrasal categories in constituent structure representations, but it seems to be the only non-transformational solution for constituent structure. A transformational solution has been proposed by Dyk (1994), discussing Biblical Hebrew participles. On this account, the parti<;ipial verb is generated under a V -node and forms a VP which is the complement of an empty nominal node, to which the participle is moved in order to get its nominal properties. This is shown in the somewhat simplified trees in (17), which represents the participial phrase m~?all~mlm ?ilummlm '(people) who are binding sheaves'. ?tilummlm I binding:PL sheaves '(people) (who are) binding sheaves' Clearly, this account is as radical a departure from ordinary assumptions about constituent structure as (16). Since the empty node N exists exclusively for the purpose of turning the verb into a noun, the structure in (17) is not really a pure constituency representation.
If one adopts a dependency representation rather than a constituency representation, these problems disappear. Word-class-changing inflection can be represented very naturally, for instance, in Tesniere's (1959) model of dependency syntax. Sentence (16) would be represented as in (18). As in any dependency structure, each word is a node of the tree, and the connection of a lower node to a higher node represents an asymmetric dependency relation. In addition to the words themselves, the word-class is given in parentheses in the tree in (18),12 The part of Tesniere' s theory that is particularly relevant for the problem at hand is his TRANSFERENCE THEORY ("th6orie de la translation"). Transference is Tesniere's term for word-class change or transposition. Words that involve a word-class change are represented in two lines separated by a horizontal dividing line, as for instance qarai-un in (18). Above the line the word-class of the whole word (i.e. the wordform word-class) is given, in this case N. Below the line the stem with the lexeme word-class and the transferring affix are shown separately. The dependency link between the governing head taiub ijizwa and qarai-un sees only the upper part of the node, so only the word-form word-class is relevant for external syntax. However, the dependency link between qarai-un and its dependents sees the lower part of the node, so only the lexeme word-class is relevant for internal syntax. Thus, Tesniere's framework expresses word-class changing inflection in a quite natural fashion, and in particular the relation between word-form word-class and external syntax, and between lexeme word-class and internal syntax is expressed succinctly. 13 The dependency trees in (19)-(20) below further illustrate Tesniere's representation system.
(19) German (= 2a): (20) Sorbian (= 9): It may well be that the phrase structure representations in (16) and (17) will tum out to be a notational variant of Tesniere's dependency trees with transference, if a number of plausible additional assumptions are adopted. However that may be, I hope to have shown that Tesniere's approach to the formalization of syntax, which has been all but forgotten even in Europe,14 has certain virtues that are not shared by the constituency-based formalisms that have come to dominate syntactic thinking in recent decades.
The most important prediction made by transference theory (and its notational variants) is that word-class-changing inflection is admitted only on the head of the construction. Structures such as those in (15a-c) cannot be represented-naturally in Tesniere's framework if we adopt the restriction that the transferring affix must be next to the transferred stem not only in the dependency trees, but also in the surface structure. Given this, the structures in (15a-c) can only be expressed in some roundabout way, reflecting the fact that they show a dispreferred morphology-syntax mismatch.

INFLECTIONIDERIVATION AND THE (NON-)PRESERVATION OF INTERNAL SYNTAX
Up to now this paper has been an exercise in formal grammar, in the sense that I have asked how grammatical structures are best described and represented. In this final section I would like to venture briefly into substantive grammatical theory by making an empirically falsifiable claim regarding a universal correlation between certain morphological and syntactic properties, concluding with the question (and a tentative answer) of why this correlation should hold. There is no space in this paper to document the supporting evidence for my claim in any detail, and a few suggestive examples must be sufficient. The phenomenon in question is the universal correlation in (21).
(21) a. In words derived by inflectional word-class-changing morphology, the internal syntax of the base tends to be preserved. b. In words derived by derivational word-class-changirig morphology, the internal syntax of the base tends to be altered and assimilated to the internal syntax of primitive members of the derived word-class.
This correlation has not gone entirely unnoticed in the literature (cf. Bauer 1988: 76-77, Rainer 1993, but it is rarely discussed in this general form. It is best illustrated with the well-known example of nominalization constructions in English: (22) a. inflectional gerund Indonesia annexing East Timor b. derivational action nominal

Indonesia's annexation of East Timor
In (22a), the subject and direct object (as well as other potential dependents) are expressed in more or less the same way as in finite clauses,15 whereas in (22b) the subject is turned into an 's-genitive and the object is turned into an of-genitive. In other words, (22a) preserves the internal syntax of the verbal base (or at least preserves it to a much greater extent than (22b», and (22b) does not preserve the internal syntax of a verb and shows the internal syntax of primitive nouns. This correlates with the inflectional vs. derivational status of the transpositional affixes. The correlation is not a hundred percent: Sometimes inflectional word-class-chang-ing affixes force a slight change in internal syntax, and part of the internal syntax may be preserved even with derivational word-class-changing affixes. More generally, the correlation can be represented as two covarying continua: (23) E -( ---more inflectional------more derivational---.;;.) «-----more preservation of internal syntax --less preservation of internal syntax -----7 Just as there is overwhelming evidence that the inflection/derivation distinction is a continuum, the distinction between preservation and non-preservation of internal syntax is not a matter of all or nothing. This is already clear from example (22a-b): As was observed in note 15, the subject of the gerund in (22a) may also be marked as genitive (Indonesia's annexing East Timor), which means that the internal syntax is not preserved completely, but still to a greater extent than in (22b). Despite this slight weakening of the correlation, it is still quite easy to see what would be a counterexample to the correlation: Whenever there are two comparable transpositional processes in a language and one is more inflectional than the other from the morphological point of vieW\ the prediction is that it will also preserve internal syntax to a greater extent, and vice versa. The example of the English gerund and action nominal is of course well-known and has often been discussed in the literature. But this is only one instance of a much more general correlation which in principle applies to any kind of transposition. The following two subsections will give two examples ofless well-known types of wordclass changes, from noun to adjective, and from verb to adjective. Corbett (1987) adduces an impressive range of data showing that the diversity of the syntactic behavior of possessive adjectives in Slavic can be expressed very well by a hierarchy of syntactic properties that possessive adjectives in different languages have to different degrees. The three most important properties are:
before us is mother's(ADJ) apartment who REL wants SUBJV it sell 'Before us is mother's apartment, who wants to sell it.' (iii) allowing attributive modifiers, as in (9)  Russian possessive adjectives can be the antecedent of personal pronouns (cf. 24), but not of relative pronouns. In Macedonian, possessive adjectives may be the antecedent of both personal pronouns (cf. 28) and relative pronouns (cf. 25), but they may not have attributive modifiers.
before us is mother's(adj) apartment she wants SUBJV it sell 'Before us is mother's apartment. She wants to sell it.' The most liberal are Upper Sorbian possessive adjectives, which allow all three processes, as is illustrated in (29) and (9). Thus, Corbett's data clearly show that the preservation of internal syntax is not a matter of all or nothing, but that different languages may preserve fewer or more properties of internal syntax. Now the crucial point for my universal (21) is that the continuum in (26) correlates with the derivationaUinflectional status of the possessive adjectives. In English, "possessive adjectives" like Shavian, Orwellian are clearly derivational. They are formed from an extremely limited set of nouns (only personal names) and show idiosyncratic formal properties. In Russian and Macedonian their formation is more regular and more general (not just person names, but also kinship terms allow possessive adjectives), and in Upper Sorbian the possessive adjective is even more regular and general (all person nouns in the singular allow it), so that it is much closer to inflectional status.

Verbal adjectives/participles
A similar hierarchy can be established for adjectives derived from verbs (cf. Haspelmath 1994: § 10). When these are derivational, they are called "verbal adjectives", and when they are inflectional, they are called "participles". If we restrict our attention to active participles/verbal adjectives, then we can contrast Italian, German and Lezgian in the following fashion: Faktum 'a fact that surprises the judge'), but not with subjects. Finally, in Lezgian the participle has no restrictions on its government properties at all (Haspelmath 1993). Now this continuum of preservation of internal syntax again correlates with the inflection/derivation continuum. In Italian, verbal adjectives are not very productive, show idiosyncratic meanings and are clearly derivational. In German and other conservative Indo-European languages, active participles are mostly considered as inflection, but not prototypical inflection. Finally, Lezgian participles are quite typical inflectional forms, in that they are as frequent as other inflectional forms and can be combined with different aspectual stems of the verb. Similar correlations between two continua would be found for all cases of transposition. There is no space here for more examples, but in the following section I would like to speculate briefly on why the correlation should hold universally.

Toward an explanation of the universal correlation
Probably the best-known attempt at explaining the correlation is Stephen Anderson's (1992). It says that inflection is part of the syntax, and derivation is part of the lexicon, and the syntax and the lexicon are each autonomous and separated from each other in a clear-cut manner (cf. also Chomsky 1970). With further assumptions it follows from the architecture of the formal grammar that syntax is preserved with inflectional affixes, while derivational affixes in the lexicon obey quite different principles. This explanation is unsatisfactory for two reasons. First, inflection and derivation form a continuum that cannot be split up into two discrete formal blocks. Generalizations across inflection and derivation, like Corbett's hierarchy in (26), cannot be captured in this model. Second, the architecture of the grammar is not a real explanation -the question is only shifted one step further: Why does the grammar have this particular architecture and not a different one?
A real explanation has to establish a relation between grammatical and extragrammatical facts, or in other words, real explanations are always functional. In the following I will speculate a little about where such an explanation might be sought. A complete account is beyond the scope of this paper.
An important property of inflectional forms that distinguishes them from derivational forms is that they are more frequent in discourse. The Lezgian participle must be more frequent than the Indo-European participle because it is the major relativeclause-forming strategy in the language, and the Sorbian possessive adjective is a major strategy for expressing possession, so it must be more· frequent than the possessive adjective in English. The more frequent a word-forming process is, the more economical it is if the syntax does not deviate from the usual, i.e. if it is identical to the syntax of the base word. Derivational formations, by contrast, are much rarer in discourse, and they are not associated as closely with their base. For instance, a derivational action noun such as arrival or detainment is more closely related to non-derived abstract nouns like power or future, which of course have nominal rather than verbal syntax. Thus, considerations of frequency and economy make it quite plausible that this correlation should be universal, and a complete explanation should be sought in this direction.

CONCLUSION
To conclude this paper, I will briefly summarize its main points: First, I have argued that word-class-changing (or transpositional) inflection exists, is widespread across languages and cannot be explained away. Second, this has a number of consequences for morphological description, in particular that several levels of word-class within a word have to be recognized, and that phrasal categories, too, must allow two-fold word-class status (or, alternatively, a dependency notation must be used). Third, the recognition of word-class-changing inflection helps us see more clearly that there is a universal correlation between the inflectional vs. derivational status of a transpositional affix and its syntactic properties.