2. The universal notion of harmony

We begin by considering the notion of harmonic word order. In the literature, a number of cross-linguistic word order tendencies have been observed. In this section I formulate these tendencies in terms of violable constraints, and propose that the notion of harmony in grammar is defined by the interaction of these independently motivated, violable, constraints in a universal ranking. In addition, I argue that these word order constraints apply only to base-generated structures, these base-generated structures showing a greater diversity, and movement being more restricted, than under certain views of syntax. Empirical evidence for the proposed ranking of constraints, and its application to the base-generated structure only, will then be provided by a study of the restrictions placed on the cross-linguistic distribution of subordinating complementisers.

Before considering the relevant constraints, we begin with the more general question of what level of syntactic representation is relevant to the notion of harmony. This is a matter that depends largely on the view adopted of the mapping of syntactic structure to linear order. One such view is expressed by Kayne's (Reference Kayne1994) Linear Correspondence Axiom, which assumes that there is a single base-generated order, and any variant of this order results from movement. As shown by Abels & Neeleman (Reference Abels, Neeleman, Brucart, Gavarró and Solà2009, Reference Abels and Neeleman2012), however, a restrictive theory of movement cannot be maintained under this hypothesis if we are to account for the full range of data. Moreover, in many cases movement must be assumed purely on the basis of the surface word order, without any independent evidence. Instead, Abels & Neeleman adopt the more traditional view, whereby there are no universal ordering – merely hierarchical – restrictions on base-generation. This approach has the advantage of being able to maintain a more restrictive theory of movement. Under this view, movement is much rarer, and need only be assumed where there is independent evidence. Given that the unmarked order is determined to a much larger extent by the base-generated structure, it is not unreasonable to assume that word order constraints refer to this same base-generated order.Footnote ⁸

The violable word order constraints concerned with harmony are formulated below:

1. (9) Head-Proximate Filter

   The highest head in a complete extended projection must be contiguous with the lexical head of its superordinate extended projection.

2. (10) Final-Clause Constraint

   A clausal dependent must follow the lexical head of its superordinate extended projection.

3. (11) Head Uniformity Constraint

   A functional head must match the lexical head of its extended projection in the direction of headedness.

The Head-Proximate Filter has its correlates in Generative Grammar in the Head-Final Filter (Williams Reference Williams1982), in Functional Grammar in Rijkhoff's (Reference Rijkhoff1984, Reference Rijkhoff1986, Reference Rijkhoff2002) Principle of Head Proximity, and in the principle of Early Immediate Constituents in Hawkins' (1990, 1994) performance theory of grammar. (See fn. 9 below for a discussion of the differences between these constraints and the Head-Proximate Filter.) The Head-Proximate Filter specifies a requirement for the lexical head of one extended projection and the highest head in the extended projection of its dependent to be linearly as close to each other as possible. For example, Rijkhoff (Reference Rijkhoff2002: 261–263) points out that the data from Hawkins' (1983: 281–290) expanded sample shows a general tendency for languages to avoid placing adjectives and possessor NPs between the noun they modify and its selecting verb; in more concrete terms, there is a tendency for a noun to precede its dependents in VO languages and to follow its dependents in OV languages, such that the noun is as close as possible to its selecting head, the verb (see Greenberg Reference Greenberg and Greenberg1963: 85–90; also (33) below). A similar concept is found in the Head-Final Filter, where, in the Germanic and Romance languages in which it applies (see Escribano Reference Escribano2004: 1, fn. 2 for references), prenominal modification is only possible where the modifier is head-final, and therefore its head is immediately adjacent to the noun phrase it modifies. Escribano (Reference Escribano2004: 2–3) shows that the effects of the Head-Final Filter in the languages in which it operates extend beyond the nominal domain to other categories. Most importantly, there is evidence that the Head-Final Filter operates exclusively over base-generated structures; structures that conform on the surface, but contain a trace to the right of the head of the modifier, such that the prenominal modifier is head-initial (as in verbal passives), are ungrammatical (Williams Reference Williams1982), while structures that contain a surface violation as a result of leftwards movement of the head-initial modifier from posthead to prehead position are grammatical (Escribano Reference Escribano2004: 4). The Head-Proximate Filter that I am proposing here simply extends the Head-Final Filter to include its mirror image. Like the Head-Final Filter, it is concerned with the base-generated structure.Footnote ⁹

The second constraint, the Final-Clause requirement, is derived from Dryer's (Reference Dryer1980) Sentential NP Position Hierarchy. Looking at the cross-linguistic distribution of clausal arguments, Dryer established the hierarchy in (12).

1. (12) Sentential NP Position Hierarchy

   clause-final position >clause-initial position >clause-internal position (Dryer Reference Dryer1980: 126)

This hierarchy states that if a language allows any argument to appear in a given position on the hierarchy, it will allow clausal arguments in that position; a position further to the left, where available, will be preferred over any position to its right. This is confirmed by the fact that clausal complements of verbs are allowed in final position in both VO and OV languages, but preverbal complement clauses are not found in VO languages, which by definition always allow arguments in final position (Dryer Reference Dryer1980; Hawkins Reference Hawkins1994: Section 5.6.1; see also the typology in (14) below). Of relevance to us here is the fact that the clause-final position is preferred over any other: from this we derive the Final-Clause requirement as a violable constraint, formulated in (10) above.

The first two constraints we have looked at have been concerned with ordering between the lexical head of a superordinate extended projection and a dependent, or subordinate, extended projection. We move now to the final constraint, Head Uniformity, which is concerned with the order within the extended projection itself. This constraint is perhaps the best known among harmonic word order constraints. There is a preference for heads in a given language or domain to be consistent in directionality of headedness: either uniformly head-initial or head-final (see Natural Serialisation Principle, Bartsch & Vennemann Reference Bartsch and Vennemann1972: 136; Head Parameter, inter aliaChomsky Reference Chomsky1981; Branching Direction Theory, Dryer Reference Dryer1992, Reference Dryer2009a).

We will see that the three relevant constraints – the Head-Proximate Filter, Final-Clause, and Head Uniformity – are not always mutually compatible. Where the three compete it is invariably the Head-Proximate Filter that takes precedence; the constraints are hierarchically ordered à la Optimality Theory in the following, universal, ranking:

1. (13) Harmonic Word Order Ranking

   Head-Proximate Filter > > Final-Clause, Head Uniformity

It is worthwhile at this point saying a word on the proposed universality of the Harmonic Word Order Ranking. While Optimality Theory generally allows for constraints to be ranked differently in different languages, language-independent, fixed rankings are also found where there are pressures from some grammar-external system – Chomsky's (Reference Chomsky2005) third factor – (e.g. the sonority hierarchy). In this case research has shown that constraints of the type that make up the Harmonic Word Order Ranking are motivated by ease and efficiency of processing (Hawkins Reference Hawkins1990 and subsequent work). Given that principles of processing do not vary from language to language, it is to be expected that the ranking of these constraints should remain uniform across languages.

The application of this ranking can be seen most clearly in the cross-linguistic distribution of subordinating complementisers (henceforth simply ‘complementiser’) heading clausal complements of verbs.Footnote ¹⁰ Typological studies have shown that the data is characterised by two striking left–right asymmetries: firstly, only initial complementisers are found in VO languages, while OV languages allow both initial and final complementisers; secondly, only postverbal complement clauses are attested in VO languages, whereas in OV languages, both preverbal and postverbal complement clauses are permitted. In addition, the complementiser must be base-generated such that it intervenes linearly between its selecting verb and the complement clause.Footnote ¹¹ This leads to a single grammatical order for VO languages, and two possible orders for OV languages (statistics from Dryer Reference Dryer2009a: 199–200; see also Hawkins Reference Hawkins1990 and subsequent work; Dryer Reference Dryer1992: 102; Bayer Reference Bayer1996 and subsequent work):Footnote ¹²

1. (14) VO languages

   V[CVO] 140 languages (100% of VO)

   *[VOC]V 0 languages

   *V[VOC] 0 languages

   *[CVO]V 0 languages

   OV languages

   V[COV] 21 languages (45% of OV)

   [OVC]V 26 languages (55% of OV)

   *V[OVC] 0 languages

   *[COV]V 0 languages

The tableaux below show that this typological pattern is precisely captured by the Harmonic Word Order Ranking. In VO languages, shown in (15), it is possible to obey all three constraints, resulting in a single optimal order – the consistently head-initial order represented by candidate (15a).

1. (15)

The dominant constraint, the Head-Proximate Filter, requires the highest head in the extended projection of the dependent – that is the complementiser – to be linearly contiguous with its superordinate lexical head – the selecting verb. Candidates (15a, b) both obey this constraint. This constraint applies only to the complementiser here, not to either verb, because the Head-Proximate Filter only applies where we are dealing with heads in separate extended projections, and then only to the highest head in the complete extended projection of the dependent. Of the candidates that obey the Head-Proximate Filter, (15a) is chosen over (15b), because (15a) also obeys Final-Clause and Head Uniformity, both of which are violated by (15b): the dependent clause in (15a) follows its superordinate lexical head, the matrix verb, and the complementiser matches the lexical head of its extended projection – the lower verb – in direction of headedness, both being head-initial.

On the other hand, in OV languages, shown in (16), there is no single order that obeys all three constraints.

1. (16)

Therefore, in order to obey the dominant constraint, the Head-Proximate Filter, either Final-Clause or Head Uniformity must be violated, resulting in two possible orders: the consistently head-final order (16b) and the supposedly ‘disharmonic’ initial-over-final order (16a). Furthermore, notice that if, as I argue here, harmony is defined by the ranking in (13), we have an explanation for why the supposedly ‘harmonic’ [[OV]C] and ‘disharmonic’ [C[OV]] are more-or-less equally common among OV languages (see (14)); according to the ranking in (13), both are equally harmonic.

We have seen then that, at least as regards subordinating complementisers, the Harmonic Word Order Ranking correctly derives both the grammatical and the ungrammatical orders. However, there is a second claim to be considered here: if it is true that the Harmonic Word Order Ranking is concerned with base-generated structures, it is predicted not only that the orders [V[CVO]], [V[COV]] and [[OVC]V] should be attested, but that these orders should occur as base-generated structures.

That the order [V[CVO]] is base-generated I take as uncontroversial. This leaves the two orders attested in OV languages: [[OVC]V] and [V[COV]]. Given that the complement clause can appear on either side of the verb, one might suppose that one order is derived from the other through movement. However, data from a number of languages suggests that this is not the case. In both cases, the complement clause, whether pre- or postverbal, is not an island for extraction. The validity of this criterion is confirmed by data following later, where extraction is impossible from a complement clause which has undergone movement (see (23)–(24) below). The lack of island effects for the orders [[OVC]V] and [V[COV]] is demonstrated below by examples from the Southern Dravidian language Malayalam, displaying the order [[OVC]V], and Persian (or Farsi), which has [V[COV]]. (Note that the complementiser, as a linker, is simply glossed lnk.)

1. (17) [aa kuLatt-il]_i ayaaL [waliya miinu-kaL t _i uNTə ennə] paraññu. Malayalam

   that pond-loc he big fish-pl have lnk said

   ‘In that pond, he said that there are big fish.’ (Bayer Reference Bayer1999: 256, ex. (35), citing p.c. from Hany Babu)

2. (18) [Un ketāb-ā = ro]_i man mi-dun-am [ke Kimiyā t _i xar-id-e]. Persian

   that book-pl=acc I impf-know-1sglnk Kimea buy-perf-3sg

   ‘As for those books, I know that Kimea has bought (them).’ (Karimi Reference Karimi2001: 84, ex. (69))

In the Malayalam example (17) an adjunct PP, and in the Persian example (18) the direct object, is extracted from the complement clause to sentence-initial position. Analogous evidence from Japanese, Afrikaans, Dutch, German, Hindi-Urdu, and Turkish can be found in Bayer (Reference Bayer1999: 256), Karimi (Reference Karimi2001), Biberauer, Newton & Sheehan (Reference Biberauer, Newton and Sheehan2009), Biberauer & Sheehan (Reference Biberauer and Sheehan2010: Section 4.2) and references cited in these works.

The fact that the complement clause may be base-generated on either side of the verb in OV languages – as long as the complementiser is adjacent to the verb – is particularly striking in certain OV languages that allow both head-initial and head-final complementisers. This phenomenon is found mainly in Indo-Aryan languages with close geographical or historical contact with Dravidian, such as Assamese, Bengali (or Bangla), Oriya and Marathi, and also the Turkic language Uzbek and the Volta-Congo language Vata (or Dida) (Bayer Reference Bayer1996 and subsequent work; Cinque Reference Cinque2005: 55). In such languages, as predicted by the Harmonic Word Order Ranking and the consequent tableau in (16) above, all complementiser-initial clauses are obligatorily postverbal, while complementiser-final clauses are uniformly preverbal. This is demonstrated in (19) and (20) by Bengali, which has two optional complementisers: bole, which is clause-final, and je, which is clause-initial. The complementiser bole is only possible where the embedded clause is preverbal, while je only appears where the embedded clause is postverbal:

1. (19)

   1. (a) chele-ta [or baba aS-be (bole)] Sune-che. Bengali

      boy-cl his father come-fut.3 lnk hear-perf.3

   2. (b) chele-ta Sune-che [or baba aS-be (*?bole)].

      boy-cl hear-perf.3 his father come-fut.3 lnk

      ‘The boy has heard that his father will come.’ (Bayer Reference Bayer1996: 255, ex. (9))

2. (20)

   1. (a) chele-ta [(*je) or baba aS-be] Sune-che. Bengali

      boy-cllnk his father come-fut.3 hear-perf.3

   2. (b) chele-ta Sune-che [(je) or baba aS-be].

      boy-cl hear-perf.3 lnk his father come-fut.3

      ‘The boy has heard that his father will come.’ (Bayer Reference Bayer1996: 255, ex. (11))

Again, there is evidence that both the preverbal clause, headed by bole, and the postverbal clause, headed by je, are base-generated; neither is an island for extraction:

1. (21) [bas theke]_i amar didi [Otogulo duronto bacca t _i laphi-ye nam-be bole] bhabe ni. Bengali

   bus from my sister so.many uncontrollable children jump-pstpart descend-fut.3 lnk think.3 neg.pst

   ‘From the bus, my sister didn't think that so many uncontrollable children would jump off.’ (Bayer Reference Bayer1999: 255–256, ex. (34b))

2. (22) kriSno mEleria-te_i bhab-che [je ram t _i mara gE-che]. Bengali

   Krishna malaria-loc think-perf.3 lnk Ram die go-perf.3

   ‘Krishna thinks that Ram died of malaria.’ (Simpson & Bhattacharya Reference Simpson and Bhattacharya2003: 130, ex. (8))

We have seen then that not only does the Harmonic Word Order Ranking successfully capture the attested cross-linguistic distribution of complementisers, but also the prediction that the three optimal candidates in (15) and (16) should be base-generated as such is borne out. One final consequence remains to be explored: while the unsuccessful candidates are predicted not to occur as base-generated structures, the possibility is left open that they may occur as surface structures following movement. If we do therefore come across any of the orders found among the unsuccessful candidates, we expect to find evidence of movement.

Certain rigid OV languages, such as Japanese and Malayalam, allow the surface order V[OVC] as a marked variant of the consistently head-final order [[OVC]V].Footnote ¹³ This is exemplified in (23a) for Malayalam and (24a) for Japanese, the postverbal clause being deaccented. Since this order is ruled out as a base-generated structure by the Harmonic Word Order Ranking (see (16)), it is predicted that the complementiser-final postverbal clause in (23)–(24) has moved from preverbal position, and will consequently be an island for extraction. The (b) examples show that this prediction is borne out.

1. (23)

   1. (a) ayaaL paraññu [waliya miinu-kaL aa kuLatt-il uNTə ennə]. Malayalam

      he said big fish-pl that pond-loc have lnk

      ‘He said that there are big fish in that pond.’

   2. (b) *[aa kuLatt-il]_i ayaaL paraññu [waliya miinu-kaL t _i uNTə ennə]

      that pond-loc he said big fish-pl have lnk

2. (24)

   1. (a) %Mary = wa it = ta, [John = ga zibunzisin = o aisiteiru to]. Japanese

      Mary = top say = pst John = nom self = acc love lnk

      ‘Mary said that John loves himself.’

   2. (b) *[Zibunzisin = o]_i Mary = wa it = ta, [John = ga t _i aisiteiru to].

      self = acc Mary = top say = pst John = nom love lnk

This clearly contrasts firstly with the parallel Malayalam example in (17) and the Bengali example in (21), and secondly with the Persian and Bengali examples in (18) and (22), where extraction from either a complementiser-final clause in preverbal position or from a postverbal complement clause with an initial complementiser poses no problem – both these structures being permitted by the Harmonic Word Order Ranking and the consequent tableau in (16).

I have argued here that the ordering properties of complementisers are determined by the Harmonic Word Order Ranking in (13) and the resulting tableaux in (15)–(16). This is not to say however that a complementiser as a lexical item can never have its own ordering statement. Indeed, we have already seen evidence to the contrary for Bengali (as well as the other OV languages with a hybrid complementiser system). Examples (19)–(20) above show that in Bengali the complementiser je can only appear in initial position, selecting a complement to its right, while bole must be final, selecting a complement to its left.Footnote ¹⁴ What is important is that any specialisation within the lexicon must remain within the confines of the principles of the grammar. The Harmonic Word Order Ranking defines what is a possible grammar in terms of the position of complementisers. An individual complementiser may therefore have its own ordering statement, as long as the result is congruous with the Harmonic Word Order Ranking. Therefore, by (16), it is possible for a complementiser in an OV language to have either an initial or a final ordering statement. On the other hand, (15) shows us that a complementiser with a final ordering statement will not be allowed in a VO language.

Finally, note that the complementiser data we have studied in this section serves as an initial piece of evidence supporting the generalisation in (2a): as a type of linker, complementisers should be base-generated such that they intervene linearly between their complement and the head (the higher verb) that selects them. The typology in (14) provides the empirical evidence, while the tableaux in (15) and (16) demonstrate that this result is predicted by the universal Harmonic Word Order Ranking in (13).

3. Disharmony

In the previous section, I suggested that the notion of harmonic word order is determined by the ranking of independently motivated harmonic word order constraints in a universal Harmonic Word Order Ranking in (13). In this section I propose that there is a second type of word order constraint made available by the grammar, which is not concerned with harmony. Constraints of this kind refer to specific syntactic features of a head, and require them to appear either initially or finally within any level of the projection of this head. There are two types of features that can be mentioned by these ordering constraints: the categorial features of a lexical head and features encoding semantics. In principle these ordering constraints can therefore refer to any head as long as it is either lexical or has syntactic features encoding semantics. There is however no expectation that the same lexical or semantic features should appear in the same position cross-linguistically. The ranking of constraints referring to specific syntactic features therefore freely varies from language to language both with respect to the constraints of the Harmonic Word Order Ranking (provided it maintains its own internal ranking) and with respect to other constraints of the same kind. Where these constraints conflict with, and override, any constraints of the Harmonic Word Order Ranking, disharmony arises. Since different languages each have their own ranking, we expect cross-linguistic diversity in terms of the extent of disharmony within individual languages.

We take the syntactic feature of negation by way of illustration, though what follows should apply to any lexical head or to any functional head with syntactic features referring to semantics. For example, disharmony is relatively common for negative markers. As operators, negative markers certainly have syntactic features referring to semantics. Therefore it is possible for ordering rules of the type described above to refer to the negative semantics of the negation head within a given language and require it to appear in a given position.Footnote ¹⁵ The cross-linguistic situation confirms this. There is a tendency to place negative markers in one of two prominent positions (see Dryer Reference Dryer, Hammond, Moravcsik and Wirth1988): initially within their projection, with the result that negation will be expressed as soon as possible, preventing a ‘semantic garden path’ (Jespersen Reference Jespersen1917; Dryer Reference Dryer, Hammond, Moravcsik and Wirth1988: 102); or finally, the position reserved for new or significant information, often the locus of sentence stress (Mazzon Reference Mazzon2004: 5). Now, if an OV language has a dominant ordering rule such that a head expressing negation must appear initially in its merger with the verb phrase, disharmony of the kind seen in (1c) above will be found: the initial-over-final order [Neg[OV]]. In the same way, a VO language may have a dominant rule requiring the negative head to appear finally in its merger with its complement, resulting in the disharmonic order in (1d): the final-over-initial order [[VO]Neg].

The tableaux below illustrate this using Optimality Theory. Suppose that Universal Grammar makes available, in addition to the Harmonic Word Order Ranking, the competing sets of constraints Neg-Initial and Neg-Final. Constraints in the former set require the negation head to appear initially in its (partial) projection, the latter finally. There will be equivalent sets of initial and final constraints for every lexical head and for every head with syntactic features encoding semantics. Note that these are sets of constraints applying to different levels or domains of projection. For example, there could be one Neg-Initial constraint applying to the domain of Neg and its complement and another applying to the maximal projection. In the case where we are dealing with a VO language, the disharmonic order [[VO]Neg] will arise whenever any Neg-Final constraint dominates both Head-Uniformity and Neg-Initial. This is shown in the tableau in (25). The Head-Proximate Filter does not apply here, since we are dealing with word order within a single extended projection (compare (9) above).

1. (25)

However, for any ranking whereby Neg-Final is ranked lower than, or equally with, either one of Head Uniformity or Neg-Initial, the optimal output will be the harmonic order [Neg[VO]]. This is illustrated by the tableaux in (26)–(27).

1. (26)

2. (27)

In (26), Neg-Final dominates Neg-Initial, but is ranked equally with Head Uniformity. The candidate displaying the harmonic order in (b). violates Neg-Final, but the disharmonic candidate in (a), which obeys this constraint, violates the equally ranked Head Uniformity, with the result that any violation of the lower ranked Neg-Initial becomes fatal. In (27), we see a similar situation, except that here it is Neg-Initial that is equally ranked with Neg-Final, and the disharmonic candidate's violation of Head Uniformity that is fatal.

Of course, precisely the same results obtain where either Head Uniformity or Neg-Initial is ranked any higher, or Neg-Final is ranked any lower. In OV languages, we will expect the inverse results: the disharmonic order [Neg[OV]] will obtain only where Neg-Initial dominates both Head Uniformity and Neg-Final; for any other ranking, the harmonic order [[OV]Neg] will be optimal. We expect exactly the same results where Neg is replaced by any other head with syntactic features encoding semantics.

Under this theory of constraints relating to specific syntactic features, the following predictions arise: firstly, disharmony is predicted to be possible cross-linguistically for any head, as long as this head is either lexical or has syntactic features encoding semantics (that is, for any head that is not a linker, see Section 1 above); secondly, for any given such head, disharmony is predicted to be comparatively infrequent, or cross-linguistically dispreferred. The conditions allowing disharmony to arise are such that there are more logically possible rankings resulting in harmony than those resulting in disharmony.Footnote ¹⁶ (For example, of the thirteen logically possible rankings of the relevant constraints Head Uniformity, Neg-Initial and Neg-Final, seven invariably result in harmony as regards the order of the negative head in relation to the verb; three result in harmony for OV languages and disharmony for VO languages; three result in harmony for VO languages and disharmony for OV languages.) These predictions in fact lead to the generalisation in (2b). The findings of broad typological studies provide an initial indication that these predictions, and hence the generalisation in (2b), are borne out. This can be seen in the data in (28) and (29), concerning the extended projection of the verb.

1. (28)

   1. (a) Initial-over-initial [Aux[V O]] 79 languages (55%)

   2. (b) Final-over-final [[O V]Aux] 30 languages (21%)

   3. (c) Initial-over-final [Aux[O V]] 19 languages (13%) (39% of OV languages)

   4. (d) Final-over-initial [[V O]Aux] 16 languages (11%) (17% of VO languages)

2. (29)

   1. (a) Initial-over-initial [Q[V O]] 75 languages (20%)

   2. (b) Final-over-final [[O V]Q] 127 languages (34%)

   3. (c) Initial-over-final [Q[O V]] 34 languages (9%) (21% of OV languages)

   4. (d) Final-over-initial [[V O]Q] 135 languages (36%) (64% of VO languages)Footnote ¹⁷

Here the disharmonic orders in the (c) and (d) examples arise where some specific constraint of the type described above outranks Head Uniformity. The data in (28), from Julien (Reference Julien2002: 330–356), shows the relative ordering of the verb and an auxiliary (independent tense, aspect or mood marker), and that in (29), from Bailey (Reference Bailey2010: 29, Table 1, using data from Dryer Reference Dryer2008a, Reference Dryerb), that of the verb and a question particle.

Data from the nominal domain, where an NP containing a head noun modified by a possessive phrase is contained within a PP, from Hawkins (Reference Hawkins1994: 259, Table 5.2, using data from Hawkins Reference Hawkins1983), shows the same, predicted, distribution:

1. (30)

   1. (a) Initial-over-initial [P[N PossP]] 134 languages (40%)

   2. (b) Final-over-final [[PossP N]P] 177 languages (53%)

   3. (c) Initial-over-final [P[PossP N]] 14 languages (4%) (7% of N-final lgs)

   4. (d) Final-over-initial [[N PossP]P] 11 languages (3%) (8% of N-initial lgs)

This kind of distribution is not limited to examples within a single extended projection. Precisely the same situation holds across extended projections (except where the relationship is marked by a linker). In this case the disharmonic orders arise due to some specific constraint referring to the semantic or lexical features of a head outranking the Head-Proximate Filter. This can be seen where the verb takes an adpositional complement, in (31), using data from Dryer (Reference Dryer2008c) and Sheehan (Reference Sheehan2008: Section 4, 2011: 47), and where it takes a nominal complement, in (32) and (33), showing data from Dryer (Reference Dryer1992: 104, Table 34 and page 91, Table 5):

1. (31)

   1. (a) Initial-over-initial [V[P NP]] 419 languages (47%)

   2. (b) Final-over-final [[NP P]V] 427 languages (48%)

   3. (c) Initial-over-final [V[NP P]] 38 languages (4%) (8% of postpositional lgs)

   4. (d) Final-over-initial [[P NP]V] 12 languages (1%) (3% of prepositional lgs)

2. (32)

   1. (a) Initial-over-initial [V[D NP]] 37 genera (44%)

   2. (b) Final-over-final [[NP D]V] 19 genera (23%)

   3. (c) Initial-over-final [V[NP D]] 15 genera (18%) (29% of VO genera)

   4. (d) Final-over-initial [[D NP]V] 13 genera (15%) (41% of OV genera)

3. (33)

   1. (a) Initial-over-initial [V[N PossP]] 63 genera (29%)

   2. (b) Final-over-final [[PossP N]V] 12 genera (52%)

   3. (c) Initial-over-final [V[PossP N]] 30 genera (14%) (21% of N-final genera)

   4. (d) Final-over-initial [[N PossP]V] 112 genera (6%) (16% of N-initial genera)

Consider now the situation with linkers. As discussed in Section 1 above, these are distinguished from other heads by their semantic vacuity. I have proposed here that disharmonic word orders arise as a result of ordering constraints within the grammar requiring a head with specified syntactic features encoding either lexicality or semantics to appear either initially or finally at some level of its projection. As we saw in the previous section, it is possible for a linker to have some kind of ordering statement as part of its lexical properties; however, this lexical ordering statement is strictly constrained by what is permitted in the grammar. What we do not expect to see applying to linkers are ordering constraints within the grammar of the type we have seen above, since these constraints refer exclusively to either lexical heads or features relating to semantics. The linker, however, by its very definition does not have such features. This explains why in the previous section, the position of subordinating complementisers (as a member of the class of linkers), and any possible lexical ordering statement, is always determined purely by the Harmonic Word Order Ranking in (13).

For example, we saw in the previous section that the complementisers je and bole in Bengali must obey the dominant constraint, the Head-Proximate Filter: complement clauses headed by initial je must be postverbal; those headed by final bole must be preverbal (see (19)–(20) above). Interestingly, it is possible for a clause headed by bole to appear postverbally – thereby violating the Head-Proximate Filter and hence the Harmonic Word Order Ranking – where it functions not as a subordinating complementiser, but heads a reason adverbial clause; in this case, bole does not mark the presence of an existing relationship (i.e. it is not a linker), but rather introduces its own relationship:

1. (34) ami ekhane eSe-chi [tomar SONge kOtha bol-bo bole]. Bengali

   I here come-pst.1 you with speech say-fut.1 because

   ‘I have come here in order to talk with you.’ (Bayer Reference Bayer1996: 255, ex. (10))

This shows that there is no ban on postverbal bole-clauses per se, but simply those where the subordinating bole is semantically vacuous.

Similarly, a brief look at the diachronic development of the subordinating complementiser kī in the now extinct East Semitic language Akkadian provides direct support that it is indeed the lack of semantics in linkers that prohibits disharmonic word order. Akkadian was an SOV language spoken in ancient Mesopotamia. The subordinating complementiser kī developed over time from the adverbial conjunction kīma. This morpheme kīma occurred in initial position within the adverbial clause, which itself overwhelmingly appeared in preverbal position. An example containing kīma from the Old Babylonian period is given in (35):

1. (35) [kīma udammiqak-kunūši] dummikā-nim. Akkadian

   as 1sg.do.favours.pst-to.you.pl do.favours.imp.pl-to.me

   ‘As I have done you favours, do me favours.’ (Deutscher Reference Deutscher2007: 40, ex. (27))

Deutscher (Reference Deutscher2007: Section 4) documents a change in this morpheme's semantics of the kind comparative >causal/purpose >factive, leading to an eventual complete semantic bleaching; that is, a semantically contentful conjunction introducing an adverbial clause (kīma) eventually became a subordinating complementiser heading a complement clause (kī). Crucial to the theory put forward here, however – and its predictions – is the result of this change on word order. Throughout the diachronic change, the morpheme kī(ma) remains consistently clause-initial. However, the semantic bleaching of the morpheme is accompanied by a shift in position of the clause it heads: by the Middle Babylonian period complement clauses headed by the subordinating complementiser have shifted to postverbal position, while adverbial clauses remain preverbal:

1. (36) bēl-ī īde [kī ultu ēlâ dilipt-u mahratan-ni]. Akkadian

   lord-my 3sg.know lnk since 1sg.arrive.pst trouble-nom 3fsg.contront.stat-me

   ‘My lord knows that since I arrived, trouble has befallen me.’ (Deutscher Reference Deutscher2007: 51, ex. (57))

In terms of the theory proposed in this paper, the shift of position for clauses headed by a subordinating complementiser, but not for those introduced by a semantically contentful adverbial conjunction, is easily understood. Where a clause is headed by a semantically vacuous subordinating complementiser, its position is determined uniquely by the Harmonic Word Order Ranking in (13) (since no ordering constraint referring to either lexical features or semantics can apply). The tableau in (16) shows us that the only grammatical option for an OV language with an initial subordinating complementiser involves a postverbal complement clause, giving the order attested in Akkadian, [V[COV]]. The final-over-initial order [[COV]V] that would result if the subordinating complementiser retained the position of its semantically contentful predecessor would be disharmonic, and hence ungrammatical. On the other hand, examples such as (35) are unproblematic for the theory put forward here, since kīma in this example is semantically contentful and therefore may have its own ordering constraints such that disharmony is possible.Footnote ¹⁸

We have seen then both the theoretical motivation for part (b) of the generalisations in (2), and some empirical indications supporting the predictions: firstly statistical evidence from broad typological studies, and secondly by comparing semantically vacuous linkers with their semantically contentful synchronic or diachronic counterparts. The rest of this section will be devoted to studying some of the relevant disharmonic structures in greater detail.

We begin by considering markers of verbal properties such as tense, aspect and mood. Cross-linguistic studies such as Cinque (Reference Cinque1999, Reference Cinque2009) and Julien (Reference Julien2002) firstly argue that these markers are functional heads in the extended verbal projection, and secondly show that these heads obey a strict, universal, hierarchy, or functional sequence (see also Grimshaw Reference Grimshaw1991, Reference Grimshaw2005 and references cited in the above works). These findings, which appear to be without exception, are based on the properties and placement of such markers both in head-initial and head-final languages.

Given, then, that markers of tense, aspect and mood – both initial and final – bear syntactic features encoding semantics, it is predicted that ordering constraints referring to these features should be able to target these heads, resulting in disharmony. We have already seen typological evidence, in (28) above, suggesting this prediction is borne out. Concrete examples given below in (37)–(39) support this evidence.

1. (37) Yi [a [yì-yé kúnúŋ]]. Supyire

   they perf they-refl be.close.by

   ‘They are (too) close to each other.’ (Carlson Reference Carlson, Traugott and Heine1991: 215, ex. (30))

The example in (37) demonstrates the initial-over-final order in the North Volta-Congo language Supyire: the head-final VP is c-commanded by an initial aspect marker, a. (See also examples from the Ge-Kaingang language Canela-Krahô in (48) and (73) below, where a head-initial TP dominates a head-final VP.)

The inverse, final-over-initial, construction, whereby final tense or aspect markers are found in VO constructions, is also relatively common (Biberauer et al. Reference Biberauer, Holmberg and Roberts2007 and subsequent work; Dryer Reference Dryer, Cyffer, Ebermann and Ziegelmeyer2009b). This is shown in examples from the Central Sudanic language Bagirmi and the North Volta-Congo language Mumuye: the Bagirmi example in (38) shows a final aspect marker, ga, c-commanding a head-initial VP, while the final tense marker, ni, in the Mumuye example in (39) c-commands a head-initial AspP:

1. (38) bɨs [sa ja tebɨre] ga. Bagirmi

   dog eat meat yesterday compl

   ‘The dog ate the meat yesterday.’ (Stevenson Reference Stevenson1969: 85, gloss Dryer Reference Dryer, Cyffer, Ebermann and Ziegelmeyer2009b: 344, ex. (104))

2. (39) Znàsọ [dé baasé Ranti] ni. Mumuye

   Znaso perf mimic Ranti immed.fut

   ‘Znaso is about to mimic Ranti.’ (Dryer Reference Dryer, Cyffer, Ebermann and Ziegelmeyer2009b: 345, ex. (106b), citing Shimizu Reference Shimizu1983: 112)

The same results can be shown for the negation head, though the situation here is less straightforward. Firstly, negative particles are not always heads, but in many languages are adverbs. Secondly, where the negative marker is an adverb in a double negation language, NegP is usually not projected (Zeijlstra Reference Zeijlstra2004). Moreover, where it is present, the position of NegP within the functional sequence varies from language to language (Laka Reference Laka1990; Ouhalla Reference Ouhalla1991; Zanuttini Reference Zanuttini1991, Reference Zanuttini1997; Cinque Reference Cinque1999; Julien Reference Julien2002). It therefore does not necessarily follow that any attested sequence of either [Neg[OV]] or [[VO]Neg] exemplifies the predicted disharmonic structure; further investigation is required. Since NegP is generally not projected in double negation languages, we restrict our discussion to negative concord languages, which under most theories of negative concord must project NegP (inter aliaZanuttini Reference Zanuttini1991, Zeijlstra Reference Zeijlstra2004). Our goal will therefore be to find negative concord languages in which the negative marker is a head, and this head differs in direction of headedness from the verb.

An example of the disharmonic initial-over-final order is found in the Northern Tungusic language Evenki, an OV language:

1. (40)

   1. (a) Bi [dukuwūn-ma duku-cā-w]. Evenki

      I letter-obj write-pst-1sg

      ‘I wrote a letter.’

   2. (b) Bi [ə-c-w [dukuwūn-ma duku-ra]].

      I neg-pst-1sg letter-obj write-part

      ‘I didn't write a letter.’ (Payne Reference Payne and Shopen1985: 213, ex. (31))

This is demonstrated in example (40b). It is clear here that the negative marker ə- is a head in the extended verbal projection, since it inflects for subject agreement and tense. Comparing the negative example in (40b) with its affirmative counterpart in (40a), we see that these markings would otherwise appear on the lexical verb; in the negative example the lexical verb has a participial form. This confirms not only that the negative marker is a head, but also that the construction is monoclausal (see Payne Reference Payne and Shopen1985: 212–214).

Turning now to the inverse disharmonic order, [[VO]Neg] is relatively common in central Africa (Dryer Reference Dryer, Cyffer, Ebermann and Ziegelmeyer2009b) and the Pacific region (Biberauer et al. Reference Biberauer, Holmberg and Roberts2010: 56, citing Reesink Reference Reesink2002). Here again it can be shown that, at least in some cases, this final negative marker is a head, and thus instantiates the disharmonic final-over-initial order. Firstly, like Evenki, the Central Sudanic language Ma'di marks negation by means of a negative auxiliary. This is shown in (41), where the final negation head is marked for tense, while the lexical verb remains unmarked. On the other hand, in the absence of negation, the same past/non-past distinction is marked on the lexical verb, as in (42).

1. (41)

   1. (a) [m′-āwí ʤótī] k. Ma'di.

      1sg-open door neg.nonpst

      ‘I won't open/am not opening the door/don't open doors.’ (Blackings & Fabb Reference Blackings and Fabb2003: 14, ex. (8))

   2. (b) [m′-āwí ʤótī] kr.

      1sg-open door neg.pst

      ‘I did not open the door.’ (Blackings & Fabb Reference Blackings and Fabb2003: 14, ex. (7))

2. (42) ká gbándà ā. Ma'di

   3sg cassava nonpst.eat

   ‘He is eating/eats cassava.’ (Blackings & Fabb Reference Blackings and Fabb2003: 13, ex. (1))

In languages where the negative marker is an uninflected particle, there may still be evidence that this particle heads NegP. An example of such a language is Lagwan (or Logone, Central Chadic), which exhibits the order T[VO]Neg:Footnote ¹⁹

1. (43) [Sà dɨ-gɨr kasku diyasɨn] sá. Lagwan

   fut 3fsg-go market tomorrow neg

   ‘She won't go to the market tomorrow.’

A number of factors indicate that the negation marker here heads the projection NegP, and therefore exemplifies the disharmonic final-over-initial order. Firstly, example (44) provides evidence that Lagwan is a negative concord language, and therefore projects a NegP: the negative marker sá is required to license the negative concord item:

1. (44) Bɨle a shima [[a lo] *(sá)]. Lagwan

   man lnk.mnci.m 3msg.perf come neg

   ‘Nobody came.’

Secondly, evidence from prohibitives shows that NegP has a fixed position in the Lagwan clause, dominating head-initial TP. Lagwan finite indicative clauses, as in (43) and (44), are marked for tense or aspect. Imperatives, on the other hand, lack any marking for tense or aspect, conforming to Kayne's (Reference Kayne, Tasmowsky and Zribi-Hertz1992) conclusion that true imperatives are truncated structures lacking TP and any higher functional projections:

1. (45) Slà a! Lagwan

   push.imp up

   ‘Get up!’

However, as is the case in many languages (see Zanuttini Reference Zanuttini1991 and subsequent work, Kayne Reference Kayne, Tasmowsky and Zribi-Hertz1992, Zeijlstra Reference Zeijlstra2004, among others), true negative imperatives are banned in Lagwan: as shown in (46a), it is not possible to attach the negative marker to an imperative clause. Instead, a surrogate is used, with the future tense head projected and marked for second person, as in (46b):

1. (46)

   1. (a) *Slà a sá! Lagwan

      push.imp up neg

   2. (b) [Sà gɨ-sla a] sá!

      fut 2sg-push up neg

      ‘Don't get up!’

Zanuttini (Reference Zanuttini1994, Reference Zanuttini, Belletti and Rizzi1996) argues that in languages banning true negative imperatives, NegP must dominate TP, thereby rendering a structure in which NegP is present but TP is absent (such as (46a)) ungrammatical. The fact that in Lagwan the ungrammatical structure is repaired by projecting TP seems to confirm this hypothesis.

Finally, there is evidence not only that NegP is projected, and that this NegP dominates head-initial TP and VP, but also that the final negative marker heads this projection, as opposed to being an adverb in its specifier. This evidence is found by applying the ‘why not’ test developed by Merchant (Reference Merchant2006). The ‘why not’ construction is analysed as a form of phrasal adjunction, and hence only permitted where the negative marker is a maximal projection. In Lagwan, no such construction is possible:

1. (47)

   1. (a) *[Age ghwani] sá? Lagwan

      because what neg

   2. (b) *Sá [age ghwani]?

      neg because what

We have seen then that heads marking tense, aspect and negation can all, as predicted, violate Head Uniformity, resulting in disharmony. The typological data in (29) above indicates that the same is true of interrogative heads. This is exemplified firstly in Canela-Krahô (Ge-Kaingang), an OV language with an initial interrogative marker, xà:

1. (48) xà capi te [po curan]? Canela-Krahô

   q Capi pst deer kill

   ‘Did Capi kill a deer?’ (Popjes & Popjes Reference Popjes, Popjes, Derbyshire and Pullum1986: 157, ex. (186))

The inverse disharmonic order, whereby a VO language has a final interrogative marker, is exhibited, among very many other languages, in Lagwan. Lagwan has a final interrogative head (ɗa), c-commanding an initial TP and VP. It occurs in both matrix and subordinate clauses, as shown in (49) and (50) respectively:

1. (49) [G-a mma ì gha] ɗa? Lagwan.

   2sg-perf leave her.acc house q

   ‘Did you leave it at home?’

2. (50) Ndalu ngwa fɨne, ki [[bɨle a shi a s-o gha] ɗa].

   1sg.prog look.at outside lnk man lnk.m some 3sg.m.perf enter-vent house q

   ‘I'm looking outside, (to see) whether someone has entered the house.’ (Aaron Shryock, p.c.)

As regards interrogative markers and other discourse-related particles assumed to head projections in a layered CP, note that the semantic explanation for disharmony accounts for an otherwise anomalous situation. We have seen in the previous section that for subordinating complementisers, the disharmonic final-over-initial order is ungrammatical (see (14)). However, this order is relatively common for other discourse-related heads assumed to belong to the CP domain, including, as we have seen, question particles (see Julien Reference Julien2002: Section 3.6.1; Biberauer et al. Reference Biberauer, Holmberg and Roberts2007 and subsequent work; Paul to appear). If the presence or absence of disharmony is related to the presence or absence of syntactic features encoding semantics rather than directly to category, the anomaly disappears.

So far we have considered examples within a single extended projection – that is, examples where Head Uniformity is violated due to some higher-ranked, specific, constraint. Before concluding this section we will take a brief look at some more complex examples from the nominal domain, spanning more than one extended projection, where the Head-Proximate Filter also comes into play.

We consider Greek and Russian, and then Germanic languages. In Greek and Russian the determiner occurs initially in its projection, while the noun follows its adjectival dependent. Since D differs in direction of headedness from the lexical head of its extended projection (the noun), the constraints D-Initial and N-Final (applying to the domain of the noun and its adjectival dependent) must both dominate Head Uniformity, as shown in the tableau in (51)–(52). In addition, N-Final must dominate the Head-Proximate Filter.Footnote ²⁰ This is shown by the sub-tableau in (52), where the AP modifying the noun is head-initial, thereby violating the Head-Proximate Filter, since the adjective (the head of the subordinate domain, or dependent) is not linearly contiguous with its superordinate head, the noun.

This contrasts with Germanic languages, where again the constraints D-Initial and N-Final must both dominate Head Uniformity, but this time N-Final is ranked below the Head-Proximate Filter. Where the noun is modified by a simple adjective, as in (53) (or indeed by any head-final AP), we have the same results as in Greek and Russian: the higher-ranked D-Initial and N-Final require a violation of Head Uniformity. However, where the noun is modified by a head-initial AP, as in (54), the results are different: in order to obey the undominated Head-Proximate Filter, N-Final must be violated, resulting in a fully harmonic, consistently head-initial structure:

4. Harmony meets disharmony

In the previous two sections we established firstly a theory of harmony, consisting of a universal ranking of harmonic word order constraints (13), and secondly a theory of disharmony, allowing divergence from the order determined by the Harmonic Word Order Ranking by higher-ranked specific ordering constraints within individual languages. We determined that this latter type of constraint cannot apply to linkers, due to their status as semantically vacuous functional heads. We have already seen the effects of this with subordinating complementisers, where only the Harmonic Word Order Ranking applies (Section 2). In this section we return to two further types of linker: syntactically independent relative clause markers and more general linkers in the complex noun phrase. These cases are of particular interest because the lexical head of the relationship marked by the linker and its dependent are of different categories, thereby enabling us to consider the outcome where harmonic and disharmonic structures interact. We will see that, even where the superordinate noun has a specific ordering constraint conflicting with the constraints of the Harmonic Word Order Ranking, irrespective of the ranking of this specific ordering constraint, the linker always obeys the Head-Proximate Filter, hence the generalisation in (2a). The results of this study will not only add empirical support to the theories put forward in Sections 3–4, but also shed some light on the well-observed near parallels between the distribution of relative clauses and complement clauses, and why such parallels fail to be realised completely.

5. Harmony, disharmony, and FOFC

6. Concluding remarks

I have proposed here that the notion of harmonic word order be defined by the ranking of the Head-Proximate Filter, Head Uniformity and the Final-Clause requirement, with the Head-Proximate Filter universally taking precedence. Evidence supporting this ranking has been given by various types of linker, in particular complementisers, relative clause markers, and linkers in the complex noun phrase. I have shown that deviation from the optimal order is possible only where specific ordering constraints target either a lexical head or syntactic features encoding semantics. The possible presence or absence of disharmony is therefore directly related to the presence or absence of certain features on a head. The combination of these theories of harmony and disharmony results in the generalisations in (2). These generalisations have been shown to be empirically superior to FOFC, in terms of both what is permitted and what is disallowed. Furthermore, our study has shown that harmony is concerned not with the surface linear order, but with the base-generated, or first-merged, order. This is suggestive that ordering at least to some extent is a product of the narrow syntax, rather than belonging exclusively to PF.