3.1 The phrase-based account: the XP trigger hypothesis

The XP trigger hypothesis has its roots in an informal suggestion made by T. J. Rhys Jones (Rhys Jones Reference Rhys Jones1977: 328, 338), and was developed within generative grammar by Harlow (Reference Harlow and Heny1981, Reference Harlow1989), Borsley & Tallerman (Reference Borsley and Tallerman1996), Tallerman (Reference Tallerman1990, Reference Tallerman2006), Borsley (Reference Borsley1997, Reference Borsley1999) and Borsley, Tallerman & Willis (Reference Borsley, Tallerman and Willis2007). The principal idea in this literature is that initial consonants undergo SM when immediately preceded by a phrase, XP. Before turning to the formalization of what Borsley & Tallerman (Reference Borsley and Tallerman1996) term the XP trigger hypothesis (XPTH), we can get an informal idea of its basic predictions by examining the data in (1) to (7). In all these examples, a phrasal category immediately precedes the element undergoing the mutation. So in (1), (2) and (3), the triggering XP is in each case y ddynes ‘the woman’, the subject of the root clause (in (1), the only clause). In (4) the trigger is also a subject, but this time the subject of the non-finite embedded clause, Mair. In (5), the trigger is again the subject, i ‘I’, and the mutation appears on the first segment of the following AP predicate. However, in (6), it is the subject itself, telyn hen iawn ‘a very old harp’, which undergoes SM, and the triggering XP is the preceding PP, ar y llwyfan ‘on the stage’. And in (7), the constituent which undergoes the mutation is itself a PP, dwy droedfedd yn nes at y ffidil ‘two feet nearer the fiddle’, with the object, y delyn ‘the harp’, as the immediately preceding XP trigger. Structural assumptions for some basic examples are discussed below.

Obviously, the exact predictions made by the XP trigger hypothesis depend on its formulation, and on what kind of structures are proposed. The version of the XPTH which I will adopt here is shown in (8), adapted from Borsley (Reference Borsley1999: 286):

1. (8) A complement bears SM if it is immediately preceded by a c-commanding phrase.

In Borsley's HPSG analysis, syntactic structures are essentially flat, so that in a simple VSO clause such as (1), the head verb, subject and object are all sisters:Footnote ³

1. (9)

   

Given such a structure, the XPTH predicts that the object, a complement, bears SM, since it is immediately preceded by a c-commanding phrase (the subject). The basic data in (1) to (5), for instance, all have the subject XP as the mutation trigger, and can thus be accounted for in a similar fashion; see Borsley (Reference Borsley1999) for full illustration. Examples such as (6) and (7) have different XPs as the mutation trigger, but are all accounted for straightforwardly under (8), given the assumptions of a flat structure analogous to that in (9).

Note that (8) predicts that only complements – and not constituents which are merely part of a complement, nor adjuncts – will undergo syntactic SM.Footnote ⁴ This restriction gives rise to an important change in the predictions made by an earlier formulation of the XPTH by Borsley & Tallerman (Reference Borsley and Tallerman1996), which proposed that the target was a constituent (not necessarily a complement) immediately preceded by XP. As Borsley's restriction is also relevant to the evaluation of the Word Grammar analysis (section 4.4.1), it is worth briefly outlining the main evidence that the target is the entire complement. One critical piece of data is shown in (10):

1. (10) Prynodd y ddynes delyn newydd, corn/*gorn a ffidil. (telyn)buy.past.3s the woman harp new horn horn(+SM) and fiddle ‘The woman bought a new harp, a horn and a fiddle.’

Here, as in (1), SM appears on the direct object, telyn newydd, corn a ffidil ‘a new harp, a horn and a fiddle’, and is triggered by the subject XP; as usual, the mutation appears on the initial segment, so telyn>delyn. Note, however, that whatever the analysis of co-ordination, it would be reasonable to expect it to involve a structure in which each conjunct c-commands the next conjunct. Since each conjunct is itself a full XP, the question then is why the first conjunct telyn newydd ‘a new harp’ does not trigger mutation of the second conjunct corn ‘horn’, and so on.Footnote ⁵ Yet if corn in (10) is mutated to *gorn, the result is ungrammatical. Crucially, when the target for SM is a series of conjoined phrases, the mutation only appears on the first conjunct, and not on any subsequent conjuncts. Borsley's solution (Reference Borsley1999: 286) is that the whole complement itself – in this case, the object of the clause, telyn newydd, corn a ffidil – is the only target for SM. Hence, the mutation appears just on its initial segment. Since each individual conjunct is not itself a complement, but only part of the complement, the formulation of the XPTH in (8) correctly predicts that only the first conjunct bears SM.

One of the questions raised by this restriction, though, concerns the definition of ‘complement’. First, given such data as (6), ‘complement’ must also include the subject of a clause, since subjects in the appropriate context are targets for syntactic SM. In most theoretical frameworks, subjects are not considered to be complements, although they are of course arguments. However, some versions of the HPSG model propose that a post-verbal subject is a member of the COMPS list, i.e. the set of complements of a head; see Borsley (Reference Borsley1989, Reference Borsley1999) and Sag, Wasow & Bender (Reference Sag, Wasow and Bender2003). As we will see in section 3.2, a rather similar assumption is made within Word Grammar (e.g. Hudson Reference Hudson2007), where subjects are also valents, and as such are included in the set of possible mutation targets under the assumptions made by this account too. Second, predicates must also be treated as complements/valents, since in the appropriate context they too are targets for syntactic SM. For instance in (4), which contains an embedded i-clause, the predicate canu'r delyn ‘play the harp’ bears the mutation. The adjectival predicate in (5) is similarly a target for SM, hence must also be a complement. Although these decisions may appear controversial in some frameworks, such a treatment allows a single generalization to be made across a wide range of mutation data. I will assume from now on that restricting mutation targets to complements/valents in the above sense is correct, that such a restriction must apply to any version of the XPTH, and must equally apply to the dependency account of syntactic SM.

In empirical terms, the phrase-based account has proved highly successful as a hypothesis predicting the occurrence of Welsh syntactic SM. Nonetheless, as a mechanism it has been criticized recently by Ian Roberts for failing to relate to other grammatical phenomena, or at least to phenomena postulated within a P&P framework, such as Case theory or checking theory (Roberts Reference Roberts2005: 76). It is far from self-evident that the XPTH must therefore be wrong, particularly as Roberts' own P&P account of syntactic SM, in terms of Case-licensing, has serious empirical flaws (Tallerman Reference Tallerman2006, Borsley et al. Reference Borsley, Tallerman and Willis2007: chapter 7). However, we can now compare the XPTH with a completely novel account of syntactic SM, which does include a proposal about the function of the mutation, and also about how this function relates to other observed grammatical phenomena. In section 3.2, then, I compare the XPTH with a new surface-based analysis, a dependency account.

3.2 Dependency distance: a Word Grammar account

The best-known syntactic models, including both P&P/Minimalism and HPSG, are based on phrase structure. Word Grammar (e.g. Hudson Reference Hudson1990, Reference Hudson2007, Creider & Hudson Reference Creider, Hudson, Sugayama and Hudson2006), on the other hand, is based on dependency structure between individual words, and has no phrasal nodes. Hudson (Reference Hudson2007: 124–129) suggests that the results of experiments on language processing show that phrase structure lacks psychological reality; instead, ‘the only relevant units seem to be words’ (Hudson Reference Hudson2007: 125). Specifically, there is a substantial body of work on what is known as ‘dependency distance’, which refers to ‘the number of words that separate a word from its parent … the greater the distance, the harder the processing’ (Hudson Reference Hudson2007: 124). Evidence from psycholinguistics and language typology indicates that difficulty in processing increases with dependency distance. Furthermore, grammatical structures cross-linguistically are generally organized in such a way as to minimize dependency distance; see, for instance, Hawkins (Reference Hawkins1990, Reference Hawkins1994, Reference Hawkins2001, Reference Hawkins2004).

Dependency distance relates to syntactic SM in the following way: the basic proposal is that the mutation has a function, namely to mark the separation of a dependent from its head. In a brief sketch of this idea, Hudson (Reference Hudson2007: 126) notes that Welsh ‘seems to have explicit markers of dependency distance. SM … applies to verb dependents under complex conditions … which may reduce to being separated from the verb’.Footnote ⁶ In fact, as we will see, cases where dependents are separated from verbal heads constitute only one instance of the overall phenomenon of syntactic SM. Stated more broadly, the proposal is that the mutation signals the separation of a dependent from any head. Interestingly, this idea reflects, and indeed extends, an environment noted in traditional Welsh grammars, known as a sangiad, which refers to cases where the normal word order is disrupted by some intervening phrase, and SM immediately follows (e.g. Morgan Reference Morgan1952: chapter 19).

If the proposal for a function for syntactic SM could be maintained – and providing the account of the mutation is empirically accurate – this would be a highly interesting result, and the analysis might even replace the XPTH account, which proposes no function for the mutation. However, we first need to establish that the Word Grammar account proves successful in accurately predicting the occurrence of syntactic SM in as many contexts as the XPTH – and as we will see, this turns out not to be the case. Moreover, if a phrase-based account such as the XPTH is actually more successful in predicting the environments for the mutation, then we will have some evidence for the psychological reality of phrase structure. In fact, it appears that an account which does not recognize the existence of the phrase requires a significantly more complex analysis than an account which is phrase-based.

The proposal can be initially formulated as in (11), which I refer to from now on as the dependency distance (DD) account:

1. (11) The dependency distance account of syntactic SM (preliminary version)

   X bears soft mutation if:

   1. (a) X is a valent of a preceding overt head H, and

   2. (b) X is separated from H by Z.

As noted above, the term ‘valent’ in Word Grammar covers both subjects and complements; see Hudson (Reference Hudson2007: 165). As we saw in section 3.1, ‘complement’ is used in this way in Borsley's (Reference Borsley1999) HPSG account. Other types of dependents, such as optional locative expressions, are not valents (Word Grammar does not recognize the category ‘adjunct’; see Hudson Reference Hudson2007: 162–164). Word Grammar thus distinguishes between valents and non-valent dependencies: only the former are predicted to be targets for syntactic SM. However, the ‘complement’ of HPSG and ‘valent’ of Word Grammar are not exact equivalents – in fact, there are critical differences. In any grammatical model which is based on phrase structure (like HPSG), a complement is an entire phrase selected by some head. Word Grammar, on the other hand, as its name suggests, relates words to each other. So an individual word can be the valent of a head, but a phrase cannot, because only individual words contract relationships in the system. Each element in a phrase is either a parent or a dependent, and each parent in turn is a dependent of some other word within a larger phrase, and so on up to the sentence root (Hudson Reference Hudson2007: 118). A phrase is just a word, its dependents, its dependents' dependents, and so on, and there are no nodes corresponding to phrases. We will see in section 4.4.1 that this crucial difference proves problematic for the DD account of co-ordination.

Before I turn to an outline of the predictions made by the DD account, note first that under (11a), the relevant head from which the target valent is separated must precede that valent, and must also be overt. The data in (12) and in (13) show why this restriction is necessary.

1. (12) Telyn/*delyn brynodd y ddynes. harp (+SM) buy.past.3s the woman ‘It was a harp that the woman bought.’

1. (13) Pa offeryn brynodd y ddynes? Telyn/*delyn. what instrument buy.past.3s the woman harp (+SM)‘What instrument did the woman buy? A harp.’

In (12) we see that a fronted object (where the head verb does not precede the valent) does not bear the mutation, thus the need for preceding in (11).Footnote ⁷ And (13) shows that sentence fragment objects (telyn) do not bear the mutation; this accounts for the need for the relevant head to be overt in (11). As we will see in section 4.4.2, however, the requirement for the valent to be separated from an overt head in order for SM to occur is problematic in light of data involving ellipsis. Hudson (Reference Hudson2007: 180) proposes that sentence fragments of this kind are ‘syntactically complete’, but with phonetically unrealized material. Note that in contrast, the XPTH analysis must regard sentence fragments as bare arguments; in other words, these do not involve the postulation of invisible structure. There is thus no covert material to trigger SM on the object in (13).

I next outline how the DD account handles both straightforward and also less obvious instances of mutation. Very often, it is the subject which separates the relevant head from its mutated valent. In (14) (=(1)), the object telyn is separated from the verb of which it is a valent by the subject, so the object bears the mutation. Example (14b) shows the dependency relations in the standard Word Grammar notation. (Note that the determiner at the start of the subject noun phrase is a dependent of the verb, with the noun itself a dependent of the determiner. I return to this point below.)

1. (14)

   1. (a) Prynodd y ddynes delyn. (telyn)buy.past. 3s the woman harp ‘The woman bought a harp.’

   2. (b)

      

The mutation in both (2) and (3) appears on the non-finite verb, and this can be handled in an identical fashion, provided that we make the reasonable assumption that this mutated verb is in each case a valent of the initial finite auxiliary or verb. Again, the valent is separated from its head by the intervening subject.

By contrast, the subject itself normally immediately follows the verb, so it typically undergoes no mutation. However, when the subject is separated from the verb, then it does indeed mutate:

1. (15)

   1. (a) Mae ar y llwyfan delyn hen iawn. (telyn)be.pres.3s on the stage harp old very ‘There is on the stage a very old harp.’ (=(6))

   2. (b)

      

Under the generalization in (11), any verbal dependents are predicted to bear syntactic SM when they are separated from their head. This accounts for the mutation of the third argument of three-argument verbs, such as the locative complement in (7): these are always separated from the head verb by both the subject and the object, so again are predicted to bear the mutation.Footnote ⁸

In (4), repeated here as (16), the mutation appears on the non-finite verb (the start of the predicate) in an embedded clause:

1. (16)

   1. (a) Dymunodd Aled i Mair ganu 'r delyn. (canu)want.past.3s Aled to Mair sing.inf the harp ‘Aled wanted Mair to play the harp.’

   2. (b)

      

This is one of the cases mentioned at the start of this section which show that the relevant head in (11) does not have to be a verb, and the target for mutation is not necessarily the dependent of a verb. Here, the Word Grammar analysis would (reasonably) consider the clause-initial functional element i ‘to’ as the head, of which the non-finite verb canu ‘sing’ is a valent. Since canu is separated from this head by Mair, it is correctly predicted to bear SM.

Before moving on to some more complex data, I complete the review of basic examples by considering the objects of non-finite verbs. These normally do not bear SM, as seen in (2), where we find telyn and not *delyn. In (17) we have another example with a periphrastic verb, where the initial element is a finite auxiliary and a non-finite lexical verb occurs later on in the clause:

1. (17) Roedd y ddynes wedi prynu telyn.be.impf.3s the woman perf buy.inf harp ‘The woman had bought a harp.’

The lack of mutation on the object is predicted under the DD account, since telyn ‘harp’ is adjacent to the non-finite verbal head prynu ‘buy’, of which it is a valent. But if some phrase intervenes between the verb and the object in such examples, then the object does mutate:

1. (18) Roedd y ddynes wedi prynu [ar eBay] delyn hen iawn. (telyn)be.impf.3s the woman perf buy.inf on eBay harp old very ‘The woman had bought on eBay a very old harp.’

Again, this is predicted by the DD account (and indeed by the XPTH), since the valent is now separated from its preceding head. These contrasts between the lack of mutation in examples like (17) and the occurrence of mutation in (18) have led proponents of the XPTH to note that unlike case-marking, to which it has sometimes been compared (Roberts Reference Roberts2005), syntactic SM in Welsh is a very superficial phenomenon; see, for instance, Borsley et al. (Reference Borsley, Tallerman and Willis2007: chapter 7).

We have now seen how a representative sample of data would be handled under the DD account. However, a conspicuous problem arises at this point. In (14), the target for mutation is the object of the verb. As a noun, it is an uncontroversial candidate for dependent of the preceding verb. But what happens in examples parallel to (14) if the element which bears the mutation is not a noun, but some pre-modifier of the noun, as in (19): could this initial element still be a valent of the separated verb? More broadly, how does the DD account work when some element other than a noun or a verb (or some other item which is clearly eligible to be a valent of a preceding head) is initial in the phrase? Can such initial elements in general be regarded as valents of a preceding head from which they are separated?

1. (19)

   1. (a) Prynodd y ddynes dri thocyn. (tri)buy.past.3s the woman three ticket ‘The woman bought three tickets.’

   2. (b) Prynodd y ddynes lawer o delynau. (llawer)buy.past.3s the woman many of harps ‘The woman bought many harps.’

   3. (c) Prynodd y ddynes wir delyn Lydewig. (gwir)buy.past.3s the woman true harp Breton ‘The woman bought a genuine Breton harp.’

In (19a), a numeral precedes the noun, and it is the numeral which bears syntactic SM, rather than the noun, which we might have expected to be a valent of the preceding verb. In Word Grammar, however, it is standard to regard a noun as a dependent of a preceding numeral (Hudson Reference Hudson1990: 303), which suggests that the head in tri thocyn ‘three tickets’ is actually the numeral. Welsh mutation provides some additional evidence for this proposal. In Word Grammar, influence is a sign of dependency, so if word X influences the form of word Y, then Y is likely to be a dependent of X. Relationships between mutation trigger and target are a good example. For instance, many prepositions (heads) trigger some mutation on the following noun (dependent). We also find such a relationship in (19a). Most of the lower numerals are triggers for some mutation on the following noun; tri, for instance, triggers aspirate mutation in formal Welsh, giving thocyn (<tocyn) in (19a). Note also that cardinal numerals can appear without a head noun, as in (20):Footnote ⁹

1. (20) Mae tri wedi marw yn y storm. be.pres.3s three perf die.inf in the storm ‘Three (people) have died in the storm.’

It seems reasonable to conclude that the numeral is indeed the head in tri thocyn, and thus that the numeral is a valent of the preceding verb, and the following noun is in turn a valent of the numeral. Given this conclusion, the fact that the numeral bears SM is predicted by the DD account.

In (19b) it is the quantifier llawer which bears the SM. Quantifiers are unlike numerals in that they do not typically trigger any mutation on the following element. However, they do subcategorize for what follows them, and in this sense are typical heads. Most quantifiers select a preposition phrase headed by o ‘of’ and containing a plural noun, as in (19b), but some select a bare nominal dependent. For instance, we find peth (‘some, a bit’), as in peth caws ‘some cheese’, or sawl ‘several, how many’, as in sawl llyfr ‘several books’ – note also that sawl selects a singular noun, hence llyfr ‘book’ and not llyfrau ‘books’. Given such selectional restrictions, it seems uncontroversial to suggest that the quantifier is the head of the string llawer o delynau ‘many harps’, and thus llawer qualifies as a valent of the preceding verb. Note also that, like numerals, most quantifiers can be used alone, as the head of the phrase. It is standard within Word Grammar to adopt what is known elsewhere as the DP analysis of noun phrases, so that in a determiner–noun pair, the determiner is regarded as the head; quantifiers can then also be regarded as heads.

In (19c), the element bearing the mutation is an adjective. Normally in Word Grammar, the noun is regarded as the parent of a preceding attributive adjective (Hudson Reference Hudson1990: 303). But in (19c), the adjective itself would have to be a valent of the initial verb for the DD account to work. There are two indications that the adjective here could be considered the head, and the following noun its dependent. First, note that adjectives are generally post-nominal in Welsh, but a small set of adjectives appear in pre-nominal position – the same position as quantifiers, numerals and determiners, all elements which are regarded as heads. Second, the direction of influence strongly suggests that a pre-nominal adjective is the head on which the noun depends, since an adjective in this position triggers SM onto the following noun: thus, in (19c), we have gwir delyn (<telyn). Note that mutation occurs in this context whatever the number and gender of the following noun. On the other hand, in the case of post-modifying adjectives, the direction of influence is the opposite: a noun (if feminine singular) triggers SM on the following adjective (e.g. telyn ddu (<du) ‘harp black+SM’=‘a black harp’). Given the direction of influence in each case, we might conclude that pre-nominal adjectives are heads, and that post-nominal adjectives are dependents, with respect to the modified noun. As a consequence, in (19c) the mutated adjective is a valent of the preceding verb, from which it is separated, and the mutation is therefore predicted.

However, unlike cardinal numerals and quantifiers, singular adjectives cannot generally be used alone (or with a determiner) to head a phrase. Plural adjectives can be used with a determiner and no noun, but there is evidence that these should actually be considered nouns, rather than adjectives; see Borsley et al. (Reference Borsley, Tallerman and Willis2007: section 5.4.5), from which these examples are taken:

1. (21)

   1. (a) *y glas the blue (‘the blue one’)

   2. (b) yr un glas the one blue ‘the blue one’

   3. (c) y Gleision the blue.p‘the Blues’

Moreover, as Bob Borsley has pointed out to me, if the adjective itself is regarded as a valent of the verb in examples like (19c), the prediction would seem to be that whenever a nominal dependent can occur, an adjectival dependent could also occur – clearly, an incorrect prediction.Footnote ¹⁰ A possible solution to this problem is to regard the set of pre-head adjectives as in some sense nominal, just as plural adjectives used as heads (as in (21c)) are nominal. Note also that Roberts (Reference Roberts2005: 91) treats pre-head adjectives as special cases, suggesting that they are adjoined to N. Assuming that such an account can be maintained, the adjective in (19c) can continue to be regarded as the valent of the initial verb. In fact, this is also consistent with the view taken by Welsh traditional grammar, which regards A+N as a compound. Without stretching credulity too far, then, we could treat A+N as a single word, labelled N.

In sum, it seems that a reasonable DD account can be provided for the data in (19) without requiring stipulations. If we consider further data, however, it is clear that the DD account must be amended. In (5), repeated here as (22), we have an adjectival predicate (bracketed) with a pre-modifier, llawn ‘full’. As this is initial in the phrase it must be the target for SM. Clearly, though, this pre-modifier is not a dependent of the preceding copula, which the DD account in (11) would require to be the case, but a dependent of the adjectival predicate which it modifies.

1. (22) Dw i [lawn mor grac â chi]. (llawn) be.pres.1s I full as angry as you ‘I'm just as angry as you.’

And in (23) (see also (7)), the preposition i ffwrdd ‘away’ is itself a valent of the preceding verb, but its pre-modifier dwy droedfedd ‘two feet’ is not; yet the mutation appears at the start of the phrase as usual: dwy>ddwy:

1. (23) Rhoddodd Elen y delyn [ddwy droedfedd i ffrwdd]. (dwy) put.past.3s Elen the harp two foot away ‘Elen put the harp two feet away.’

Such instances, where some pre-modifier of the actual valent occurs at the start of the phrase, are in fact common, but they cannot be accounted for by the DD account under the formulation in (11). The account evidently needs to be revised, but this revision involves a significant complication to the grammar. It introduces a new relationship, which we can call first, to link the head word in a phrase (which is the actual valent) to the first word of its phrase.Footnote ^11,Footnote ¹² Welsh is a strongly head-initial language, so in many contexts the head of a phrase and the first of the phrase are the same; therefore first(X)=X. This is the case in all the instances of syntactic SM in (14) to (19). But if X has a pre-modifier of some kind, as in (22) and (23), then this pre-dependent is the first: first(X)=pre-dependent(X). The DD account in (11) must then be revised as in (24):

1. (24) The dependency distance account of syntactic SM (revised and final version)

   X bears soft mutation if:

   1. (a) X is a first of a valent of a preceding overt head H, and

   2. (b) X is separated from H by Z.

This can be indicated schematically as in (25):

1. (25) H Z [X ....]

   [+SM]

Note that only the first is the target for SM in (24), so we avoid a situation in which both a pre-dependent and the actual valent itself are predicted to bear SM. Under (24), the words llawn in (22) and dwy in (23) and in (7) are now predicted to bear the SM as firsts. As already proposed with respect to examples like (19a), given a Numeral+Noun sequence, the numeral is the head – and therefore it is the numeral which is a pre-dependent of the preposition i ffwrdd ‘away’ in (23). However, the notion of first is actually recursive, and will also apply to a pre-dependent of a pre-dependent of X.

The revised DD account in (24) can now handle the data shown so far.Footnote ¹³ It also presents a possible alternative solution to the issues posed by pre-modifying adjectives, as in (19c): if the idea above that such adjectives are the head in an Adj–Noun string is not accepted, then the data will in any case be accounted for under (24). From now on I adopt (24) as the DD account, though for clarity I generally continue to refer to a separated ‘valent’, rather than a ‘first’.

Note, though, that the complications which led to the revision of the DD account do not arise at all under the XPTH. In all the contexts examined so far, the SM is predicted simply because the target constituent is in a post-XP position; which particular element occurs at the start of the mutated constituent is irrelevant to the account. It also seems that in a substantive sense, the notion of ‘phrase’ – which, as noted above, is not a recognized concept in the Word Grammar machinery – comes in by the back door in (24). What the DD account attempts to say is this: the target for SM is whatever comes at the start of the phrase following some separating sequence of words. The formulation using the concept of first is a way of saying this without referring to phrases. But the XPTH is more straightforward, and does not have to invoke any special concepts such as first in order to deal with the full range of data. The phrase-based account is in fact simpler, and an analysis utilizing the notion ‘phrase’ seems to make exactly the correct predictions about syntactic SM.Footnote ¹⁴Section 4.2 below presents further problems arising from the lack of the notion ‘phrase’.

More problematically, although (24) is now empirically accurate, the DD account itself appears far less motivated. As noted in section 3.2, Hudson's original idea was that syntactic SM serves to signal the separation of a dependent from its head. Now we have to say that the mutation may signal the separation of a dependent of that dependent from its head – or even a dependent of a dependent of the actual valent of the separated head – a relationship which is intuitively far less satisfying. It seems, then, that the extension to the rule, though vital to capture the data, results in the loss of much of the functional motivation which made the DD account seem appealing in the first place.

4.1 The mutation effects of unrealized elements

Before looking at the specific details of unrealized elements in subject position, we need to understand something of the usual behaviour of unrealized elements in the mutation system. Mutation in general is regulated by a principle of strict adjacency between trigger and target (see Lieber Reference Lieber, Barlow, Ferguson, Flickinger and Wescoat1983, Zwicky Reference Zwicky, Testen, Mishra and Drogo1984, Borsley & Tallerman Reference Borsley and Tallerman1996: 8). For instance, in ei phum telyn ‘her five harp(s)’, we have a third person feminine singular possessive proclitic ei, which triggers aspirate mutation on the following numeral: pum ‘five’ >phum. It cannot, however, trigger aspirate mutation on the non-adjacent noun telyn ‘harp’: *ei phum thelyn. In spoken Welsh, it is very common for certain small functional elements to be phonetically unrealized, in an appropriate context. Typically, such elements are not invisible for mutation purposes, but behave as if they were overtly present, in the sense that they both block mutation arising from some previous potential trigger, and continue to trigger mutations themselves even when unrealized. In this sense, they respect the principle of strict adjacency.

On the first point, blocking, consider (26), where we have the third person plural agreement proclitic eu, a morpheme which is itself not a mutation trigger. Like other agreement proclitics, it is often unrealized in the spoken language (hence the strikeout). However, even when eu is only covertly present, for many speakers it blocks mutation from occurring across it:

1. (26) Gwnaeth Aled (eu) gwerthu nhw. do.past.3s Aled 3p sell.inf them ‘Aled sold them.’

So under the XPTH, the subject XP Aled does not trigger syntactic SM on the verb (giving werthu), because Aled does not immediately precede the verb: the mutation is blocked by the intervening proclitic, whether or not eu is overt. An analogous effect must be recognized in the DD account: even if eu is unrealized in (26), it continues to ‘count’ as the relevant first, since otherwise the prediction would be that the verb gwerthu would undergo SM, as it is separated from its head, the initial auxiliary gwnaeth.

The second point concerns mutation triggers which are phonetically unrealized. These are in fact quite common, but they typically continue to trigger the same mutation as if they were overt. For instance, toponyms are frequently preceded by the definite article, which, as noted earlier, triggers SM on feminine singular nouns: so, for instance, y waun (<gwaun) ‘the meadow’, y foel (<moel) ‘the bare hilltop’. Often, the triggering article itself is absent in the placename, but, crucially, the mutation remains: Waun, Foel. Similarly, clause-initial finite verbs may bear SM, generally considered to be triggered by the null counterpart of the (optional) affirmative particles fe or mi, which themselves trigger SM.

In general, then, phonetically unrealized elements play a full part in the mutation system, as the principle of strict adjacency predicts. This means that ‘immediate precedence’ in the statement of the XPTH in (8) does not refer just to immediately preceding overt items, but includes syntactically present but phonetically unrealized items. We can conclude that in any framework which tries to predict the occurrence of syntactic SM under either the XPTH or the DD account, some empty categories are crucially present. They block a mutation which might otherwise be triggered by some earlier item, and they trigger mutation themselves, just as if overt. This is quite a surprising finding for two very surface-oriented models of grammar, but the conclusion seems inescapable.

Consider now the unrealized subjects shown in (27) and (28):

1. (27) Pwy_i brynodd t _idelyn? (telyn) who buy.past.3s harp ‘Who bought a harp?’

1. (28) Prynodd pro delyn. (telyn)buy.past.3s harp ‘He/she bought a harp.’

Under the XPTH, the categories wh-trace and pro, shown in (27) and (28) respectively, are the XP triggers for the mutation on the direct objects.Footnote ^15, Footnote ¹⁶ A null pronominal subject can also be assumed in imperatives (and indeed can be realized as an overt pronoun in informal speech):

1. (29) Rho pro dipyn o egni iddi. (tipyn) put.imperative.2s bit of energy to.3fs ‘Put a bit of energy into it.’

Again, given the presence of pro, the mutation in (29) is predicted by the XPTH.

Within Word Grammar, unrealized words were until recently considered to play no part in the syntax (Hudson Reference Hudson2007: 172). In more recent work (e.g. Hudson Reference Hudson2003, Creider & Hudson Reference Creider, Hudson, Sugayama and Hudson2006, Hudson Reference Hudson2007: 172–181) this position has been overturned in favour of an approach in which unrealized or covert words are a legitimate part of the syntax; extensive evidence for covert subjects is cited from Ancient Greek, Icelandic, and other languages. Null subject constructions like (28) are now also assumed to have a covert subject pronoun (Hudson Reference Hudson2007: 174f.), and the same holds for imperatives, thus accounting for the mutation triggered by the separation of the verb from its valent tipyn ‘a bit’, in (29).

However, wh-traces like that in (27) are highly problematic for the DD account. Within Word Grammar, wh-traces are not amongst the set of unrealized words; instead, the dependency is shown by structure-sharing, and there is no empty category (or anything else) intervening between the verb and its object in (27). Citing a similar example from Welsh, Hudson (Reference Hudson2007: 127) suggests that ‘extracted subjects seem to behave like unrealized pronouns separating the object from the verb’. Indeed, unless some element intervenes between verb and object, the SM is not predicted to occur. In fact, though, there is good evidence that here the unrealized element (contra Hudson) cannot be pronominal. First, note that although a resumptive pronoun strategy is available in wh-extractions from certain lower positions on the NP accessibility hierarchy (Willis Reference Willis2000, Borsley et al. Reference Borsley, Tallerman and Willis2007: chapter 4, Willis Reference Willis2008), it is absolutely not available for extractions from subject and direct object position, thus ruling out the possibility that the unrealized subject in (27) is in any sense pronominal.Footnote ¹⁷ Second, if there were an unrealized pronoun in subject position in (27), one would expect the verb to be plural when the wh-phrase is plural, since finite verbs agree in number with pronominal but not lexical subjects in Welsh. But this is ungrammatical:

1. (30) Pa ddynesau welodd / *welson y ddraig? which women see.past.3s see.past.3p the dragon ‘Which women saw the dragon?’

We are therefore left with a problem for the DD account which has no apparent solution. Various empty categories must clearly ‘count’ in the calculation of dependency distance generally, but if wh-traces are not recognized, then the DD account cannot correctly predict the occurrence of mutation in (27).Footnote ¹⁸

4.2 X but not XP separates head and valent

The XPTH distinguishes between heads and phrasal categories, predicting that only the latter trigger syntactic SM. In Word Grammar, on the other hand, there is no inherent distinction between words and phrases. Hudson (Reference Hudson1990: 104) notes that ‘the grammar refers only to the relations between pairs of words’, and ‘the notion “phrase” can be defined in W[ord] G[rammar] in terms of dependency relations, but it is probably never needed in the grammar – i.e. no rules refer to “phrase”, “noun-phrase”, “verb-phrase” etc.’. In fact, ‘[g]rammars make no reference to any unit longer than the word’ (Hudson Reference Hudson1990: 10). So nothing in the Word Grammar model suggests that words and phrases should behave differently for syntactic purposes. Evidence from finite negative clauses in Welsh, however, shows that this is problematic for the DD account. Finite clauses are negated using the negative adverbial element ddim:

1. (31) Gwnaeth Megan ddim cysgu / *gysgu. do.past.3s Megan neg sleep.inf (+SM)‘Megan didn't sleep.’

Adverbial ddim cannot be modified, which suggests that it is a word but not a phrasal category; Borsley & Jones (Reference Borsley and Jones2005) also analyse adverbial ddim in this way. As such, it is not predicted to be a trigger for the mutation under the XPTH.

Under the DD account, it is reasonable to assume that the non-finite verb cysgu ‘sleep’ is a valent of the initial finite auxiliary gwnaeth ‘did’. Since cysgu is separated from gwnaeth, it is predicted to undergo syntactic SM. If ddim is missing (i.e. the clause is affirmative), then the prediction is correct, and the separated valent bears SM:

1. (32) Gwnaeth Megan gysgu. (cysgu)do.past.3s Megan sleep.inf ‘Megan slept.’

On the other hand, this prediction is incorrect in (31): *gysgu. The contrast between (31) and (32) is apparently inexplicable under the DD account (see also Hudson Reference Hudson2007: 127f.), but is predicted under the XPTH: Megan is the XP trigger in (32), whereas ddim, as it is not phrasal, is not a trigger in (31).

An obvious question is whether ddim in (31) might form a constituent with the following non-finite verb cysgu ‘sleep’, in which case it could be argued that it is ddim cysgu which is separated from the head gwnaeth, and ddim could perhaps be seen as the target for mutation as a first.Footnote ¹⁹ According to Borsley & Jones (Reference Borsley and Jones2005: section 5.2.1), post-subject ddim does not behave like a pre-modifier of the following verb and does not form a constituent with it. However, a referee for JL notes that there are cases where ddim does form a constituent with the following material, as shown by co-ordination, and by the occurrence of ddim with following material in sentence fragments, as Borsley & Jones (Reference Borsley and Jones2005: 104) also illustrate:

1. (33) Ddim yn gweithio.negprog work.inf‘Out of order’ (lit. ‘Not working’, e.g. as a note on faulty machines)

It seems, then, that ddim can be a pre-modifier, which might rescue the DD account. But the apparent conflict here arises because there are actually a number of distinct negative elements with the form (d)dim, perhaps as many as six (Borsley & Jones Reference Borsley and Jones2005: section 6.6 outline this in full). The two elements that concern us here are the post-subject adverb ddim, and the distinct negative element ddim which pre-modifies aspect phrases (as in (33)) and predicate phrases. Whereas the adverb ddim in (31) does not form a constituent with the following material, pre-modifying ddim in (33) does. In case there is any doubt that these are distinct elements, Borsley & Jones (2005: 133) show that they can co-occur:

1. (34) Dw i ddim [ddim yn poeni].be.pres.1s I negnegprog worry.inf‘I don't not worry.’

It seems, then, that ddim in (31) does not pre-modify cysgu – which means that the DD account is left with a separation between a head, gwnaeth, and its valent, cysgu, but with no SM occurring on the valent. I conclude that this lack of mutation on cysgu in (31) is indeed problematic for the DD account. By contrast, the XPTH predicts correctly that syntactic SM does not follow the negative adverb ddim.

As noted above, the XPTH predicts that phrases, but not words, will be triggers for syntactic SM, while the DD account predicts that there should be no difference: any element separating a head from its valent would be followed by SM. Ideally, we would need other test cases in which some mutable element is immediately preceded by an X and not an XP.Footnote ²⁰ There are certainly other adverbial elements which, like ddim, cannot be modified, suggesting that they might also be non-phrasal. Two examples are hefyd ‘too, also’ and erioed ‘never’. Both of these can be followed by SM, as shown for erioed in (35):

1. (35) Weles i erioed gymaint o bobl mewn un stafell. (cymaint) see.past.1s I never so.many of people in one room ‘I never saw so many people in one room.’

Are these adverbial elements then problematic for the XPTH (though not, of course, for the DD account)? There are three possible responses. The first is that elements like erioed and hefyd might be lexical triggers for SM. This seems very unlikely, however: we expect lexical triggers to form a constituent with the target of mutation (such as a triggering head preposition and its mutated object), which these adverbial elements do not. Moreover, they are never listed as lexical triggers by traditional Welsh grammars, such as the highly reliable Morgan (Reference Morgan1952).

The second response is that elements like erioed and hefyd might be ‘transparent’ in some way, so that they are invisible as far as a preceding trigger is concerned; if this were the case, then the trigger for SM would be the subject XP in (35). However, we have already seen in section 4.1 that in general, mutation operates according to a strict adjacency principle. The idea of elements which are phonetically overt, but invisible for mutation purposes, would be both unprecedented and stipulative.

The third possibility is that – unlike ddim – these adverbials actually are phrasal categories, and as such are predicted to trigger SM under the XPTH. Evidence that this third position is correct comes from the fact that these adverbials can either be expanded into full phrases, as in (36), or replaced by a full phrase with the same meaning, as in (37):

1. (36) Welais i [erioed yn fy mywyd] gymaint o bobl.see.past.1s I never in 1s life so.many of people ‘I never in my life saw so many people.’

1. (37) Gallwch hefyd/ ynogystal e-bostio. can.fut.2p also pred as.well e-mail.inf ‘You can also e-mail/You can e-mail as well.’

This behaviour contrasts on both counts with that of the post-subject negative adverb ddim. I conclude, then, that the adverb ddim truly is non-phrasal, and is thus correctly predicted not to be a trigger for SM under the XPTH, while remaining problematic for the DD account, which predicts incorrectly that it should be followed by SM.

In section 3.2, the non-availability of phrasal categories led to the complication of the DD account using firsts. Here, the notion ‘phrase’ seems to make a crucial contribution to the analysis of syntactic SM. Thus, we have further evidence that ‘phrase’ is a category required in a grammatical model.

4.3 Head fronting

In this section I examine the effects of a construction in which a non-finite verb is fronted. The data in (38) illustrate some basic facts involving the fronting of an entire aspectual verb phrase:

1. (38)

   1. (a) Mae Gwyn *(yn) canu emyn. be.pres.3s Gwyn prog sing.inf hymn ‘Gwyn is singing a hymn.’

   2. (b) (*Yn) canu emyn mae Gwyn.prog sing.inf hymn be.pres.3s Gwyn ‘Gwyn is singing a hymn.’

In (38a) we see that the progressive aspect marker yn is obligatorily present; but in the focussed version (38b), the progressive aspect marker is – idiosyncratically – obligatorily absent: other aspect markers do not disappear under fronting.

Now consider the following data ((39) and (40) are from Thomas Reference Thomas1996: 517, and (41) and (42) are from Morgan Reference Morgan1952: 432). The normal word orders are shown in (39) and (41), and two very formal examples with fronting are shown in (40) and (42). As in (38b), the progressive marker yn disappears under fronting. But unlike in the earlier examples, here the head non-finite verb alone is fronted, leaving the complements in situ in (40) and (42).

1. (39) Yr oeddent pro yn cofio [gweld un o 'r gorymdeithiau].prt be.impf.3pprog remember.inf see.inf one of the processions ‘They remembered seeing one of the processions.’

1. (40) Cofio yr oeddent pro weld un o 'r gorymdeithiau. (gweld)remember.infprt be.impf.3p see.inf one of the processions ‘They remembered seeing one of the processions.’

1. (41) Y maent pro yn gwrthod [profedigaeth Crist a 'i Groes].prt be.pres.3pprog reject.inf tribulation Christ and 3ms cross ‘They are rejecting the tribulation of Christ and his cross.’

1. (42) Gwrthod y maent pro brofedigaeth Crist a 'i Groes. (profedigaeth)reject.infprt be.pres.3p tribulation Christ and 3ms cross ‘They are rejecting the tribulation of Christ and his cross.’

In these examples, we have a non-finite verb with a complement: cofio ‘remember’ in (39) takes the (bracketed) non-finite VP or clausal complement (depending on the analysis) and gwrthod ‘reject’ in (41) takes an object, also bracketed. In the unmarked word orders, the complements do not bear SM, but in the fronting constructions in (40) and (42), the complement in each case does bear SM. Why does syntactic SM occur? Here an analysis can be given in terms of both the XPTH and the DD account.

Under the XPTH, the trigger is the null subject pro, shown immediately preceding the complement bearing the mutation in (40) and (42). However, we need to assume that when the non-finite verbal head fronts, the progressive aspect marker yn is not ‘left behind’, even covertly, since if it were it would intervene between the subject and the complement, and should block the mutation triggered by pro.Footnote ²¹ I suggest that the aspectual particle fronts along with the verb but is syntactically unrealized, and phonetically null.Footnote ²² There is, though, no reason to think that the fronted non-finite verb itself is associated with a gap in the syntax (i.e. a verb trace). In HPSG, the unbounded dependency gaps created by XP-fronting and other processes are generally analysed as empty categories, and these can indeed intervene between a mutation trigger and a potential target. Conversely, the head-fronting seen in this section is not an unbounded process, but rather appears to be clause-bound, just like so-called long head movement in Middle Welsh and Breton (e.g. Borsley, Rivero & Stephens Reference Borsley, Rivero, Stephens, Borsley and Roberts1996).

A similar fronting construction also occurs in clauses without aspect markers. Again in very formal Welsh, a non-finite verb may front, leaving its complement(s) behind. An example is given in (43), with normal word order, and (44), with marked word order, from Thomas (Reference Thomas1996: 505):Footnote ²³

1. (43) Gwnaeth yr arbenigwyr brofi cemigion eraill ddoe. do.past.3s the specialists test.inf chemicals other yesterday ‘The specialists tested other chemicals yesterday.’

1. (44) Profi a wnaeth yr arbenigwyr gemigion eraill ddoe. (cemigion)test.inf prt do.past.3s the specialists chemicals other yesterday ‘The specialists tested other chemicals yesterday.’

Here, the object of the non-finite verb, cemigion ‘chemicals’, bears SM in (44) only. This is predicted by the XPTH because only in (44) does an XP, yr arbenigwyr ‘the specialists’, immediately precede the object.

Compare now the DD account of the verb-fronting contexts for mutation. Under (24), this is straightforward, since a valent in each case is clearly separated from its parent, the fronted head verb. So in (40), gweld ‘see’ is separated from the fronted head cofio; in (42), profedigaeth ‘tribulation’ is separated from the fronted head gwrthod; and in (44) cemigion ‘chemicals’ is separated from the fronted head profi. Each of these valents is correctly predicted to bear SM. The slight complication for the XPTH, concerning progressive marker yn in the first two cases, is not relevant for the DD account.

In sum, both the XPTH and the DD account make the correct predictions in this section, but the XPTH clearly requires the additional assumption that there is no unrealized aspectual yn to block the mutation in (40) and (42).

4.4 Co-ordination

4.5 Predicative complements and quotative verbs

In this final brief data section, I indicate some directions for future research by examining two constructions which prove challenging to both the XPTH and the DD account. Verbs with predicative complements are normally not useful for testing hypotheses about syntactic SM, because the initial word of the complement is typically a functional element with no mutable initial consonant, such as predicative yn (see note 22) or fel ‘as’. However, a ‘bare’ predicative complement can also occur in naming contexts, but unexpectedly, this does not undergo syntactic SM:

1. (50) Gelwir y dref Caerdydd/*Gaerdydd. call.pres.impers the town Cardiff (+SM)‘The town is called Cardiff.’ (Morgan Reference Morgan1952: 250)

1. (51) Gelwir y lle caeëdig cyntaf cae/*gae. call.pres.impers the place enclosed first field (+SM)‘The first enclosure is called a “cae” [field].’

1. (52) Gelwir hwy llywodraethwyr/*lywodraethwyr. call.pres.impers them governors (+SM)‘They are called governors.’ (Morgan Reference Morgan1952: 251)

Regarding (50), it might be thought that a proper noun would not mutate, but in fact it is standard to mutate native placenames following a mutation trigger, e.g. i Gaerdydd ‘to Cardiff’. In any case, this explanation would be unavailable for (51) and (52). These data are problematic both for the DD approach and for the XPTH. The latter predicts mutation after the preceding XP (the object of the impersonal verb), and the DD approach predicts mutation because the valents in bold are separated from the head verbs.

We are left, then, with an unforeseen absence of mutation in (50) to (52). However, this context is not entirely unique. A referee for JL asks if the same absence of mutation occurs in reported speech, which indeed it does:

1. (53) Dywedodd yr Athro Davies: ‘Pleser o 'r mwyaf i ni yw … ’.say.past.3s the Professor Davies pleasure of the greatest to us be.pres.3s ‘Professor Davies said “It's the greatest pleasure for us … ”.’

1. (54) Dywedodd y cadeirydd: ‘Cyfrinach y llwyddiant yw … ’.say.past.3s the chairman secret the success be.pres.3s ‘The chairman said “The secret of success is … ”.’

Thus, we find no mutation at the start of the reported speech in these examples: pleser, not *bleser, in (53), and cyfrinach, not *gyfrinach, in (54), despite the fact that these words occur in what is normally a triggering context (i.e. following an XP under the phrase-based account).

It appears, then, that these contexts in which syntactic SM unexpectedly fails to occur – both (50)–(52) and (53)–(54) – have a common (quotative) thread, and thus constitute a natural class of exceptions to the general principle. The fact that neither of the analyses under discussion predicts this environment shows that further work on mutation contexts is still required.

4.6 Summary

We have seen throughout this section that the XPTH is able to predict the correct outcome in all the data analysed (apart from section 4.5), whereas the DD account has been shown to make incorrect predictions in a variety of contexts, involving unrealized (wh-trace) subjects, negative finite clauses, co-ordination and ellipsis. In just one context, verb fronting, the two accounts made the same predictions, though the XPTH required a small additional assumption regarding aspectual yn which was not needed by the DD account. At this point, it seems reasonable to conclude that, in empirical terms, the XPTH is the superior analysis. We have also seen that the availability of the notion ‘phrase’ is crucial in the analysis of syntactic SM. A phrase-based account turns out both to be empirically more accurate and also to encounter fewer theoretical problems than a dependency account.