Non-Distributive Features in Welsh Coordination Louisa Sadler
University of Essex
louisa@essex.ac.ukProceedings of the LFG99 Conference
The University of Manchester
Miriam Butt and Tracy Holloway King (Editors)
1999
CSLI Publications
http://www-csli.stanford.edu/publications/
We consider various analyses of certain asymmetries concerning morphosyntactic features in coordinate structures in Welsh in the light of the theories of agreement and of coordination in LFG. In classic LFG, a very simple view is taken of agreement phenomena such as person, number and gender agreement between finite verbs and their subjects, or gender and number agreement between determiners, adjectives and nominals and similar phenomena. Agreement is generally modelled by means of constraints stated over the grammatical features PER, NUM and GEN of the controller argument: since these features belong to the vocabulary of f-structure, agreement is viewed as a surface syntactic matter. It is well known that agreement with coordinate structures may require some computation of controller agreement features and [Dalrymple and Kaplan 1998] show how the LFG formalism may be extended to express such feature resolution principles, again treating agreement at f-structure. In this paper we focus on a different pattern of agreement under coordination, that of single conjunct agreement, and consider a range of analyses of this phenomenon in Welsh. We argue that these data suggest the need for a more sophisticated view of agreement and sketch several possible analyses. Section One presents the relevant data. Section Two reviews the theories of coordination and agreement in LFG. In Section Three it is argued that the Welsh constructions under consideration really are coordinate structures. Finally, Section Four presents three "solutions" to this puzzle.
In Welsh, a rigidly head initial language, we find that arguments control (person, number and gender) head agreement not just on finite verbs, but also on nominals and prepositions. Only pronominal arguments are agreement controllers, however: heads do not agree with their non-pronominal arguments. Thus finite verbs agree with pronominal subjects in person and number, and take the unmarked third person singular form with all non-pronominal subjects. This is illustrated in the examples in (1) below.
A form inflected for the person, number and (sometimes) gender of a (pronominal) argument, such as darllenasant `read-3PL', arnoch `on-2PL' or dy dy `2S house' is in fact ambiguous between an agreement form and a pronominal incorporating form. That is, agreement with a pronominal argument is obligatory in Welsh and the pronominal argument itself is optional.1
Under coordination, agreeing heads exhibit what we might call an asymmetrical agreement pattern, agreeing with the first conjunct of a coordinate subject, so long as it is pronominal. The examples below illustrate. In (2a) and (2b) the verb appears in the `unmarked' 3rd singular form with a plural coordinate subject where the first conjunct is non-pronominal, while in (2c) it agrees with the pronominal first conjunct. Precisely the same pattern is illustrated in (2d) and (2e).
An identical agreement pattern shows up in nominal structures containing possessor phrases. In Welsh, nominal heads take a proclitic agreeing with pronominal (but not non-pronominal) possessors (the canonical position for possessors is post-head). This is illustrated in (3).
If the possessor phrase is a coordinate structure, the nominal head agrees with the first conjunct, just in case it is pronominal (4).2
The majority of prepositions in the language have a full inflectional paradigm, and inflect to agree with their pronominal (but not non-pronominal) objects. Again, where there is a coordinate argument, the preposition inflects to agree with the first (closest) argument, if it is pronominal, as illustrated below for the inflecting preposition am `about'.
Recalling our earlier remark about `agreement morphology' alternating between an agreement reading and a pronominal incorporation, with the `doubling pronoun' being optional, we should note that there is one significant difference between the agreement pattern found in coordinate structures and that found with simple arguments. With coordinate structures, the pronominal argument must always be independently expressed, despite the presence of agreement morphology on the head.
A similar pattern of asymmetrical agreement is found in Irish and discussed in [McCloskey 1986], from whom the following data is taken. As in Welsh, the finite verb agrees with a leftmost (i.e. closest) pronominal within a coordinate subject, (6), a preposition with a leftmost pronominal within a coordinate object (7), and a nominal with a leftmost pronominal within a coordinate possessor (8). The general agreement pattern in Irish differs systematically from that in Welsh in one respect however. Whereas in Welsh, as we have seen in (1b) and (3b), agreement morphology and proclitics may be optionally doubled by overt pronominals, in Irish such overt (doubling) pronominals are not possible, although emphatic or constrastive nominal particles may occur in the relevant argument position. In similar fashion, just as in Welsh the doubling pronominals are required under coordination, so too the doubling emphatic or contrastive particles are required under coordination in Irish (while the leftmost pronominal itself is obligatorily absent). This is illustrated in the examples below from [McCloskey 1986].
Notice that if the leftmost conjunct is not pronominal the head does not bear agreement features, and as in Welsh, a full pronominal may occur as the non-initial conjunct.
Though not as common as "resolved" agreement under coordination, similarly asymmtrical agreement patterns are found in other languages outside the Celtic family. [Corbett 1983] gives data for a range of languages in which one single conjunct controls person, number and gender (index) agreement. He observes that crosslinguistically, agreement with the nearer conjunct is more common when the predicate precedes the subject than when the subject precedes the predicate, giving the following examples from Czech and Latin respectively, while [Johannessen 1989] cites the Palestinian Arabic data in (12) (from [van Oirsouw 1987]).
In the following example from Swahili the conjunct nearest to the agreeing head is the final conjunct and it controls gender (noun classifier) agreement on the (following) head [Corbett 1991]:
Although less frequent crosslinguistically, [Corbett 1983] shows that agreement may also be controlled by the most distant conjunct. This occurs in Latin (which also exhibits agreement controlled by the nearest conjunct, as shown above), Serbo-Croat and the following predicate agreement example from Slovene:
To summarise, there is robust crosslinguistic data illustrating the phenomenon of single conjunct agreement. The more common asymmetrical pattern appears to be that in which the closest conjunct to the head controls agreement (but distant agreement is also attested), and this pattern is itself more common where the predicate (agreeing head) precedes the coordinate argument.
In this example, both the subject and the finite verb, are associated with defining equations over the same f-structure. In some analyses, agreement targets introduce instead non-monotonic constraining equations over the values of the controller's agreement features. This captures the intuition that the relationship is asymmetric, and builds in a distinction between realizing a feature and requiring a feature. For example, [Andrews 1982] provides the following entry for the 2nd person plural form of the present tense of the Icelandic verb elska `to love':
We are not concerned here with the differences between these methods, though they can be rather significant (for some very thought-provoking discussion see [Johnson 1997]). What they have in common are that both model agreement in terms of constraints over one f-structure (that of the controller). This differs from the way in which morphologists tend to think about agreement phenomena from a morphological viewpoint as involving both target and controller being specified for an inherent set of person, number and gender features, which are then required (in the syntax) to match.
A treatment of coordination in LFG is outlined in [Kaplan and Maxwell III 1988], from which the following example is taken. Coordinate structures are analysed by means of phrasal expansions such as that shown in (17) and are modelled as sets at f-structure. To account for the distribution of elements across the set, such as the f-structure of the subject John and that of the object apples, the definition of function-application is extended to hold of sets of functions.
Given this treatment of coordination, we expect constraints over the subject associated with the finite verb, for example, to distribute over the set of f-structures corresponding to a coordinate subject: in fact, the normal situation is for such features to resolve, as in John and Mary aren't happy, or (21) below. In a recent paper, [Dalrymple and Kaplan 1998] consider agreement and feature resolution in coordinate structures. They discuss essentially two sets of cases, those in which a somewhat indeterminate feature is checked against each and every conjunct, and those in which a feature value is checked against a (resolved) value for the coordinate structure as a whole.
The former is exemplified in (20): intuitively, this example is grammatical because the wordform kogo is indeterminate enough to be able to satisfy both the requirement that it is ACC (imposed by lubi) and the requirement that it is GEN (imposed by nienawidzi).
Cases of indeterminacy involve what Dalrymple and Kaplan call distributive features:3 they argue that Case, Vform and Nounclass are all distributive features. To accommodate indeterminacy in feature values, the LFG f-description notation is extended to include set designation (giving an exhaustive enumeration of the set in question), so that a feature value for a given wordform may be a set (and `case checking' constraints check for set membership, not equality).
The second set of cases, that of feature resolution under coordination is exemplified by (21) below. Here the agreeing element introduces agreement features which are distinct from, but related by some simple computation to, those of each conjunct. Dalrymple and Kaplan propose that a subset of features are special in being non-distributive (here Number, Gender and Person). They represent coordinate structures as hybrid structures having both elements and (non-distributive) features as illustrated below.
The definition of function application is refined to take into account this rudimentary typing of features or properties ([Dalrymple and Kaplan 1998] page 26):
[Dalrymple and Kaplan 1998] introduce a representation for person and gender features which enables a simple statement of the computation involved in resolution (number resolution is essentially semantic in nature). Resolving features are expressed by means of marker sets encoding complex values, as illustrated in (24) ([Dalrymple and Kaplan 1998] page 27).4 Given this representation, resolution involves simply set union, and "resolution rules" are simply stated as annotations.
With this background, consider now the data on asymmetric agreement with coordinate structures presented in Section 1. It should be clear that these data do not fit comfortably with this combined analysis of agreement and coordination in LFG. Coordinate structures are sets at f-structure: properties holding of a set of f-structures are either distributed over the members of that set (distributive properties) or they hold of the set itself (non-distributive properties). In languages showing asymmetric agreement under coordination, however, neither assumption about the agreement features is correct, however. It seems that either some different treatment of agreement is required, or some different treatment of the (coordinate) structures themselves. In the following section, we consider whether it is appropriate to consider the Welsh structures as coordinate structures in the sense of the c-structure and f-structure analysis outlined above.
If these constructions do not map to a set of f-structures, the alternative is that they might correspond to some sort of head - dependent structure, with the first "conjunct" mapping to the relevant grammatical function and the rest of the coordinate structure being an adjunct. There are two logical possibilities. The first is some sort conjunct union analysis [] in which the rest of the coordinate structure is an adjunct "upstairs" and the second is an analysis under which the non-initial "conjuncts" are adjuncts "downstairs" to the grammatical function associated with the initial "conjunct". These are shown schematically below for a case like (2c) (in each case, we finesse the matter of how the conjunction is represented at f-structure under this hypothesis):
What these proposals have in common is that they treat non-initial conjuncts as adjuncts. But as we will see, the non-initial conjunct instead has the properties of the grammatical function it would instantiate as a member of a set of f-structures under the coordinate structure analysis.
[1]. For example the coordinate structure as a whole (that is, the set of f-structures) serves as controller in the examples (2d,e) repeated here for convenience:
[2]. Note also that the a/ac phrase does not have the sort of mobility we associate with adjuncts, but appears in an absolutely fixed position with respect to the head. Interestingly, there is a subordinating use of the conjunction a/ac, introducing absolute clauses, and the adverbial clause so introduced can precede, interrupt or follow the clause which it modifies (examples from ( [Thorne 1993] pages 382- 383).
[3]. Semantically, the a/ac phrase has the sort of status that is associated with a conjunct within a coordinate structure. Consider examples of coordination in possessive constructions, such as (35).
The `inside' and `outside' adjunct analyses associate the following structures with (35) respectively, where the grammatical function borne by the NP is represented simply as GF:
The interpretation under which the property of having green eyes is associated with hair is simply incoherent, and is certainly not the interpretation associated with (35), but this is the sort of interpretation that we would expect for the sort of f-structure schematized in (36). Similarly, the structure in (37) may be expected to map to similarly unhelpful semantics under which green eyes is associated with the dominating (presumably verbal) predicate.
Crucially, the possessor Mair is in fact interpreted as a semantic argument of both hair and eyes, precisely as one would expect if the f-structure were a set, with the possessor distributed over the members of the sert, that is, if the f-structure representation of (35) were that of a coordinate structure.
[4]. Pronominal coordinate structures (that is, those showing asymmetric or initial conjunct agreement, for which we are considering a head-adjunct f-structure representation) do not differ from non-pronominal coordinate structures in terms of their interaction with other syntactic phenomena. Anaphoric pronouns and pronominal clitics show precisely the same pattern of concord with an asymmetric (pronominal initial) coordinate structure as with other coordinate structures. This can be exemplified with the Welsh personal passive construction, which involves the verb cael, `get' used as a passive auxiliary and the main verb in non-finite, VN form. The non-finite main verb is preceded by a pronominal clitic agreeing with the passive subject. The pair of examples below exemplify a passive, the first involving a pronominal initial coordinate subject. In each case the (passive) pronominal clitic agrees with the coordinate structure as a whole, that is, as though feature resolution has occurred in the coordinate subject.
Finally, [McCloskey 1986] notes an incorrect prediction of the conjunct union analysis, under which the first conjunct is the SUBJect or OBJect, and so on, while the other conjunct takes on some sort of `upstairs' function (such as ADJunct). He observes that the initial (subject) conjunct does not behave like a SUBJect. In Irish, a relative clause formed on the immediately dominated subject position obligatorily involves a gap on subject position, rather than a (null) pronominal, the presence of the latter being signalled by verb agreement. This restriction does not extend to coordinate subjects, and in particular to those which are pronominal initial, thus (41) is grammatical. This suggests that the pronominal conjunct is not itself the SUBJect (in our terms, it is a member of the set of f-structures which together provide the SUBJect function).
An analagous argument may be made for Welsh. A relative clause on prepositional object position requires the use of agreement morphology on inflecting prepositions and the absence of the pronominal itself - the latter condition is suspended in the case of a coordinate object:
The evidence, then, is that all coordinate structures, whether they have pronominal conjuncts or not, are represented as sets at f-structure. Coordinate structures with a pronominal initial conjunct differ from other coordinate structures only as far as the head-argument agreement between a finite verb and a subject, or a prepositional head and its object, or a nominal head and its possessor is concerned. This strongly suggests that they should be treated equivalently. Note further that predicate agreement and prononinal anaphora show that semantic feature resolution operates with both pronominal and non-pronominal coordinate structures, a fact which will be of significance when we consider analyses in the following section.
To conclude this section, we summarise the puzzle that pronominal coordinate structures represent. We have good reason to conclude that coordinate structures in Welsh, whether or not they contain pronominals, involve multiply-headed c-structures which map to sets of f-structures. On the one hand, evidence from anaphora and predicate agreement suggests that the coordinate structure bears semantically resolved person and number agreement features, while on the other hand evidence from head-argument agreement suggests that the coordinate structure bears the agreement features associated with an initial, pronominal conjunct.
The advantage of this approach is that it permits us to maintain the notion that the agreement features PER, NUM and GEN are non-distributive. However, there are several aspects which make this analysis unattractive. For one thing, the intuition that the target really does agree with the first conjunct is captured only indirectly, by means of a feature passing mechanism of the sort which is generally eschewed in LFG. Secondly, the approach is perversely at odds with the intent of the [Dalrymple and Kaplan 1998] proposal. That proposal extended the formalism to permit the grammar to express what is essentially semantic resolution in a syntactic agreement environment, but in the approach here, the agreement features associated with the coordinate structure as a whole are precisely not those required for "semantic" agreement for cases such as (38), (39), (44), (45) and the following examples:
The definition of f-precedence is:
This approach also treats agreement as a matter of co-specification of a single f-structure rather than as feature matching. Note that it is also consistent with semantic resolution: it combines unproblematically with the computation of `resolved' features on the f-structure corresponding to the coordinate structure as a whole. The particuliarity of Welsh is simply that agreeing heads do not state constraints over this structure. So, for example, the subject in (54) may be represented as in (55).
This approach captures directly the asymmetry in head-argument agreement, but at a severe cost for the lexicon. On the other hand, there is in any case an irreducible aspect to this lexical proliferation, since, as we have seen, pronominal incorporation is impossible from coordinate structures.6
A third approach to the problem departs from the view that all agreement is at f-structure and distinguishes between a set of semantically motivated agreement features (at f-stucture) and a set of purely morphosyntactic agreement features, which are represented at m-structure.7 We assume that the default relationship between these two sets (outside of certain syntactic constructions) is an identity mapping.8 The behaviour of the (semantically motivated) PER and NUM features at f-structure is captured by the analysis of [Dalrymple and Kaplan 1998] outlined above, but morphosyntactically the coordinate structure is associated with the agreement features of the first conjunct - and it is these features which the agreeing head specifies constraints over. The essence of this idea is captured in the coordinate schema and lexical entry below:
In this approach, then, the f-structure of the coordinate structure in (54) is indeed as in (55), while the m-structure of both f1 and f3 is as follows:
