The previous examples indicate that binding constraints cannot be expressed in terms of SUBCAT, if the grammar makes use of argument inheritance. As an alternative account we use the feature ARG-S (``argument-structure'') [11,17].
The feature ARG-S will be the focus of control and binding. Typically the value of ARG-S is a list of syntactic dependents, ordered with the least oblique element (often the subject) left.
Note that in our analysis only finite verbs contain a subject on the SUBCAT list. This is in line with our assumption that control relations are established on the value of ARG-S. In (35) the verb kussen has a SUBCAT list only containing its accusative object. The verb probeert has a SUBCAT list containing a VP and its nominative subject. Hence control cannot be established on SUBCAT.
The value of ARG-S can be used, as can be seen from the lexical specifications of the two verbs:
Control theory should establish that the index of the nominative subject of proberen is structure-shared with the first element on the ARG-S list of its VP-complement.
Note that the value of ARG-S of verbs generally must contain a subject. Even subject-raising verbs have to contain a subject on ARG-S because otherwise control relations cannot be established in examples such as:
Here the subject-raiser zullen has a subject on its ARG-S even
though it does not assign a semantic role to that noun-phrase.
We can then define o-command in terms of the structure of ARG-S; i.e., a referential sign X locally o-commands Y iff X precedes Y on ARG-S. The general notion o-command then holds between X and Y iff X locally o-commands a Z, and Z dominates Y.5
Let us now return to the examples (33). As we will explain below it is necessary to assume a subject-to-object raising analysis for the ARG-S value of verbs such as zien. The value of ARG-S of the verb ziet is therefore as follows for the first four examples of (33).
The embedded ARG-S value is in these cases the relevant binding domain both for principle A as for principle B. (33a) is well-formed because the reflexive is locally bound. (33b) is ruled-out because the pronominal is not locally free. In (33c) the reflexive is locally o-commanded but not locally bound, hence (33c) is ruled out. (33d) is fine because the pronominal is not locally bound.
If the anaphor is the raised object (33e-h) then the ARG-S of the verb ziet will be:
(33e-f) are ruled out because the nonpronoun Piet is not o-free. The remaining two examples are more interesting, and form the motivation for the raising-to-object structure of the value of ARG-S. Given such an analysis, the reflexive in (33g) is not o-commanded in the embedded ARG-S structure, hence it need not be o-bound there. It is locally o-commanded in the topmost ARG-S structure, and it indeed is o-bound there. In (33h) the pronominal is not locally free in the topmost ARG-S structure. Note that the structure of ARG-S is similar to the functional structures advocated in LFG for English raising-to-object verbs, cf. for example .
However, note that we now still have to face the question whether the binding constraints are interpreted existentially or universally, because on ARG-S we also have (a limited form of) structure-sharing. We argue (against ) for the universal interpretation. The universal interpretation seems to be the formally most attractive assumption because it allows for a ``local'' check (binding constraints can be implemented as constraints on argument structures). In the existential interpretation this can only be done in relation with global information of all other argument structures that a pronoun might appear in.
The universal interpretation is also forced by sentence (33h). This sentence should be ruled out. Note though that the pronominal occurs both in the argument structure of the matrix verb and in the argument structure of the embedded verb. The sentence is ruled out in the universal interpretation, because the pronoun is not free in the argument structure of the matrix verb. The existential interpretation, on the other hand, predicts that the sentence is well-formed, because the pronoun is free in the argument structure of the embedded verb.
The universal interpretation is also neccessary in order to rule out anaphors in the subject-to-object-raising position that don't find an antencedent:
In an existential interpretation this sentence would be accepted because the anaphor is not o-commanded in the embedded ARG-S hence principle A is satisfied in the embedded ARG-S.
If we follow the suggestion in [24, chapter 6.8.3] that local o-command should include subjects of embedded (infinite) VP's then we could adopt an analysis for raising-to-object verbs in which there is no structure-sharing on ARG-S. This would complicate the definition of (local) o-command, but it would avoid the choice between the existential and universal interpretation of binding constraints.
It is interesting to compare the current analysis to the LFG analysis of . As indicated above the functional structures she assumes are similar to our ARG-S. Anaphors and pronouns lexically constrain the (functional) environment in which they are allowed to occur. In such an approach it is quite natural to assume that constraints that are expressed in a positive way (``my environment should contain ...'') need only be satisfied once. On the other hand, constraints that are expressed in a negative way (``my environment should not contain ...'') will get a kind of universal interpretation. Dalrymple's analysis predicts the set of facts given above. However, in order to be able to give such a lexical account of binding she needs a special type of functional uncertainty (``inside-out functional uncertainty''). Even if functional uncertainty were added to the formal machinery of HPSG it would be unclear how the ``inside-out'' type of functional uncertainty could be expressed in HPSG ().