Problems with Verb-clusters

There are numerous non-transformational accounts of the cross-serial dependency construction.¹ All these accounts assume that the clause-final sequence of verbs is a constituent (the ``verb cluster'' or ``verbal complex''). Furthermore, with the exception of [16], who argue for a left-branching verb cluster, it is assumed that this cluster has a right-branching structure, as illustrated in (2).

The assumption that the verb cluster is right-branching is problematic for categorial accounts as well as for HPSG accounts using argument inheritance.

Within Categorial Grammar, it has been proposed to derive verb clusters by means of (disharmonic versions of) composition or division. Composition can be used to combine a functor with its argument, even if this argument is ``unsaturated''. Division derives a category (A/C)/(B/C) (or (C \A)/(C \B) if used disharmonically) from a category A/B. The effect of such a rule is that a functor ``inherits'' the arguments of its argument. Using disharmonic division, one can for instance derive the verb cluster in (2) as follows:

A problem for categorial accounts has been the fact that cross-serial word order is obligatory, that is, if the governing verb is a verb-raising verb, it must be followed by the governed verb and may not be followed by any of the non-verbal arguments of the governed VP. Sequences in which the governed verb is followed by a full or partial VP including non-verbal complements are ungrammatical:

In a categorial grammar, such word orders are derived quite easily, however, using application instead of division or composition:

To eliminate these ungrammatical derivations, it has been proposed that the verbal argument of a verb-raising verb must be ``lexical''. However, example (2) illustrates that under the assumption that verb clusters are right-branching, the proper requirement cannot be that the verbal argument must be a single verb. Rather, complex phrases consisting of a sequence of verbs must be allowed as arguments, whereas phrases containing one or more non-verbal complements must be excluded. As such information cannot be read off the categories of the constituents involved, this implies that the categorial formalism needs to be extended with a feature distinguishing ``verbal complexes'' from other verbal constituents and a method for assigning this feature to derived constituents. Although such a system can be designed (see [1], for instance), the result remains unsatisfactory as it requires an ad-hoc system of feature passing, which cannot be subsumed by general methods such as head-feature passing.

The problem is not restricted to CG, but surfaces in HPSG accounts as well. Hinrichs and Nakazawa [8], for instance, use an argument-inheritance mechanism, which lets the SUBCAT-list of a modal or auxiliary be determined in part by the SUBCAT-list of its verbal complement. Essentially, this is a restatement of the categorial rule of division as a constraint on SUBCAT-lists. They account for ``auxiliary-flip'' in German by assuming that an auxiliary may either follow or precede a verbal complex. The latter case gives rise to ``flipped'' word order (note that this order is the standard situation in Dutch). To avoid spurious ambiguity and ungrammatical word orders, they must introduce a feature NPCOMP, distinguishing phrases containing NP-complements from phrases not containing such complements. That is, as in categorial accounts, NPCOMP must distinguish between verb clusters and other verbal constituents. The percolation of NPCOMP does not follow from general principles of feature-percolation, but instead is stipulated in the relevant rule schemata.²

In this paper, we argue that the problem of characterizing verb clusters can be avoided if one assumes that clause-final verb sequences do not constitute a constituent.