Complex predicates: Structure, potential structure and underspecification

This paper compares a recent TAG-based analysis of complex predicates in Hindi/Urdu with its HPSG analog. It points out that TAG combines actual structure while HPSG (and Categorial Grammar and other valence-based frameworks) specify valence of lexical items and hence potential structure. This makes it possible to have light verbs decide which arguments of embedded heads get realized, somthing that is not possible in TAG. TAG has to retreat to disjunctions instead. While this allows straight-forward analyses of active/passive alternations based on the light verb in valence-based frameworks, such an option does not exist for TAG and it has to be assumed that preverbs come with different sets of arguments.


Introduction
While comparing current syntactic theories (Hagemann and Staffeldt, 2014, Müller, 2010a, 2018, Kertész et al., 2019, one may get the impression that the frameworks are rather similar and easily translatable into each other. One reason for this is that we all deal with the same data and provided that we made similar categorial distinctions in the raw analysis of the data it would be a surprise if the analyses were radically different. However, it is not all the same. It matters which formalism is chosen and some are able to express intuitions rather directly while others are not.
This brief discussion note deals with the analysis of complex predicates consisting of a preverb and a light verb. Preverbs often have an argument structure of their own. They describe an event and the light verb can be used to realize either the full number of arguments or a reduced set of arguments.
( Similar examples can of course be found in other languages making heavy use of complex predicates (Müller, 2010b). Ashwini et al. (2019) assume that the structures for the examples in (1) are composed of elementary trees for tareef 'praise' and the respective light verbs. This is shown in Figure 1 and Figure 2 respectively. The TAG analysis is only sketched here. The authors use feature-based TAG, which makes it possible to enforce obligatory adjunction: the elementary tree for tareef is specified in a way that makes it necessary to take the tree apart and insert nodes of another tree. This way it can be ensured that the preverb has to be augmented by a light verb. This results in XP f being inserted at XP 1 in Figure 1 and Figure 2.
What the analysis clearly shows is that TAG assumes two lexical items for the preverb: one with two arguments for the active case and one with just one argument for the passive. In general one would say that tareef is a noun that describes the praising event, that is, one person praises another one. Now this noun can be combined with a light verb and depending on which light verb is used we get an active The tree of the light verb is adjoined into the tree of the preverb, into the XP1 position sentence with both arguments realized or a passive sentence with the agent of the eventive noun suppressed. There is no morphological reflex of this active/passive alternation at the noun. It is just the same noun tareef : in an active sentence in (1a) and in a passive one in (1b). And here we see a real difference between the frameworks: TAG (Joshi et al., 1975) is a framework in which structure is assembled: the basic operations are substitution and adjunction. The lexicon consists of ready-made building blocks that are combined to yield the trees we want to have in the end. This differs from Categorial Grammar (Ajdukiewicz, 1935) and HPSG (Pollard andSag, 1994, Sag, 1997) where lexical items do not encode real structure to be used in an analysis, but potential structure: lexical items come with a list of their arguments, that is, items that are required for the lexical element under consideration to project to a full phrase. However, lexical heads may enter relations with their valents and form NPs, APs, VPs, PPs or other phrases, but they do not have to. Geach (1970) developed a technique that is called functional composition or argument composition within the framework of Categorial Grammar and this was transferred to HPSG by Nakazawa (1989, 1994). Since the 90ies this technique is used for the analysis of complex predicates in HPSG for German (Hinrichs and Nakazawa, 1994, Kiss, 1995, Meurers, 2000, Müller, 1999, 2002, Romance (Miller and Sag, 1997, p. 600;Monachesi, 1998;Abeillé and Godard, 2002), Korean (Chung, 1998), and Persian (Müller, 2010b). See Godard and Samvelian, 2019 for an overview. For instance Müller (2010b, p. 642) analyzes the light verbs kardan 'do' and odan 'become' this way: both raise the subject of the embedded predicate and make it their own argument but kardan introduces an additional argument while odan does not do so.
Applying the argument composition technique to our example, we get the following lexical item for tareef : (2) Sketch of lexical item for tareef 'praise':  The arg-st list contains all arguments of a head. The arguments are liked to the semantic representation and are mapped to valence features like specifier and complements. Depending on the langauge and the realizationability of subjects within projections, that subject may be mapped to a separate feature, which is a head feature (Kiss, 1995,  The verb kar 'do' selects for a noun and takes whatever the subj value of this noun is ( 1 ) and concatenates the list of complements the noun takes ( 2 ) with the value of subj. The result is 1 ⊕ 2 and it is a prefix of the arg-st list of the light verb. The lexical item for ho 'be' is similar, the difference being that the subject of the embedded verb is not attracted to the higher arg-st list, only the complements ( 1 ) are.
For finite verbs it is assumed that all arguments are mapped to the comps list of the verb, so the comps list is identical to the arg-st list.
The analysis of our example sentences is shown in Figures 3 and 4. The conclusion is that HPSG has a representation of potential structure. When light verbs are present, they can take over valents and "ex- hu-ii be.part-perf.f.sg/be.pres ecute" them according to their own preferences. This is not possible in TAG since once structure is assembled it cannot be changed. We may insert items into the middle of an already assembled structure but we cannot take out arguments or reorder them. This is possible in Categorial Grammar and in HPSG: the governing head may choose which arguments to take over and in which order they should be represented in the valence repsresentations of the governing head.
LFG is somewhere in the middle between TAG and HPSG: the phrase structural configurations are not fully determined as in TAG since LFG does not store and manipulate phrase markers. But lexical items are associated with f-structures and these f-structures are responsible for which elements are realized in syntax. As complex predicates are assumed to be monoclausal it is not sufficient to embed the f-structure of the preverb within the f-structure of the light verb (Butt et al., 2003). Since the grammatical functions that are ultimately realized in the clause do not depend on the preverb alone the light verb may have to determine grammatical functions contributed by the preverb. In order to be able to do this Butt et al. (2003) use the restriction operator (Kaplan and Wedekind, 1993), which restricts out certain features or path equations provided by the preverb's and the light verb's f-structures. This is another instance of too strict specifications: once specified, it is difficult to get rid of it and special means like partial copying via restriction are needed. The alternative not relying on restriction was suggested by Butt (1997): embedding relations can be specified on the a-structure representation and then a mapping is defined that maps the complex a-structure to the desired f-structure. Mapping between several levels of representation is a general tool that is also used in HPSG: for instance, Bouma, Malouf, and Sag (2001) used arg-st, deps, and comps in the treatment of nonlocal dependencies. See also Koenig (1999) on the introduction of arguments via additional auxiliary features. As I showed in Müller (2007, Section 7.5.2.2), one would need an extra feature for every kind of argument alternation that is to be modeled this way.
Summing up, I showed that there are indeed differences between the frameworks that are due to the basic representational formalisms they assume. While TAG assumes that the lexicon contains trees with a certain structure, HPSG assumes that lexical items come with valence specifications, that is, they have descriptions that together with dominance schemata (grammar rules) that are separate from the lexical items determine how possible structures look like. Since valence representations can be composed by superordinate predicates there is enough flexibility to deal with various light verb phenomena. LFG is a bit more constrained due to the use of f-structures, but using a restriction operator unwanted information about grammatical functions can be keept out of f-structures of matrix predicates.