Complex Predicates

Complex predicates are instances in which two or more predicational elements combine to form a single monoclausal predication. The definition proposed in Butt (1995) is as follows.

> Complex predicates are formed when two or more predicational elements enter > into a relationship of co-predication. Each predicational element adds arguments > to a monoclausal predication. Unlike what happens with control/raising, there are > no embedded arguments and no embedded predicates at the level of syntax.

Classic examples in the LFG literature come from Romance and Urdu/Hindi (Alsina 1993,1996;Butt 1995; Mohanan 1994). The basic idea is illustrated here with respect to a type of N-V and a type of V-V complex predicate from Urdu/Hindi. In the N-V example below, the complex predicate consists of the predicate `do', which licenses two arguments: an agent and a predicate that represents the thing done, namely `memory'. The `memory' in turn licenses the NP `story'. This NP functions as the direct object of the clause yet is not licensed by the finite verb `do'. The standard analysis is that `memory' and `do' combine at the level of a(rgument)-structure and that their arguments are combined into one monoclausal predication and linked to GFs (grammatical functions) as being in a single argument domain.

nadya=ne   kahani  yad   k-i
Nadya.F.Sg=Erg  story.F.Sg.Nom memory    do-Perf.F.Sg
`Nadya remembered a/the story.' (lit.: `Nadya did memory of the story.')

The same analysis holds for the V-V complex predicate below, where `give' licenses three arguments: the agent (`Nadya'), the beneficiary/goal (`Yassin') and the predicate (event/proposition) that is being allowed to be done (`cut'). This predicate in turn licenses the theme `plant'. Again, the arguments separately contributed by `give' and `cut' are combined at a-structure and linked to GFs as a monoclausal predicate so that you have one composed predicate `give-cut', a SUBJ `Nadya', an OBJ `plant' and an OBJ_theta `Yassin'.

nadya=ne yAssin=ko pAoda kaT-ne di-ya th-a 
Nadya.F.Sg=Erg Yassin.M.Sg=Dat plant.M.Sg.Nom cut-Inf.Obl give-Perf.M.Sg be.Past-M.Sg
‘Nadya had let Yassin cut the plant.’

Determining whether a given structure is monoclausal or not must be done via tests for monoclausality. These tend to be language specific, but can include tests involving control, agreement, anaphora resolution or the distribution of negative polarity items (see Butt 1995,2010).

For example, the sentence below is almost identical to the sentence above, but Butt (1995) shows the sentence below is structurally biclausal, involving an XCOMP `to cut the plant' whose SUBJ is controlled by the matrix OBJ_theta `Yassin'.

nadya=ne yAssin=ko [pAoda kaT-ne]=ko kah-a th-a
Nadya.F.Sg=Erg Yassin.M.Sg=Dat plant.M.Sg.Nom cut-Inf.Obl=Acc say-Perf.M.Sg be.Past-M.Sg
‘Nadya had told Yassin to cut the plant.’ 

A-structure and LFG's _Linking_ or _Mapping Theory_ is not implemented within XLE. In the absence of a-structure and Linking Theory, the question faced within ParGram was as to how one could treat complex predicates in as linguistically satisfying a manner as possible. The solution arrived at ultimately involves the use of the _Restriction Operator_ (Kaplan and Wedekind 1993). The notation for the Restriction Operator is "/" and it allows for the definition of a given f-structure that corresponded to another f-structure but without some of the information contained in the original f-structure. For example, if one wanted to restrict out the CASE feature from an f-structure, then one can say ^/CASE. The correspoding restricted f-structure will be identical to the original one, except for the CASE feature.

As originally conceived, the Restriction Operator applied at the lexical level. However, as complex predicate formation of the type described above applies in the syntax, the Restriction Operator was extended to apply in the syntax. Details on how to use the Restriction Operator are provided in the XLE documentation (http://ling.uni-konstanz.de/pages/xle/doc/notations.html#N4.1.11). Butt et al. (2003) provides a very detailed explanation and sample rules for an application of the Restriction Operator to complex predicates in Urdu. Butt and King (2006) show how to extend this basic treatment to morphological causatives and discuss interactions between passives and syntactically formed complex predicates. Readers are referred to those two papers plus the XLE documentation if they want to being using the Restriction Operator to model complex predicates.

The intuition behind the application of the Restriction Operator to model complex predication is as follows. Although XLE does not include a model of a-structure and linking, the system can make reference directly to arguments on the argument list of a PRED via variables. The arguments of a PRED are ordered and are numbered from left to right. In a subcategorization frame like `verb<(^SUBJ)(^OBJ)>, the SUBJ is Arg1 and the OBJ is Arg2. These argument positions can be referred to as: : ARG1, ARG2, etc. and can thus then also be manipulated via f-structure annotation. Now, if one has a light verb like `do' which specifies that it has a SUBJ argument, but where the nature of the second argument (%ARG2) is unclear because it needs to be determined dynamically as part of the syntactic analysis, then the lexical entry would look as follows.

do LV * (^ PRED)='do<(^ SUBJ) %ARG2>'
"%ARG2 will be filled in by a predicate".

The "filling in" of the second argument can be done via f-structure annotations on a c-structure as follows. The effect is that the Noun (N) "loses" its own PRED, but that this PRED becomes substituted in for the ARG2 of the light verb. The noun thus becomes a part of the verb and all arguments and other information it might have also had specified is now part of the verb's f-structure. For example, if the N is specified to have an OBJ, then that OBJ will now be the OBJ of the joint predication.

 VPcp ->  N:    !\PRED = ^\PRED
                        (! PRED) = (^ PRED ARG2)
                 LV "light verb". 

The above is a very quick characterization. In practice, working with a grammar that contains the Restriction Operator is difficult in comparison to the other grammatical devices that XLE offers. Debugging and grammar maintenance becomes much more complicated because the Restriction Operator does not "lose" information. Rather, it creates more f-structures and any constraining equations or other checks for well-formedness can now in principle apply to both the "original" version of the f-structure as well as the restricted out version. To solve the problems this created, the Urdu grammar, for example, added a feature RESTRICTED + to the grammar to "mark" the f-structures that had been restricted out. Wellformedness constraints are then engineered to only apply to f-structures that are not restricted out.

Generation with Restriction also poses major problems. However, these have recently been addressed in Kaplan and Wedekind (2012).

Finally, Butt and King (2006) note that there are problems in the interaction between passives and complex predicates with the use of the Restriction Operator. That paper did not fully understand the problem, but this was later solved by Özlem Cetinoglu with respect to the Turkish grammar. The problem is essentially one of ordering and is due to the fact that passives were treated via Lexical Rules (http://ling.uni-konstanz.de/pages/xle/doc/notations.html#N5) as per the classic idea in LFG. In the Turkish and Urdu grammars the Passive Lexical Rule was invoked as part of the lexical entry of a stem. For example, something like `break' would specify via the application of the lexical rule that it could be either active and have a SUBJ and an OBJ, or be passive and have a SUBJ and a NULL or ADJUNCT. This information was passed on through the morphology which takes care of inflection and derivation and then enter the syntax as part of the fully inflected verb. However, when you add causative morphology to the mix, you have the problem that the causative morpheme serves to augment the basic a-structure of a stem (by adding a causer, usually) and that this augmentation needs to happen BEFORE any application of passivization. That is, the causative information would also apply to the subcategorization frame that said: `break<SUBJ>' and thus produce unexpected solutions.

The current solution has been to move passivization out of the lexicon listing verb stems and have it apply within the sublexical rules that govern word formation. The ordering problem is thus solved.

Link Files from ParGram meetings

References

• Ahmed, Tafseer and Miriam Butt. 2011. Discovering Semantic Classes for Urdu N-V Complex Predicates. In _Proceedings of the International Conference on Computational Semantics (IWCS 2011)_, Oxford.
• Alsina, Alex. 1993. _Predicate Composition: A Theory of Syntactic Function Alternations_. PhD thesis, Stanford University.
• Alex Alsina. 1996. _The Role of Argument Structure in Grammar_. CSLI Publications.
• Butt, Miriam. 1995. _The Structure of Complex Predicates in Urdu_. CSLI Publications.
• Butt, Miriam. 1998. Constraining argument merger through aspect. In: Hinrichs E, Kathol A, Nakazawa T (eds) _Complex Predicates in Nonderivational Syntax_, Academic Press, pp 73‚Äì113.
• Butt, Miriam and Tracy King. 2006. Restriction for Morphological Valency Alternations: The Urdu Causative. In M. Butt, M. Dalrymple and T.H. King (eds.) _Intelligent Linguistic Architectures: Variations on_ _Themes by Ronald M. Kaplan._ CSLI Publications, 235-258. http://ling.uni-konstanz.de/pages/home/butt/main/papers/ron-fest.pdf
• Butt, Miriam. 2013. Control vs. Complex Predication. _Natural Language and Linguistic Theory_ 32(1):155--190. DOI 10.1007/s11049-013-9217-5
• Butt, Miriam. 2010. The Light Verb Jungle: Still Hacking Away, In M. Amberber, M. Harvey and B. Baker (eds.) _Complex Predicates in Cross-Linguistic Perspective_, 48‚Äì78. Cambridge University Press.
• Butt, Miriam, Tracy Holloway King, and John T. Maxwell III. 2003. Complex Predicates via Restriction. _Proceedings of LFG03._ CSLI Publications.
• Kaplan, Ron, and J√ºrgen Wedekind. 1993. Restriction and Correspondence-based Translation. _Proceedings of the 6th European Conference of the Association of Computational Linguistics_, pp. 193‚Äì202.
• Kaplan, Ron, and J√ºrgen Wedekind. 2012. LFG Generation by Grammar Specialization. _Computational Linguistics_ 38(4):1-49. DOI: 10.1162/COLI_a_00113
• Mohanan, Tara. 1994. Argument Structure in Hindi. CSLI Publications.