Verb Phrase MWEs

Overview

This discusses idioms that lexicalize (at least) a verb and a complement noun. In the literature, various subcategories of VP idioms are known as Verb-Noun Constructions (VNCs), Light Verb Constructions (LVCs), Support Verb Constructions, and phrasal idioms. Such constructions possess varying degrees of fixedness/productivity.

LVCs: a simple approach

Light verb constructions, in which the noun is eventive and the verb adds no or little meaning about that event, are quite frequent, so it makes sense to tackle them first. Issues:

How to constrain the combination of allowed verb and eventive noun?

Some light verbs are productive: e.g., it seems there is a frequent function of ‘take’ that can work with almost any single-argument activity noun. Ideally, a CCG lexicon would license light verb–noun combinations based on their semantic compatibility, but this does not look simple to accomplish, either theoretically or practically with broad coverage.

As a first step, let us consider the following two alternative simplifications:

In the lexicon, a fixed list of light verbs and a fixed list of eventive nouns. Any LV is allowed to combine with any eventive noun.
In the lexicon, a fixed list of licensed LVCs.

Option 1 would produce analyses like:

Nobody paid attention to me , so I took a long bath .

Nobody : NP
paid : ((S\NP)/PP)/NP[+evt]
attention : NP[+evt]
to : PP/NP
me : NP
I : NP
took : (S\NP)/NP[+evt]
a : NP[+evt]/N[+evt]
long : N[+evt]/N[+evt]
bath : N[+evt]

paid attention : paid > attention => (S\NP)/PP
to me : to > me => PP
paid attention to me : paid attention > to me => S\NP
Nobody paid attention to me : Nobody < paid attention to me => S
long bath : long > bath => N[+evt]
a long bath : a > long bath => NP[+evt]
took a long bath : took > a long bath => S\NP
I took a long bath : I < took a long bath => S

where [+evt] marks the noun’s potential to appear in an LVC. Determiners and adjectives would have to be designed so as to propagate this feature up to the full NP.

If we are doing parsing, it may not matter too much if the grammar overgenerates LVCs (*make a bath, etc.).

Option 2 would constrain LVCs to hard-coded lexical combinations:

Nobody paid attention to me , so I took a long bath .

Nobody : NP
paid : ((S\NP)/PP[to])/NP[attention]
attention : NP
to : PP/NP
me : NP
I : NP
took : (S\NP)/NP[bath]
a : NP/N
long : N/N
bath : N

paid attention : paid > attention => (S\NP)/PP[to]
to me : to > me => PP
paid attention to me : paid attention > to me => S\NP
Nobody paid attention to me : Nobody < paid attention to me => S
long bath : long > bath => N
a long bath : a > long bath => NP
took a long bath : took > a long bath => S\NP
I took a long bath : I < took a long bath => S

Option 2 would disallow *paid attention of me, *paid a bath, and *made attention. It is thus higher precision, lower recall.

Both solutions assume there is a one-to-one correspondence between light verbs and eventive nouns. It does not consider, e.g., I took a walk, then a nap, or the zeugmatic ??I took a picture, then a nap.

How to make the LVC semantics equivalent to its heavy verb paraphrase?

From a semantic parsing perspective, the value in treating LVCs specially is to reduce sparsity by assigning the same semantics to an LVC (take a bath) as to its paraphrase with a heavy verb (bathe).

The simplest approach is for the light verb to simply take on the semantics of the NP. To actually reduce sparsity, the semantic predicate associated with the noun should match that of a heavy verb in the lexicon. In many cases this will be orthographically identical (walk, nap, lecture, etc.), but in some cases the noun will be different (attention → attend, bath → bathe, decision → decide, speech → speak, etc.). There are existing resources from which we can extract lists of these correspondences.

As long as light verbs are distinguished from heavy verbs in the lexicon, this will work equally well for Options 1 and 2 above. Here is the Option 2 example with GraphParser-style logical forms suggested by Siva:

Nobody paid attention to me , so I took a long bath .

Nobody : NP : λz. [z=nobody]
paid : ((S\NP)/PP[to])/NP[attention] : λf1,f2,f0,x. ∃y,z. f1(x) ∧ f2(y) ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y)
attention : NP : λx. [x=attend]
to : PP/NP : λf1,x. f1(x)
me : NP : λy. [y=me]
I : NP : λy. [y=i]
took : (S\NP)/NP[bath] : λf1,f0,x. ∃y. f1(x) ∧ f0(y) ∧ arg1(x_e, y)
a : NP/N
long : N/N : λf1,x. f1(x) ∧ long(x)
bath : N : λx. [x=bathe]

paid attention : paid > attention => (S\NP)/PP[to] : λf2,f0,x. ∃y,z. [x=attend] ∧ f2(y) ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y)
to me : to > me => PP : λy. [y=me]
paid attention to me : paid attention > to me => S\NP : λf0,x. ∃y,z. [x=attend] ∧ [y=me] ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y)
Nobody paid attention to me : Nobody < paid attention to me => S : λx. ∃y,z. [x=attend] ∧ [y=me] ∧ [z=nobody] ∧ arg1(x_e, z) ∧ arg2(x_e, y)
long bath : long > bath => N : λx. [x=bathe] ∧ long(x)
a long bath : a > long bath => NP : λx. [x=bathe] ∧ long(x)
took a long bath : took > a long bath => S\NP : λf0,x. ∃y. [x=bathe] ∧ long(x) ∧ f0(y) ∧ arg1(x_e, y)
I took a long bath : I < took a long bath => S : λx. ∃y. [x=bathe] ∧ long(x) ∧ [y=i] ∧ arg1(x_e, y)

A few notes about the above:

The semantics is neo-Davidsonian: in the logical form, event-participant relations are denoted with predicates arg1(evt, argfiller), arg2(evt, argfiller), etc.
x_e indicates the event reading of an instance whose variable is x. The non-eventive sense of bath will not permit this coercion, so it will not work semantically with light verb take.
N and NP categories are typed as functions that take an instance variable as an argument. This is a trick that GraphParser uses to postpone issues of quantification. The notation [x=bathe] means that x denotes an instance of bathing.
In the ultimate logical form, there is a single argument—the instance of the head predicate. Entities mentioned in the logical form are introduced with existential quantifiers.

TODO: Ensure the formulation works for syntactic constructions such as passive and cluster coordination:

give Sam a handshake and Alice a high-five

give : (VP/NP)/NP : λf1,f2,f0,x. ∃y,z. f1(x) ∧ f2(y) ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y)
Sam   : NP : λx. [x=Sam]
      : <T => (VP/NP)\((VP/NP)/NP) : λu. u(λx. [x=Sam])
a handshake : NP : λy. [y=shake-hands-with]
            : <T => VP\(VP/NP)     : λv. v(λy. [y=shake-hands-with])
and : CONJ
Alice : NP : λx. [x=Alice]
      : <T => (VP/NP)\((VP/NP)/NP)
a high-five : NP : λy. [y=high-five]
            : <T => VP\(VP/NP)

Sam a handshake   :  Sam <B a handshake   => VP\((VP/NP)/NP) : λz. (z(λx. [x=Sam]))(λy. [y=shake-hands-with])
Alice a high-five :  Alice <B a high-five => VP\((VP/NP)/NP) : λz. (z(λx. [x=Alice]))(λy. [y=high-five])
and Alice a high-five : and <&> Alice a high-five => VP\((VP/NP)/NP)
Sam a handshake and Alice a high-five : Sam a handshake <&> and Alice a high-five => VP\((VP/NP)/NP) : λz. (z(λx. [x=Sam]))(λy. [y=shake-hands-with]) ∧ (z(λx. [x=Alice]))(λy. [y=high-five])
give Sam a handshake and Alice a high-five : give < Sam a handshake and Alice a high-five => VP : λf0,x. (∃y,z. [x=Sam] ∧ [y=shake-hands-with] ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y)) ∧ (∃y,z. [x=Alice] ∧ [y=high-five] ∧ f0(z) ∧ arg1(x_e, z) ∧ arg2(x_e, y))

Further examples

take LVCs

take a walk (swim, shower, break, …): any eventive noun for a durative activity with one (non-denoted) participant. AmE. (TODO: check CU work on ‘take’ LVCs)
take a {picture,photo}
take s.t. for granted/take for granted s.t.: similar to a VPC?

attention SVCs

pay attention (to NP)
give one’s attention to NP
give NP one’s attention
pay no {heed, mind}

give ditransitives

give s.o. their due
- possessive pronoun’s antecedent is the indirect object
give s.t. one’s best shot
- possessive pronoun’s antecedent is the subject

phrasal idioms

kick the bucket
spill the beans [decomposable]

References

Regarding take LVCs: Claire Bonial’s dissertation (in prep.)
Regarding idiom decomposability: Nunberg, Gibbs