Automatic translation of support verb constructions 
Morris SALKOFF 
Universit~ Paris-7, L.A.D.L. 
2, Place Jussieu, 75005 PARIS, FRANCE 
Verb,; like have, do, take, be, etc., appear with verb 
nominalisations in sentences like the following: 
(1)a Max alluded to the crime ~> b Max made 
an allusion to the crime 
(2)a Max harmed the child ~> b Max did harm 
to the child 
(3)aMax walked along the street~> b Max 
took a walk along the street 
(4)a Jim's story contradicts Max's 
confession ~> b Jim's story is in 
contradiction with Max's confession 
M. Gross (1981) calls such verbs 'support verbs', and 
I shall adopt his terminologLv. These verbs exhibit 
many interesting properties which have been studied 
systematically for several French support verbs: faire 
(make), avoir (have), prendre (take), etre (be), etc. An 
examination of the results indicates that support 
verb,; must be taken into account in the parser and in 
the lexicon of a program of automatic translation. I 
shall discuss the problem in the context of a possible 
program for French- English translation. 
In what follows, Vsup designates the support verb, 
and Nsup stands for the 'supported' noun associated 
with Vsup. NO and N2 designate the subject noun and 
the indirect object noun, respectively. The symbol'+' 
is used to indicate a choice. 
I. The support verb construction 
In order to be able to parse and then translate 
correctly such sequences as make an allusion, 
do harm, etc., the nominalizations allusion and harm 
must be marked in the lexicon as appearing with the 
support verbs make and do. There are several reasons 
for this: 
(i) The translation of a simple verb (e.g., allude) or of 
a support verb construction may or may not be an 
equivalent simple verb or support construction in the 
target language. For example, (1)a cannot be 
translated into French with a simple verb: 
(5) Max alluded to the crime = Max (*alludait + 
a fait une allusion) au crime 
Conversely, there are French support constructions 
which am more easily be translated into English by a 
simple verb than by a support cotrstruction : 
(6) Max fait une beuverie chaque samedi' = Max 
carouses every Saturday 
(ii) A support verb can be zeroed, leaving behind an 
elliptical construction that cannot be translated as it 
stands. Consider the following sentences : 
(7) People who are on drugs (are..) = Les gens 
qui prennent des drogues (sont..) 
The English support construction be on drugs is 
related to the sentence Ahz~c drugs h#nself. In French, 
the support verb must be prendre (take). When 
who are is deleted in (7), we obtain: 
(8) People on drugs (are...) 
But this sentence can only be translated into French 
as (7) was, since the French support quiprennent 
(who take) is not deletable. Note that it is not a 
question here of on having a different meaning from 
its more common locative sense (usually translated as 
sur). Rather, the support verb be can be deleted in 
English, and the elliptical construction (8) has no 
direct equivalent in French where the support verb is 
not zeroable. 
(iii) More complicated cases of support verb deletion 
occur. Certain support verb constructions contain a 
predicate noun that is not the nominalization of a 
verb; when the support verb is deleted, a problem in 
translation may arise. 
(9)a L'attention que Max a prSt6 au probl~me 
(est..) = The care that Max gave to the 
problem (is..) 
(9)b L'attention de Max au probl~me 
(est..) = *Max's care to the problem (is..) 
(10)a L'entorse que Max a fait aux r6glements 
(est..) = The violence that Max did to the 
rules (is..) 
(10)b L'entorse de Max aux r6glements 
(est..) = *Max's violence to the rules (is..) 
I 243 
The French support verb constructions 
prater attention in (9)a and faire une entorse in (10)a 
translate without difficulty into English. However, the 
support verb can be deleted in French, as in (9)b and 
(10)b, but not in English. The latter sentences can 
only be translated by re-introducing the deleted 
support verb. 
To handle all these difficulties, a link must be 
established in the lexicon between the support verb 
and all the nouns associated with it (since each 
support verb can appear typically with many hundreds 
of nouns). Thus, beuverie (in 6) and entorse (in 10) 
must be marked as appearing with faire. In English, 
drug, s must be marked for be on. Then, after parsing, 
faire une beuverie will be a sequence of marked, hence 
linked, categories and the French-English transfer 
lexicon furnishes a translation specific to that 
sequence. Similarly in (8), (9)b and (10)b the mark 
associated with the nouns drugs, attention and entorse 
will enable the program to reconstitute the deleted 
support verb. 
Cases like (10)b require more lexical markers than 
the simpler cases of (1) to (4). The problem is that 
there are sentences similar to (9)b and (10)b which 
yet do not require the addition of the deleted support 
verb in order for a translation to be possible. This is 
the case in such sentences as the following: 
(ll)a L'acquiescement que Max donne au projet 
(est..) = The consent that Max gives to the 
project (is...) 
(ll)bL'acquiescement de Max au projet 
(est..) = Max's consent to the project (is..) 
In this case, it suffices to indicate that acquiescement 
can be supported by donner, so that (ll)a can be 
parsed correctly; then the transfer lexicon contains 
the translation of donner acquiescementgl = give 
consent to. This entry further indicates that 
acquicscenlent, without donner, translates as consent. 
This part of the entry is sufficient for the translation 
of acquiescement ?t in (ll)b. 
In (9)b and (10)b, the nouns attention and entorse are 
marked for the Vsup (prater, faire) with which they 
can appear, just as acquiescement is marked for 
donner. However, if the sequence de NO ~ N2 
appears after the nominalizations, it has to be 
transformed into que NO Vsup ~ N2 before it can 
be translated correctly. Therefore, the lexical entries 
of attention and entorse must contain more than just 
the translation of faire attention and faire entorse: they 
must additionally be marked as requiring this 
transformation of a prepositional phrase into a 
relative for the purpose of translation. 
II. Parsing support verb constructions 
The principal feature of the support verb construction 
is that the verbal slot in the sentence is occupied by 
the combination of a verb, Vsup, plus a noun, Nsup. 
Thus, if I schematize a sentence as follows: 
(12) Sent--> S V O 
then the verb slot V, when parsing (1), can contain 
either allude or Vsup Nsup = make an allusion. This 
means that the verb V must have at least two options 
in its definition: a simple verb, or a support verb 
construction. This yields the following kinds of 
analyses: 
Sent 
S V 0 
NP PP 
NO all es Prep NP2 I l I 
Max to the crime 
FIG 1 
Sent 
S V 0 1 I 1 
NP VSUP OSUP I \ 
NO Vsup NP 
Max makes Nsup Prep NP2 I / I 
an attusion to crime 
FIG 2 
The symbol VSUP in Fig. 2 designates the whole 
support construction. In these two sentences, the 
object O and the support verb object OSUP contain 
the same prepositional phrase to the ctbne. In other 
cases they may differ, e.g., as in (2), where O is an NP 
(child) and OSUP is a PP (to the child). 
In order to obtain just the two analyses O f Figs. 1 and 
2, many details must be attended to: 
(1) The noun Nsup associated with Vsup in Fig. 2 
(allusion) must be analyzed as part of a complete 
noun phrase NP, and not just as a bare noun N. In 
effect, left and right modifiers may appear with this 
noun: 
Max makes a brief allusion in Latin to the crime 
In this sentence, allusion takes an adjective left 
modifier, and a prepositional phrase as a right 
modifier. 
(2) Since the NP appearing after Vsup has a slot 
for the prepositional phrase as a right modifier, the 
parser will always try to place the sequence 
244 2 
OSUP = to the crime in that slot, as a modifier of 
allusion. This can be prevented by attaching the 
following test to the set of right modifiers in the NP: a 
prepositional phrase is not allowed as a right modifier 
of the noun Nsup if the preposition is precisely the 
one associated with Vsup. 
(3) Conversely, the parser will try to analyze the 
sentence of Fig. 2 with the schema of Fig. 1. That is, it 
tries Nsup = allusion as a simple direct object of 
makes. This should be disallowed, even if the 
sentence is such that OSUP is empty, e.g., Max made 
a strange allusion. Whenever a Vsup appears with an 
Nsup it should be analyzed as a verb support 
construction, and not as an ordinary verb followed by 
a direct object. 
This is accomplished by adding a test which examines 
the markers attached to the verb and object in their 
lexical entries. If the test can verify that the verb 
carries the mark Vsup, as make does, and the noun 
object carries the mark of an Nsup, as aUusion does, 
then the option 'ordinary vcrb' of V is disallowed. 
This forces the parser to choose the option VSUP for 
the verb. All these tests and constraints on the 
parsing can easily be incorporated into the string 
analyzer of French that I have developped for the 
IBM 9370 (1973, 1979), and which is currently being 
~'ewritten in ADA. 
IlL Additional problems 
Various difficulties have to be takcn care of when a 
~,;upport verb is added to the source sentence for the 
purposes of a smoother translation. 
(i) We have seen in (10)b and (ll)b that some 
modifiers of an Nsup of the form de NO gi N2 
require the addition of the deleted support verb if a 
correct translation is to be obtained. Adding the 
support verb implies that a tense must be supplied for 
it. This tense can be derived, but not unambiguously, 
from the tenses of the verbs in which the support 
construction is embedded. The problem is that more 
than one tense may be acceptable, even in the source 
language. \]Hence, only an approximate solution can 
be given. 
When the support verb construction is the subject, a 
tense can be added to the support verb as follows. 
a) If the tense of the principal verb is present, future 
or conditional, then the support verb can take the 
present tense: 
(13) L'entorse de Max aux r6glements 
(est + sera + serait)nuisible -- The violence 
Max does to the rules (is+ +will be + would 
be) harmful 
b) If the tense of the principal verb is some form of 
the past, then the support verb takes a past tense: 
(14) L'entorse de Max aux r6glemeuts (6tait + a 
6t6 + aurait 6t6) nuisible = The violence 
Max did to the rules (was + has 
been + would have been) harmful 
When the support verb construction is in tile direct or 
indirect object, then either the past or present tense 
can be added. The French construction is ambiguous: 
(15) (Je remarque + J'avais remarqu6) l'entorse 
de Max aux r6glements = (I notice + I had 
noticed) the violence Max (does + did) to 
the rules 
(ii) Some support constructions, like (ll)b, do not 
require the addition of the deleted support verb. The 
difference between those that do, as in (13) and (14) 
above, and those that do not must appear as a mark 
in the lexical entry of the supported noun. 
Thus, the Nsup entorse (violence) has a first mark 
indicating that its Vsup is faire (do). A second mark 
indicates that the following transformation must be 
effected in order to obtain a correct translation: 
(16) Nsup (entorse) de N0hum ~ N--> Nsup 
(entorse) que N0hum fait ~t N 
In addition to checking for the presence of this 
second mark, the program must also verify that the 
transformation of (16) is applicable, i.e., that the 
noun appearing after de is a 'human' noun that can be 
an acceptable subject of faire une entorse. 
(iii) In many support constructions, the noun phrase 
consists of a frozen expression, i.e, an invariable 
sequence of words: 
(17)a Max fait grand cas de cela- Max sets a 
high store on that 
(17)b Max fait fi de vos conseils = Max scorns 
your advice 
There are many hundreds, or thousands, of 
expressions of this kind in which no substitutions can 
be made, nor any words added. 
To parse this type of support construction, the string 
VSUP, schematized in Fig. 2, must contain a second 
option which is not a noun phrase NP, but which 
consists only of a sequence of fLxed words. Then, 
another test must be added to VSUP to ensure that 
grand cas orfi is not analyzed as a variable NP, i.e., as 
an occurrence of the first option. A mark can be 
added to the lexical entry of cas and fi to indicate 
invariability. In addition, one distinguished entry must 
contain the exact sequence of words constituting the 
3 245 
supported expression. For example, the entry for cas 
contains the sequence grand, standing for the 
supported expression grand cas. 
The passive transformation is possible for some of 
these support constructions, but its applicability 
cannot be predicted: 
(18)a Grand cas a 6t6 fait de cela = High store 
was set on that 
(18)b *Fi a 6t6 fait de vos conseils 
Hence, this possibility must also be marked in the 
lexical entry, so that the absence of an article in the 
subject grand cas can be taken into account by the 
test for the well-formedness of the subject. 
IV. Conclusions 
Systematic studies of French support verbs have 
shown that there are many thousands of support 
constructions of the type discussed here, for each 
such verb (cf. J. Giry-Schneider, 1987; G. Gross, 
1989). The translations of such expressions is 
irregular, but this is to be expected, and such 
irregularity will be confined to the lexicon. The 
regularities of support verb constructions can be 
handled in the manner discussed here, and this 
treatment simplifies considerably the parsing module 
of a program for automatic translation. 

References 

J. Giry-Schneider, "Les pr6dicats nominaux en 
franqais", Droz, Gen6ve, 1987 

G. Gross, "Les constructions converses du fran~ais", 
Droz, Gen6ve, 1989 

M. Gross, "Les bases empiriques de la notion de 
pr6dicat s6mantique", in Langages, No. 63, 
Sept. 1981, pp. 7-51, Larousse, Paris 

M. Salkoff, "Une grammaire en cha~ne du fran~ais", 
Dunod, Paris, 1973 (out of print) 

M. Salkoff, "Analyse syntaxique du fran~ais", 
J. Benjamins, Amsterdam, 1979 
