LEXICAL FUNCTIONS AND MACHINE TRANSLATION 
Dirk Heylen, Kerry G. Maxwell and Marc Verhagen 
OTS, Trans 10, 3512 JK Utrecht, Netherlands 
CLMT Group, Essex University, Colchester, Essex CO4 3SQ, England 
email: heylen@let.ruu.nl, kerry@essex.ac.uk, verhm@essex.ac.uk 
This paper discusses the lexicographical concept of 
lexical functions (Mel'~uk and Zolkovsky, 1984) and 
their potential exploitation in the development of a ma- 
chine translation lexicon designed to handle colloca- 
tions. We show how lexical functions can be thought 
to reflect cross-linguistic meaning concepts for colloca- 
tional structures and their translational equivalents, and 
therefore suggest themselves as some kind of language- 
independent semantic primitives from which transla- 
tion strategies can be developed. ~ 
1 Description of the Problem 
Collocations present specific problems in translation, 
both in human and automatic contexts. If we take 
the construction heavy smoker in English and attempt 
to translate it into French and German, we find that 
a literal translation of heavy yields the wrong result, 
since the concept expressed by the adjective (something 
like ' excess:i_ve ' ) is translated by grand (large) in 
French and stark (strong) in German. We observe then 
that in some sense the adjectives stark, grand and heavy 
are equivalent in the collocational context, but that this 
is of course not typically the case in otber contexts, ef 
grande boite, starke Schachtel and heavy box, where 
the adjectives could hardly be viewed as equiwdent. It 
seems then that adjectives which are not literal trans- 
lations of one another may share meaning properties 
specifically in the collocational context. 
How then can we specify this special equivalence in 
the machine translation dictionary? The answer seems 
to lie in addressing the concept which underlies the 
union of adjective and noun in these three cases, i.e., 
intensification, and hence establish a single meaning 
representation tbr the adjectives which can be viewed 
as an interlingual pivot for translation. 
Collocations have been studied by computational lin- 
guists in different contexts. For instance, there is a 
substantial body of papers on the extraction of "fre- 
quently co-occurring words" from corpora using sta- 
tistical methods (e.g., (Choueka et al., 1983), (Church 
and Hanks, 1989), (Smadja, 1993) to list only a few). 
These authors focus on techniques for providing mate- 
rial that can be used in other processing tasks such as 
x The research rcpmlcd in this paper was undmtaken as the project 
"Collocations and the Lexicalisation of Semantic Operations" (ET- 
10/75). Financial contributions weir by the Commission of the 
European Community, Association Suissetra (Geneva) and Oxford 
University Press. 
word sense disambiguation, information retrieval, natu- 
ral language generation and so on. Also, the use of col- 
locations in different applications has been discussed 
by various authors ((McRoy, 1992), (Pnstejovsky et 
al., 1992), (Smadja and McKeown, 1990) etc.). How- 
ever, collocations are not only considered usefnl, but 
also a problem both in certain applications (e.g. gen- 
eration, (Nirenburg et al., 1988), machine translation, 
(Heid and Raab, 1989)) and fiom a more theoreti- 
cal point of view (e.g. (Abeill6 and Schabes, 1989), 
(Krenn and Erbach, to appear)). 
We have been concerned with investigating the 
lexical .\['unctions (IJTs) of Mel'0,uk (Mel'6uk and 
Zolkovsky, 1984) as a candidate interllngual device for 
tbe translation of adjectival and verbal collocates. Our 
work is related to research by (Heid and Raab, /989). 
In some respects it is an extension of some of their 
suggestions. Our work differs fi'om theirs in scope and 
also in the exploration of wtrious other directions. 
2 Representation 
The use we make of lexical functions as interlin- 
gual representations, does not respect their original 
Mel'~.ukian interpretation. Furthermore, we have trans- 
ferred them from their context in the Meaning-Text 
Theory to a different theoretical setting. We have em- 
bedded the concept in an HPSG-like grammar theory? 
In this section we review this operation. First we con- 
sider the features of Mel'~:nk's treatment that we have 
wanted to preserve. Next we show how they have been 
imported into the HPSG fi'amework. 
2.1 Collocations and LFs 
In Mel'~nk's Explanatory Combinatory Dictionary 
(ECD, see (Mel'~uk et al., 1984)), expressions such as 
uneJerme intention, une rdsistance acharnde, un argu- 
ment de poids, un bruit it~fernal and donner une lefon, 
faire un pas, commetre un crime are described in the 
lexical combinatorics zone. These "expressions plus 
ou moins fig6es" will be called 'collocations'. They 
are considered to consist of two parts -- the base and 
the collocate. In the examples above, the nouns are the 
bases and the adjectives and the verbs are the collocates. 
The idea that all adjective collocates and all the verb 
2Head Drivt~n Phrase SlltlCItlrc granllllar, see (Pollard and Sag, 
1987), (Pollard and Sag, to appear). For another treatment ,:ff collo 
cations in HPSG, see (Krenn and ltrbach, to appear). 
1240 
collocates share an important meaning component --- 
roughly paraphrasable as intense and do respectively -- 
and the fact that the adjectives and verbs are not inter- 
changeable but are restricted with this meaning to the 
accompanying nouns, is coded in the dictionary using 
lcxical functions (in this case Magn and Oper). 
Each article in the ECD describes what is called a 
'lexeme': a word in some specflic reading. In the lex- 
ical combinatorics zone, we lind a list of the lexical 
funclions that are relevant to this particular lexeme. 
Each lexical function is followed by one or more lex- 
crees (the result or value of the function applied to tile 
head word). The idea is that each combination of the 
argument with one of the values of the function forms 
a collocation in our terminology. The argument corre- 
sponds to tile base and each value is a collocate. The 
|ollowing fcatures of this representation are important 
to us. 
• l,exical functions are used to represent an impor- 
tant syutactico-semantic relation between tile base 
and the collocate. 
• The restricted combinatorial potential of the col- 
locate lexcme is accounted for by listing it at each 
base with which it can occur. 
The secund of these characteristics points out that 
the collocational restriction is seen as a purely lexical, 
idiosyncratic one: all collocations are explicitly listed. 
One other aspect of collocations which we have to 
deal with is the relation between the collocate lexeme 
and its freely occurring counterpart. Collocate lcxemes 
often differ in some respects from their literal variants 
while sharing other properties. Mel'6uk deals with this 
by including in the ECD an entry for the free variant and 
putting tile collocate-specific information in the entry 
for the base (with the result of the lexical functions). 
The fill entry of the eolh)cate is the result of taking 
the entry for the free variant and overwriting it with the 
information provided at the base. 
2.2 Collocations in HPSG 
The three aspects of Mel'6uk's analysis we wanted to 
encode in HPSG were the following. 
Coding the base-collocate relation in the lexicon. 
• Choosing the level at which \[cxical functions will 
be situated. 
• Relating the collocate information to the free wni- 
ant entry. 
We have provided straightforward solutions to these 
problems. For tire first problem we have taken over 
the ECD architecture rather directly, by creating a ded- 
icated 'collocates' field in the entry t)r bases which 
contains all the relevant collocates. As far as the sec- 
ond problem is concerned, the obvious place to put 
lcxical funclions is in the semantic reprcscntation pro- 
vided by HPSG. There are wtrious reasons for this. One 
is that 13;s arc used in lhe deep syntax level in Mel'6uk's 
model, a level oriented towards meaning. Another rea- 
son is that this level seems most appropriate to be used 
in transfer/translation and because we want to use lex- 
ical functions in transfer, this is where they should be. 
In contrast to the ECD, the meaning of the collocate is 
represented by the lexical function only. 
The following is an example of the entry for criti- 
cism with the encoding of strong as a collocate, a We 
use SEM_IND as all abbreviation for the feature path 
SEM.CONT.INI). 
PHON criticism 
REST {cl'iticislll(\[~)} 
COLLS { SEM_IND VAIl \[~' } 
REST {Magn(~l~)} 
Just as in the ECD the base contains a specific zone in 
which the collocates are listed. In our case, the feature 
'COLLS' has a set of lexical entries as its value. 
Each collocate subentry bears the value of the lexical 
function in its semantics field. In this representation the 
lexical function is chosen as the real semantic value of 
the collocate. One should read the feature structure as 
specifying that the semantics of strong (as a collocate) 
is the predicate Magn(\[~). 
The collocate subentry only provides partial infor- 
mation. In fitct, it provides only the intbrmation that 
is specific to the occurrence of strong in its combina- 
tion with criticism. In this case only the semantics is 
given. We further assume that the lexicon also con- 
tains a 'super-entry' which provides all the information 
that is shared by all the diflerent occurrences of strong. 
This entry is where the variable Sstrong points to. Of 
course, other architectures that try to avoid redundant 
specification of information are equally possible. For 
instance if one assumes a mechanism of default uni- 
fication, one can have Sstrong refer to the full entry 
describing 'strong' in say its ordinary use, and have 
the values that are particular to the collocational strong 
overwrite the values provided in the ordinary entry, as 
in Mel'~uk's proposal. 
Collocations, Rules and Principles So far, we have 
not specified in what way one gets flom the lexical en- 
tries for the base and the collocate to the representation 
of the collocational expression. 
ill HPSG, tile descriptions of complex expressions 
arc constrained by principles. We will assume that 
collocations are snbject to the same constraints. The 
ordinary rules of combination (combining adjectives 
and nouns, for instance) thus account for lnost of the 
properties of the collocational combination. However, 
we are still left with the typical 'collocational restric- 
tion' which nceds to be accounted for. 
We havc therefore addcd a principle which says that 
constructions that are analysed as collocations (indi 
cated by tile type COLI.OCATION) are either head-adjunct 
structure or head-complement structures with specific 
rcstrietions holding between the head anti the adjunct or 
aNoticc that hcrc we use a simple VCl'Sion of HPSG based 
on (Pollard and Sag, 1987) whereas the actual ilnplmncntation was 
based on (Pollard and Sag, to appear). 
1241 
the head and the complement respectively. Let's con- 
sider the former case 4, illustrated by the heavy smoker 
example, The adjunct daughter will contain the adjec- 
tive collocate. In such collocational constructions the 
collocate adjuncts have to be 'licensed' by the noun or 
the head daughter. This is implemented by requiring 
that the collocates field (C'OI,LS) of the head daughter 
contains a reference to a lexical entry that is compatible 
with the adjunct daughter. In the literal reading of an 
expression such as heavy smoker, the phrase will not 
be analysed as a COLL.OCATION and the principle does 
not apply. 
COLLOCATION -~e 
I H; jA Ii))7 I,)R'IR ~ CELLS {...~1~\]...} 1 
...v~OLLOCATE .., > \] 
COMP_DTRS < ...\[CELLS{... 1\[~..}\]... > 
3 Issues in Translation 
The project has tried to investigate the use of lexical 
functions as an interlingual device, i.e., one which is 
shared by the semantic representations of collocations 
in the language pairs ~. 
The typing of a collocation with such aflmction 
opens up the way to a treatment of collocations inside 
a given language module and hence to a substantial re- 
duction in the number of collocations explicitly handled 
in the multilingual transfer dictionary. The existence 
of a collocation function is established during analysis. 
This infi)rmation is used to generate the correct trans- 
lation in the target hmguage. To illustrate, the English 
analysis modnle might analyse (1) as (2). The transfer 
module maps (2) onto (3) which is then synthesised by 
the French module to (4). 
(l) heavy smoker-)~ (2) Magn(smoker) -r 
(3) Magn(fumeur) -4 (4) grand fumeur 
The exmnple points out that the translation strategy 
is a mixture of transfer and interlingua. The bases 
arc transferred but the representation of the collocate is 
shared between the source and the target representation. 
This treatment of collocations rests, among others, on 
the assmnptions that there are only a limited number of 
lexical functions, that lexical functions can be assigned 
consistently, that all (or a signilicant nmnber ot) collo- 
cations realise a lexical function, that lexical functions 
are not restricted to particular languages~ etc. In the 
following paragraph we present an outline of the trans- 
lation process. Next, we discuss some of the problems 
which follow flom our approach and we propose some 
ways to solve them. 
4"lb illustrate tile case of huad-conlplenlent structures olle 
coukl lake some support verb construction (also called ligh! verb 
consh'uction). 
t;For another application of LFs in a muhilingual NLP context 
see (Held and l/aab, 1989). For other Imatlnents of collocations in 
language generation see (Nirenburg et al., 1988) and (Smadja and 
McKeown, 1990). 
1242 
3°1 Lexical Functions as Interlingua 
it was assumed that the starting point for transfer is the 
semantic representation of the phrase. Using a semantic 
representation as input to transfer implies that we relate 
semantic values of wm'ds and phrases. For our purposes 
this is very satisfying since we will now be using the 
semantics of collocates instead of their orthography, 
in other words: we use lcxical flmctions and abstract 
away fl'om the particular realisation of a collocate in a 
particular language. 
We now state the relation between the semantic rep- 
resentations of the source language and target language. 
The semantic relation between the phrase heavy smoker 
and its French counterpart can be made explicit in the 
tbllowing bilingual sign: 
ENISEM-IND \[ VAR ~ \] RUST { smoker(~l-b,Magn(\[~)} 
FRISEM_INI) \[ VAR \[~ 1 REST {f.,,,~arI\[-~ M.~/~I)} J 
Typically, the lexicon will contain a bilingual sign 
for each possible value of RELN. Thus, for translat- 
ing heavy smoker into grandfumeur we will need the 
obvious entry tot smoker-fumeur plus the entry below: 
ENISEM-IND \[ VAR \[~ \] 
REST {M'lgn(~)} 
The interlingual status of the lexical function is self- 
evident. Any occurrence of Magn will be left intact 
during transfer and it will be the generation component 
that ultimately assigns a monolingual lexical entry to 
the LF. 6 
3.2 Problems 
l,exical Functions abstract away from certain nuances 
in meaning and from different syntactic realizations. 
We discuss some of the problems raised by this ab- 
stractkm in this section. 
Overgenerality An important problem stems fiom 
the interpretation of LFs implied by their use as an in- 
terlingua- namcly that the meaning of the collocate in 
some ways reduces to the meaning implied by the lex- 
ical./unction. This interpretation is trouble-free if we 
assume that LFs always deliver unique values; tmlb,'- 
tunately cases to the contrary can be readily observed. 
An example attested fiom our corpus was the range of 
adverbial constructions possible with the verbal head 
oppose: adamantly, bitterl3; consistently, steadjastly, 
strongly, vehementl); vigorously, deeply, resolutely, etc. 
The ftmction Magn is an appropriate descriptor in all 
cases since each adverb functions as a typical intensi- 
tier in this context. However each adverb also denotes 
6p'or more details we refer the reader to (Hcylen, 1993). There 
we also discuss our implementation in Alep, the C.E.C.'s unification. 
based glTHillllilr writing environment. 
some other nleaning aspect(s). 'file inlprecisio|l of I,l;s 
will nlean that we have no means of distinguishing be- 
tween the vmious intensifiers possible it| tile context 
of a given keyword, and hence will not have sufticient 
in forination to choose the most appropriate translation 
where, correspondingly, nntltiple possibilities exist in 
tile target language. All important question here is how 
dramatic this loss of translation quality really is. 
It is essentially ill addressing the issue of ovelgener- 
ality that Mel'~:uk introduces sub- and superscripts to 
lexical functions, enhancing their precision and mak- 
ing them sensitive to meaning aspects of tile lcxical 
items over which they operate. Superscripts are ill- 
tended to make the nleaning of tile I,F nlore precise 
and he|me |nero likely to imply unary inappings be- 
tween argu|nents and vahlcs, subscripts a|e used to ref- 
erence a particular semautic COlllpOUellt of a keyword. 
The introduclion of such devices into tile account of 
l,Fs demtmstrates hoth the need tk)r precision and the 
fact lbat it does seeul necessary to address semantic 
aspects of lexemes stand| ng it| co-occurrence relatio|ls. 
Ill fact it has been asserted by sonm (e.g., (Anick and 
Pustciovsky, 1990), (lteid and Raab, 1989)) that col- 
locational systems are systematically predictable from 
the lexical Selllantics Of nt)tUlS, it) till atteln\]Jt to explore 
this notion furthel; we have investigated the appr(lach to 
nolninal semantics known as Qualia structure (Puste- 
jovsky, 1991) and conside|ed how this lnay ct)tnple- 
u|ent the LF notion to inlprove its descriptive powe| r. 
alnoDg tile prolnising avenues that occur to tlS are, 
firstly, tile postulation of I,F subscripts based on the 
four Qualia roles (assuming thal these are tim lexically 
hies) relevant aspects of noun selnantics) and, secondly, 
the application of l,Fs to senlaulic (Qualia) structures 
rather titan monolithic lexenles; cg: tile I ,l; Ibm is used 
in delivering evahlative qualitiers which are standard 
expressions of praise or approval. One could ilnagine 
application of the ftmctio|| over the Coustitttlivc and 
Agentive tolcs of file noun lecture, to deliver: 
Flon(Const : lecture) =intormative 
Ben(Agent-. : lecture) = clear 
In both cases tile idea is that tile precision of tile 
lexical function is essentially enhanced by appealing to 
tile semantic facets of ils argunlcnt. 
Syntaetic Divergences Allother issue that has lo be 
raised conccl'lls tile trat|slation el'collocations into non- 
colh)cational constructions. It' we are to ulaintain a 
co||sistent interlingual approach to tile hanslation of 
these cases, we illUSt cXlelld our l,F-bascd approach 
accordingly. We consider o|/e case brielly. 
Cross-lingtlistic analysis reveals many cases where 
nonlinal-based collocational ctmstructs are real|sod as 
conlpot|nds in Gernmnic hulguages, e.g., hunch of keys 
sleutelbos. A possible account of such phenom- 
ena nlay be developed fronl the coucept (11: merged I,Fs 
(Mel'Suk and 7,olkovsky, \]970), Mcrged i,Fs are ill. 
tended It) be used ill cases where a value lexeme exists 
)'For i/ COlllllat'i,'-;on belw¢c/1 aspects of Ou.'l\[ia slrtlCllll'CS alld lcx 
ical ftmclions see (I lcylcn, to appear). 
which appears to effectively reduce ("|ncrge") an LF 
meani|lg and its specitied a|gu|nent to a single lcxi-. 
calised for|n, rather than projecting a syntagmatic unit. 
We could argue that ill cases of compound lbrtnation, 
exactly tile same process is to lie accounted tbr, siuce 
the Ctl|nl)ound embodies both the concept mediated by 
tile LF and its argument lexetne. We coukl therefore 
allow compounds to be delivered as values of merged 
I,F's, eg: //Mult(sleutel)= steutelbos. 
These observations are uscful in the MT COl|text 
if we assmne that we cau effect a |nappiug betweeu 
merged and unmerged lhSs and thcrefore capture tile 
correspondence between distinct structural realisations 
of tile same concept. One way to ennflate such a 
Inappi|~g |night be through the use of Mel'~:uk's lex- 
ical paraphrasing rules. For instance, one could 
conceive ol' a lexical paraphtasiug rule as follows~: 
W-}- Mult(W) e, >//Mull(W). 
If we assurne that ill our |uouolingual English lexi- 
cou, we assign tile collocate bunch as tile Mult value of 
keyword kt?.,, and that accordingly in tile Dutch iexical 
entry Ior sleutel we instant|ate sleutelbos as the vahle of 
tile nlerged 1,1;//Mult, then we can use the paraphras- 
lug rule to effect a nlaplfing between tile two 13;'s and 
hence arrive at an iuterlingual approach to tile traus- 
lation of tile example, despite structural |nisu\]atches, 
i.e., 
key + bunch\[ Mult(key)\] 
sleutel bos\[ llMuit( sleutel ) l 
l;u~lher examples exist where productive nlorpho- 
logical processes (e.g., affixation 'q) lead to tile lexicali 
sat|tin in one language of concepts that exist as syntag- 
nla|ic constructs ill another. Again, we suggest tile usc 
ot'|nerged l,Fs and corresptmding nlappings via lexical 
paraphrasing rule,; as a possible Iranshttion strategy in 
these cases, 
4 Summary and Conclusions 
hi ihis paper we have discussed how the lexicographi- 
cal concept of h,xical./iulctionx, introduced by Mel'~,uk 
h) describe collocations, can lie used as till intmlingual 
device in the machine mlnslation of such structures. 
We have shown how the essentials of the E(33 analy- 
sis can be embedded ill the lexicou and gramnlar of a 
unit|cation based theory of language. 
Our use of lexical functions as an intcrlingua as- 
sunles thai the relevant aspects of tile meaning of the 
colhleate are fully captured by the LK The 1,1: there.- 
fore determ i lies tl m accu racy of )ran s l at| on s, whi chm ay 
s This is our own |nit|alive it seems lo hc the case as we examine 
the literal urc lhat neither l,Fs such as Magll, BOll etc (i.c., those repro 
sent|rig slandard qualifiers/attributes) nor indeed metxed 13:s tt'aluic 
in lexical imralflnasing rides, Wc wouhl argue thai cross-linguistic 
analysis stiggt~Ms Ihal Ihcy should enter this donmin; COlnpotmd lot 
lnation alld other lypcs of leg|ell|sat|on appear Io bc ICe)liar pal)of rig 
of II'aUshllioll tR;ross lllally o:~llocatioll,'d COllStrHctS, as WE ilhlslralc 
here. 
9()no could Ihink of an example such as mis-ituerl)ret. 
1243 
be impoverished due to the generalised nature of ba- 
sic LFs. We have suggested some ways in which LFs 
can be enriched with lexical semantic intbrmation to 
improve translation quality. 
The interlingua level reflects what is semantically 
common to expressions which form translational equiv- 
alents. It abstracts away from specific syntactic re- 
alisations. Given that collocations may translate as 
non-collocations, we also have to provide a way to rep- 
resent these expressions using lexical functions. We 
have provided an illustration on how to proceed in one 
such case. 
Acknowledgements We would like to thank the fol- 
lowing partners and colleagues: Susan Armstrong- 
Warwick, Laura Bloksma, Nicoletta Calzolari, R. Lee 
Humphreys, Simon Murison-Bowie and Andr6 Schenk. 
References 
\[Abeill6 and Schabes, 1989\] A. Abeill6 and Y. Sch- 
abes. Parsing idioms in lexicalized tags. In EACL/89, 
Manchester, 1989. 
\[Anick and Pustejovsky, 1990\] E Anick and J. Puste- 
jovsky. An application of lexical semantics to 
knowledge acquisition from corpora. In Coling/90, 
Helsinki, 1990. 
\[Chouekaetal., 1983\] K. Choueka, S.T. Klein, and 
E. Neuwitz. Automatic retrieval of frequent id- 
iomatic and collocational expressions in a large cor- 
pus. ALLC Journal, pages 34-38, 1983. 
\[Church and Hanks, 1989\] K. W. Church and 
E Hanks. Word association norms, mutual infor- 
mation and lexicography. In ACL/89, Vancouver, 
1989. 
\[Heid and Raab, 1989\] U. Heid and S. Raab. Colloca- 
tions in multilingual generation. In EACL/89, pages 
130-136, Manchester, 1989. 
\[Heylen, 1993\] Dirk Heylen. Collocations and the lex- 
icalisation of semantic operations. Technical report, 
OTS, 1993. 
\[Heylen, to appear\] D. Heylen. Lexical functions and 
knowledge representation. In E Saint-Dizier and 
E. Viegas, editors, Computational Lexical Seman- 
tics. CUP, to appear. 
\[Krenn and Erbach, to appear\] B. Krenn and G. Et= 
bach. Idioms and support verb constructions. In 
J. Nerbonne, K. Netter, and C. Pollard, editors, Ger- 
matt Grammar in HPSG. CSLI Lecture Notes, to 
appear. 
\[McRoy, 1992\] S. W. McRoy. Using multiple knowl- 
edge sources for word sense discrimination. Com- 
putational Linguistics, 18( 1 ): 1-30, 1992. 
\[Mel'Suk and Zolkovsky, 1970\] I.A. Mel'~:uk and 
A.K. Zolkovsky. Sur la synthSse sdmantique. T.A. 
lnformations, 2:1-85, 1970. 
1244 
\[Mel'5uk and 2;olkovsky, 1984\] I.A. Mel'~uk and 
A.K. Zolkovsky. Explanatory Combinatorial Dic- 
tionary of Modern Russian. Wiener Slawistischer 
Almanach Sonderband 14, Vienna, 1984. 
\[Mel'6uk et al., 1984\] I. A. Mel'duk, N. Arbatchew- 
sky-Jumarie, L. Elnitsky, L. Iordanskaja, and 
A. Lessard. Dictionnaire explicatif et combina- 
toire du frangais contemporain. Les Presses de 
l'Universit6 de MontrEal, Montreal, 1984. 
\[Nireuburg et al., 1988\] S. Nirenburg, R. McCardell, 
E. Nyberg, S. Huffman, E. Kenschaft, and I. Niren- 
burg. Lexical realization in natural language gener- 
ation. In Second International Conference on Theo- 
retical and Methodological Issues in Machine Trans- 
lation of Natural Languages, Pittsburgh, 1988. 
\[Pollard and Sag, 1987\] C. Pollard and I. Sag. b~for- 
mation Based Syntax and Semantics. CSLI, Stan- 
ford, 1987. 
\[Pollard and Sag, to appear\] C. Pollard and I. Sag. 
Head driven phrase structure grammar, to appear. 
\[Pustejovsky etaL, 1992\] J. Pustejovsky, S. Bergler, 
and E Anick. Lexical semantic techniques for corpus 
analysis. Computational Linguistics, 19(2):331- 
358, 1992. 
\[Pustejovsky, 1991\] J. Pustejovsky. The generative 
lexicon. Computational Linguistics, 17(4), 199 l. 
\[Smadja and McKeown, 1990\] Frank Smadja and 
Kathleen R. McKeown. Automatically extracting 
and representing collocations for language genera- 
tion. In 28th Annual Meeting of the Association for 
ComputationalLinguistics, Pittsburgh, PA, 1990. 
\[Smadja, 1993\] F. Smadja. Retrieving collocations 
ti'om text: Xtract. Computational Linguistics, 
19(1): 143-177, 1993. 
