SYNTACTIC CHOICE IN LANGUAGE GENERATION 
r 
Lee Fedder, 
University of Cambridge, Computer Laboratory, Pembroke Street, 
Cambridge CB2 3QG, England. 
lf~uk.ac.cam.cl 
1 Abstract 
A considerable amount of study has recently 
been concentrated on the use of linguistically 
motivated unification based grammars for sen- 
tence generation (see for example Appelt 1987, 
Calder 1989, Shieber 1988, Shieber et al. 1989). 
The grammars state correspondences between 
semantic structures and syntactic ones. In 
most grammars, several syntactic structures 
will correspond to each semantic one. It was 
suggested at a fairly early stage that control 
over which sentence is generated could be ap- 
plied by adding "functional" features to the 
grammar. 
This idea has been extended in the work 
presented here. The use of feature threading 
techniques allows control over a wide range 
of syntactic structures, in a fairly sophisti- 
cated grammar, while avoiding the need for 
rule duplication even when generating un- 
bounded dependencies. However, we will see 
that the feature system required quickly be- 
comes complex, and: may be difficult to ex- 
tend to more comprehensive grammars. 
2 Introduction 
"A thematic system is one where 
corresponding members of the con- 
trasting terms normally have the 
same propositional meaning, and 
the same illocufionary potential." 
(Huddleston 1984:p437 ). 
Most phrase structure or categorial unifi- 
cation based grammars encode some form of 
thematic system. The simplest would involve 
the inclusion of both active and passive voice. 
Typically, the grammar defines the syntactic 
structure of each form, but does not include 
the pragmatic information that distinguishes 
one from another. When using such a gram- 
mar for parsing, this is not important, so 
long as the information is not required by the 
system using the parser's output. However, 
there has recently been an upsurge in the use 
of these grammars for generation. The lack 
of pragmatic data now becomes important: 
the generator is under-constrained, being ca- 
pable of producing any of the available the- 
matic forms. One way of applying the neces- 
sary constraints, is to introduce a system of 
"functional" features into the feature struc- 
ture of the grammar itself. These features are 
so called because they refer to the function 
of the various parts of the sentence in a dis- 
course. McKeown suggested the use of func- 
tional features for the TEXT system (McKe- 
own 1985) in which the grammar was based 
on the FUG formalism (Kay 1979). The func- 
tional features were defined as part of the ini- 
tial specification of the sentence, which was 
then filled out by traversing the grammar in 
a "Top Down" fashion. For example, the fol- 
lowing was given by McKeown as an initial 
sentence specification. 
PROT = 
VERB = 
GOAL = 
TOPIC = 
\[N === DOC\] 
\[V === BITE\] 
\[TENSE === PAST\] 
.\[ADJ === OLD\] 
\[N === MAN\] 
\[PROT\] 
45 
The functional feature is "TOPIC", and 
is specified as being the agent (or PROTago- 
nist) of the semantic structure. The feature 
value controls wether an active or passive sen- 
tence will be produced. 
The work reported in this paper extends 
this technique to a grammar which encodes a 
richer thematic system than just active and 
passive. We use a unification based gram- 
mar with a phrase structure backbone, which 
was originally developed to provide a simple 
computational description of current linguis- 
tic theories (mainly GPSG, Gazdar 1985). As 
in the example above, a system of functional 
features is introduced. A bottom-up genera- 
tion algorithm allows the production of sen- 
tences given an initial semantic form. The 
assignment of some initial values to the func- 
tional features constrains the structures gen- 
erated, and typically just one sentence will 
be generated for each semantic input. 
This work was done in the context of a 
database enquiery system with single sentence 
output. We assume there is a discourse man- 
ager which initiates generation by passing the 
generator a "message". This message con- 
sists of the propositional content of the out- 
put required, and some pragmatic informa- 
tion. 
The rest of this paper is in three main 
parts. The first is the definition of a co- 
herent set of discourse parameters that de- 
scribe the behaviour in discourse of the var- 
ious elements of a sentence. The second sec- 
tion describes the thematic system used, and 
how each member relates to the discourse pa- 
rameters. Finally, we see how the grammar 
can be augmented with functional features to 
provide filtering during generation consistent 
with the discourse parameters. 
3 Discourse Parameters 
'The members of the thematic system to be 
described below behave differently in discourse. 
In the linguistics literature, there is a long 
tradition, of assigning labels to various clause 
constituents in order to describe this behaviour. 
Labels such as "given" and "new", "topic" 
and "comment" ,"theme" and "rheme" and so 
46 
on (a summary can be found in Quirk 1985, 
18.9). We have adopted a set which allows a 
distinction between the members of the the- 
matic system we use. 
3.1 Speech Act Type 
This parameter conveys information about 
the sentence as a whole. Something similar 
is to be found in most grammars, but prece- 
dents in generation can be found in Appelt 
1985, and Bunt 1987. Values are :- 
. 
2. 
. 
Declarative - John gave a book to Mary 
yes-no question - Did John give a book 
to Mary 
wh-question - Which book did John give 
to Mary 
3.2 Theme 
The theme is :- 
"... somehow an element seman- 
tically crucial to the clause ... the 
communicative point of departure 
for the rest of the clause" - Quirk 
1985, 
In general, the theme is the established or 
given part of a message, and lays the ground 
for the rest of the communication. So, when 
it occurs in its expected or unmarked form, 
it will tend to be the first element of the sen- 
tence. 
3.3 Focus 
The label "focus" has been widely used in the 
linguistics and A.I. to name a whole range of 
concepts. We use the following definition :- 
"The focus ... indicates where the 
new information lies" - Quirk 1985. 
This definition is easy to assimilate in terms 
of a database enquiry system where the new 
data is easily identified. As to where the fo- 
cus occurs in the sentence 
i 
"The neutral position of focus is 
what we may call END-FOCUS, 
that is (generally speaking) chief 
prominence on the last open-class 
item or proper noun in the clause" 
- Quirk 1985 '. 
There may be several elements in the gen- 
erator's input whitch are given, and several 
which are new. For simplicity, we assume 
the discourse manager is able to specify one 
as the most thematic, and one as the most 
focussed. 
3.4 Emphasis 
The emphasis parameter indicates that some 
stress is to be laid on the indicated sentence 
element, above that supplied by an unmarked 
sentence, as when correcting a false presup- 
position. Emphasis is associated with par- 
ticular marked sentence constructions, as we 
will see below. Either the topic or the focus 
may be emphasised: other sentence elements 
may not. 
4 Discourse parameters and 
the thematic system 
We can now move on to see how the discourse 
parameters relate to the thematic system in 
the grammar. In general, guided by Quirk's 
definitions, we have~ adopted the simple rule 
that the theme is the first NP in the sentence, 
and the focus is the last. 
4.1 Active 
The active sentence is considered as "unmarked" 
form in which the parameters adopt their de- 
fault or neutral values. Thus the subject NP 
will be the theme, and the focus will be on the 
verb, direct object, indirect object, or verb 
modifier, whichever comes last. 
4. John slept in the garden. \[theme = 
John, focus = the garden\] 
4.2 Passive 
Creider (1979) classifies the passive as prin- 
cipally a topicalising structure, whilst Quirk 
47 
(1985) discusses the focussing effect. 
We have modeled these effects as follows. 
With transitive verbs, the subject is focused 
and the object becomes theme. If the subject 
is omitted, the verb itself can be focused, but 
in addition, this produces some emphasis. If 
the subject is not omitted, the verb can still 
be focussed and emphasised by fronting the 
object, which then becomes the theme (see 
fronting). Modifiers may take the emphasis. 
5. Mary was loved by Jim. \[theme = Mary, 
focus = Jim\] 
For bi-transitive verbs, the direct or indi- 
rect object can be thematised. 
6. Mary was sold a book by Jim. \[theme 
= Mary, focus = Jim\] 
4.3 The indirect object transforma- 
tion 
Creider (1979) classifies this transformation 
as having a thematising function. 
Q. What did you give to George? 
A. I gave George a pennywhistle. 
A. ?I gave a pennywhistle to George. 
This is modeled by transferring theme to 
the indirect object, and focus to the direct 
object. 
7. I gave George a pennywhistle. \[theme 
= George, focus = a pennywhistle\] 
The transformation can be combined with 
class II passivisation. The result is treated as 
a passive :- 
8. A book was given by John to Mary. 
\[theme = a book, focus = Mary\] 
4.4 Fronting 
This construction is generally accepted as es: 
tablishing the theme (see Creider 1979 - he 
calls theme "topic", and fronting "topicali- 
sation"). The fronted item is not new data, 
and seems to be associated with some form 
of contrast. This shows up in examples like 
9. John I like, but Mary I hate. 
This is modeled by assigning both the 
"theme" and "emphasis" parameters to the 
fronted item, the focus being at the end of 
the sentence as usual. 
10. To Mary John gave a book. \[theme 
= Mary, focus = a book, emphasis = 
Mary\] 
4.5 Clefts 
These constructions introduce the cleffed el- 
ement as new data, and apply special empha- 
sis, as when correcting a presupposition :- 
Q : Was it John who robbed the 
bank? 
A : No, it was Arther 
Usually, the other entities in the sentence 
are given, and uncontested. As we saw in the 
description of the grammar above, any NP or 
modifier in as sentence can be clefted. So, the 
clefted item is in focus, and the theme now 
moves to the end of the sentence. 
11. It was to Mary that John gave a book. 
\[theme = a book, focus = Mary, em- 
phasis = Mary\] 
4.6 Intonation 
The intonational centre is assumed to be at 
the end of the phrase, except in cleft forms, 
where it falls at the end of the first clause. 
If the theme or focus is realised as a relative 
clause, the intonational centre comes at the 
end of that clause. These are important as- 
sumptions since non-standard intonation can 
serve to shift the emphasis or focus to almost 
any part of a sentence. 
5 The Grammar Formalism 
The grammar is encoded in a framework built 
as part of the Alvey natural language tools 
project, and known as the GDE (Grammar 
Development Environment). The syntactic 
analyses are based on those developed by Pul- 
man 1987, with extensions to cover all the 
48 
thematic forms mentioned in the last section. 
They are couched within a simple unification- 
enriched phrase structure formalism. Seman- 
tic rules are associated with the syntactic rules 
on a rule-to-rule basis. The semantic rules 
are instructions for building logical forms of 
a typed higher order logic. The semantic 
translation of an expression is assembled us- 
ing function application and composition, and 
by using beta-reduction. The logical forms 
the rules built are a type of "intensionless 
Montague", similar to PTQ (Dowty 1981), 
but without the intension and extension op- 
erators. Here, we are only interested in the 
syntactic part of the rules, so the semantics 
can be omitted. The following rules couched 
in GDE notation will serve as an illustration 
R1. 
R2. 
S\[type decl\] ~ NP\[agr @x\] VP\[agr @x\] 
NP\[agr ~a\] 
Det\[agr @a\] Nbar\[agr @a\] 
R3. Nbar\[agr @x\] ~ N\[agr @x\] 
R4. VP\[agr @a\] ~ V\[agr @a, subcat np\] NP 
Here, the prefix "@" denotes a variable. 
NP's are type rMsed. Syntactic categories, 
subcategorisation, and unbounded dependen- 
cies, are treated similarly to GPSG (Gaz- 
dar 1985). Topicalisation, cleft forms, and 
relatives are all treated as problems of un- 
bounded dependency, using gap threading tech- 
niques. The tricky problems of passives and 
dative shift are covered by a version of the 
neat treatment presented in Pulman 1987. 
This involves the construction of passive and 
dative shifted versions of verbs, before inclu- 
sion in the rules which combine them with 
noun phrases, such as R4. No special struc- 
ture rules for passives are needed. 
6 The generation algorithm 
The current GDE generation system uses a 
chart based bottom-up grammar traversal al- 
gorithm, similar to that described in Shieber 
1988). 
The starting point for generation is a log- 
ical form involving symbols which represent 
entities in the discourse model of the applica- According to the description of the the- 
tion program. For example "LOVE(ENT1,ENT2)"matic system above, a plain active sentence 
The referring expressions for these entities 
are pre-generated and entered in the chart, 
along with all the lexical items compatible 
with the rest of the logical form. 
During generation, chart entries are re- 
peatedly combined into larger constituents 
via the grammar rules. A semantic filter blocks 
any constituents whose semantic formulae are 
incompatible withthe goal logical form. 
t 
7 How the:discourse param- 
eters are encoded in the 
grammar 
So, how can the discourse parameters be em- 
bodied in in the feature system of the gram- 
mar. 
The speech ac t type of the sentence is in- 
troduced at the sentence level using the fea- 
tures "sentence-type" and "wh". Assignments 
are as follows :- 
Declarative S\[type decl\] 
Question - S\[type quest, wh -\] 
WH-Question - S\[type quest, wh +\] 
The other parameters, theme, focus, and 
emphasis, are conflected with entities in the 
application program's discourse model. For 
generation, they are added to the initial chart 
entries for those entities. Assume, to be- 
gin with, that wehave a functional feature 
for each discourse parameter, "thin", "foc" 
and "emp", which take the values + or - 
as appropriate. Then, given the start logical 
form above, assume ENT1 is pre-generated 
as "John" and ENT2 as "Mary". From the 
discourse model, we discover that ENT1 is to 
be the theme, ENT2 the focus, and that nei- 
ther is to receive emphasis. This gives us an 
initial chart with the following entries for the 
referring expressions :- 
John:NP\[thm +,foc -,emp -\] 
Mary:NP\[tllm -,foc +,emp -\] 
would be suitable. 
12. John loves Mary 
We could constrain the generator to pro- 
duce just the active form by augmenting the 
grammar rules as follows (irrelevant features 
will be omitted from the rules; altered rules 
retain their original numbers, augmented with 
a,b,c ... and so on) :- 
Rla S ~ NP\[thm +,foc -,emp -\] VP. 
R4a VP =>V\[foc -,emp -\] 
NP\[thm -,foc +,emp -\]. 
Functional features on the verb will be in- 
cluded for completeness, but are not actually 
used in the current system. 
Here, the NP of R4a is assumed to be the 
last constituent in the sentence. Our treat- 
ment of passives means that these rules would 
generate passive sentences correctly as well, 
since there is no separate passive transfor- 
nlation rule. Rules for intransitive and bi- 
transitive verbs could be handled in the same 
way. However, the system breaks down when 
we introduce VP modifiers. Now, we no longer 
know which NP will be last until the VP 
has been incorporated into a sentence. This 
can be handled by making the focus value of 
the NP dependent on a similar feature in the 
mother VP, as follows :- 
Rlb S => 
NP\[thm +,foc -,emp -\] VP\[foc +1 
R4b VP\[foc @f\] 
V\[foc -\] NP\[thm -,foc @f, emp -\] 
R5 VP\[foc @fl ~ VP\[foc -1 
VPMOD\[thm -, foc @f, emp -\] 
This, however, only works if there are no 
gaps.If the NP of rule R4b were a gap, and 
there were no modifiers, the V would then 
carry the focus. This can be handled by thread- 
ing the focus feature through the NP. If the 
NP turns out to be a trace (that is, the cre- 
ation of a gap), the focus value is threaded 
through to the V, but if it is a real NP, it 
keeps the focus value for itself, and passes 
49 
the value "foc -" to the V. The "foe" feature R0a 
is now replaced by "fin" and "fout" features. 
This allows a gap in the VPMOD as well. If Rld 
there is a fronted NP, the theme shifts to it, 
from the subject NP. This can be accounted 
for by linking the value of "thin" to the sen- R2b 
tence. If a fronted element takes the theme, 
this is set to -, otherwise it is set to + . Be- 
low, the topicalisation rule assigns + to the R4d 
thm of the fronted NP, and - to the thm 
of the subsequent sentence. The thematised 
NP receives emphasis as well. Transitive or 
bitransitive verbs which ends up as the focus 
also receives emphasis. So, we also link the R5b 
emp value of such a verb to its "fout" value. 
K0 SFIN ~ S\[thm +\] ; Top level 
Rlc S\[thm @t\] ~ R6b 
NP\[thm @t,fin -,emp -\] VP\[fin +\] 
R2a NP\[thm @t, fin @fi, fout -, emp @e\] ~ R7b 
Det Nbar 
R4c VP\[fin @fi\] =~. V\[fin @fo,emp -\] R9 
NP\[thm -, fin @fi, fout @fo, emp -\] 
R5a VP\[fin Qfi\] ~ VP\[fin @fo\] VPMOD\[thm 
-,fin @fi, fout @fo,emp -\] 
R6 NP\[thm -, fin @f, fout @f, emp -\] 
\[\] ; trace NP 
R7 S =3- NP\[thm +,fin -,emp +\] S\[thm-\] 
; Topicalisation 
R8 VP\[fin @fi\] =~ V\[fin @fol,emp fol\] 
NP\[thm -, fin @fo, fout @fl, emp -\] 
PP\[thm -, fin @fi,fout @fo,emp -\] 
; Datives 
Now we need to deal with clefting. In this 
construction, the theme isshifted from the 
front of the sentence to the end, and the fo- 
cus shifts to the clefted element, which is also ROb 
emphasised. In response to this, we need to 
introduce a "shifted theme" feature, "sthm", Rle 
and link the fin feature up to the sentence 
category. Once shifted, the theme needs to 
be treated just like the focus - landing at the 
end of the sentence. That means it needs R4e 
threading, and we replace thm with the fea- 
tures "tin" and "tout". Treatment of clefting, 
then, causes the following alterations :- 
50 
SFIN ~ S\[tin +,fin +\] 
S\[tin @t,sthm @s,fin @f\] 
NP\[tin @t,fin -,emp-\] VP\[tin @s,fin 
@f\] 
NP\[tout -, fout -, emp @e\] :=~ Det 
Nbar 
VP\[tin @ti,fin @fi\] 
V\[tin @to,fin @fo,emp -\] 
NP\[tin @ti, tout @to, fin @fi, fout @fo, 
emp -\] 
VP\[tin @ti,fin @fi\] =~ 
VP\[tin @to,fin @fo\] 
VPMOD\[tin @to,tout @to,fin @fi,fout 
@fo,emp -\] 
NP\[tin @t, tout @t, fin @f, fout @f, emp 
-\] =v \[\] ; trace NP 
S ~ NP\[tin +,fin -,emp +\] 
S\[tin -,sthm -,fin +\] 
S ~ Pro Aux NP\[tin -,fin +,emp +\] 
S\[type tel,tin -,sthm +,fin -\] 
; Clefting rule 
Finally, for dative movement, focus stays 
at the end of the sentence, (unless a cleft from 
is used) but the theme moves to the indirect 
object. This can happen if the theme has al- 
ready been shifted by a cleft, or if it hasn't. 
This is treated by introducing one final fea- 
ture "normal shifted theme" or "nst". This 
feature is set to - if there is a dative shift, 
and + otherwise. Then, wherever tin used 
to be set to +, it is now takes its value from 
the nst feature. The exception is topicalisa- 
tion, when dative movement is prevented by 
setting nst to -. The rules changes that im- 
plement this are as follows :- 
SFIN ~ S\[tin @d,nst @d,fin +\] 
S\[tin @t,sthm @s,nst 
NP\[tin @t,fin -,emp 
VP\[tin @s,nst @d,fin 
@d,fin @f\] -\] 
@f\] 
VP\[tin @ti,nst +,fin @fi\] 
V\[tin @to,fin @fo,emp -\] 
NP\[tin @ti, tout @to, fin @fi, fout @fo, 
emp -\] 
R5c 
R7b 
R8a 
R9a 
RiO 
VP\[tin @ti,nst @d,fin @fi\] =~ 
VP\[tin @to,nst @d,fin @fo\] 
VPMOD\[tin @to,tout @to,fin @fi,fout 
~fo,emp -\] 
S ~ NP\[tin ++,fin -,emp +\] 
S\[tin -,sthm -,nst +,fin +\] 
VP\[tin @ti,nst +,fin @fi\]~ 
V\[tin @tol,fin @fol,emp fol\] 
NP\[tin ~to, tout ~tol,fin @fo, lout ~fl, 
emp -\] 
PP\[tin @ti, tout @to, fin @fi,fout @fo,emp 
-\] ; threading as normal 
S ~ Pro AuX NP\[tin -,fin +,emp +\] 
S\[type rel,tin:-,sthm @d,nst @d,fin -\] 
; Clefting rule 
VP\[tin +,tout -,nst -,fin ~fi\] 
V\[tin -,fin @fol,emp fol\] 
NP\[tin +,fin @fo, lout @fl, emp -\] 
NP\[tin-,fin @fi,fout @fo,emp ,\] ; da- 
tive movement 
7.1 Initial feature values 
An NP now carries five functional features, as 
opposed to the three we assumed at the start. 
They are initially set as follows. If the entity 
is theme, we have \[tin +,tout -\]. If the entity 
is focus, we have fin +, four -\]. Otherwise, 
theme and focus values are threaded, as in 
\[tin @t, tout @t, fin:@f, lout @f\]. 
8 A Simple  Example 
Let the message be :- 
LF- LOVE(ENT1,ENT2) 
Speech-Act-Type - Declarative 
Theme- ENT1 
Focus - ENT2 
ENT1 is pre-generated as the NP "John" 
and ENT2 as "Mary!', and this gives the fol- 
lowing initial chart entries :- 
C1. John: 
NP\[tin +, tout -,fin @f, lout @f, emp -\] 
53. 
C2. Mary: 
NP\[tin @t, tout ~t,fin +, fout -, emp 
-\] 
C3. loves:V 
C3 represents the entries in the lexicon 
which are compatible with the initial logical 
form. 
From this position, C2 and C3 can be 
combined via rule R4e to give the new chart 
entry :- 
C4 loves Mary:VP\[tin @t, nst +, fin +\] 
Then, C1 and C4 can be combined via 
rule le to give :- 
C5 John loves Mary:S\[type decl, tin +, nst 
+, fin +\] 
Other sentence forms are blocked by the 
functional features. If the NP "Mary" were 
originally assigned "emp +", the generation 
would only be able to succeed by using the 
cleft form "It was Mary who was loved by 
John". If "John" were emphasised, genera- 
tion would fail: the current system has no 
way of emphasising a thematised agent. It 
would be necessary to use a different verb, or 
use prosodic stress. Neither of these methods 
is available in the current system. 
9 Discussion 
The functional feature system is clearly be- 
coming rather complex, a problem which will 
only increase with the inclusion of more elab- 
orate thematic forms. Further research would 
show if this becomes completely unmanagable. 
A possible solution to this problem is dis- 
cussed in McKeown 1987 in which pragmatic 
constraints are added to a FUG grammer. 
They show how the constraints can be stated 
in one place, rather than duplicated through- 
out the feature system. 
Certain combinations of initial feature as- 
signments cause failure to generate. In these 
cases, some form of constraints relaxation will 
be necessary. 
10 Conclusions 
The addition of "functional" features to a 
unification grammar is used to provide con- 
trol of syntactic variants during generation. 
The use of threading avoids the need for du- 
plication of rules, whilst allowing a fairly wide 
range of thematic variants. However, the fea- 
ture system required quickly becomes com- 
plex. 
11 Acknowledgements 
This work was made possible by funding from 
the Science and Engineering Research Coun- 
cil, and Logica U.K. I would like to thank 
Marianne McKormick and Steve Pulman for 
the insights underlying this work, and John 
Levine for much discussion and collaboration. 

References 
\[1\] Appelt, Douglas E. 1987. Bidirectional 
grammars. TINLAP-3, position pa- 
pers. New Mexico State University. 
University Press, Cambridge, England. 
\[2\] Bunt, H. 1987. Utterance Generation 
from semantic representations aug- 
mented with pragmatic information. In 
Natural Language Generation by Kem- 
pen (Ed.). Martinus Nijhoff. 
\[3\] Calder et al. Unification Categorial 
Grammar. The Centre for Cognitive 
Science, University of Edinburgh. 
\[4\] Creider, C. 1979. On the explanation 
of transformations. Syntax and Seman- 
tics, Vol 12. By Talmy Givon (Ed.). 
Academic Press NY. 
\[5\] Dowty, D, R Wall, S Peters. 1981. 
Introduction to Montague Semantics. 
Reidel, Dordrecht. 
\[6\] Gazdar G, E. Klein, K. Pullum, I. 
Sag. 1985. Generalized Phrase Struc- 
ture Grammar. Basil Blackwell. 
\[7\] Huddleston 1984. Introduction to the 
grammar of English. Cambridge Uni- 
versity Press. 
\[8\] Kay, Martin. 1979. Functional Gram- 
mar. Proceedings of the Annual Meet- 
ing of the Berkley Linguistics Society. 
\[9\] McKeown K.1985. Text Generation. 
Cambridge University Press. 
\[10\] McKeown K, and Cecile Paris. 1987. 
Functional Unification Grammar Re- 
visited. ACL Proceedings. 
\[11\] Pulman, S. 1987. Passives. The pro- 
ceedings of the European ACL - Copen- 
hagen. 
\[12\] Quirk, R., Greenbaum, S., Leech, G. 
and Svartvik, J. (1985) A Comprehen- 
sive Grammar of the English Language. 
Longman. 
\[13\] Shieber, Stuart M. 1988. A uniform ar- 
chitecture for parsing and generation. 
Coling proceedings. 
\[14\] Shieber et al. 1989.A semantic head 
driven generation algorithm. ACL Pro- 
ceedings - Vancouver. 
