ON LEXICALLY BIASED DISCOURSE ORGANIZATION 
IN TEXT GENERATION 
Leo Wanner 
Ulfiversity of ,qtuttgart 
\](eplerstr. 17, 70174 Stnl;tgart, (~erlllany 
Summary 
In this paper, we show that Reileralion and Collo- 
cation relations as introduced I)y Ilalliday and llasan 
may function as lexieally I)iased discourse structure 
relations and that these relations are well represented 
by sequences of Mel'&flds Lexical Funclions (Ll,'s). We 
propose to use Lie sequences for tl,e final determina- 
tion and realization of discourse organization during 
lexical choice in text generation. 
1 LEXICAL PHENOMENA IN DIS- 
COURSE 
1.1 The Problem 
In text generation, the tmsk of content selection and 
discourse organization, i.e. text phuming, has often 
been opposed to the task of linguistic realization ()1" the 
information selected and organized by the text plan- 
ning process (el., e.g., MeKeown and Swartout, 1987). 
However, discourse organization is not possibh', with- 
out taking into account linguistic means that are avail- 
able to express a particular meaning (el., e.g., Meteer, 
1992; Rubinoff, 1992). Esl)e.eially the \['allure to in- 
tegrate lexical choice into the phuming process may 
lead to monotonous, awkward, or even ungrammati- 
cal text (note that when used separately, the clauses 
in (Ix) and (2a) are fully acceptable):' 
(I) a. ? Alle bewahrten Ruhe; nut llans holt- 
rite keine lbtlw/(sie nieht) \]~ew~bE,Lt! 
lit. 'All kept cabnness; only l\[ans could 
not keep c,~hnness/ it'. 
VS. 
b. Alle bewahrten Ruhe; nnr llans konnte 
nicht rnhiq blciben 
lit. 'All kept calmness; only I lans eouhl 
not kee l) cahn'. 
(2) a. ? l)er langgephmte AusJlu9 fired a,n 
Sonntag start; wir unternahmen ihn 
mit der 9anze;, 1'braille 
lit. 'The long-pl~tnned trip took platte 
on Sunday; we undertook it with the 
entire family'. 
VS. 
b. Der hmggephmte Attsfln 9 fand ant 
Sonntag stalt; die (mnze l'hrnilie 
nahm daran teil 
lit. 'The iong-phuumd trip took pl~tcc 
on Sunday; the entire family took part 
in it'. 
1 In the following exatnples~ the inal)prol)viate lexieal expres- 
sion in the (a) sentences and its IllOl'e alq)rol)riatc alternat, ive in 
the (b) sentences are underlined. 
(3) a, *lhtns muel, te clue Entdeckun9; diese 
Entdccktmg war wirklich 
lit. 'll;uts m~ule ~ discovery; this dis- 
covery was real'. 
VS. 
b. llans maehte eine lrnldeckun.q; diese 
lCntdeekun9 war eine I')ntdeekung i m 
wahrslen £'inne des I'Vorles 
lit,. 'Ilans made ~ discovery; this dis- 
covery was a discovery ill the real SellSe of the word'. 
These examples show thai, lexicaI constraints are of 
a sl)e('ial relewmee to discourse org;mization if related 
discourse segments conmmnicate illfOrlll&tiOll on lilt(! 
same or related object, event, process, etc. While in 
the past, considerable work has been done on the real- 
ization of anaphoric links between related entities via 
referring expressions (ef., e.g., Tutin and Kittredge, 
1092; Dale, 1980; Relier, 1091), only a few pro\[)osals 
eml)hasize the relewmee of lexical means for the real- 
ization o1" discourse struel:nre relations such ;m CON- 
TItAST in (11)) and ELABORA'PION ill (2t)) and (3b), 2 
It is iml>ori.ant to note that the actual realization of 
a discourse relation mw vary with the semantics of 
the lexemes involved. For examl)le , ill (4), the second 
ebmse is an IN'I'I,HU'II.I,~TATION or CONSI,IQUENCE of the 
lit'st; despite the analogous syntactic construction in 
(5), tim second clause is a JUSTIFI(:ATION or an EXPLA- 
NATION of the first rather than an tNTI,',It.I'RETATION 
or CONSEQU\[,INCI ,l . 
(4) l/e t,'avels ,, lot he is a '1,,'ofeasional' 
traveller. 
(5) Ile jlics a lot - he is a professional Jlier. 
11.2 The Proposal 
.quch relations as those between Ruhe bcwahren '\[to\] 
kee l) eahnness' and ruhi9 bleiben '\[to\] keep cMm' (ill 1); 
between Ausfl'ng finde! slall 'trip t~tkcs place' and am 
Ausflug leilnehmen '\[to\] take part in the trip' (in 2); and 
bet, ween eine Enldeckung 'discovery' and im wahrslen 
Sinne des Worles % the real sense of the word' (in 3) 
have been introduced by (l\[alliday and Ilasan, 1976) 
as Reiteration and Collocalion relations, a 
21n this paper, we use the \]llllll{!S of diSCollrse S~,l'lletitlr¢! tel;t- 
lions as they are known from the Rhetorical Sh'uclttre 7'heovy 
(Mann and Thompson, 1987'). 
3All.hough preferably used so flu' to describe dlscom'se links 
hetween infiwmatiol~ segments realized by norms, Reiteration 
and Collocation relations may well hohl between segments which 
m'e realized by other parts of speech and evell })y multiple word 
12xl)ressi()lIS, 
,?69 
Reiteration stands for a strict repetition on a lex- 
ieal expression in related discourse segments; for a 
substitution of a lexical expression by a synonym, or 
for a substitution by a snperordinate. Consider the 
following examples, which illustrate the three differ- 
ent reiteration relations (strict repetition in (7a), 
synonymy in (7b), and superordination in (7c)): 
(6) Last summer, Monica flew to Italy, 
(7) a. while Daniela flew to Norway. 
b. while Daniela took the plane to Nor- 
way. 
c. while Daniela travelled to Norway. 
Further substitution relations such as metaphoric 
repetition (7d), negated az~tonyray (7e), etc. can 
be added: 
(7) d. while Daniela wafted away to Norway. 
e. Daniela also did not stay at home. 
Collocation stands for 'any recognizable lezicose- 
mastic relation' between lexicM expressions in related 
discourse segments. Examples of collocation relations 
are attribution (7f), partition (7g), means (7h), 
etc.: 
(7) f. it was a very pleasant journey. 
g. making a stop over in Munich. 
It. it was one of those ~ifl aircrafts. 
As our examples show, reiteration and collocation 
relations help to ensure not only cohesion, but also 
coherence in texts. Therefore, a text generator has to 
provide an organization of lexical resources that tailors 
discourse structure relations to reiteration and collo- 
cation relations. This presupposes, on the one hand, 
a precise picture of which reiteration and collocation 
relations are available in language and how they are 
realizable; and, on the other hand, a fine-grained dis- 
course model that contains these relations. 
To make allowance for the global discourse orga- 
nization, which is performed independently from lex- 
ical resources, we suggest a two level text planning 
task implementation, with the first level realized by a 
Rhetorical Struclure 7'heory (its'r) (M ann and Thomp- 
son, 1987) style text planner and the second level -- 
by a separate lexical choice module. Then, the dis- 
course organization of a text is done in two steps: 
in the first step, the text planner predetermines the 
discourse structure relations; in the second step, the 
lexical choice module provides, in accordance with lin- 
guistic constraints, the final determination and the re- 
alization of these discourse structure relations. 
The present paper reports an attempt to define lex- 
ieally biased discourse structure relations used in a 
partially implemented lexical choice module. I)ue to 
the lack of space, we do not discuss the module it- 
self; it is described in detail in (Wanner, 1992, 1994). 
Here, we demonstrate how discourse organization for 
text generation can be refined by lexically biased dis- 
course structure relations and how these relations are 
related to global discourse relations specified in the 
output of an RST style text planner. 
In contrast to the most discourse models (cf., e.g., 
McKeown, 1985; Ch'osz and Sidner, 1986; Mann and 
Thompson, 1987), which take the clause ~ts the min- 
imal discourse segment, we consider as discourse seg- 
ments "perspectives" (cf. McCoy, 1989)- specific 
views taken on a semantic entity (an object, an ew~'nt, 
etc.). A perspective is a wording which is tailored to 
the lexical repertoire of an entity; it is realizable as a 
clause, a phrase, or as a single lexeme. Each of the 
clauses in the examples above can be considered as a 
realization of a single perspective; ~md the reiteration 
and collocation relations that hold between the clauses 
as well-defined perspective pairs. 
In our model, a single perspective is represented as a 
composition of Mel'&fl~'s Lexical l,~unctions (hereafter 
LFs) (Mel'&fl{ and Polgu~re, 11987); perspective pairs 
are represented as LI~' sequences. 
The lbllowing distinctive features characterize our 
model: 
• it makes snre that all relations defined ~tre ex- 
pressible in language, 
• it allows for a realization of lexicM relations as 
subclansal relations between discourse segments, 
• it is sensitive to lexical and syntactic wtriations \['or 
the realization of discourse structure relations. 
2 LEXICAL FUNCTIONS IN DIS- 
COURSE 
2.1 The Basics 
Formally speaking, an L~ f is a standard semantico- 
lexical relation which holds between a lexeme 1,1 (the 
keyword of f ) and a set of lexemes f(t,) (the wdue of 
f). Examples on LFs are: 
~yll: ~synollynl ~ 
Anti: 'antonym' 
Gener: 'hyperonym' 
Figm': 'metaph. rep.' 
C(}IIV2I ', tcotlvel.SiOll~ 
So: 'situation' 
A0: 'situational ~tdj.' 
Vo: 'action' 
Magn: 'intens@y)' 
Operl : 'perform' 
Incep: 'beginning' 
Fin: %rid' 
anus: ~C~llsa tioll ~ 
Manlf: 'manifest~ttion' 
Syn(bible) = God's Book 
Anti(victory) = defeat 
Gener(lamb) = meat 
Figur(fog) = wall \[of f<,9\] 
Cony71 (to include) 
: l/o\] belonfl 
So(to teach) : teachin9 
$1 (lie) = liar 
Ao(sun) = solar 
V0(aeal) = \[to\] ,leal 
Magn(beautg) = real, 
stunnin 9 
Operl (cry) 
= \[to\] let ,,,,t {a ,:,'y) 
Ineep(to sleep) 
= \[to\] fall asleep 
Vln(to sleep) = \[t,,\] ,,,ak~ ,,i, 
Cans(to sleep) 
= \[to\] put to sleep 
Manif( happy) 
: \[to\] beam with joy 
Mel'&lk distinguishes about sixty simple t,l,'s on the 
above Idnd. Simple LI,'s can further be combined with 
370 
Opert A ()l+erl 
Vo A SynV0 
Vo A GenerS0 
So A FigurSo 
V0 A No'r AntiVo 
Vo A Conv2tVo 
V0 AS: 
(\]aus A Vo 
\[ncep A Vo 
Ao A Magn oSo 
V0 A Manif 
strict repetition (l/o\] ha,w 
a look A \[lot have a look) 
synonymy 
(\[t,,\] di.wJp,~o," A \["4 v,,r,m,) 
superordination 
(l/o\] searclt (a Jtal) A reprisal) 
metaph, rel)etition 
(\]o:1A ,..H '4 \]+::J) 
neg. atttonymy 
(close A not far away) 
conversion (I/o\] sell A \[lot I, uy) 
process-actor (\[to\] lie A li.,') 
cause-pFocess 
(\[lot v,,l lo .~l,..j, A\[M .+t,,.p) 
init ializat ion-process (\[to\] 
fldl ash:ep A \[to\] .~leep 
attribution ( beaulifld A 
real beattlg) 
manifestation (\[lot bc happy A 
\[1o\] be,,m ,,:ill, jog) 
'Pabh'. 1: q'he realization o1" reil.eration and collocation 
relations by LI,' sequences 
each other; l.he meaning of such comphs: LFS is, as 
a rule, a eombitmtion of the meanings of tim partic- 
ipating l.Fs. q'hus, Ant.iMagn lneans 'slightly' (e.g., 
AntiMagn(i,0ury ) = minor); and Incel)()lmrl 'start 
performing' (e.g., htcepOpert( delmle) = \[lot shirt (a 
debate)). 4 
In text generation, the bettellts from l.t.'s are I, hree- 
fold: (i) they provide subclausal collocational con- 
straints l)etween the keywor<ls aiid the wdues (<:f. h:r: 
danskaja el al., 199 l) as, e.g., between narrow and ma- 
jo,'ity (with AntiMagn(majoritv) = narrot,) i,, Ilo,,- 
dini won with a narrow majority; (ii) they provide in- 
terclausal cooccurrenee links (el. 'l'utin and Kittredge, 
1992) between the keywords and ,,he wdues as, e.g., 
between spaghetti and pasla (with Gener(spaghetli) = 
pasla) in Let's take spaghetti; pasta is not bad he.re; and 
(iii) they allow \['or explMt statement:s on tit('+ coo('eur- 
fence between values orw~rious LFs ill relal;ed discourse 
segments, as, e.g., between the values of Vil(sler'p) - 
\[to\] sleep at,d h, cep(steep) -- \[to\] si,k i,lo sl,,ep i,t 
llardly in bed, 7bny sank into sleep and slept all the 
night till the morning. 
In our work, we use (i) for single perslmet.ive reatiza- 
l.ions (ef. Wanner and Bateman, 1990); (ii)and (iii) 
serve for the representation of perspective sequences, 
i.e. reiteration and collocation relations in discourse. 
One such relation is giwm by all pairs Lt.'t A 1,1:2 ('I.F 
sequences') which show the same eooecurrenee behav- 
ior (e.g., the sequences ()pert A Operl and Vib A V0 
show l, he same eooeeurrenee behavior; both stand for 
strict repetition). Consider Tabh'. 1. 
LF sequences are dlre.cted, i.e. l,Ft A I,1,'2 ~ I,F2 A 
4 \[f sevet'g:tl (silnp\[e 07" conll)lex) l,Fs COltll)OS;l! a phrase ()r ;t 
clause (as, e.g., AntiMagn itlld SO COlllpt)Se ?7t{'I~OT {Tlju~'y), W(! 
separate these LF8 by a 'o' sign, For the theoretical background 
a.Itd furl.her deL,~tils of how IA,'s can |:,e ct-~lnposed wil.h each other, 
see the literature on Mea'ni'a 9 "l'e:ct 7'heory; e.g., (Mel'(mk and 
F'olgub..re, L987). 
1,1,' 1 . Moreover, Lhe I~xig(,(?llCe O~ I,F l A I,F 2 ill a lan- 
guage does not mean that 1,1"2 A I.F I ix also awdlabhL 
'\['hererore, in LF sequences, one argument, is t, he 'hu}>' 
(.he point o\[" dep+u'ture (or l, he expanded LI,') and the 
other at'gtllllelaL ix Lho 'hul) expattder'. Ilow a specilic 
LF e.an |)e cxpallde(I, i.e. which I,F Se(ltlellces a.re possi- 
ble, depends individually on this 1,1.', and on whMl Ll:s 
are fm'ther awdlM>le for the enl:it.y the LFs are apl>lied 
I,O. (~Ollll)al'e , e,g., L\]le I,F seqtletlces t;llpJ, instant,ial;e 
the negated antonymy reiteration for Vo(forgelling) 
and l,he l,wo, whMt insl,a.nt, iat,e t.he same relal, i,:nt for 
Vo(lie): 
Forgetting (the Vu clause is in all examl>h's realized 
ns l forgol; to ahbrm,,ial.e, we write '.+.' instead): 
V 0 A N(YI' Q',onv2lAlltiV 0 .,. ; it, does not 
",'emi'nd me of anylhi'n!l. 
V 0 A (Magn o) NOT A,giVo ... ; can (absohttel9) 
"not think of it (now). 
Vo A (Magn o) I"inOp,,rl o Am, iSo ... ; it (totally) 
has slip?ed my 'mind. 
Vo A (Magn o) NOT AnliSyn~o ... ; (absohtlclp) no idea. 
Lh': ('...' stands here rot He is lying): 
V 0 A NOT ()perl o SoAnl,IV0 ... ; (simply) does not 
tell lh~ truth. 
Vii A (hmerV0 o NOT AntiA0 ... ; 
what he says is not h'Pte. 
Apart fi'om the reiteration or collocal,km relation 
it st.ands for, an I.F s,~(lllellce is rurther characl;erized 
by its possible syntactic roalizal, ions and its functlonM 
COIll,elIL 
2.2 Syntactic Realizations of I,F Se- 
quences 
As a ruhL an LI,' sequence is reMizahle by severM ¢li\[l~r- 
ent syntactic constructions. Ilow these eonst.rucl.ions 
can look like ix predel, crmined by each IA" sequence 
individually (and by the information to be comnumi- 
cared). I,'or example, Opert A Opert (more precisely, 
strict repetition) is in general realizable only as a 
parataclic eomplca: clause; of. Ilave a look at it; please 
have a look. In cont.rast., \['or examph!, ()perl A Magn 
o Su ix realizable when applied to, e.g., decision 
by nil synt.act.ic const.rucl.io,ls possibh'., of.: 
(8) a. ,Iohn made a decisim*; this deci,qiou 
was important to him (parata(:t.ic cf,,l,- 
plex clause); 
b. John made a decision, which was im- 
portanl to him (hypota.ctic complex 
obtuse); 
c. The decision, which John made, was 
important to him (embedded clause); 
d. Join+ made an important decisim~ 
(siml)le cla.use); 
e. John's recently made importanl dcci- 
,don (phrase). 
The reh!wmee of syntactic varial,ions for the realiza,- 
lion of discourse sl.ructure relations is well known, ef., 
e..g., (l\[ovy, 1!)93). 
371 
2.3 Functional Content of LF Se- 
quences 
Semantics, lexis, and syntax of LF sequences do not 
provide sufficient criteria for the choice of one sequence 
over all other comparable ones. These criteria nmst be 
provided by tile functional con lent we associate with 
each sequence (or reiteration and collocation relation, 
respectively). The flmctional content of tile reiteration 
and collocation relations listed in Table 1 is presented 
in Table 2. '~ 
strict repetition issisting~ restatement 
synonymy clarifying restatement 
superordination generalizing restatement, 
clarifying restatement, 
class-referencing 
metaphor, illustrative restatement, 
repetition pictoresque restatement, 
intensifying restatement 
negated antonymy contrastiw~ restatement 
conversion clarifying restatement, 
constituent enhancement, 
pespective shirting 
process-actor iden tiiication, 
actor-lntroduction 
cause-process i)rocessu al ellhallCeulellt~ 
causal enhancement, 
causer introduction 
initiMiz~ttion- process processual extension, 
beginning extension 
attribution attributive refinenmnt 
manifestation predicative relinement 
manifestatlon enhalleelnent 
Table 2: Functional content of some reiteration and 
collocation relations 
3 TOWARDS LEXICALLY BIASED 
DISCOURSE RELATIONS 
Due to their functional content, El,' sequences serve. 
~s instantiations of individual discourse structure re- 
lations. In our work, we suggest that these individual 
discourse structure relations can be organized cohe,'- 
ently in terms of the fimctions anti semantic distinc- 
tions they represent. In accordance with the claim 
that the availability of specific LF sequences is depen- 
dent on the entities the LFs are applied to, we furthex 
suggest that this organization must be done individu- 
ally for each predicative entity (el. Wanner, 1994). 
Based on this, we define taxonomies (one for 
each predicative entity) which have been inspired by 
tlalliday's proposal for grouping interclausal logico- 
semantic relations (el. Ilalliday, 1985). l\[ow such an 
organization carl be realized efficiently using inheri- 
tance teclmiques is described in (Wanner, 1992). 
SThls is not to say that these functions are the only ones 
that are possible 
Although our model is not restricted to interclausal 
relations, two features o\[' lIalliday's proposal are wtlu- 
able to us: (i) that a Iogico-semantic relation 'expands' 
one wording by an another one rather than connecting 
two given wordings and (ii) that a logico-semantic re- 
lation can be further decomposed with respect, to its: 
1. semantics, 2. syntactic realization, 3. communica- 
tive structure, and 4. with respect to the speaker's 
intention, which motiw~tes the selection of this rela- 
tion during the text production process. 
In what follows, we discuss first the general taxon- 
omy of our 'expanding' discourse structure relations 
for processes and then the decomposition of the reh> 
tions along these four dimensions, l?ollowing the con- 
ventions in R.ST, we call the expanded part 'nucleus' 
and the expanding o,e 'satellite'. 
3.1 Taxonomy of Lexlcal Discourse Re- 
lations 
A taxonomy of lexical discourse, structure relations is 
to be understood as a hierarchy of alternative choices 
of increasingly delicate relations. The most delicate 
relations are I,F sequences represented by their func- 
tional content. The top level of the. taxonomy repre- 
sents, thus, the most global types of exI)ansion. In 
accordance with (llalliday, 1985), these are F, Lalm- 
RATION, MXTENSION, and I,\]NIIANCEMENT. ELAIt()I1.A- 
TION subsumes all those, expansions which ensure a 
deeper understanding of t,he meaning comnmnicated 
by the nucleus wording. A deeper understanding or 
the nucleus wording is ensured by reslaling, refining, 
or clarifying it (the next level of ELABORATION in the 
taxonomy). For example, all reiteration relations are 
of the I,;LABORATION type. 
The EXTENSION expansions extend the meaning 
comnmnicated by tile nucleus wording. This can be 
(lone by introducing a new constituent that is related 
to what has linen said ill the nucleus, by adding a 
new action of I.he known constitueuts, etc. Beginning 
extension is, e.g., an EXTENSION. 
The ENIIANCIi:MI,'NT expansions qualify the mean- 
lug conmmnieal.ed I)y tim nucleus wording by adding 
a reference o\[" causation, time., location, l;qanne.r, 
mode, etc. An example of F, NIIANCEMENT iv causal 
enhancement. 
figure 1 shows ill lllOl'e detail tile ELABORATION 
fragment of the taxonomy in network tbrm. Ac- 
cording to this figure, ILESTA'PI,3MENT can be real- 
ized as a conlraslive, a generalized, or as a repeal- 
ing reslalemenl, respectively. As shown ill Table 2, 
CONTII.ASTIVE RESTATEMENT corresponds to tile reit- 
eration negated antonymy, GENEItAL1ZING II.ESTATE- 
MENT to superordination, respectively. ItEPl,',ATING 
RESTATI,;MENT is further insisting, clarifying, illustra- 
live, pictoresque, etc. (see again Table 2 for corre- 
sponding reiteration relations). 
372 
• Rlil;INEMEIql .... 
TOP 
ELABORA'rlON - 
EXTENSION '" 
• CI,ARII~ICATISN ,,, 
UONSTRAST. .., 
RESTATIibltiNT 
GENERAl ..... 
\[\[~S'I'ATliM I~Nrl ' 
- RESTA'IliMENT 
" PI(YI'ORESOUII. 
RtiS I'A'I. 
• IN~;IST 
S'lRlCr RIiS fAT,. 
RiiS I'ATEMI~I' 
('I,ARIIL 
R E"; I'AT.. 
. I1 LOSTRA F, 
I(I!STA'I'. 
IN'lENS, ., 
RI{S'rATIiMENT 
- gNI IANCI~MEN'I .... 
Figure 1: A fragment of a lexieal discourse sl.rtlcture 
relation taxonomy 
3.2 Decomposition of Lexical Dis- 
course Structure Relations 
As presented in Figure 1, the relations are still too 
global to be useful for lexical choice. Consider, e.g. 
ATTI:tlBUT1ON - it subtype of the R1,;HNI,',MF, NT reliL- 
lion; it allows for various decompositions with resi)ect 
to all h)ur dimensions mentioned al)(')ve: 
• S(,.manLi('.s; thus, A'PTI\[If~U'PION call llle~%ll A'I'- 
TItlBUTION, e.g., of a process, of one of the par- 
ticipants of this process, or of one of the cir- 
cumstances of this process; if ATTIUBUTION of ;t 
participant (let's slty tim ACTOR) is llleltllt, it is 
still undetermined what lind of attribution this 
is (e.g., a one which enables the actor to engage 
in the process, a one which l)rew~nts him front 
engaging in this process, etc.). 
• Syntactic realization; how the wu'ious ATTILI- 
BUTIONs Call be realized syntactically del)en(Is on 
the senlantic and \]exical properties of the infer 
marion to be eommunical, ed. For exanll~h!, 111o71- 
ica flew 1o Ilaly; it was a very pleasant journey is 
also realizable as a subordinated clause (Mortice 
flew 1o Italy, which was very pleasant); its ;t sim- 
ple clause (Mortice had a very pleasant journey Io 
llaly); and as a. phrase (Mortice's pleasant journey 
lo Italy). 
• Communicative structure; the corrlnl,micatiw~ 
structure of ATTRIBUTION Varies depending on 
the order in wl,ieh m|ch.'us and satellite, are re- 
alized. Cf., e.g.: Moniea flew 1o \[Zaly; it was a 
very pleasant journey vs. II was very pleasant, 
Monica's journey Io Italy. 
• Speaker's intention; selecting the A'I'TILIIIU- 
TION relation the speaker is assumed to intend, 
(RI/CONSEtlUENCE 
:action (LYING/ SITUATION 
:actor (PERSON 
: in-focus + 
: ~,eX lllai(~ 
:name #unknown#) 
: obligatory-roles ( : actor, 
: situation) ) 
:consequence (LYING/ CI,ASS-ASCRIPTION 
: domain (PERSON 
:in-focus + 
:sex lllale 
: name #unknownS) 
:range (:actor 
:situation l.Y INCl) 
: obligatory-roles ( :domain, 
: range) ) ) 
Figure 2: The text plan for the text with the meaning 
'The man is lying; the co||sequence of this is that this 
man is a liar' 
e.g., ;t justification of what has bee.u (-omnmni- 
cared in tile nuchms as in John failed the exam; 
it was very dijlicull; a consequence of it John has 
beeTz shot - he ist dead, el, c. 
'Fhe increasingly delicate specifications achieved by 
decomposition are also represented hierarchically in 
network fl)rm; one network for each dimension. 
4 GETTINC, TIIE RELATIONS AC- 
CROSS 
The lexieal choice process, which makes use or the 
discussed discourse structure relation taxonomies, and 
tile representation or lexical resources arc described in 
detail in (Wanner, 1992, 1994). lh, re we focus on the 
interface between the first, level text planning and the 
lexical choice n|odule; and ou the output as prodt|ced 
hy the lexieal choice n|odule. 
The ecmllmlatioual fr;mu~work in which our model 
has partially been implemeI|tcd, is the systet||ic text 
gc|lerator KOMI,:'I' (Ibtl.eman ct al., 1991 ). One source 
o\['constrainl.s l'or the. first hwel text orgm,ization comes 
ill KOMI"T fl'Olll all It.S'l'-l)ased pla|Hler, t; The outpul, of 
this Iflanner is a collection of case frame.s with RSq" 
relations hohling between them as shown in Figure 2. 
Starting from a text phu, of this lind, I,he lexical 
choice module traverses a multilayered collection of 
networks (one of these layers is given by ;~ ta×onomy of 
lexical discourse structure relations disc||ssed), l)ur- 
ing the I,r;~versal, the text plan is transformed into n 
lexicalized Partial (;','a~,matieal /~truclure (pf~s); 7 it 
is called 'partial' because it contains precisely that 
an|ou||t ofgran|matic.al infornmtion which is necessary 
(;lh:ceut, dewdopmcuts of this phmtmr are described in (l levy 
el eL, 19,q2). 
7A l'(~s corrcsponds~ roughly speaking, to the Partial Sur- 
face l,'~tnctional Description (I'SFD) specilicatiml in the COMET 
syst, mn (McKcown ct el., 1990). 
373 
process: LiE\[spelling: "lle"\] "1 LEX- \[verl': + \] / 
\[\[l Pl~I MARY ~saye* MAN \[spelling: "n.m" \] - .': \[\]LEX- \[ ........ + J/ 
ttheme: \[\] 3 
g \[spelling: "be"l "1 /p ....... LEX-BE\[ve,,t,: + J 
\[glSECONDARY\] t°ken: \[\] \[spelling' "liar"\] / |valuo: L.o.n..+ 11 
ttheme: \[\] 
Figure 3: PGS structure for the senteuce The man is 
lying; he is a liar 
for lexical choice. The PGS iS passed to tile grammar 
(a systemic grammar of Gerlnan; el. Teich, 1992) for 
final syntactic realization. Figure 3 shows a sample 
pos encoded ;m a Typed Features Structure (el. Bate- 
man el al., 1992). 
The tirst and the most important task in tailoring 
the text plan to linguistic resources is to find a lex- 
ieally biased discourse structure relation for the Rs'r 
relation specified in the text l)lan. The search is done 
in accordance with the filnctional content, the inten- 
tion of the speaker, and the contents of the arguments 
of the RST relation. If tile RST relation connects unre- 
lated case frames 8 (as, e.g., EVIDENCE in In winter, the 
days are short. It is getting light late and eaHy dark.) 
these case frames are realized independently without 
being connected by a lexical discourse structure rela- 
tion. If the cruse frames are related, tile following three 
variations are possible: 
(i) An RST relation instantiation coincides with a lex- 
ieal discourse structure relation; as, e.g., the instanti- 
ation of RESTATEMENT in the following rudimentary 
text plan coinsides with our RESTATEMENT: 
(R2/RI.~STaTI~Mtr NT 
:statement1 (SAYING/ SYI'UATION 
:sayer Sveta/ person 
: verbiage #unknown# 
:manner (quiet 
:scale #minima\]#) 
: obligatory-roles ( : sayer, 
:verbiage, :manner, 
: situation) ) 
:statement2 (SAYINO/ SITUATION 
:sayer greta/ person 
:manner (quiet 
:scale ~minimal#) 
:obligatory-roles (:sayer, 
:verbiage, :manner, 
, , .o :~ituatiQn))) . If so, tim suDciassmeatlon of the lextcal discourse 
structure relation deter,nines its final realization. For 
example, the above text plan could be realized as 
a GENERALIZED II.I,~STATF, MI,~NT: S'vela fl'l'islerle; sic 
sayle elwas .qanz leise lit. 'Sveta whispered; she said 
something very quietly'; an INTENSIFYING RESTATE- 
MlgN'r: Svela sagle ehvas sehr leise; sic hauchte es 
kaum hb'rbar hin llt. 'Sveta said something very quietly; 
she breathed it hardly audible', etc. 
(ii) An RST relation instantiation subsumes several dis- 
s Case frames are considered to be unrelated if between them 
or one of their roles no identity, is-a, causer, location, etc. rela- 
tion holds. 
tinct classes of lexical discourse structure relations; as, 
e.g., the instantiation of the aST relation CONTRAST 
in (this plan is also highly simplified): 
(R2/OONTRAST 
:actimn (OCCUPATION/ SITUATION 
:actor Roman/ nation 
:actoe Gaul/ state 
:obligatory-roles (:actor, 
actee, :situation)) 
:action2 (DeCUPATION/ SITUATION 
:actor Roman/ nation 
:actee (village /location 
:part-of : Gaul) 
:negation + 
:obligatory-roles (:actor, 
actee, :situation))) 
may be realized either as CONTI~ASTIVE CI,ARIFICA- 
TION (ga) or as CONTR.ASTIVE ENIIANCEMENT (,Oh): 
(9) a. (2aul is entirehj occupied by the I~o- 
marts; well, not enti~vly ...one small 
village still hohls out. 
b. Gaul is uhnost entirehj oecvpied by 
the l?omans; but one small village still 
hohls out. 
In this case, tim taxonomy of lexical discourse struc- 
ture relations is entered at a relatively general level 
(in the worst case at TOP). 
(iii) An lIST relation is not captured by our taxonomy 
(as, e.g., CONCI,ISSION). Then, the corresponding ease 
fi'ames are treated as unrelated (see above). 
5 RELATED WORK 
Our proposal for tile description of lexically biased 
discourse structure relations resembles l)anlos' work 
(Danlos, 1987), who presented acceptable ehmse pat- 
tern sequences explMtly in a Discourse Grammar. 
The. basic difference between Danlos' work and ours 
is that in the Discourse (h'amrnar, clause pattern se- 
qtlences are represented as concreLe valency sehelllatlt 
while in our model they are represented as functional 
distinctions that encode sequcndes o\[' I,FS. As a result, 
we do I/el. \['aee the probhml of being restricted to a 
concrete small domahl as l)anlos does. 
Metoer's text phumer (Meteer, 1992) is another 
prol)osa.l for the realization of lexically biased dis- 
course strucl.ure relations. \[/ut while we argue that 
lexically biased (liscourse structure relal;ions are to 
I)e realized by a functionally motiwd,ed h'.xical choice. 
model, Meteer sugggests a single struetundly moti 
rated model For text planning, which also sul)sumes 
lexical choice. This is different from, e.g., (R.ubi- 
neff, 1992), who ensures the expressil)ility o1" discourse 
structure relations provide(l I)y a conventional text 
l)hmner by annotating linguistic structures. 
Elhadad's prol)osal (Elhadad, 1992)to use 7bpoi 
(inference ,'ules that encode relations between t)ropo - 
sitk)ns incorporating lexical material) as discourse 
structure relations is aimed at exploiting lexical phe- 
nomena for discourse organization. Elhadad focuses, 
however, on the 'argumentative potential' of lexieal 
items r;~t.her t.llan on lexica.lly biased discourse struc- 
ture relations. 
374 
6 CONCLUSIONS AND FUTURE 
WORK 
In this paper, we argue<l that it, is useful Lo distin- 
guish between two levels of discourse organization: a 
global discourse organization, which is not afli~cted by 
linguistic means; and a finer discourse orgauizal,iou, 
which is built tip in accordance with the linguistic ma- 
terial that is availal>le for the meaning communicated. 
We have shown that reiteration aud collocatiott re- 
lations may function as discourse stru<:ture relatious 
and that these relations are well represented by Lez'- 
ical Function sequences. We presented a taxonomy 
of lexically biased discourse strucl:ure relations, which 
is related to lIalliday's proposal for grouI>ing inter- 
clausal logico-semantic relations aud suggested l.o use 
this taxonomy ill a lexical ch<>ice niodule. 
One of the ope=n problenis we face is how sulli<:iently 
detailed COlltP, Xtllal COllstrailits Cttll be aC(lllircd ill or: 
der to guide the choice of one discourse structllre re- 
lation over others. '.\['his will ccrtaillly be olle of the 
topics we wilI have to address in the future. 
ACKNOWLEDGEMENTS 
'I'his work has been carried out willie tile author was 
affiliated with the Inleqratcd P+tblicalion and lnforma- 
lion Systems \[nslilule of the (IMI), I)armsi.adt. Many 
thanks are due i,o John \]laternan, l, idija Iordansk~\ia, 
and Igor Mel'~.uk for valuable COllllUelits Oll ail earlier 
draft of this pal>or. 
References 
Batcman, J.A. ct al. (11991). 'lbwardsanArchitccture fro' 
Situated Text Generation. In I6'UICL, Penang, Malaysia. 
Bateman, J. A. et aL (1992). The Nondirectional I{cpre- 
sentation of Systemic Functional Grammars and Se- 
mantics its Typed Feature Structnres. h, I>rocccdinqs of 
UOLlNG-92, vohune IlI, F,;~tges 9 I g-. \[12(I. 
Dale, R. (1989). aeuerclti.q Referring l;':rprcssions in ~l 
Domain of OI,jccl.~ and Pr<wcssc..¢. P h I) thesis, I I n ivelsil.y 
of Edinburgh. 
l)anlos, L. (1987). 77~e Linguistic Ba,~is of 7)':rt (,'cncra- 
lion. Cambridge: Cambridge \[hliversity Press. 
Elhadad, M. (1992). (\]enerating Coherent Argunienl.atiw! 
Paragral>hs. In \]~roceedili¢ls oS COLIN(; '92, vohllile I I, 
pages 638 - 644. 
Grosz, B. J. and L. Sidner (:1986). Attcntimh Intentions 
and the Structure of Discourse. In Computational Lin- 
guistics Jotirnal, 12(3):175-204. 
tlalliday, M. A. K. and R. llasan (1976). CoheMon in 
English. \[,omlon: Loll l~lllB.n. 
Ilalliday, M. A. K. (1985). An Introduction to l"unclion+d 
Grammar. Lolidoli: Edward Arnold. 
IIovy, E. II. (1993). Autmnatcd Discourse Generatim~ Us- 
ing Discourse ,qtructure Relations. In Artifici<d htlclli- 
.qerice, 63(1-2):341-386. 
lfovy, E. et al. (1992). Employing I(uowlcdge I\[csources 
in it New Text Planner Architecture. In I>rocecdings of 
the 6th. Natural Language Gcnerntion Work.shop, Trento, 
italy. 
Iordanskaj;L, L. N. at ,L (1991). I,exical Belectiml 
and Paraphrase in a Meaning-Text General.ion Model. ht 
('. L. Paris ct al. (cdltors), Nat+tt'<ll L,nquage Generation 
in ArtiJici,l Intelligence and Computational Linguistics. 
I)ordrcchl: Khtwer Academic Pul~lishers. 
Mann, W. C. and S. A. Thmnl>son (1987). 
l/helorica\[ Strn<:ture '\['heory: A Theory of "Pcxt Organi- 
zation. In L. Polanyi (editor), 7'he Strncture of Discourse. 
Norwood, New Jersey: Ablex Publishing Corporation. 
McCoy, K. F. (1989). Generating Context Sensitive Re- 
sponses to Object-related Misconceptions. In Artificial 
Intelligence, 41 : 157 195. 
McKeawn, K. R. and W. R. Swartout (\[987). 
Language Generation and Explanation. In Anmml Re- 
+~iews in Computer Science. 
McKeown, K. R.., e't al. (I 99 l). Na.tural \],itllguage (-lener- 
ation in COMET. In R. I)ale, et al. (editors), Current 
l~ese, rch in Nalnral L<lngu~tgc (,'encration. \[,ondon: Aca- 
demi<: Press. 
McKeown, K. 1/. (1985). 7'c:rt Gencr(ltion: Using Dis- 
course Slt¥llegies and Focus Uo~zslrctints lo Generale Nat- 
ural Lctnqltnge 7'e:rl. (~ambridge: Cambridge University 
Press. 
Mel'<Nik, \[. A. and A. Polgu~re (1987). A |"orntaJ Lexicon 
in ~,he Mcaning/l'cxt Theory (or Iiow to Do Lexica with 
Words). In (/ornpul(llionaI Linguistics , 13(3-4):276 289. 
Meteer, M. W. (1992). l':t:pressibility and the l>roblem o.\[ 
l'~Jli'cicnt 7'c:vt l'hmnin9. London: Pinter Publishers. 
l/citer, E. (1991). A New Model of Lcxical Choi¢:e for 
Nouns. In Cotnputalional lnh,lligcncc , 7(4). 
Ilubinoff, II. (1!392). Integrating 'l'cxtldanning and I,in- 
guistic ('hoicc by AnmmLI.ing hinguistic Structures. In 
l~roceedings (>f the 6lb. Natm'al Language (,'cner, limt 
Workshop, '\['rento, Ihdy. 
Tcich, E. (1!192). I(omet: Grammar 1)ocu- 
mentation. '\['echnical reporl,, C~M1)/Institut fiir \]ntegri- 
erte lhllflikations- und htformationssysteme, l);trmstadl., 
(lermany. 
Tutin, A. and R. Kittredge (;19!12). Lexical Choice in 
Context: (\]cnerating lh'ocedural 'Fexts. In P1vcccdings 
of COL\[N(I '92, w~htme If, pages 763 - 769. 
~vVanner, l,. and J. A. lla.teman (19911). LexlcM Ctmc<:ur- 
l'(~llC(~ llelaiions in Text. (Icncration. In l'roce:cdings of 
the 51h. Nalm'¢d L<mquage (lcncrcllion Workshop, l)aw- 
son, PA. 
Wanlicr, L+ (1!J92). l,cxical (Jhoi<:e and the Orga,tization 
of Lexical l/csour<:es in '\]'cxt (leneration. h, Proceedings 
of Ihc I'\]+trop¢:alt (;oltfer¢'m'c on Artificial httclligetwe, Vi- 
enna, Austria. 
Wauner, I,+ (1!/9,t). Building Another Bridge <wer the 
(leneratiml (lap. In l>rocccdings of the 7th. Nalural Lnn. 
(lUOgC (~cncration Workshop, l(cnnehnnkl>ort, Maine. 
375 

Parsing 

