NON-CONSTIT UENT CO()t:{.I)IN ATION: TIII!;OII.Y AND PRAC'I'ICI,;* 
David Milward 
(Jentre I'or Cognitive Scieiice, \[Jniversity of gdinlmrgh 
2, Iluccleuch Place, t",dhflmrgh, ElI8 9I,W, ,qcotland, davidmCc~cogsci.cd.ac.uk 
ABSTR.AC'I? 
l)espite tile la.rge ainounl, of theoretical work done 
on non-coastituent coordination chu:ing the last t.wo 
decades, lrlany co\[npitt, atiollal systems still treat co 
or(lination using ada.pted parsing st, rategies, in a si- 
rlriilar fashion to the SYSCON,I system develol)ed tbr 
A!I'Ns. This 1)a.per reviews the i.heoretical literal;ure, 
a.nd shows why IIla.liy of I, he theoretical ;u:couu(.s tic- 
tualiy \]lave worse coverage than ac(;Otllt\[;s based on 
l)ro(:e.ssing. IPiimlly> it shows how l)rocessiug a.ceounts 
(:{IAi he described fornmlly and dcclara.tively in terlns 
o1' I)yna.mic (', ramma.rs. 
INTRODUCTION 
This pN)er is concerned with .sTlmmel.'rical coordina- 
I;ioil, whore the order of the con.iuncts (the items being 
coordinated by a eonju n('tioil such as amt el: o'r) ea.n be 
alt(;red without affecting a(:cel)tahility. (k)o,:dim~tion 
of this kind is traditionally split iuto co~slituent cool 
diwttimt, Wllel:e ('.llch (:olljltlict \[()rnls a consl, ii,ueul a(;- 
col:(lillg 1,o 'sl,anda,r(\[' pltrase strtnetm:c gl:i/.ITllll&f,'q, alKl 
~wn-couslitucut coo.rdinaliou. (Mnstitueni; and non- 
coustituent coordination have heen treated as entirely 
separate phenonmna. (se.e van Oirsouw, 1987 For dis- 
cussiols), ;rod dill~renl, mech:misnls have heen prol)o- 
secl for ea(:h. I lowcw~r, I)y considering grammat.i(:ality 
judgements Mona, there seems little .iustifiea.ion \[b," 
such a. division. ~lk) illustrate this, considel' the, sen- 
tCIICC: 
l) John gave Mary some books 
Ea<;h of the tina.I proper substrings ol" the selttellce (i.e. 
so'm( booh:.s, Ma'r~/ .some books etc.) cml I)e used as a 
conjunct e.g. 
2) a 3ohn g~*v(; Mai, y \[som(. hooks\] ~uid \[soirie pa- 
\[)ers\] 
b .Jolut gave \[Miu:y sonic books\] alld \[Peter soln~a 
impers\] 
e 3ohli \[gltv(; M:~/.i'y sortie books\] mid \[lent I'eter 
soliio pal)ers \] 
Siniila.l:ly, each of t, he inil.i+fl suhstrings ca.n l)e ilsed as 
a, (;onjllll(:t e.g. 
:~) a \[,Iol,n gave\] a,,d \[l'(:ter le,,l;\] Mary sonic I,ooks 
*This r(!sr~m'ch was SUlq)Oi'l;((I I~y. t.he U.i'{. El:it!ucc 
mid Engineering Itesem'ch Council (llcsearch I"cllowrdlll ~ 
Ig/90/TTl"/288). 
h \[John gave Mt~ry\] ~md \[Peter lent George\] some 
books 
c \[John gave Ma.ry seine\] mid \[P(;ter h;nt George 
many\] books 
and so can each of the middle substdngs e.g. 
4) a .lohn \[gave Mary SOllle\] a.lld \[lel|( Peter many\] 
books 
1, a(,im \[>~,,(; mary\] an(l \[/e.t Peter\] .~,,~y l:,ooks 
e Jol:,, gay,.. \[V<y so,:,e\] ,:,,(l \[fete,' ,.a.y} books 
Only exa.mples (2a) and (2c) are constituent coordi- 
n~tions. Example (3(:) seems slightly mmal;ural, but 
it is much improved if we replace books by a heavier 
string such as books about 9arde~ing. '\]'hus, for this 
example, any substring of the sentence I can tbrm a 
viable conjuuct. 
DELETION ACCOUNTS 
In (,he last twenty to thirty yea.rs there have. been a 
sm'ies oP ae('ounts of coordina.tion involving various de- 
letion me(:halfisms (FIYOHI (~.g. Gldtman, 1965 to vm~ 
()i\]:sOUW, 1987). For exmnple, \['rom the following 'an- 
tee:edellt' se.\[ltellCe.~ 
5) Sue gaw: l"rcd a book by Chomsky and Sue gave 
Peter ~ paper I)y Chomsky 
van Oirsouw ~dlows deletion of words to the left and 
to t\]le right of the conjunction, 
g~8 ~,~gl)eter a i)apcr t)y Chomslcy 
resulting in the sentence: 
6) Sue gave Fred ~ book aud Peter a. l)a.per by (Jhom- 
sky 
Most deletion ac(:otmts assume that, deletion is l)erfol> 
reed under idctitity of words, httt don't amdyse what 
it means for two words to I)e i(lentleal (an ex('cption 
is van Oirsouw who diseussc.s 1)honological, lnorpholo- 
gic~d and referential identity). (Jonsider the following 
example oF deletion. 
I'l'hc cx;mll)h:s Mmvc only consider substritlgs containing 
lllOl'O thl.tll ()1/(~ WOI'(I. Coordhiaiahm ~3f the individual words 
(whi,:h is necessarily consi.ituenl: coordination) is ~dso possible. 
Nat, ural exmlll)les involving lhe dei;erlnlncr, some, are di\[llcult 
to ~u:hlcw!, howew!r deiern'~htcr coordhla.tion is possible (cc)ll- 
sider: \] (t{d~l'~ /~t0tu ~u\]zclh,?r to ea:\])cc~ Jt:tu o~" ~na~zl/ peopl{ to 
co~Nc). 
935 
7) John will drive and Mary built the drive 
* \[John will\] and \[Mary built the\] drive 
Here the two cases of drive are phonologically identi- 
cal, but have different syntactic categories. Now con- 
sider: 
8) a * John bored \[the new hole\] and \[his f>llow wor- 
kers\] 
1) * Mary came in \[a hurry\] and \[a taxi\] 
These are cases of 'zeugma' and are unacceptable ex- 
cept as jokes. It therelbre seems that the deleted 
words must have the same major syntactic category, 
and the same lexical meamng. 
lIowever, even if we fix both syntactic (:ategory and 
lexieal meaning, we still get some weird coordinations. 
For example, consider: 
9) a * Sue sawi the manj \[through the telescope\]/ 
and \[with the troublesome kid\]j 
b *1 saw \[a friend of\] and \[the manuthcturer oil 
Mary's handbag 
In example (3) the two prepositions are attached dif- 
ferently, one to the verb saw, the other to the noun, 
man. In exmnple (b), attributed to Paul I)ekker, the 
two coRjunets require Mary's handbag to have a dif 
ferent syntactic structure: the bracketing appropriate 
\['or the first coajunct is \[\[a friend of Marv/' s handbag\]. 
The unacceptability of these examples suggests that 
word by word identity is insufficient, and that deleted 
material must have identicM syntactic structure, as 
well as identical lexieal meanings. 
Some of the most compelling arguments against de- 
letion have been semantic. For examqfle, Lakoff and 
Peters (1969) argued that deletion accounts are inap 
propriate tbr certain constituent coordinations such 
gtS : 
10) John and Mary are alike 
since the 'antecedent' sentence John are alike and 
Mary arc alike is nonsensical (it is also ungrannna- 
tieal if we consider number agreement). 
However, semantically inappropriate or nonsensi- 
cM 'antecedents' are also possible when we consider 
non-constituent coordination. For example, consider 
'antecedents' for the lbllowing: 
\] 1) *t \[The man who buys\] and \[the woman who sells\] 
rattlesnakes met outside 
b Many former \[soldiers living in England\] and 
\[resistance members living in N:an(:e\] have. si- 
milar memories 
c John sold dill>rent dealers \[a vase using his in- 
tensive sales technique\] and \[a bookcase using 
his market-stM/ technique\] 
(llb) is non-constituent coordination under the pri- 
mary reading where the scope offouner does not con- 
tain living in England i.e. where the semantic bracke- 
ting is: 
12) \[\[former soldiers\] living in England\] 
Examples (a) and (b) could be expanded out at the 
NP level, but not at the S level, l\]owever (e) can- 
not be expanded out at any constituent level, whitsL 
retaining an a.ppropriate semantics. For exan~ple, ex- 
pansion at the V1 ) level gives: 
13) John sold different dealers a vase using his inten- 
sive sales technique and different dealers a book- 
ease using his market-staJ1 technique 
'\['lms, although I,akoff and Peters' argmnenls count 
against standard deletion analyses, they do not coullt 
as general arguments against a unified treatment of 
constituent a,nd non-constituent coordination. 
SHARED STR.UCTURE 
Consider the sentence: 
14) John gave Mary a book m\]d Peter a. paper by 
Chomsky 
Instead of thinking of John 9ave and b.!\] Uhomsky as 
deleted, we can also think of them as shared 1)y the 
two conjuncts. This strncture can be represented as 
follows: 
Mary a bo{}k 
John gave a~M l)y Chomsky 
Pet, m" a paper 
From the result of the previous section, cac.h (-onjm~ct 
must share not just the phonological materiM, but 
Mso the syntactic structure and the lexieal meanings. 
There are three main methods by which this sharing 
of structure can be achieved: phrasal coordination, 3- 
D coordination, and processing stra.tegies. 
PHRASAL COORDINATION 
At first sight, analysing nou-constitnent coordination 
using phra.sal (i.e. constituent) coordilmtion seerns 
nonsensical. This is not the case. (2)or(linations are 
classified as non-constituent coordination if' the con- 
juncts fail to be constituents in a 'sta.n&~rd' phrase 
structure grammar. However, they may well be con- 
stituents in other grammm:s. For ex~ml)le, it has \])een 
argued that the weaker notiol, of constituency i)rovi- 
ded by (Jategorial (\]rammars is {'xa(:tly what is requi- 
red Bo allow all con.imml;s to be treated as constituents 
(Steedmal\] 1985). 
Phrasal coordination is exernplitied by the schema 
2. 
X -~ X Conj X 
2'\['here \]lave \])CCll vgLriolls argtlIllell(;S (st;elnllling \]'1"13111 Hoss 
1967) for the adopl;ion of a wtri~m~ of Ihls s,:heln~L, in which 
the coordinaling conjunctions is assoclatcd sMely with tJle lasl 
conjulR:t. 'Phe schema is revised as l;-)l\]ows: 
X -+ X X\[CoRi\] 
X\[Conj\] --+ C.m,j X 
936 
The sh;~red ma.t('+rial is necessaa:ily l,reated ident, ically 
\['or 0a,(:h COil>lille|; sine0 t;here is only a single COl)y: L\]le 
conjunct;ion is euibeddcd in a. siligle synt, ax Lrce. 
The phrasa.| coordhm, l, iou scheuia, requires each con-. 
junct, t,O be given a single l;yl)c, and for the ColijUlicl;.s 
and the +;olijllilc(,ion i/,s ;-i whole to I)c of ,;tic st'i, ilie I,yi)e. 
Prol)letns with the la.tl;er r(;(ltth:enicnt, WCi'c i)ointe(l 
OUl; I)5' Sa,g (',t M. (1985), who gawe l, he following (:cmii- 
t;('re xlm~ pies: 
15) a We wMked slowly and with great care 
I> I)a,L is ;~ II.cpul)lica.n and l)iX)ud of it, 
c I am }l(>l)ing~ Lo gel; all hiviLi~tion all(\[ oI)i,iulisti(: 
abOtll, lily Cl>a, llces 
Sag ei ;d. dca, I wil, h these e, xau\]ples by t,t'ei~t, ing care-+ 
gories as feala\]re Imndl0s, and allowing i:oordina,(,iou 
in (:as(;s where ,:her0 ~lrO t'Cil.(,tlrO~ in COtlllYlO\[l, I?o1: ox- 
ample, the l;WO colljtlllcl, s hi lisa,) .<;ha.re I,he f'ei~ture 
I-MANN\["J~ 3. As it, sta, nds, Lho ;_l.('(;()tlili, dot~s IIOg &'al 
with ex~t\[ill)les Sltch as the \['ollowing, 
16) TNT deliver cf\[ic\]eid, ly aud on ,qliti(\[ay.'-; 
I\[er(; t.he a.(lverbia.\] I>lu!a,se wou\[(I l)re.suuud~ly I)e 
+MANN#ell and I, hc l>rel>OSi(,ional i)hras(~, t-'/'\]';M/'. 
I;ill'l,ll(~r eXa,liil)l(~,s which a,r(', prol>leula,(,ic fo\]' Sag (!1. a,l. 
a£(': given I>y ,Iorg(~l\]SOll a.iid AI)cilld> (i!)92). 
All aAL(;l:ila, f, ivc, siiggeM, ed I>y Morrill (\]990) ;_,ii(l ~i- 
lni\[ar Lo .Ioi'~c~l.son at>(I A l>cil|(i (\]!){)2)> is tx) ,is(': t,ho 
followhtg <:ooMinaLion ,'-.cholna: 
XVY -+ X (~onj Y 
'l'hi.s ,:loes not hupos(! auy (:onditiou i.iia(, i.h(; i,wo (:aix: 
<e)Ji;ics \]Yl~ au(I Y ,sharo alLyl,hing iu COliliilOli. I lowev(~l:> 
,,he now ca.{.cgory .XVY is il,<,e(l tx) (!liStlr(~ Lhal. boi.h 
cui;(;gories ilr(! approl)ria.(x; in ,;he C('lltl,(~Xl;. For oxaiil- 
I>le, (15b) is a c(:(3)t;d>\[e since the co(>rdhla, gion Ioyl)e 
is NI'VAP, a.ud i.s sul)cat, egoriscs lbr Iml;li Nl's a.lld 
it Ps. 
A i;i/(;llCl* tnorc (li\[ficitlt l)rob\[oln is l;liat, (>f pix)vidh~g 
t;yi>es for all l)ossiblo (:olLjuncLs. C, onsi(ler tll(; lbllo+ 
wing: 
17) it Sue ga.ve Fred a bool<: and Pel;er a I)itpcr 
I~ Mmy a(htdrcs aud ,%c thhA,:~ ~hl~ likes Pctx.' 
(~l) i:+ ,. <:(,,tJ,i,,<:t+i<,,, or t,,,,(, >i~-~ or ~>,,,, ~,l,,:~,~(,~. (i>) i~ 
a, (:ase of ~ili\]l)otil,(l(x\[ I/ight,-No(.le Raishtg' whcro l, he 
llOllll i)hrasc \])~l(r is onibc(ld(',(I ai, (lill'er(,,nt (l(;i'>iJis ill 
I, he I,wo (:oujiuicLs. 
There ha.vo I>ceii I.vvo iuai\]i iq>proa(;hes l;o (lea, thlg 
wii, h exalll\])les Stlch as (a,) liSillg pht'+~sal (:oor(ti11~'d;iOll. 
'\['he lirsL is 1,o hlLro(hlce ;'l.,l explicit l)l'()(ltlc:l; op(;ra, t, or 
(e.g. Wood 1!)88), allowing 1;y\])es ()\[' i~he l'ortrl NI'*NP+ 
The :-;econ(I is t.o use a cldculus hi whi(-h t;yl)eS c'~/ll 
iill(\[Cl'gO 'Lype-raishig' (c4.' ;. l)owi;y 1988), <, c<"ui Im 
\['OlHll(~d \])y a,l>sl,l;a,('l;ioli (as ill t;h0 I,;tnll>ek ()aJ('tl\[llS, 
I,a.tul)cl<: 1958). The ell'e(q, is i,o l;l'eal. \[@(d (l book as a 
verb l)\]lr~se missi'llg ii, s verl>. 
:"~al+g cL ;~.|. also suggest; ill, ii.ll,(~rnatlve li'(!~ltliiClll; ilSillg ;Lll 
~Li)l\]arcn|ly <,l.hcrwlse unm,~tlw0.ed gr~tlnlll:Lr rule Ally\[ ~ - )- PI ~, 
The advantage of t~doi>ting ii genera\] al)sLractio\]l 
nteclmnism, its in the l,amlmk Ca.h'uhls, is thaL this 
a.lso provides a, l;l;(;ag\[/lelll, o\[e+xalnl)les such as (I)). Un- 
fortuna,|;cly, the abilit, y Lo i)e, rfot:ln M~stracl.ion of caLe- 
I,;ories will, fmtcgio\]ml Lypes (which is required \['or (a)) 
a.lso allows shared m~d;eria.1 Lo get dift+erent, syuta.ct, ic 
aatalyses, restdt, iug in acccpta+nce of +dl the scntenrcs 
predict;ed by delel;ioIt account, s where identil;y of lcxi- 
cal raLegories and lexicM senw, ntics is resl)ccl;e(l , bu(, 
not idetltity of syni,acti(: stru(:tur< l/,econsidcr: 
Is) *\] ..+~.w \[+,. r,:i(md or, +m(t lib(, ,.~umru+(-t.m'(;r or} 
Mary's handbag 
We can ol>tain identi(:al sy, lta(:l;ic tyl)es for a j:ric'nd 
q\]' and the man+@+ct.m+cr of by sulfl.racl;ing the lexi- 
(:al types of I, saw, Mary, 's, and handbag from (he 
s(;ntcnl;e Lype S 4. Since (,he tyl>eS a.i:e idcu(,icM, (:o- 
ordination can Lhe.u I,ake pla(:e. Thus Lhe ability to 
'~qtll)I;ra(:(;' otto t;yl)e from anol;lte\]' allows I;he Lambek 
(',atcutus tXl replica+to a. delet, ion account, and it, the- 
rel)y suli'ers froru the sanlc 1)robt(>u,s. 
'l'h(;re have 1)eet~ SOlllO l)rol>osa.ls t;o l'esLri(:L \[.he 
I,a,nbel,~ (Jah:uhls ill oi:der t.o pr(~v(mg such overg('.ue 
raLion, l:larry and Pi(:keriug (I 993) proposc it i:ah:ulus 
in whi(;h (17~+) is d(.alt with using a. produ(:L opera 
Lion, +rod absLra.ction is limited to (:atcgories which do 
l~ ()t ;~r(:l' ~; ~t functiou in tile dcriw, tiolL This a(;COUlll; 
nmlles reasot\]ably good crop,riced prcdi(:t;ions, t+hough 
it, does fa.il I'or t;1>(+' following exaniples: 
19) tL You !:a, call ml+ directly or aft;el: :ll>ln I,hrough 
Lily S(:(;I'(~{,a ry 
I) Sue mr, a. huup ou the tarl)lc irn(l ott (,he ledge 
a la.\]'gc a.uti(lue I)Ulmh )owl 
lit (a,l, (!at:h (:()ll, jllllCl, COlltariltS dil'l'Cl:CU(; iilIIllh(ws ()f 
rood,tiers o1' di\[l'er(;nt types (a,u adverb, all 1)hrase wil, h 
two i)reposiLiona\] 1)hrases). In (I)) the sul)l:a+tegorisa- 
(:ion order is SWal)ped in the two (:on.bm(:t;s. 
Sttcces,.sful grea(,ill(;ilI, ()f llOli constil.uent (:oor(lim~- 
t, ion using phrasal (:oor(lination se('ms I,o re(ILL,re ela.- 
bor~d;e encs)(ling in (,h(; (;otljllll(;l; type of at siln\[->\[(? ge- 
n('.ralisat;iou: (:Otl.iltllCLq (:all (:oordina.te l)rovidcd Lhey 
a, re ac(:Cl~tM)h~ wi|;hin Lhe 8a.l.ite synt, act.ic conLext. ','he 
3-1) al)l)roa(:hcs a.nd i)rocessing stra, t(:gies nse synt, ac- 
ti(: (:ont.cxt more dircl:t;ly, and it is t;o these ,nel;hods 
whi(:\]l we llOW tllrll. 
3-1) CoordinaLion 
l+et; us briefly re.consider otu: cxplmlat.iou o1' ddct.ion. 
I",x+mqfle (6) was exl)lainc(l by saying t.hat t.he two 
sta'ings by Cho.msky and Sue galm are deleted iltldcr 
SOil,(+ IlOI,ioIt 0\[' ideutity. Ilowever, we could e(lmdly 
',veil hame descrilled this as a process whereby the first; 
in,'-+La,nl:e of by Clwm.sk9 is ,nerged with the second (tlll- 
der some noLion of idenLity), and the second insLance 
of ,9'uc 9av+! is merged wit,h the \[irst+ 
'1The t;yl)e given to boLh <:OIIjllII(:I~S 1 using r<!asona+l)ly sl;a.n- 
d:Ll'd I;yp<! a.qsiglmlenLs and +L;~+llllmk' lI('d;tl,ioll~ would be: 
(((NP\((NI+\B)/NP\S))/NI+)/(NP\NI>)/NP)/NP 
937 
Merging word strings instead of deleting them does 
not help with the problems of deletion acconnts which 
we outlined earlier. In particular, it does not help to 
exclude examples (9a) and (gb) which suggest shared 
material must have identical syntactic structnre. Ho- 
wever, once we have started to think in terms of mer- 
ging, there is an obvious next step, which is to move 
f'rom inerging of word strings to rnerging of synt;ax 
trees. This is the move made by Goodall (1!)87), who 
advocates treating coordination as a. union of phrase 
markers: '% 'pasting together' one or, top of the other 
of two trees, with any identical nodes tnergil~g to- 
gether" ((;oo(hdl, 1987, p.20). We can visualise the 
result in terms o\[' a three-dimensional tree structure, 
where the merged material is on one plane, and the 
syntax trees for each eonjmlct are on two other pla- 
nes. For example, consider the a-l) tree for example 
(17a.) given in Fig. 1. 
np vp 
StlO / t-. ~'~ "" np 
W',. ~\ " ....... ll 
~ " ,1 t) , " paper 
v ~ Peter np ciet 
gave ~ /, a ' I x x 
I \ np n 
det book Fred a 
Fig. I 
The merged part of' the tree inch,des all the nodes 
which dominate the shared matel:ial Sue gave. The 
coujmlets retain separate pla.nes (denoted here \] U 
using dot, ted and dashed lines respectively). 
(k)odall's aeco,mt does not deal with examples such 
as (171)), which he argues to I)e examples of'a different 
phenomcaon, llowew;r flmse can be incorporated into 
a ad) aeco,nt (e.g. Moltmamh 1992). 
There are various technical difliculties with Gee- 
dali's account (see e.g. van Oirsouw, 1987, and Molt- 
nmnn, 11992). There is also a f'undanmnl,al l)lro- 
blem COl,cerning semantic interpretation of coordina- 
ted structures (see Moltmanu, 1992 which provides 
a revised a.nd more complex 3-1) account based on 
Muadz, 1991). 
\["ot: coordinatioll el' unlike categories, as in the ex- 
amples in (15), Goodatl proposes a treatment some 
what similar to Sag el; al. (1985). I\[owe.w~r there is 
still a problem in dealing with examples where there 
are clit\[>rent nurnbers of modifiers, such as (19a) or 
the lbllowing: 
20) a. We can meet at the office or in London outside 
the theaure 
b 'FNT deliver efficiently and after 5pro in Edin- 
burgh 
(~onsider example (b). The synl;ac\[~ic structure appro- 
priate for "FNT' deliver eJficicnthd has one S node and 
two VP nodes, llowever, the structure fbr 7'N7' dell- 
vet after 5pro in l','dinbm~h requires one S node and 
three VP nodes (or three S nodes and one VP node). 
The two structures therefore fail to merge since the 
structure dominating the shared material 7'N7' de- 
liver must be identical. The nse of ordered phrase 
structure trees also excludes examples such as (19b). 
\]n summary, the /I-l) approaches correctly enforce 
identity of synta.ctic structure for shared material. 
ltowever, the way of characterising syntactic struc- 
ture using (parts of) st, andard phrase structure trees 
results in a.n overly strict requirement of parallelism 
between the conjnncts. We will now consider pro- 
cessiug strategies, where syntactic structure of shared 
material is eharacterised more indirectly by the state 
of the parser. 
PROCESSING STRATEGIES 
There have been several attempts to trea.t coordi- 
nation by adapting pre-existing parsing strategies. 
For example, NFNs were adapted by Woods (1973), 
I)CGs by Dahl and MeCord (198:1), and chm:t parsers 
by lfaugeneder (\[992). Woods and Dahl & McCord's 
system are similar, ltaugeneder's system has very li- 
mited coverage. 
In Wood's SYSCONJ system, the pa.rscr can back 
ap to various points in the history of the parse, and 
parse the second conjunct according to the configura.- 
lion fouud. I"or example, in parsing, 
21) John gave some books to l)eter and some papers 
I;o George 
at the point after encountering and, the parser can 
reaecess the configuration after parsing John gave i.e. 
a stack consisting of a sentence and a. verb phrase, and 
an arc traversal by the verb. The second eoldnnct is 
then parsed according to this contlguration. 
SYSCON./ does not imlnediately merge the two 
stack configurations after completing the second con- 
.iunct, bnt, ins,ca.d, separa.tely parses both conjuncts 
in para.lM until a constituent is completed. For ex- 
ample, on parsing the sentence, 
22) John gave Mary a book and Peter a paper about 
subj acen cy 
tim SYSCONJ systean sepa.rateIy parses Peter a paper 
about subjaceney and Mar9 a book about s.abjacencg 
before conjoining at the level of some enclosing con- 
stituent (for example \[.he verb phase). The result is 
theretbre similar to starting with the sentence: 
23) .Jolm gave Mary a I)ook about, subjaeency and 
gave Peter a paper about subjacency 
As noted by Dahl and McCord, this meehanisrn means 
that SYSCONJ inherits the problems of nonsensical 
semantics which plague the deletion accounts, since 
John and Mart arc alike is treated the same as ,Jo/u~ 
a~v alike a~d Mart arc alike. The mechanism adso 
causes problen,s \[br dealing with nested coordination. 
938 
Consider the sentence: 
2d) John wanted to study medicine when he was elc-. 
ven, law when he was twc'lve, and to study no- 
thing at all when he was eighteen 
The smallest constituenL containing to siudg medicine 
when he was eleven is the verb phase wanted to study 
medicine 'wh~m hc 'It;as clcve:n. \]lowevcr, if coordina- 
tion of the first two conjunets occurs at tl~is level, it; 
is diflicull, to see how to deal with the final coidunet. 
Both Woods and l)ahl ,% McOord usc stack based 
configurations rather than ~ lifll parsing history. '\]'hus 
once something is popped off the stack its internal 
structure cannot bc accessed by the coordination rot,- 
tint. This rules ont exmnplcs such as the following, 
25) John gaw'~ some books to Mary and papers to Gc- 
orgjc 
where the NP, some books is COlnpleted prior to the 
conjunction being reached. 
Although I)rocessing a(;cOltlltS call provide reason 
al)le coverage of the coordination data, the exact 
predictions often require detaile.d examination of the 
code. This suggests a need for the more absl, ract level 
of description which dynamic grammars I)rovidc. 
DYNAMIC GRAMMARS 
I)ynamics is ,lust the study of states and transiLions 
between sta~es. It can be used to specify the states of 
a left; to right parser and the possible mappings 10et- 
ween states, l"or example, Milward (:1992b) provides 
a dynamic description o\[' a shift reduce l)arscr, and a 
dynamic description of a fully incremental parser I)a- 
sod on dependency grammar. Suitable languages for 
dynamics arc both forma.l and declarative., and are 
therefore also appropriate to exl)ress linguistic gcne- 
ralisations. 
liu a l)yumnic Grammar (Milward 199210), each 
wood is regarded as an aetion whi(:h I)erforms some 
change in tim syntactic and semantic context, For 
cxaml)lc, a parse of the seutcnce John likes Mary be- 
tomes a mapping between an initial state, ci, through 
some intermediate states, c,, el, to a final state c/ i.e. 
.Ioh~z likes Mar9 
(;i ~-> C a t--} C b g-} (;j 
I f we use a dynamic grammar to dcseril)c a shift re- 
duce parser, states encode the current stack configu- 
ration, and are related by rules which corrcsl)ond to 
shilTting and reducing a Since there are arl)itl:arily 
large, numbers of different stack configurations (the 
stack can I:)e of a.rbitrary size), the dynamics for shift 
reduce l)arsing involves the use of all infinite l/Ulrl- 
her o\[ states. It thus differs from, say ATNs (Woods 
SShlft corresl)onds to: L ~-,' <X> *L on input of a word, 
W, where L is a wtrlablc stmMing for a list of cat;thin'its, ',*' is 
list COllC~tl:enation, mid X is the category for W. l~(!duce co,'- 
rcsl)onds to <(;n ... Cl 2> ol, ~ <:('0 :> el, on (unpty inl)ul;, 
where C0 -} Cl ... On is a phrase sl.l'u(:lAire rule of I, he gramm~w. 
1973), which have a tinite nmnber of states, augmen- 
ted by an explicit recursion mechanism. 
Dynamic grammars can t)e presented as rewrite 
grammars by using transition types instead o\[' the 
more usual S or NP. For example, to get the parse 
above we need the lexieal entries: (~ 
John:ciF-~ca likes:caP+% MaJ:y:cbr-Ocf 
aud a single combination rule. schema which states 
that, 
l)k)r any C1, C2, C3, 
(?D-+C3 -4 C1~C2 C2~->(J3 
A string of words is a sentence if it has the tyl)e , 
ci ~-->cf 
where ci and c/ are al)propriate initial and final states 
for a parse r. 
\[n a dynalnic grammar, any substring of a sentence 
<;all be assigned a type. \['%r example, likes and Ma W 
can 10e combined to get the type ('aF~rcf. Thus we 
have an appropriate hwel to perform substring cool 
dination. Dynamic grammars may be extended using 
the following combination rule (arm and or are both 
given the. special transition type CONe): 
l"or any (;1, C2, 
C'1~}C2 --~ CIF+C2 CONJ C1~-~C2 
Similar to SYSCON,1, this allows coordination when 
two conjmlcts map between the salue pairs of states. 
l'roecssing is also similar, with the eneountering of a 
conjtmetion causing back-up to an earlier stage in the 
parsing history. Howew~,r, since there is no popping 
of a stack, the fnll parsing history is awdlable s. I"or 
example, lieu gave some books to ,flue has the transi- 
tions: 
IJe,z fla~e sonle books to ,guc Ci --} (;k'-- Cl --O Cnt ---} C n -~" e o --} (;/" 
WC can then parse papers to Joc nsing the transitions: 
pal~rs to Joe C m (;~q 
---)- e o ~-)- Cf 
Since the final state c/ matches the stab; immediately 
1)efbre the conj/inction, the two strings caa combine. 
'Fhe resnlting transit, ion diagram is as \['ollows: 
c i -o e h c/ _s 
books to ,%c and pollers to .loc ('m --~ C.f 
Sl,'or exmnple, for tile shift reduce parser, Ihe word .foh*t 
would gel the \[.ype, L ~+ <rip;> o\],, corresponding t. a shifting 
of tile NP onto the stack. The mnl)ty siring gets the I, yl)e, <Cn 
... (;t > ol, F+ -<O0 > ol, where Co -+ (;1 .., (7,, is a rule of 
tile grammar, ,:orrespondlng to reduction. 
7For the shift reduce parser, the initiM st~tc is tile ('Alll)ty 
list, <>, the final state is <s>. 
aSomethlng parallel to pOl)l)ing occm's tufty after tt c~)ordi- 
nation. \[Iowever this is exactly what is required since we do 
not want (~verlappingcoordinal;i(m as in The gill ~md the or th~ 
bO~.l and the adult came. 
939 
Iterated coordination (e.g. for examples such as Mary, 
.Petc~" and Sue) can be treated ill tilt same way as 
iterated constituent coordination is treated in phrase 
strncture grammars. For example, each transitio~ 
type can be augmented with a feature (+/-) deno- 
ting whether or not; that transition has been iterated. 
The eoordination rule becomes: 
For miy C1, C2, 
(-JD+-C2 _,v Clio 1~/-(32 OONJ (Jt~-C2 
Iterated types are formed as follows: 
\]"or any O1, (J2, 
C1~0+O2 -o CI ~-FF/- C'2 O1~-+-(72 
The precise gl:amnlaticality predictions lnacle 1oy 
the dynamic approach depend npon the characteri- 
sation of the states, and hence depend oil the parti- 
eulal: l)arsing strategy which is specified by the (ly- 
naniics, tlowever there are some general predictions 
which can be inade. Firstly, consider conjuncts which 
correspond one t;o one ill the categories of the eel 
responding words. Here the eonjmletS must provide 
the same transitions, and heuce lIlnst be able to coo> 
dinate (this is a reflection of tile fact that processing 
can back np to any point in the parsing history). This 
predicts that ally substring of a sentence can eoordi- 
ilal;e with itself, and heuce that any snbstring of a 
sentence can act as a eol/junet. I"or eonvel/ience we 
will call this the substrmg h;l/pothesis. Tills hypothesis 
has l)een In:oadly adopted in the work of van Oirsouw 
1987, Barry and l)ickering 1993, and l)y work on the 
1,anlbek Calculus (e.g. Moortgat 1988). 
Apparent counterexamples are at fbllows: 
213) a * The woman spoke to George aild nlan to Pe- 
ter 
b * ,John told \[Mary Bill\] and \[Fred Slle\] was co- 
ming (Barry and Pickering 1993) 
I Iowever it; is dillicult to exchlde these using syntaetie 
constraints, without also exclnding the 1note accepta- 
ble: 
27) a Every woman spoke to (.leorge and lnan spoke 
to Peter 
b ,Iohn told the niotllers that their daughters and 
the fathers that their sons were all at the party 9 
More natural examples where conj/lncts are forlned 
by \[raglnents l'rorrl different COllStit, uent.s are the folio- 
wing: 
28) a The police fonnd sonic \[cars inside\] and \[lorries 
outside\] the warehouse 
b Everyone who I \[admire most came\] and \[ad- 
mire least stayed away\] 
c Mary showed litany \[h'iends the weird books\] 
and \[eolleagues tilt lnore respectable pal)ors \] 
written by her mother 
The relative UlmCCel)tability of the examples in (26) 
is perhaps best explained as dne to violations of into- 
9This exmnple is al.l.ributed by Barry and lqckerlng (1993) 
I.o Jmme Johmulessell. 
national requirements, rather than syntactic require- 
lnents (cf. Steedman, 1989), 
One case where the dynamic grammars correctly 
violate tile substring hypothesis is when a string al- 
ready involves a coordination, lIere, the internal st;a- 
tes arc not aecessible, so we can't get interleaving of 
two coordinations, as in: 
29) * The girl and tile or the boy and the adult came 
'\]'here may be an argument for similarly blocking co- 
ordination in cases which would involve the breaking 
apart of idioms or ot her structures which are nol; Stall- 
dm:d eases of lexical Sld)categorisation. An example 
(due to Mark Steedman), which may be sneh a case, 
is the following, 
30) * One lnan ill \[ten spoke against and twenty ac- 
t ually protested\] 
As noted above, the precise granunaticalil;y predic- 
tions depend on the kind of parsing lnodel which is en- 
coded in the states. In Milward (1992a), the dynarnics 
specifies a word-by-word incremental parser lbr a lexi- 
calised version of dependency gramlnar. Fach state is 
a recursively defined category, similar to ~/cal;egory in 
Categorial Gramlnar. For exarnple, after parsing You 
can call me one possible state is a sentence missing a 
sentence lnoditier 1°. This state is appropriate as the 
initial state R~r a parse of both dirccll?/, or of after 
,Tpm thro,ugh my sccTvtary, resulting in a final state 
of category sentence. Thus examples Stlch as (19a) 
are dealt with, since the syntactic context after You 
can call me dots not distingnish between one or lnore 
than one subsequent modilier. This lack of distinc- 
tion as to whethel: one or lnore modilier is expected 
is actually a necessary prereqnisite \['or performing de- 
cidable fl,lly word-by-word incremental interpretation 
(see Milward and Cooper, 1994, in these proceedings). 
Some of the problelns with eategorial granlmar ac- 
counts of coordination do reoccur with a dynamic ac- 
connt based on the parser used ill Milward (1992a). 
For exaniple, 
31) \[,/ohn\] and \[Mary thought that l'eter\] slept 
is predicted to 1oe a(x:eptable, as are the following, 
32) a \['lbd~ w John\] and \[mary thonght that Peter\] 
slept 
I/ I he.ard \[that\] and \[that no-one else knew that\] 
Fred won the seholarship 
This second batch of examples is particularly dillicult 
to exclude withont malting changes to the eharacteri- 
sation of the states. A feature I>lus or lninns tensed 
vel% on each conjlmet does block them, but is difl-icnlt 
to motivate. 
I)ynanlie grammars can be regarded Imrely as for- 
real systems, as direct representations of proeessing, 
or as something inbetween (for example, ill the packed 
l°l)ependency grammar does not have VP modifiers 
940 
im.rallel parser described in Milward (1{)921)), the ac- 
l;ua.l pro'sing sl,~tes ~tro plwked w;i'SiOllS o\[' i;hc ,qt,al, es in 
/,he gr3.,llilliil, r). ;If + we consider I;he dyiiauiics t,o Im a di- 
reel, rol>reselil;~ti, ioll 0\[' in:ocessing, I,\[lell ;t dependence 
o1" linguistic (lal;a upoii I)al',<dl~.g sl;~i;es woliid ou\[y Seelll 
/)bmsibl0 i\[' the parsing i)roces.<, corresl)olids, ~1; leasl; 
1,0 S()l\[iO exl, ont,, wigli acl;iutl hunmn hulgnage IJrocos - 
sing,, 'l'his brings it 1) Llle intriguing possilfility 1;ha, l, we 
c~_/,ll predict coordin;fl, ioli I)tcl;s l'roD;l kiiOWll 1)IX)('.(',SSillg 
(\[a,l;a,> a.nd vice versa,, bkn- ex3.niple, consider I;iie well 
kuown oxa.\[liDle Of gaxden i)ai, hing: 
33) 'I'll(! horse ra.cecl lmSl, llhe ba.rii fell 
'\]'he ci~oic.e bctweeu l, ho Ilse o\[' facc(l as I,\[ie iiiaili verb, 
or ~l,s pi~rt o\[' i;ho i'edllCC(t relative is tb<-Jtla.lly IISSlllI'I(;(\[ 
I,(/ l)e within 1,he \]'lX'i, gliielll, 1,\]l,'? \]lol'.sfs v(Iccd, .quggestittg 
I;lii~l; I;here m'e I;wo disi,inguished parsing stai,cs ~flTter 
raced. 'l'hus this correcl, ly predicts l;he uiuux'el)l,alfility 
o\[" I, he following: 
3d) * The horse raced \[lmsi, 1,1,,:~ Imru \['ell\[ and \[heside 
\[,he hedge\] 
CONCLUS\[ON 
This pa.per ha, s skol;ched wlriotls problents wil, h solne 
o\[ the lhtguisth: &cl;Olllil;s O\[ (:Ool;dhl~tl, iou. II. Sltggc.. 
sled tlia~ l;his was prinuu'ily duo l,o dilliculty ill el\[co-. 
(ling it. I)rOl)er nol, ion o\[ syul;ac{,ic COlil,ext,. 'l'he l>al)er 
i;\[l(,.ll cousidered v~rious processing (I,CC()II\[IIIS, Wiiel:O 
I,he synt0acl, ic context is e.ucoded wit, hill I.he sl.ai;e (/2 
l, he i)al:ser. I"imdly it sliowe.d how dyltalliic,s Call \])e 
tLsed aS ~l l'orili&\[ dosc.rip{,ion o\[" processiu<~,, i/cc, Olliil.,s 
which iiso ;i full imrsing hisl,ory, and how \[lie cliltrlic 
l.erisations ()\[' I)arshig sl;a.l,es C~/li I)e chosell i,o elil;rsl'ce 
i,il<; requisite deglx, e o\[' i)a, ralh!l{mn Im(,weeu (:olijliltCl,s, 
RI'\]I,'EItEN (31~8 
Ibm'y, (I. mid M. I)ickering (1993). I)epell(len('y 
(jal, egorinl (',ra.iDiiiar a, lid (~ool'din~tlion. l,ingui,s1,ios, 
<71(5), p.a55-{)02. 
I)ahl, V. a.nd M.C. Mc(',ord (1983). 'l'reai;ing (k)Ol'-- 
dinaflion ill f,ogic (Jralmilars. Compu1,a1,iolla\[ \[)iilgul- 
s1,ics, 5)-2, I).09-4)1. 
I)owl.y, 1). (1988). Type I{aising, I,'unctiollal Con> 
posit, ion ~md Non-(ik)nsl,i\[,ueni, (<.on.junctiou. In 
I/. Oehrle elL. al. Eds., (;'a1,cgorial (h'ammm'.s' <rod sVa- 
hu'al Language ,5'truc1,m'~es. I).lLeidel. 
( ',leit;tnan, I,. I{. (19(i5). ( JoordhlaA,hig~ ( Jonj unct.ions ill 
English. Langufl\[l(!, /11, 1/.260-2"93" 
(k)od~*ll, (:. (I.C187). Parallel ,h'lrttt:lm'cs ill ,5'!lttlela:: 
Coordination (\]ausa1,it~cs and l~c<s1,rltc1,uri,yl, (~ain- 
bridge t hi\[ w;rsil;y Pre.qs. 
Ilmlgenedcr, II. (\[{){)2). A Conipul;al,iolial Model for 
I)roces;silil,j (',oordim~te S\[,rllCtliros: t'arsing (Joorditia 
l;ion wil;holll; (Jl'allllll3.r. \[11 I'roceedings of I,\]()AI {12. 
,\]or?jcIlS(:ll, II. and A. sXl)eill6 (1992). (k)or(lhial,\]oli 
of "Uulike" Categories in TA(I. In Proc.c.cdi,.qs (~.1" 1,1a 
2,d TA(7 Workshop, I)hiladell)hh/. 
hal(off, G. and S. Peters (1969). Phrasal (\mimic 
l.iou and Symmetric Ih:edicates. hi B. I{.eibcl and 
S. Schane, I,',ds., Modern ,Vtudics in b'wflish. Eagle- 
wood (Jlifl's: Preutice-Ilall. 
I,ambek, ,I. (19,38). '/'he Mal, hema.tics of Seutence 
Strucl.ure. /lmcrican Mathematical Mon1,1Ug, 6"5, 
i>. 15d 170. 
Milward, 1).1{. (1992a). Dynamics, I)elmit(lency 
(ara,nmnr and \[m:rcmcn'0al lul:erpretation, lu Pr, x'cc- 
din qs of COLING 92, Nante.s, p. 1095-1()9{L 
M \[1ward, I). IL (1992b). I)ynamic I)el)endem:y (a I'll.lit-- 
mar. Paper l)resented a(, t.he :h'd Meeting on Ma.l, he- 
uml;ics (/\[ l,~uiguag,'.. To a.ppeax iu the proc{;e(lil~gs, or 
in gdinburgh Working I>al)ers in CogniLive Sciem:e. 
Mi\]w~u'd, I).FL and \]:I.. (Jooper (I 9{)4). \[ncremenl;a.\[ \[n- 
l.erl)rel;atk)n: AI)plical, ions, 'Pheory, n.u(I \[I.(qi~l, icmshil) 
t.o I)ymmiic Sctrmnl;icm lu I'rocc~:ding.~ of (.'OLIN(/ 
04, I(yoto, ,lapall. 
X'loll.tmmu, I". (1992). Coordi*mZion and Compelrati- 
i,,.s. Pll. I). disserl.i~l.ion, M IT, ( Jand)ridge Ma. 
Moortga.t, M. (1!)88). Catcgorial htl,,c,slig<t1,io,t.s: Lo- 
.qical and l,in.q,i.~lic /\[.sl)ccts o J Ih~ l.ambf /c Cnh:,hts, 
l)ordrecht: I,bris. 
Morrill, (i. (1!)90). (\]Iraillll\]al; a,n.:\[ Logical 'l'ypcs. In 
IJ'rocLcdil~g,,~ of lhc 71,1, tlm.s1,erdam (;olloq,ium. l'l'hl, 
Universii, y of Aiiisl;erda.Hi. 
Muadz, II. (1991). (,'oo'rdina1,c 5'1,r,c1,,rf: A I~htnar 
lb:lJrCs<:*daiion. Iqi. I). disserl, n.t,iou, tJniw'.rsil,y o\[' ari 
Z()lH{ s 'l'llCSon. 
\qtll OirsOllW, I/.IL. (1{}8'2). 'lTw ,b'g~lh,:,: ~tl (,'oordiua- 
1,ioi'l. (JiX)Olii-| loin\[. 
I/.oss, ,J.IX.. (1967). Con.strais~1,.s <m l/ariabl,.s ill ,<;gn.. 
/am, Ph.D. disserl.al.ion, MI'I', (lanihridge Ma. 
S;~g, J./\., (~. (;a,z(hl.r, '1'. Wasow and ,<J. Weisle, r (1 {185). 
(Joor(linM;ion aud I low t;o I)istiuguish Ca.l,(w)ries. Na.. 
1,ural l, an qluujc and 1,,ingui.~'tic 771cm'g, <7, p.117-171. 
SI,e(;dlliali, M.,I. (1985). I)el)en(leucy and (k)c/relhla- 
I, ioll hi },lie (Irall/lluIr O\[' I)utc.h a.nd I,higlish. I,all.quage , 
ill, 1).52:.t-5(J8. 
Si;eodliuu,, M.J. (198{I). hil.Olla(,iOll altd Syut;a× iu 
~l>oken \],allgumgje ,~y.ql,elli8. 'l'ecllllicai rel)l)rf, , ~lV- 
(Jl,C-J-89-20, I)ept. o\[' (Joluplller ;lli(l hlForiliat0icJli 
~cieilce> Universil;y o\[' Penli~ylva.nia. 
Wood, M.M. (1988). A (;'a1,qlorial S;llnl, a.~: dbr Coordi- 
~za1,c Conslrlu:lions. I~h.l). Thesis, University ('.o\[lee;c 
I,ondon. l\wfilal)h~ ;is 'li~chni(:a.l ILeport, \[JM(\]S-89-2 
I, I)ept. oF (;~.OlUl)ul, er Science, thliver:di, y o1' M a.lu.Jite-. 
sl,er. 
Wood:~, W. (1973). Au I';xperhiienl, al I>axsing S.y- 
sl, ein lT)t' 'lYan,~il, ion Nel,work (~rannnars. lu H. ILu 
sl;in, I';(I., Nattu'al I,al*quay/c lZroc~ :<,mlfh Algoril.huiics 
I'ress, New York. 
941 
