Appficafio~ of the Direct Memory Access paradigm to 
~,ot~ °~a~ ~a~gl~age interfaces to knowledge-based systems 
~I~deto '.COMABECHI and Masaru TOMITA 
Center for Machine Translation 
Carnegie Mellon University 
Pittsbm'gh, PA 15213 U.S.A. 
&bstract 
't'h~s pa~r describes file use of the Direct Memory Ac~ 
~:css (DIdA) pmadig~n hi a practical ltatu~tl lmlguage in- 
~:c~f~tec, Advaaltages and disadvantages of DMA in such 
~pplications art~ discussed. 'ihe DMA natural language 
inteffa~x~ 'I)M-.COMMAND ~ described in this paper is be.. 
tug u.~;c ¢l tk~r development of a knowledge-based machine 
translation system at rite Center for Machine 'lYanslation 
((;NIT) ~t Ciancgie Melloli University. 
L )~ntroduction 
The Dh'ect Memoiy Access (DMA) paradignr has been re.- 
searched as a new model fox' natural language processing 
(Riesl~eck&M~fin\[1985\] and Riesbeck\[1986\], Tomabechi- 
\[1987a\]). h this paradigm, natural language understanding 
is viewed a; an effort to recognize input sentences by using 
p~c-cxisting knowledge in memory, which is often experien.- 
tial aud cpi:;odic. It is contrasted with traditional models of 
parsing in which syntactic and semantic representations itre 
built as the result of parsing and are normally lost after each 
parse. In the DMA model, input sentences are identified with 
the memory sU'ucture which represents the input, and are in- 
stantiatcd to represent flint specific input. Since understanding 
is performed as recognition through the memoly network, the 
result of understanding is ~aot lost after each sentence is pro- 
cessed. Also, since parsing and memory-based inferences are 
integrated, ~arious memory~based activities emt be triggered 
directly through natural language understanding without sep- 
arate inferential processes. 
As on~' application of DMA, at the Center for Machine 
"l?anslatioil (CMT) at Carnegie Mellon University, we have 
developed a natucal language interface for our large-scale 
knowledge-based machine translation system t called DM- 
(~OMMAND. This application of DMA demonstrates the power 
of this m(~lcl, since direct access to memory during parsing al- 
lows dyna,~dc evaluation of input commands and question an~ 
~we~ing without running separate inferential processes, while 
dynanfic~d~y utilizing the MT system's already existing do- 
,~lahl 13~owlcdge sonrces~ The implementation of the DMA 
aTtte CM!3-MT systcin which is the target system fo~ tic DMo 
COiV~r¢~AND sy:~tem described in this paper is described in detail in Tomita- 
&Cmbtmell\[1987\] and Mitamufa, et a/t1988\]. 
natural language system has been completed and is used for 
development of actual grammars, domain knowledge-bases, 
and syntax/semantic mapping rules by the researchers at CMT. 
This system has been demonstrated to be highly effective as 
a MT developmental support system, since researchers who 
develop these individual knowledge sources are otherwise un- 
knowledgeable about the internal implementation of the MT 
system. The DMA natural language interface can provide 
access (currently English and Japanese) to the system's inter- 
nal functions through natural language command and query 
inputs. This use of the DMA model for natural language iw 
terfaces demonstrates that it is an effective alternative to other 
natural language interface schemes. 
II. A background of DMA 
The Direct Memoly Access method of parsing originated in 
Quillian~s\[1968\] notion of semantic memory, which was nsed 
in his TLC (Quillian\[1969\]) which led to further research in 
semantic network-based processing 2. TLC used breadth-first 
spreading marker-passing as an intersection search of two lex- 
ically pointed nodes in a semantic memory, leaving interpreta- 
tion of text as an intersection of the paths. Thus, interpretation 
of input text was directly performed on semantic memory. Al- 
though TLC was the first DMA system, DMA had not been 
explored as a model of parsing until the DMAP0 system of 
Riesbeek&Martin, except as a scheme for disambiguations. 
I)MAP0 used a guided marker-passing algorithm to avoid the 
problem of an explosion of search paths, from which a dumb 3 
(not guided) marker passing mechanism inherently suffers. 
DMAP0 used P~markers (Prediction markers) and A-markers 
(Activation markers) as markers passed around in memory, 
adopting the notion of concept sequence to represent linear 
ordering of concepts as linguistic knowledge, which guides 
linear predictions of concepts sending P-markers in memory. 
ZSueh as Fahhnan\[1979\], Hirst&Charniak\[1982\]. Chamiak\[1983\], 
Haun&Reimer\[1983\], Hirst\[1984\], Charniak\[1986\]. Norvig\[1987\], and 
eonneetionist and distributed parallel models iimluding Small. et al\[1982\], 
Granger&Eiselt\[1984\], Waltz&Pollack\[t984\], Waltz&Pollack\[1985\], Berg- 
\[1987\], and Bookman\[1987\]. 
aWe call it 'dumb' when markers are passed everywhere (through all 
lix~ks) from a node. In a 'guided' scheme, markers are passed through 
specific links only. 
661 
(~ c q ' .,o~leept s(Y,l~:itmoes, ~ W~.ch (~r~colnpassee phrasal pattc~:~l;,, are 
attached to nodes in memory timt ~elwemnt some specific ex,, 
perientiaI memory strta:tm-eo ~a DMAP0~ Aomad~e~s at~:~ sent 
above in the abstractimi hierarchy fiom the lexically acfivaied 
m~de in memory, and P~ma'kers ale se~t to the ~'~e×.~ elema~t 
of the concept sequence only after the Aom?rker from t~\]ow 
' , is already Pqnarked, Concept • ~ is lfit~ aaodc m~,t ~ elllleffleitt 
_ . (~ r • performed uxi~g c(mcept refinernent ~inlo; (xe~>hnks) when 
a whole concept sequence is activated. Co~cept ~'eti~emc',w~ 
locates the most specific node in memory, below t!~e activated 
root node~ which represe~is the speeific instance of ~he inpm 
text. DMTRAN~; ('/bmabechi\[1987al) evolved the DMA into 
a theory of crossqinguisfic ~ranslations and added mectmism,,; 
of explanatory generati(m, (;--Marker pass\]us (for further con-. 
t~:xual disambiguation,~;)~ and a revised scheme of concept :~'o.. 
finement while performing English/Japanese t~anslaiions. 
~t!~o DM-Com~nand 
The DIVL-COMMAND system which wc describe in tiffs pa-~ 
per is a l-~atural language interface developed for grantmar, 
knowtcdgeobase, aud synlax/sema~tic mapping rule writers at 
CM'I; which enables these researcher,~ to access the MT sys-. 
tem's internal functions for their development aud de/ragging 
proposes. The DM--COMMAND parser borrows the bask- alto-. 
rithm from the DMTRAN,~; machine translation system, which 
performs recognition of inp~t via the guided spreading acdva. 
tion marker-passing of A-.mm'kers, P-mm'kers and C-markers ~' 
in memory. 
As a brief example~ let us consider the process\]lit the 
input command "show me *HAVt%A-PAIN', where *HAVE-A ~ 
PAIN i,~ au actual name of a concept definition in oar fl'ame 
system 0~'.RAMEKIT~ Nybergl\] 988\]). Independent of the se- 
mantic network of domain knowledge used by the MT sys.. 
tern, the DM-COMMAND has separate memory network repo 
resenting concepts involved in performing various actions in 
the MT system. Among such concepts is the concept 'show- 
frame', which represents the aetiou of pretty-printing bSRAME- 
KIT definitions stored as domain knowledge. This concept 
has the concept sequence <mtrans-word person *CONCEPI"> 
attached to it. This concept sequence prcdicts that the first 
input word may point to au instance of 'mtraus-word ~ (such 
as 'show~), followed by an instance of person fi)llowed by 
some concept in the font~ of a }~'RAME\]\[{'YF name, Whe~ the 
first input wonl "Show" (:ome~ in, it activates (puts au A- 
marker on) the lexieN node %how', which in turu sends ae- 
4C~markers (Conte~ual.-markers) were imrodnced in DM'i~IANa, ~md m'e 
propagated to mak eontexually highlighted concepts in memory. DMTRAN3 
used Comarkers lor word-scn~ disambigaations through ¢ontexual mark- 
in 3. DM'rRANS also added an explanatory geueration m~hanism whielt 
generates sentences in Ihe target langt~age :for concepts thqt did not hav¢~ a 
lexieal entry i~ the target language, by explaining the concept in that target 
language. 
fiW~,fio~ (A-mmker) above i~ fl~ abst~acficm hie~acchy mid 
hits ':mt~ans-.word'o At d.~e v(~ry begimfi~ng of pmxi~?g, ait the 
first efeme~ts of cmicept scqucx~ces are predicted (P-~?~arke.:i), 
dherc%re~ whe~ ar~ A=mark~ ~ is sent fl=o~:, %how" nard idi~ 
A--mark~" m~d P.marker coil\]de at °m~-a~s-.WoaF,. Wheaa t~fia 
eollisiot~ of two ma~kers happens, the P.maft,~,,~ ' },~; ~;c~,~a te~ itr, 
~e×t elcmc~t of concept scque~~ce, which is ~c~:qoC, 'i!t'i~c:~ 
am! then aefivat~?s 'person' (an A..mm'ker is ~ent above :~ ~ 
3:~e abstainer\]on hierarchy)° Since 'person ~ was ~'-maked. at 
a p~'evio~s marker collision at 'mtra~s-wordL m~oihe~' c,~iJ.. 
sloe (~ccurf; here. Therefore, a P-.marker ix agaif~ ~,:e~t to th~ 
~)ex~ cleme~t of the concept seq~_~ence~ which ix '*CON(:~:,',l"r ~ 
}~'i~atly, ~:~qiAVE.*A-f'AI.N" co~n6x i~L Now, ~:~;' spreading ac- 
tivation occurs not in the command liciet~tO.t-y netwod~, b~:~ 
in the domain Nmwledge network (dcetor/paficn~ diatog do.- 
main) activating '*t/AVE~-A4,AIN ~ initiall~cand thc,~ activai.i~g 
the coucepts above it (e.g, ~"".lqAVE~A-SYMPTOM') until fl,e ac. 
dvafion hits the concept ':::CONCI'2t~I '' which wax P.ma?kt)d vt 
the p~'evious coil\]stem Since it is the final element of d~c 
concept uequence <rattans-word t~-rson ::~CONCIiPE>, t:his corf- 
cept sequence is accepted wlmn this collision of A-.~nart~er 
and P-mm'ker happens. When a whole concept sequence is 
accepted, we ~mfivated the root node for the seque~ee, which 
in tiffs case is the concept 'show--fl'amC. Also, in addido~ 
to activating this coneept~ we perform concept refinemem6~ 
which searches for a specific node ia ~he eomma~M netwofl, 
ihat represents our input sentence° Since it does not exis~ i~: 
this first parse, DM~COMMAND creates that concept 7, This 
newly created concept is mi instance of 'ratransqi'amC~ m~d 
its object slot is now filled not by genetic ~*CONCt?!~r ~ N~t in- 
stead by '*HAVE.-A--PATN'~ specific to Ollr input so~t_e~i:eo '\['h~', 
final concept.-mfined concept is the result of the pacse?'o 
5One firing to note here is that the concept 'aIIAVE-.AopAIN' Nat i~ acti- 
vate, d by input "*IIAVE-A-IPAIN" is not part of the memo W Petwock tb~ Ne 
I)M-COMMAND~S MT system eonnnanding coaeepls, instead it is a menany 
unit that is a paa of lhe MT systenxs domain knowledge, in o~her wof&~ 
'*t/~VEoAq'AN' belongs to a different memory network fi'om 'showofiv_me', 
'mtransoword', and 'person'. This doe, s not causo a problem ~o !he DM-, 
COMMAND, and actually, it cmi utilize any number of indepe~Moat s~aaa~ltic. 
networks shnnltam~usly, as long as concept ~quences g~fide pa%iug oi! 
P-marker from one network ~o anoiher. For example, l!~e '*I'|q-',S¢)N ~ i)i 
ihe domai~ knnwlexlge ~manfic network rep~s~mt scum generic pe~:o;~ 
whereas 'person' in DM--COMMAND c9~'fnaand knowledge t~etwofl~_ rcp:c-- 
,~enfs t~:rsoos involved in the us~ of the DM..COMMAND Syste~'ll. 
~'Lytinen\[19841 has a discussion of 'concept-refinement' with his M(X'~ 
T~L~d,~S parser. 
71in DM~INANS~ when stleh eread(m of conc~;pts ~ce~t~rcd th,~ /~.~" wa; 
a~ed to provide the vocabulary, and tla~s s~:xved as a model ibr w~e~N'~day 
acquisition as well as concept creation, h~ ).)~V't-'COL~IVJAND, WO rm~don~iy 
generac names lbr such newly created iustames ~:(td ~ser does ~o\[ ~3appiy 
names tot the newly c~eated concepts. 
aAct~M inputs to DM.-COMMAblO m'c ~o~mNly much longest aid ~.c 
eompa.uy multiple concept sequences; however, d~e basic mechad:~m tiff 
r,,:cog~lifio~ of input is a~ explained here. Also, DMoCo~.~V~ND baadle~ 
662 
pw¢,Kyf'tt,sk~:: (lr~.~:/,,Mt2~{W"s flmcfic~.~ A)r N~ti.y..pfilii;hG a :hm~,:.c: 
7~:~,;,ttltlti~.~.ttJ,J. £0 \[~t'~ :Tt'A'VE--A.Pt~IN C:¢,k~.~;i~: iS a Stlbb~aSL~ tff :~c.©rJ. 
~:s)i;.<<'~: ' a~d if; ih~ objcc~, of pfi~ti,g 7~ oi~r exanlp\]m :hq)ltL 
{Qtq\]dt !~ t,~; C)~Vsl/~, ~no.~.(;~ ttatltl'al htngl;lage, tlildo:tsi.~itid~iqr~ ~,{; 
{~:p ~:i~cf;7, ;4dO~!{J~:\]fiit~ i~q~t with the Si}'~:Ci\]il;, t;OtlOO!)t S¢;~tiO;~CO 
~ii~{!{ i:::.;~ ~:{'~:Di:;{,t/;~; a, i'¢,;;;:)1i, GG~iGO~)\[9+ alld l:iJ~;;~l pOr~.:OlJ~i~(!g ()i(),t)G(}~j~7: 
la~gtmgo h~cor fa(;m h~r hJgg¢;rhlg sys/ra~ ~i~ilt~Cdo~ is {~lo}f~ ai:cd 
h~to i\]~m nK.m~;~y ~;ea~t;~ ,~icdvity u;ml\]o~' tht~ :()l;/J:~ pu'¢a,:iigm, 
avid fi6s w;~y, i~te:~ce is iul~:gratod h~to nat~iral \]auguag~ 
,':m&~staadh~go 
;{".7, :i! J)~ac,~t.,ss~i,~r.; 
~IAt~ to s{x:og~dz(~ tho 7ap~.d: ba~:ed oil what ~t ~fit'oi~dy :!tao'ws 
"~s doma, i~ ~.;pecitic know\]odgo i~ tho m'oa. 4~t b:alls\]athJ~ and 
tho ~;ya~m's owr~ implomG~tafi~m, Whe~i somo aci,ion if.; re ~ 
(\]~GSI;G{i, \[tlG i~\[(31-~l.(j~3 IHIISi. alKt6rsl.~!(~d thc i'cqi:lgst a~ld lgspon(i 
acc(gdin E to what is r~;q~t(;si:fxt, at~d d~crefore it \]s ~coessary to 
~:s~i\].C ~i~(l ,':l!Y~O)ll~; fi\[iSOOO~:,~;~ alld i,o ~l-J/Jty{)l ' th{i :-;y:-;t{~rn<"s ~.. 
iol.~t:-;il {\]~.l~{;fions appiroprJah3\]y, iTof tlxamp\](% ~": ~ i~:nowlodgt:- 
ba<,;(~ dove, lopo~' i~tplttf; "Ehow me all the mapping lu\[c~; on 
::flIPq)t)WIq.),EVER '~' i~ o~d¢3y to debhg somo conccpt~!al bug 
pm~m~m ~efc~ci-tcc rcSrMufion, e, llipscs, and <~ome types of anaphora (c,x-- 
a~aplcs arc ilich',dcd in ihc Appcndix)o Al,co, ~)Mo(JOMMAND Istilizcs C- 
~aakcr propai~#-ion ~t) di~?aiabJgtmic sfmi~, ~ of ik~J co~tt,xually diflicolt x~:n.- 
tcncc:4. Toi~'~a:)echi\[1987b\] gives a dc~oiled dcscdptkm (if Ibis dis~uid)igua. 
d,u.. ~edutM,, sao 
!):I,: ~t:~y b,:.',~l~ tbFA f{~c Jh~fitafiOll Of tiffs ~itieihoLi is ~}iat t}ii. P.,'.;iiicHco Ca,; 
,:;:,: }t~?Slt{\]:t( t.~ ~)iI~:~' "~7!1C4t {hey fi~, a p<v,:~p~:cificd concept a.quv~tt:u (pattcm); 
J!i)4/TG'~?(:i~ ~)GG ltlGU \[llk. ll{~\[V$Ol'k IS ;l.tl i:ll|lG\[itNIGl)-llG\[l WG (;~-ill GitC'J{~G tz{J~3r 
L,,,~;::,h~ ~:txt~c:nce:.; which ~_)re like ~<:ynu~,ciic 9,;niplato::. ~:ur c:Kaiap\]% th*.; 
t:~/4~c~cc~ <':'fi:~torc;"' *fJhysical.-obj~:ct> (*tbatoie':' i,~ reitvr~fl~) att?tJh~.TJ. £0 
d~ c.t,ac~pt ::)K,y~;~caf.object ~ ,<;imiha rto rCpl~.JNtmfill\['; ;-X fq P a~L, calagodzix~l~ 
~;.~r /~..'i).~t: 'Jl:hm:, we c.axt e~co&: absiract cor.<:ept <~t;,l~;rlcc~: ~_imt ~!<<;l ~:; 
\[J~rasM ~c>;icc~k 
"l{i~'~'~!\[lJ;:i!iiO~t ifJ iiTl\[JttOl'l\]l(\]\[liC{t ill \[i\]{tgtJlT.\[~,i(\]!'li' ,~;ys~c~il aS I\]1~ idggCfi~g 
{~:' d:_'~:~tt)~iG tVJfit:h 7s ctnllparablc to i~les~ag~ \]pa,.;sh~g i~i obi,Jc\[-odcntcd 
i:o YiD{:o~c~,~ wha~ "~app~,~ ~:~=d~/~ m~:anb I~ ~h~ ~.<~-,~ ~ ~. /( 
~i.L,;c ~) d~;q: ':',;u~.d ~' ~.,,o:~.us i.u sc,;J via ~h&: ...;ill ',,iiiiy; ;.~d ?) 
Also,, d~c rcm.dt ........ ~; O, ~ Lo. At IK~.i',';~; {,~; Hoi. \[CG~ \])i~. z:.c(;i!\[~H~J:.6:;{ i~ 
:~{;{;,"IS i:O ~:<JOogniZO ~\]~(~ i~p~J\[ E(;c.qrdi~i~ {'o Wb=.qi i~;~: E":' :/:; 
proset I:z 'w~H r(xpfirc :.m ¢~);te~'~ia_l i~ffc~'{~niiai p:{/:c;g-; dmL wit\] 
his -hi <,:umchhlg fo~' flio .>.q~pcop,daic a(;iio, rL<;a>mha;d ~b: d~<. 
,~:>)/~',i:~;)~l~t; \]t;)i{\[iJ/ll~h' \]xli.ol~.Kti a~c~i\]i:cc~.t!i't;. 7d~:u,l!.y, fi,~m \]~iU; i, ~;r 
\[Y!O~;I{!(~O ,~ltl.\] t.!t{; ~KIY~;tbl ill04;l; :ii!{C,t~!Ct: {li~;Jil{ ~, jKI~:~JU-I'~ (I~iG)l) ;!.,'Z 
{;o~lf',h-ailfls ptl{ ca~ i!lt" t~lldt~f<<~tali(ih~g o{ i\]i(; <gy::;'~;;~;:, ~,,:~\[i~i,.~ !~,_';: 
co.tcKt c;smblishcd by l|it; knowk;d<v,o dom~fi~4 alld d~,i; ;.~y,C;i:i;H! '; 
~ifG'~+)J;:IH(D:(II~3IIOH. ~\]O'7713VOf~ illft~i;SN El'lO\]!iRJlty i)Hii. Jil\]'G\[~D,qL:{: ~HC= 
h~tel~'~atcd, S|ICh ll~i Jni:(;:cacfior~ k; d:ilficldt i~) ;c;;;tt)tl.t ", ~:O(i 
WJThotl£ s~lch \]nlci'aotions, parsitig tim Oc.~ 4.;7{\]itJi 'q£;ty F.i{)Vi 
or fail hi oontc)alally difiicul~ sere{aloe:,- b.'_~cr,~r.c ~>J 1i~+: h,. 
tci'dc:t;;D~ldoncic,<~ of cono(;pt l~loaliin/;s o)qncs,<$:d i~ ti~v, \]7~pm 
J~;l dt ,l~:/~agc: 
\]ilJ i:\[3-G \])\[V~-{:()iVIM,Adql) Sygt(:'ill~ lil~;H~(;ty il; I)(7~:L~iZ.;:; V,(; 
i'lbmat.~oclfi \[ 19g'tb\] -, a(~d ih{; ( :iv" ~/. CM!' tcc;L~ ii(;~A i cpo i v ~ ;~ ,-;i<)J ~ (: i l/il; 
lmp~Tl c.ot~tah~s the detailed dh~ot~ssi(ms as well ;;,: ,q~!tiipit ~dw.-; t)i hi~itfJt;o { 
lilog ~ {y1~;~,£4 (Jr Nt;tl/i~,llb'iL~, 
l'ZJTy hih;l.,XaiCd pal'f:ez~ V,'(: re, Jail ~.{ pm,-;t:l Iila { lsc,=ibP.s~; ;;~)it:. ,,j:~;t~? :eli 
i~3td ~;,J),lra~¢k; allaly~ll$s in bJl)iDI} l~itegra/ed lil'llltlOtt 
t3)~or {;xamplG i;~1~7l (Clflli~gtbrd & }fo,.\]th\[l!JgS\]) n~Ji:d it J,;ql~i;:;\[,b~v.:fJd 
¢o~lC~:;phlat illtaiy:~,c/(it/os,'Jcck\[19'/:/\]) ibr pm;~hG 7)ilJiX{ ~{1 I~K; ii{li:llit ! i;~1. 
F)IIW,(; htlei J~l(:O v/i~ich ,<:upplkxi mu.,"!llilq; lt;pri> {:!l{;tik~12 tu \[iii ,~: p~ < ;:{; i\[~ 
~D2'(;~C;{; ~tG{ll.lkT. TJ'L~; NOl)~tl;~ill!,;h O1" !lit: tWO iil(i(h~|; ,*; W;~S \]',.'~4;\')Utl)! ; 7~ gU(;\]: 
i! b~yt~i.t;?il<~ l;{X;ill.t~3 CO~lCt=;i{)hLa.t ~Td~aiyZ{3L~i W'.:lt~ t,V\[i~ttiii'~ ~tIXt#, {t;~tt U:i 7)1(:J'1% 
that the concept which represents the request for action is di- 
rectly connected to the concept that represents the action that 
is requested. Likewise, the direct memory access recognition 
of a question means that the concept which is identified by the 
input is directly connected to the concept that represents the 
answer, as long as the system knows (or potentially knows) 
the answer. In oilier words, in the DMA model, recognition of 
a request for action is a triggering of the action requested and 
recognition of a question is knowing the answer (i.e., as soon 
as we understand the question, either we know the answer, or 
we know the inferences to be performed (or functions to be 
evaluated) to get the answer) as long as memory contains the 
action and the answer. 'lb reiterate the literature on the DMA 
paradigm, in this model, memory is organized in the hierarchi- 
cal network of concepts which are related by links that define 
the concepts. Thus, as soon as we identify the input with a 
certain concept in the memory, we can trigger the action (if 
this is a concept that represents some action (or request for ac- 
tion)), or answer the question (if the concept represents some 
knowledge (or request for some knowledge)). Thus, parsing 
and inference are integrated in the memory search process, 
and no separate inferential modules are necessary. It should 
be understood; however, that it is not our claim that we can 
eliminate inference altogether. Our claim is that 1) the mem- 
ory search through concept refinement itself is an inference 
which is normally performed by separate inference modules- 
(such as eontexual inference and discourse analyses modules) 
in other parsing paradigm; and 2) whenever further inference 
is necessary, such inference can be directly triggered after 
concept refinement from the result of parse (for example, as a 
daemon stored in the abstraction of the refined concept) and 
therefore, the inference is integrated in the memory activity. 
C. Ellispsis and anaphora 
In a practical natural language interface, the capacity to han- 
dle elliptic and anaphorie expressions is important. DM- 
COMMAND is capable of handling these phenomena, be- 
cause under the DMA paradigm (which is typically called 
"recognize-and-record paradigm"), the result of each parse is 
not lost after each sentence, but instead remains as part of the 
contexual knowledge in the memory network. On the other 
hand, in the traditional parsing paradigm (we call it "build- 
and~store" paradigm), since the result of the parse is lost af- 
ter each sentence, the parsers can at best handle indexicality 
within a sentence. Specifically, 1) ellipses are handled by 
DM-CoMMAND; since ellipses are characterized as the lack of 
elements in a concept sequence, and these are recoverable as 
long as the elements or their descendants had been activated 
in previous parsesl4; 2) anaphoric and pronoun references are 
~4For example, with the input "jgt92.gra o uchidase, sem.tst mo." (Print 
jgr92.gra. Sem.tst also). Second senteuce has the object droped; however, 
resolved by utilization of both semantic knowledge (repre-. 
sented as restrictions on possible types of resolutions) and 
also by the context left from the previous parses in memory 
similar to the way,that the elliptic expressions am handled. 
Finding a contexually salient NP corresponding to sotr~e NP 
means, in DMA, searching for a concept in memory which 
is previously activated and can be contexually substit~te fox' 
currently active concept sequencetS o 
Do DMA and syntax 
One weakness of current implementations of th~ If)MA 
paradigm is that the concept sequence is the sole syntactic 
knowledge for parsing 16. Therefore, a DMA system needs 
deliberate preparation of concept sequences to handle syntac- 
tically complex sentences (such as deeply embedded clauses~ 
small cauls, many types of sentential adjuncts, etc.). This 
does not mean that it is incapable of handling syntactically 
complex sentences, instead it means that concept sequences 
at some level of abstraction (at syntactic template level down 
to phrasal lexicon (Becker\[1975\]) level) must be prepared for 
each type of complex sentence. In other words, although 
such sentences can be handled by the combination of con- 
cept sequences, designing such sequences can be complex and 
less general than using external syntactic knowledge 17. Thus, 
current reliance upon a linear sequence of concepts causes 
limitations on the types of sentences that can be realistically 
handled in DM-COMMAND. Of course, there is nothing to pre- 
vent DMA paradigm to integrate syntactic knowledge othea' 
than a linear sequence of concepts. Actually, we have already 
implemented two alternative schemes for integrating phrase- 
structure rules into DMA. One method we used was having 
syntactic nodes as part of the memory and writing phrase- 
structure rules as concept sequences 18. Another method was 
to integrate the DMA memory activity into an augnrnented 
context-free grammar unification in a generalized LR parsing. 
Second method used in a continuous speech understanding 
is described in Tomabeetti&Tomita\[ms\]. We will not discuss 
these schemes in this paper. 
While handling syntactically complex sentences is rather 
expensive for DM-COMMAND, since it relies solely on linear 
concept sequences, natural language interfaces are one appli- 
tiffs can be supplied since the memory activity after the first sentence is ~ot 
lost and the memory can supply the missing object. 
15Fur example in "Pretty-print dm.lisp. Send it to mt@.nr'~ "it" can be 
identified with the concept in memory that represents din.lisp which was 
activated in memory during the understanding of the first sentence. 
t~Although generation is normally helped by external syntactic katowl- 
edge snch as in file case of DM'I'RANS. 
17Also, pronoun and anaphora resolution is based upon contexual knowl- 
edge alone; however, use of syntactic knowledge (such as rite governing 
category of an anaphora) would help such efforts. 
18Due to recursive nature of phrase-strncture rules, we did not find tiffs 
method appealing, urtiess we obtain a truly parallel machine. 
664 
cation area where the capacity to handle phenomena such as 
ellipsis, auaphora, pronoun resolution, and contexual disam- 
bignation is more valuable than handling syntactically com- 
plex sentences. It seems that DMA is one ideal paradigm in 
this axeao This is evident if we consider the fact that input 
to a natui'al language interface is normally in a form of di- 
alog and users tend to input short, elliptic, ambiguous aud 
even ungrammatical sentences to the interface. Our experi- 
ence shows that an increase in the size and complexity of 
the system ~n order to integrate full syntaciic processing, en- 
hancing the DMA's capacity to handle syntactically complex 
sentences, has so far outweighed the need for such capacity 19. 
Eo 1V~fip~e ~e~anti~: ~etworks and portability 
\])M-COMt,.~AND utilizes two types of semantic networks. One 
is the semantic network that is developed under the MT system 
as domain knowledge that DM-COMMAND utilizes. The other 
as the network of memoxy which is unique to DM-COMMAND. 
This memory represents a hierarchy of concepts involved in 
commanding and question-answering necessary for the devel- 
opment of machine translation systems. This memory network 
is written with generic concepts for development of MT sys- 
tems, so that this memory we have developed at CMT should 
be portable to other systems 2°. 
The control mechanism (i.e., spreading activation guided 
marker-passing algorithm) and the actual functions for per- 
forming actions are separate (actu',d functions ale integrated 
iuto the \]D~/~-COMMAND memory network). This separation 
makes the system highly portable, first because virtually no 
change is necessary in the control mecharfism for iranspolting 
to other systems, and second because the size of the whole 
system can be trimmed or expanded according to the ma- 
chine's aw61able virtual memory space simply by changing 
the size of the DM-COMMAND memory network 21. 
Thus, ~mder DMA, a natural language interface can 1) 
directly spr,~ad markers on the target system's already ex- 
isting semautic network 22, utilizing the existing knowledge 
19Although, we have seen that it is effective in parsing noisy continuous 
sl)eech input (Tomabechi&Tomita\[ms\]). 
~Of conrse~ we will need to change the specific functions that are stored 
in some of the nodes and perhaps some of the specific (lower in the hier- 
archy) concepls need to be modified for each specific system. 
21if only a l)asic command natural language interface is required, then 
we can trim |h,~ pints of memory used for adwmced interface and question- 
answering. (h~ the other hand, if machine's memory is of no concern, we 
can write memory-net and concept-sequences fbr all the system functions of 
ltie tin'get MT .,;ystem. Also, note thai due to the spreading activation guided 
mal'ker..passing algorifllm of the DM-CoMMAND recognizer, the speed of 
the system is ndnimally affected by an increase in the size of the memory for 
commanding and qnestion-mlswering. It is because spreading activation is 
local to each concept and its packaged nodes under guided marker-passing 
that even if the size of the whole memory network increased, the amount 
of computation for each concept should not inerea~ accordingly. 
~-:'~As long ~L~ semantic nets are implemented in a general frame language 
or object oriented systems. 
for understanding input texts! 2) utilize a command and 
query conceptual network developed elsewhere (such as DM- 
COMMAND), with minimum ~todifications in the functions 
stored in the root nodes that ~h-igger the actions; 3) be ported 
to different systems with virtually no change in the control 
mechanism since it is a guided spreading activation marker- 
passing mechanism and no system specific functions are in- 
eluded (those functions are included in the comand/query se- 
mantic net). 
V. Conclusion 
DM-.COMMAND is the first practical application of the DMA 
paradigm of natural language understanding, in which pars- 
ing and memory-based inference is integrated. This system 
has been proven to be highly effective in knowledge-based 
MT development. It is due to the complexity of system im- 
plementations in a large scale MT project that grammar and 
knowledge base writers axe not expected to have expertise ou 
the internals of the translation system, whereas it is necessary 
for such a group of project members to access the system in- 
ternal functions. DM-COMMAND makes this access possible 
through a natural language command and question answering 
interface. Since DM-COMMAND uses the spreading activa- 
tion guided marker-passing algorithm, in a memory access 
parser which directly accesses the MT system's already exist~ 
ing network of concepts, inference is integrated into memory 
activity. Since there is a separate memory network for con- 
cepts representing commanding and question-answering that 
are generic to MT system development, the system is highly 
portable. The DM-COMMAND system demonstrates the power 
of a direct memory access paradigm as a model for a natu- 
ral language interface, since understanding in this model is a 
l'ecognition of the input sentence with the existing knowledge 
in memory, and as soon as such understanding is done, the 
desired command can be directly triggered (or the question 
directly answered). 
With DMA's ability to handle extra-sentential phenomena 
(including ellipsis, anaphora, pronoun reference, and word- 
sense ambiguity), which are typical in a practical natural lan- 
guage commaud/query inputs, DMA is one ideal paradigm for 
natural language interfaces as shown in our DM-COMMAND 
system. Also, DMA's integration of parsing and inference into 
an unified semantic memory search has proven to be highly 
effective in this application. 
Appendix: Implementation 
ql~e DM-COMMAND system has been implemented on the 
IBM-RT ~3 and HP9000 AI workstations, both running 
23Due to the space limitation, the actual sample output of the system is 
not included in this proceedings paper. The tectmical report from CMU- 
.665 
CommonLisp. The. sys~:em directly utilizes the I?RAMEKIT -. 
represented domain kn~wA~xlge (currently in the area of cent- 
purer manuals and doctor/patient cot~w.rsations) ~f the CIVIl.\]-. 
MT k~mwledge-based large-scale iuachine lratl.%ttion ~ystem~ 
it handles inpals in both English and Japanese, '~'he current 
:-:ize of the t)IM~-(~'OMIVIAND ~ystem is roughly 5,(X)0 li~e~; of 
ifisp ~;ode (this does riot irtchtde the MT system fimctions 
and the \[?I/AMEK/T l)'ame system, parts of whicii must also 
be loaded into memory) and is not expected to increase, tinct 
the fntam variety in types of commands and questions thai the 
system will ha~dle wilt be 7~ltegrated into the network of mere-. 
try that represents concepts for ~:ommanding and qt~esticm/ 
ailsworhtg aild not iiilo the system code il:self pz. Compiled 
code on IBM-.R'7's and l\[/\[Pg0l?0s is fast enough that parsing and 
l~erforming commanded action happens virtually in ~eal-fimeo 
We are expecti~g to increase die variety in types of system 
fimctions arm grammar/rule development fimctions; however, 
as noted above, since such increases will occur in the mem- 
ory network, as a system implementation, I)M-.COMMAND is 
a completed syslem. 
Ac~ow~edgments 
The autlto~s would like to thank members of the Center for 
Machine Translation for frtfitfal discussions. Erie Nyberg and 
~l~ruko Mitamura were especially helpful in prel~aring the final 
version of ~his paper° 
CMT under the same title c(mtains the sample runs oi Ihc DM-.COMMAND 
on an IBM-RT nnming CMIJ-CommonLisp for development of CMIJ-MT 
project's conceptual entity definitions ~md syntax/semantic mapping r0_les. 
The example input sentences in Japanese include s(nr~e of the ellipses han- 
dlings in discourse lhat are typically problematic ti~r natural la, igtlage inter- 
faces. The system ,also accepts English as the input language. Some of the 
input sentences m'e "*have-a-pain no zenbu no m~pping role o misenasai"; 
"SO fie oya i11o"; "koremade. no o zcmbu nlisenas~i'; and "so no shtlnllyokn 
o takcda san ni okare". 
~'~()ne advantage of DM-COMMAND is that the whole system is only 
5,0(X) lines king and we need not load the whole MT system (which is 
quite largo) for developing grammar and concept entity definitions lind 
writing syntaxNemantics mapping roles. 

References 

\[1\] Becket, J.D. (1975) The phrasal lexicon In 'q-laeoretical issues 
in natural language processing'. Proceedings of the workshop 
of the ACL. Eds. Schank, R.C. and Nash-Webber, B.N. 

12\] Berg, G (1987) A Parallel Natural Language Processing Ar- 
ch#ecture with Distributed Control. In 'Proceedings of the 
CogSci-87'. 

\[31 Bookman, I,.A. (1987) A Microfeature Based Scheme for Mod- 
elling Semantic'.,. In 'Proceedings of the iJCAI-87'. 

\[41 Charniak, E. (1983) Passing Markers': A theory of Contexual 
Ir~Tuence in Language Comprehension. Cognitive Seince 7. 

115\] Charniak, E and Santos, E. (1987) A Connectionist Context- 
Free Pars'er Which is not Context-Free, But Then It is Not 
Really Connectionist Either. In 'Proceedings of the CogSci-87'. 

\[6\] Char~fiak, E (1986)A neat theory of marker passing, h~ q~ro. 
ceedings of the AAAl-86L 

BTi Cullingford, R.tL a~ld B(mtti, S.L (1985) How to mf~'e a 
t-~.a.¢ural..language intetface robust. G~T-\]\[CS.-85/27, Georgia ~1~ 
stimte of Technology. 

\[g\] t~ahhnan, S.E. (1983) NEIL: A system fiJ~" represe~¢bsg (~nd 
t~.sing real-worM knowledge. The MIT P~ess. 

191\] Granger, R.H., Eiselt, K.E (1984) The parallel orguedzagion oj 
lexical, s~vntactic, and pt~gmatie inference proces'ses i,~ '~'~-~: 
ceedings of the First Annual Workshop on Ttaeoi'etica! {sm~c~-; 
in Concc:ptual \]{nfol~iiation l:~ro~essing ', 

\[10\] }{aim, U. and Reimer LI. (1983) Wor#l experg pa,,'sing: .Ae~ 
approach to text parsing with a distribuNd le~ical grt.tmmar. 
Technical Repc:rq Universitat Konstanz, Wear Germaity'. 

\[11\] Hirst, G. and Charniak, E. (1982) }ford Sense a~,d Slot/)is-. 
~,mbiguation. In 'Proceedings of AAAI--82'~ 

\[12\] Hirst, G. (1984.) A Semamie PtvJcess for SyntucUc L)isamb,'gl~u-. 
lion. In 'lh-oceedings of AAAl-84'o 

\[1311 Lytineu S. (1984) The organization of knowledge in a muisi. 
lingual, integrawd parser. Ph.D. thesis Yale Univelzsity. 

1114\] Mitamura, ~Ii, Musha, It., Kee, M. (1988) 7_'lie Generalized \]ff 
Parser/Compiler Version 8.1: User's Guide, ed. "lfbndta~ M.., 
CMU-.CMT-88-MEMO. Carnegie Mellon University. 

\[151 Norvig, R (1987) Ir~'erence in Text Understanding. In 'h'o.- 
ceedings of the AAAI-87'. 

1116\] Nyberg, E. H~ (1988) The t~AMF, Krr User's Guide Versioe~ 
2.0. CMU.-CMqt-MEMO, Carnegie Mellon University. 

\[17\] Quillimi, M.R. (1968) Semantic Memory. In 'Semantic lnfi)r- 
marion Processing', ed. Minsky, M. MIT Press. 

11:181 Qlfillian, M.R. (1969) The teachable language comptwhender. 
BBN Scientific Report 10. 

\[191 Riesbeck, C. (1975) Conceptual Analysis. in 'Conceptual In-. 
formation Processing ~ ed. Schank, R. C. North Holland. 

\[2011 Riesbeck, C. (1986) From Conceptual Analyzer to Direct 
Memory Access Par~'ing: An Overview. In 'Advances in Cog- 
nilive Science 1' ed. Shm'key, N.E. Ellis Itorwo~xt. 

\[21\] Riesbeck, C. and Martin, C. (1985) Direct Memory Access 
Parsing. Yale University Report 354. 

\[22\] Small, S., Cottrell, G. and Shastri, L. (1982) "lbward comtec.. 
tionist parsing. In 'Proceedings of the AAAI-82'. 

1123\] 'Ibmabeehi, H. (1987a) Direct Memory Access Trar~latio~. In 
'Proceedings of the IJCAI~87'. 

\[124\] "lbmabechi, It. (1987b) Direct Memory Access Translation: A 
Theory of Translation. CMU-CMT-87-105, Carnegie Mellon 
University. 

\[25\] Tomabechi, H. and Tomita, M.. Mmmscript. The lmegra 
tion of Unification-based Syntax~Semantics aml Memory-based 
Pragmatics for Real-Time Understanding of Noisy ConUnuous 
Speech Input. 

\[26\] "lbmita, M. mid Carbonell. ,L (1987) The Universal Purser Ar.. 
chitecture for Knowledge-Based Machine Tram'lation. In 'Pro.. 
ceedings of the I.ICAI-87'. 

\[271 Waltz, D.L. and Pollack, J.B. (1984) PhenomenologieaUy 
plausible parsing. In 'Proceedings of the AAAIo84'. 

\[28\] Waltz, D.L. and Pollack, J.B. (1985) Massively Parallel Pars. 
ing: A So'ongly Interactive Model of Natural Language tT~te~. 
pretation. Cognitive Science 9. 
