Reconstructing Spatial Image from Natural Language Texts

This 1)al)(,r drs( 'ribrs tit(' un(h,rstanding l)ro('('ss of the spatial descriptions in ,lal)anese. In order to understatt(I tlw described worhl, tit(' a, it |hors 113" to r('('Ollstru('t tit(' gc(nm,tric model of tit(' gh)bal s('en(' frmn tlw scenic descriptions drawing a spaco. It is done by an experimental ('Omlmter itrogranl SPR INT. whiclt lakes natural language texts att,l l)roduces it nmdel of the described win'hi. To reconstruct the modvl, the altthors extract the qualitative ~patial constraints front the text, and represent Ihel l t ;IS tlw nunierical constraints on the spatial a t t r ibutes of the entities, This makes it lmssibh' to eXln'ess the vaguent,ss of tit(, spatial concepts attd to derive the ntaxinlally plausible interl)retation froltl it ('hllnk of illforltl&tloil ;t('('llnmlated ;m tit(, constraints. The int('rln'ehdi(m r('fleets the tentlmrary 1)clief about the world. 1 I n t r o d u c t i o n This p~ttter (!OltCelltl'3I,(!s (ill l;h(! tllld(wst3.1tdillg proc(!ss of l;he verba l (!xtn'essiolls ('OlWerllillg it[)olt|; space. ()lit, c an easi ly inla.gilw ~11(} (h> sc r ibed worM fl'om the vcrbM express ions . We l'(!gltl'([ the iltgerpl'('.~0.Lit)ll of ({(!s(rriptiolts ;US ;~11 ;tcLiv(! process , t;IIM, is t}m proc(~ss of I'(R:OlIS|il'lt{'t,ion ()flt s i t ua t i on which tim st)eak(w in tended . In this process , tree will use m a n y k inds of infi t rmaLion. T h e natm'a} bulgu~ge desc r ip t i tms c o n t a i n st)Ill(} of th(~lll, itltd it, is vm'y illl])ort, allL to extl ' i tct a n d use t}mnl, b u t they are no t enough . A m o n g t}tem, in fo rn ta t ion a b o u t the configm'at, ion of the wor}d ill (rim's illlltge p lays ;tit inl[)Ol'tltltt, rote. W(! h3,v(. ~ lllltd(! itIl (iXpCl'itil(?ll[,itl (!Olllplltcl" in 'ogrmn S P H I N T (fl,r " 'S}'atial Repres(,'ntI~tion INTtwttrt~ter ') , whi( 'h takes spa t ia l (tescrii)tions wrigtelt ill ,]itl)itll(!,q(L I'{',COllS|.I'/ICt,8 3-dillt(~Itsit)lllt]. nto(l(~l of the wm'}d, and t)llI;plltS thc ct)rl'(}Spollding inlitge on tht~ g r aph i c disp}ay. In this pal ter . We (h~scrib(~ t, lw overvi(,*w of th is syst, enl. 2 T i l e A p p r o a c h T h e esscn('(~ of our a p p r o a c h is its fo}}ows: a ~[(qillillg of thP llal(l(';ll lktll{4111tg(' rxl)rrssions ILs the constraints among tit(' sl)atiM era;ties • lmagv r('l>r('srntation of Ill(, worhl as a collection of tit(' lmram('trrizvd entili(,s • int(,rl)rrting tit(, qualit;ttiv,, r(qations as tit(' mlttWt'il'a[ ('ottstr~lilHs KltlOtlg lilt' p;tr;tttt(qt)t's • Pot(qttbfl (ql(.'rgy fun('tions for lhr vague const raitll ~, • l'~xtracting, th r procedur(, of the ro('onstrll('tiol; froltt lit(' itatttral lallgtlag(' (,xl)t't,ssiotls • Su('('essive refinement mul modification of tit(' worhl model. We r ega rd the w o r m its ml i~ssembly of the spa t ia l ent, iti(~s, a n d r(q)resent each en t i t y a.s the (:onfltimd;i[m of its p l 'o to ty lm lind the real wdues t)}" il:s t)itl';ttlltlt,(!l',q. W(! [n'eplu'e the gl'ap}tic object, s cor r ( !sponding to t im p ro to types . E a c h g r a p h i c (d)ject is r e p r e s e n t e d by the p a r a m e t e r s p resc r ib ing the de ta i l s of it. T h e pa, r a m e t c r s prescr ibe its }ot:ation, o r i en t a t i on , a n d extmd,. Now the t~sk becomes to gen(, 'rate the g r a p h i c ob jec t s co r r e spond ing to the descr ibed ent i t ies }tll(| to (|(!tCl'lllill(? th(~ p}Ll.'itltl(!tCl' values pres(tl'ibing thmn. AcrEs DE COLING-92, NANTES, 23-28 Aour 1992 1 2 7 9 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 *ynt~etic rule diction ~ry object dictlonary ~emal~tic rub" @ 1 7 2 : ...... patlal co'nst raint~s of objel cls }pltranletric reprenel~ ~ (Ioeal) min imum '1 ...... ' ............ I ~)t~ t pu t Image Figure 1: The Overview of the Exper i lnenta l System S P R I N T I t is difficult to deter ,nine the tmrameter values direct ly front the na tura l l anguage descript ions. beeattse of the I)art ial i ty of the informat ion and tim vagueness about the st tat ial rehtt ions alnollg tile entities. So. a t first, we ex t rac t such informat ion as the qua l i ta t ive spa t ia l cons t ra in t s among the spat ia l a t t r ibu tes of the ent i t ies , and then. in te rpre t these COllstrailltS alld calcuhtte the t>a: ran le te r values. This process is shown in figure 1. Given ~ text , S P R I N T makes a sm'faee ease s t ruc tu re using the lexical information. Each ent i ty is descr ibed ms a noltlt . Next, S P R I N T ext rac ts spa t ia l cons t ra in ts abou t the ent i t ies by analyzing the related words in the case. s tructure. At this t ime. S P R I N T also ex t rac t s the sequence of the informat ion references fl'om the lexical mtbrmat ion as dependencies, which are used as cues in the ca lcula t ion of the paranwters. At the nex t stage, the ex t rac ted qua l i t a t ive cons t ra in t s are intertn 'eted as the nmner ica l cons t ra in t s among the e, l t i ty paranletc.rs. These nul l le r ica l COllstrailltS :4re rep l ' e sen ted }kS the COlllbinatiol t of the pr imi t ive constra ints . Tit(! tmten t ial energy function is one of such pr imit ives , and this is an efficient me thod to t rea t the vagllelteSS in the constra ints . Other pr imi t ives are the tot mlogical cons t ra in t s and the regions. The po ten t i a l energy flmction is a kind of the cost fllnetiolls which t~tkes all re la ted paranmters and o u t p u t the ('()st. The less the wthm of the poten t ia l (mergy flmction, the more credit the combina t iml of the gemnetr ic pa ramete r s gains. Using the gradient descendent method, the solut ion wi th n l in innnn cost is calculated. The t)otcntial (mc.rgy flmction 1)rovides a means for accmnula t ing f l 'om fragnmntal 'y infornuttion. (Tit(; b a s i c i ( te~ of the pote , l t ia l ene.rgy fm|e t ion is repor ted in [4].1