File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/96/c96-2161_metho.xml
Size: 19,954 bytes
Last Modified: 2025-10-06 14:14:19
<?xml version="1.0" standalone="yes"?> <Paper uid="C96-2161"> <Title>Positioning Unknown Words in a Thesaurus by Using Information Extracted from a Corpus</Title> <Section position="3" start_page="956" end_page="956" type="metho"> <SectionTitle> 2 Knowledge Sources </SectionTitle> <Paragraph position="0"> This sectlo. (lescrihes the thesaurus and stal islical da.l;a used in l.his pa.lmr. A Jalm.U(,se uoult t;h('salJrus ca.lh~.<l 1SAMAI' is a. set <)f' IS=A r,qa.lio,shil,s. 1l contains at)out 4,000 nou.s wil, h a.lu)ut tel, h,vels.</Paragraph> <Paragraph position="1"> Each node of \[SAMAI ) is a. woM or a woM a.l,l it, s (one or two) synonyms. Figure 1 shows the t,o 1) categories of ISAMAP. Some words are 1)la,:ed al.</Paragraph> <Paragraph position="2"> mult, iph~ l>osil,i(ms iu l, he thesa,rus, tg),&quot; ,>xantlde , SENSUIKAN (sulmla.ri,e)&quot; is cla,ssilied a.s &quot;wa.l,er wJdcle&quot; and &quot;weapon.&quot; To extract viewpoints For the ,,xlsling sl, ructure o\[' t,}le 1;hesa, llrtls ill oMer to position u,klmwu words in i% a collect, io. of pairs of words a,d t.heir rela~ t;ion markers, togel.her wil.h l.}mir \['requ,mcy, was ex Lra.(:l;ed from acorpus. The source of the words was artMes pu blished in a. aa.p.umse , ewspape r (Ni kkei St,inbun) iu 1993. The a.rl.icles were mOrld,,)h)gi tally analyzed a, utoma.tica,lly, a,,d t t,,', stored i, t,}te following form: oc(wordl, rel, word2) = n This mea.ns 1;ha,l; ~ordl a.nd word2 occur n 1:i rues wil;h a relation marker rel. I(elati<)u markers (:(resist of cause markers such as &quot;G A&quot;, &quot;WO&quot;, alld ~&quot;N I'~'~ and adn<)minal Forms of adjecl,ives a.d a(ljefq,ive nouns. The statisl, ical (lal,a \[',)r each relationshi l) are shown in Figure 2.</Paragraph> <Paragraph position="3"> We use restrictive relatiolmhil)s with 1,he mark ers, rather than wor<l 2-grains, for two reaso,s. In the tluknowli-word=sense disa, ml)iguatiotl task, the n u tuber of possi hi,; ca.di(late word-s(mses (l),)sitions iu the thesaurus, in this paper) is very larg% a,.d thus it; is iml)ortant t,o re(luc(~ noises t\[la, l, l,r(wenl, the output of a result. Secou(\[~ 1;hese case relati,)nships <'an I>e used l,o i(h.ntify classilical, iolt viewpoints for thesauruses. For examph~, suppose that,</Paragraph> </Section> <Section position="4" start_page="956" end_page="958" type="metho"> <SectionTitle> 3 Positioning Unknown Words </SectionTitle> <Paragraph position="0"> 'l'his sectio, descrilms the procedure fi)r positioning ()\[&quot; words iu \[SAMAI ). In this taM<, tim inlmt is a woM to t)e l~la,('ed somewhere in ISAMAI ). 'l'h(~ goal is 1,o (letermine tim most suil, a,t)le area for the woM.</Paragraph> <Paragraph position="1"> The procedure consists (,1&quot; 1he following three steps: Step l: I';xtraclh)n of viewl)ohgs I'or each node in ISAMAP.</Paragraph> <Paragraph position="2"> Step 2: I<xtra,ction or('andidat,,~ areas f.r the input, WOF(\[.</Paragraph> <Paragraph position="3"> Step 3: F, wLhm.tion ol&quot; the ca,,ldldates and selection <)f' the rooM. preferabh+ area+.</Paragraph> <Section position="1" start_page="956" end_page="956" type="sub_section"> <SectionTitle> 3.1 Basic Idea </SectionTitle> <Paragraph position="0"> The I)asic idea. is very simple. For a, unk.ow.</Paragraph> <Paragraph position="1"> wor(l~ I,\[le word l,o-.word rela.l.ions\[lips l.ha.L (~oltl;a, in it, are exl.racte(l. 'l'h,, similaril;y between the word a.nd each ,,,)de in ISAMAP is calculated. The nodes for w hich tim similarity exc(mds a predefi ned l,h reshold ar(' Illal'k(~(t a,u(1 cOIIllCCl;('d ill the l;}H~Sallrtl.S. '\['\[m left tree in Figure 3 shows nodes in 1.he t.hesa.urus.</Paragraph> <Paragraph position="2"> 'l'h(, ma.rke, I nodes are represented I)y hlack circh's. For st raightfi)rwa.rd statistical similarity caJcula.l.i(ms, there are ma W similar words, inclu(tiltg ,()isy words. \[n this l)aimr) the followi,g three hytmtheses a, re used to resolw~ l, he probhmt. First, the marked words \[})rm cerl;aiu areas (connect(~d nodes) <)fwords in tlmt.hesaurus. Tim areas tha.l, occupy a large sl~ace are preferred. The right tree in Figure 3 ~;\[IOWS a,i'c~aS of words.</Paragraph> <Paragraph position="3"> Se(:o,t<l, specific words, that is to sa,y~ words a,t h)wer hwels of trees a, re preferred. In Figure 3, areal is pro.ferre(l to area2.</Paragraph> <Paragraph position="4"> Third, ea.ch node in the thesaurus has viewpoi'nts that distinguish it: from ol.her nodes. The viewlmi.ts fi)r ca,el, node are ext,ract,ed 1)y using case aml modilication relati(mships t;i~a,t contain sta, tis1,teal data extracted \['rom the corpus, lfa, n unknowll word has the sa, me viewpoints as a certain ,ode, t,\[le simila,ril, y tbr 1,he .ode is weighted. 'rh(~ next sub secti(m (h~s(:ril)es how viewlmints axe exl, ra,cted.</Paragraph> </Section> <Section position="2" start_page="956" end_page="957" type="sub_section"> <SectionTitle> 3.2 Extraction of Viewpoints </SectionTitle> <Paragraph position="0"> A viewpoi.t is a set of disl, i.guishing M,,tures \['.r each node i. a thesaurus. The viewpoint of a no(le node is defined ~s a. list, O~od<, marker, word). '\['hough Stl(:h features are implicitly used in the creation of most existing thesauruses according to hilRiaJi in-tuition, they axe lost wheii the constructe(l t\]~le-sara'uses are used. An exception is the Wor(1Net~ in which the distinguishing \[ea.tures a.re nlP~nually listed. In tliis l)aper~ the distinguishing fea.tllres aa'e extracted automaticMly, reflecting the characteristics of the corpus to be used.</Paragraph> <Paragraph position="1"> For example, Figure ,5 shows a f)a.rt of ISAMAP.</Paragraph> <Paragraph position="2"> The viewpoint of a. node in the the.sa.urus is estimated by using a certa.in l)rocedure. Suppose we want to extract the viewpoint o\[' the noun &quot;HF, I{IKOP-UTAA&quot; (helicopter). Tile wor(t occurs 131 1.iines in our corpus. Figure 3.2 shows exa.mples of the rela_ tlonshil)s.</Paragraph> <Paragraph position="3"> For each rela.tionship~ a sea.tell is nlade h)r nodes that have the sa.me relationship. In tile c~se of the pattern &quot;TUKAU&quot; (use), 385 nodes with the s~mm relationship ~tre extra.cted Ijrom a.reas, scattered throughout ISAMAP. On the other ha.nd~ the pattern &quot;TOBU&quot; (fly) shares only two nodes, h.elicopter aaid aiwlaue. The nodes have direct IS-A relationships; in other words, the nodes are c.an be connected in the hierareDy of nodes. Since the viewpoints of a node are inherited by its children in many cases, the existence of the connected nodes that include ISA relationships is strong evidence for the viewpoints. In this case., (fly, SUB) is a view-point for the node &quot;airpla'n% &quot; which is the topmost of the connected nodes.</Paragraph> <Paragraph position="4"> Viewpoints are extracted by calculating th.c typicMness of word-to-wo,Y=t relationsh@s. Given a node nd a'nd its candidate viewpoi.n.t (a pair of a relation marker rel and a wo,d w), the typicalness of the viewpoint is calculated as</Paragraph> <Paragraph position="6"> where N is a set of no&s i'n \]b'AMAP, and C is a set of conuected no&s that contain the: word w. Examples o i the vie'tvpoi,nts (whose typicalness exceeds 0.5.) are as follows: I flying vehicle land vehicle water vehicle / / iic(~ket balloon /% a ear train coach s I air plane helicopter cargo ship patrol boat</Paragraph> </Section> <Section position="3" start_page="957" end_page="958" type="sub_section"> <SectionTitle> 3.3 Example for Positioning Words </SectionTitle> <Paragraph position="0"> in ISAMAP Let us consider ~n exainph~ to see how Mgorithm works. Suppose the word &quot;SEN'I'OUKI&quot; (fighter a) is to be placed in the thesaurus.</Paragraph> <Paragraph position="1"> First;, for each node in the 1SAMAP, the slmil~rity between the word and the node is cMculated. The similarity is ca.lculated according to the rollowing formula.:</Paragraph> <Paragraph position="3"> and the argument &quot;_&quot; can be any words. If the simib~rity va.lue exceeds a pre-delined threshold, the node is marked. Figure 6 shows marked nodes that h~ve high similarity.</Paragraph> <Paragraph position="4"> aIn English, a fighter meltns 1)oth a t)lane and a, person; however, the original Jetpturese word SENTOUKI means only a pla,ne.</Paragraph> <Paragraph position="5"> Areas tha,t co)d;a,i)! ma, rked !moles ~r(: (m,hmla,ti~d. The results ~Lre given in Figure 7. 'l'h(~ )Host suite.hie a.rea, for tim word &quot;fighter&quot; must be s(de(:ted from mull.!pie c~mdida.te sets of con imcth)l!s.</Paragraph> <Paragraph position="6"> The I\]uM liha~se is tim evaJua.thm of tim ca.udi (hm,s. E~L(:h (:aii!(lida.te is (wa.hlail;ed aic('(irding (,o t.l)e fl)lh)wing t'o(!r criteria!.</Paragraph> <Paragraph position="7"> Criterion 1: The size of the ca.,dida.t(~. Giv(,~ ~ri inl2UL word w (in this ca,se, &quot;fight(.'&quot;), a, ud a, nod(: (,b~t is conta,ined i, the ca,ndida, te C, CI \]C,<,<~,;c c' &quot;~i&quot;&quot; (&quot;&quot;, &quot;,od,:). Criterion 2: q'he h(:ight ()f&quot; l.hc (:a.)!dida.t('. C2 is th(; number of levels in the ca,,dida,l,e. For cxa.mI)lc, in the c,~ndidaA;e (a.) in Figure 7, C2 = 2.</Paragraph> <Paragraph position="8"> Criterion 3: The a.ver~ge det)l.h of the nodes. For exa.mph,, the depth of the node &quot;a.irpla.ne', whose. node-id is 0.0.O.0.0.0. l.0.2.0, is 10.</Paragraph> <Paragraph position="9"> Criterion 4: The nund)er of viewpoints. For exa.inple, c~L),lida,lc (a) (whose top node is &quot;human&quot;) ha,s the l~rgest imml-)er (if no(h,s. However, ~s show. in Figure 6, the ma.tclmd rehd, ionshilis (&quot;ba.d hume.u/fight ~r ~nd &quot;l)rotect hulua.u/fighter&quot;) are not typica.l (~xpressions h)r (he word &quot;fighteF'; th~( is, the r(:h~.tii)nsltips a.rc not vii:wpoints. On !hi&quot; other ha.,d, &quot;a.iriih~rm&quot; in cn.ndi(hLte (c) sha.res t\]!(, &quot;fighter (a.irpla),') fly&quot;, which is the viewpi)int of th('. n<)<le &quot;~Lirphm(C' C,4 is the numl)i~r (if ma.tclmd rela,tionships tha.t a.re ('l))tsider(~d as viewpoints of the node ill (,h(~ ca.)ldid~Lte.</Paragraph> <Paragraph position="10"> ) :') Tim totaJ i)refi~r(mcc P(word) is/'I (71 + it 2c,- + p3C3+p4C4, wh(:re p!, p2:/>3, amd P4 axe weights for ea.(:h crit(~rion. Intuitively, ~u!d according to a prelimhl~!'y (,Xl)Crime.t ~ 1,he conl.ril)ution of C3 sho!!ld ca.try more weight tha.n the ol.her criteria. (in our ex-Imrhtte)d,, p) : I, P2 --= I ~ P3 --= 0.4, ~nd P4 = 3). 'l'h(~ mosl. l)r(ff(~ra.ble ca.ndid~Ll.e for l.h(, word &quot;fighttC' is (c); tha.t is, &quot;fighter&quot; is I)la,ced in the +~)'ea whose top node is &quot;tlyi+!g vehich:.&quot;</Paragraph> </Section> </Section> <Section position="5" start_page="958" end_page="959" type="metho"> <SectionTitle> 4 Experiment and Discussion </SectionTitle> <Paragraph position="0"> This section describes some (,Xl)eriments for l)()si )ioning woMs in ISAMAP. Figure 4 shows pa, rt of the rest!Its. In the eXl)erhnent ~ 2,000 nodes with the root &quot;physical obje(:t&quot; in ISAMAt' were used.</Paragraph> <Paragraph position="1"> tionships and the a.ccura,('y (ff 1)ositil)ning of the node. For example, the likely area of &quot;heavy oir' is &quot;((,I)jeet (lnaterla.l (ruel (g~s, l)em)let.~O)))&quot; , whose top node is very a,bsl, ract. \[towew'~r, the relat.ionshi 1) &quot;heavy oil and gas ~i'' suggests the position of &quot;heavy oil.&quot; By ,sing the proposed method, the existing thesa, urus was expanded to cower a large quantity el text. Though ISAMAI ) was designed for general purposes, the method alh)ws it to reflect a specific domain through tile use of a. dt)ma.in-dependent col pus. One o1&quot; our gems is m develop a corpus-based thesaurus, c(msisting I)f' a c()re thesaurus such as ISAMAP and a eorl)us that reflects domain knowledge. When a thesaurus is used for NLI ) applications, such as a.n information retrlewd a,nd disam: ....... , ..... * ,,. bigu~tion system there is no need for it 1.o have 1 lie exDerlll\]ellL Vlel(l(~(1 SeVOI'~LI ODserv;LLIOIIS. VIeV~:- * * &quot; * % ~ , . * .... , well-defined tree-like strtlcl, tlre The system ca31 rise pOlllLS aJ'o sbrollg Cl/les lot (lel,er\[llllllll~ Llle Still,- ....... al)le positions in the thesa, rus for unknown words, the thesaurus a~ a, bl~ck box via certain functions. In co-occurrence based similarity calcula.tion, words with strong similarities but whose relatioashit)s seem st;r~nge to huma.n intuitioa reduce the a.ceuracy of the proposed method. Howew~r, ill trla, lly cases~ these strong similarities are caused by less typica.l co-occurrences. In Figure 6, the words &quot;buy, .... tmr chase,&quot; aim &quot;h~ve&quot; convey less informa.tiw, relation ships than viewpoint relationships.</Paragraph> <Paragraph position="2"> If ttmre are many role.lion,ships \[br an unkuown word, the possibility of the existence of viewpoints will increase. However, some relationships ma.y be noisy. Figure 9 shows the relationships between the mtmber of relationships a.nd the ~ccuraey of positioning. In this case, the a.cctlra.cy mea.ns the percenta,ge of words tot which the most preh~ra.lde area estimated hy the proposed method contained the node theft the word really belonged to. As shown in Figure 9, 50-100 relationshil~s are needed to est.im~te the nodes. On the other }ran(l, too ma, ny relationships prevent the ext.raction of useful view points.</Paragraph> <Paragraph position="3"> It is very di\[\[ictdt 1.o position a, word with pin point accuracy. Experiment showed lhat the fol lowing heuristic is usefll\], If all ullkltoWIl word lies conjunctive relationships with a node (word) in a, pa,rtieular area. it can be positioned a.s a sibling I%)r examph', the following f/l.ctions are needed fi)r a corpus based thesa, tlrtls system: 'positiou(w): ret;,rns the position (or pa.th) of the word w.</Paragraph> <Paragraph position="4"> supcvordi,~al, e(@: returns the superordinate words of the word w.</Paragraph> <Paragraph position="5"> subovdinale(w): returns the subordinate words of tim word w.</Paragraph> <Paragraph position="6"> simihrr(,~): returns the words similar to w.</Paragraph> <Paragraph position="7"> dista'nc~-(wJ, w~): returns the distance between wl and w2.</Paragraph> <Paragraph position="8"> It is important that the return vMues of the %nctions should be depend on the corpus and the locM context of words, s q'he proposed method can t)e used to reMize these functions. Viewpoints make it po.ssible to realize ;t, dynamic interpretation of disLance. null</Paragraph> </Section> <Section position="6" start_page="959" end_page="960" type="metho"> <SectionTitle> 5 Related Work </SectionTitle> <Paragraph position="0"> The method proposed here is rela.ted to two topics in the litera.ture: ~utomatic construction of a thesaurus and word-sense d isa.mbiguation.</Paragraph> <Paragraph position="1"> For this lmrpose., the functions can be exp~nded to contaiu tit(', local conlexl o\[ the word as augmeutations of the functions (e.g. posltion(w, context(w))).</Paragraph> <Paragraph position="2"> Tltl;re tiave I)eeli sevei'a,l st u{lies of' the a, ti I,litnatit + ct}iisl;t:llt;lJ{}tl (}\[' lJtt+sa, tirlises or ,'-;e(,s 1}1 ~ IS-A i'olaJJiin sllil)s \[7, 1\]. In i,ttese si, udies, l tl{&quot; c{)nMruclx,<l rela, tionshi i}s :-;Oitl{~{,inie,'-; {Io ii(il; liiaJ,t2h ti tl itl gt, ll in l, llilio it. IS-A re, la.tionshil)s {1o ilol, a,t)tiea, r itl {;}tc~ {tt)rl)()l';,i, oxt}li{:itly+ a.iid it, is t,}iert,f'{}r{! dillicult 1,(} ('xl, ril.tq t heiti witlitiul, including tltii:-;y re/a, t iliiishilis, Ill otlr ;'t. l) I)roa>ch~ a, (2()1&quot;{2 lJL('S~t, III'IIS iS /IN('~(\] |,O iIIL(B~I'~L(,{+ ~Itltll+l.il int uitit)n with cor l}tis- l ia,sed co occu I'l'{+li{T' in fl)rtita_ l,ion.</Paragraph> <Paragraph position="3"> ga, rowsky i}roiJ()s(~(I a, iilC, l, hod f'{ir word {lisa.tit t}igua,tion using |(o<get's The,~a, urus \[9\]. In his at)f}r()a,ch} a wor(\[ wtlose sr, ilSO:-; ;i,r(' \](il()Wll (;i, word itla,y have severa,\] ,'-;c'iiso,w,) is disa,tiltiigual,e(l I)y liB-Dig &quot;sa.lienl, words&quot; for oa.ch word s(qise. A <~et of. sa, lieiit words is a, llsl, {)f wilt'lib with rio r{qati()nshit}s. I n {lilt&quot; ;i,t>l)roa{'li } wlir{I- 1,o word rela,i,ionsilil}S wii, h iti:-t.rkers a,re used, in ()t'{l(~t&quot; {,{i r{~tIIiC{~ li{}ises a,n{I to (;xl, ra,{;L viewl}oints. 8(}ili{~ ol, ller ineth()ds tip w~ir{lseiisc~ {lisa~litl}igtla,l, iOli liSiiigj W{irdN(% Ira,re t){;+~il pro l}{}scd \[7> 6} 3\]. Their a,l}l)roa,t:hes are siinila, r tl} ()u i':% wiLh~ the difl'{u'ence that tile s(~ltst~ {}f ~L w{}r(I to I}l&quot; placed ill 1;h0, 1,he,,-ia, ill'llS is tlIIkiLoW/t. T}lolts;t, ll(ls (}\[ lit)dos ht lJlo. l;h(~sa, tlrllS ;/,pc: candidat,es, a, tt{l Ll,,re fore} liloro sut'll,le knowh'dge is ne(~<l(~(l. /Is,' (}1' a, core Lhe, sa>tlrLl8 a,tld vlew|}oili{\]s tha, t, is~ word 1,o word relati()nships wiLh rclaLio, n~a,rkcrs makes it l}{)ssible 1,o (~stima,te a. suit,a,I}h~ area for a,i u, known word.</Paragraph> </Section> <Section position="7" start_page="960" end_page="960" type="metho"> <SectionTitle> 6 Conclusion </SectionTitle> <Paragraph position="0"> This tin,per has (lescri\]led a. m<g, ll(id for l)osili(}lt iltg tltlkilowii w(ir(ts ht a, tl exisiJtig Lh(,,'-;a.llrll.'-; /}y t\[,',.;ilig wor(l-l,o-wllr{I re\]a,lJons}iil}S with relaJ, i()n tna.rloers extracl;e(l frotl\] ;t, la,rge {!(ll'\]}tlS. SuiLM>le a,rl)~ks in the/,\]l(;~l+ill;tlS f<}r liliktiOWll WOl'dS wl~l'e es/iuiai,e{l I}y iul,(~p;r;/,1,ing hiiltia,ii iiil, uil,lon ll/irh+{I in Lhe l,}io~a.l\[rils with ~l,atistical (laA, a, ext, ra,{%ed flOiti /Jie {~()rpu:-;. Ex l}l~t'itll(~ti{,s :-;h()w<~{l thai; ;t,s,'-;igllillp~ '&quot;viewl)oinl,s&quot; rt)r e,%ch nolle gives i lilf}orl,ani; in forin al,i(}tl (~a, tl l ic, iise{t 1,(} csCiina.te Stlital}h~ t}osition,~ iti tile lhesauru~ ~t)l' tl~ikiloWii word,<.4. Tile fi}ll(}wiliPPS lX)l}ics ,<-;hoti hi \])e i ti vestigai;ed in ful, tlre w(}rk: ill W\]ioti a,il tlttkiiOWtl wor(l \]las sevoral wordse\[isos, derivative iii(&quot;a, iiilig~ {)f it {oii{I 1,() lie t)tiried a~lii()ii~ the (',a.n{lidal,e,~. if&quot; we take the exli~llllllc ()\[ the wor{I &quot;fi<ghl,cr&quot; use(I in 1,tlis l}a, i)l% &quot;w{~a4}oii&quot; is r(;cl}~itiT,,(~d ;~s a, {~a,,{lillalx~ a, reP~} but is not g;ive~i a sl, r(}ll~ sinlilarity. Otie r(;a,,<-;{)il for tltc prol}hqii is the \]ack o(' view l}()iitt,,~. Morc~ lo{'al {'ontexts o\[&quot; 1,he word ai'(~ tit;t;{led to s\[}o(:ify ~ii{'li lil('~a,iiiit~:-;.</Paragraph> <Paragraph position="1"> I The sitiiila,riLy v~i,lll('~ and viewpoints (;;t,tt t}t ~ u><-;e(l to refine l,h(~ sl\[,l'llCl,/lr{~ o\[&quot; the 1,ht~silAli'ilS. They rna, ke il, l}ossillie 1,o l'}l;t,lL~(': dy\]tami(;a.lly the rela,1;ionshil)s luq,ween words in the the sa, tlrti:-; a,{x:or{ling to d()ina, in-sensitiv(~ cor\[}us.</Paragraph> <Paragraph position="2"> WI~ ~l't~ il()W (hwehll)ittg Lhe getiera, l \['uiiclJ{ins described iii 1Jie i)r(wious se<%ion to realize a large-scale thosatir/is t(}l' NIJ > systoillS.</Paragraph> <Paragraph position="3"> * I\[&quot; tho lllllil\]){~l&quot; of o{~Cllrl'~llC/t~,~ of a,lt ullklloWll wor\[Is is low, the proposed method t, en{ls to (ml,t)ul, larger a.rl,a.+<+ ;i.s<:pO~ith)ns. (_)th('+r con.+ l.ra.ints 8llt~}l it.s rise o/&quot; local (:OlIl,l~Xt ~l,ro rl+.</Paragraph> <Paragraph position="4"> qtl ired.</Paragraph> </Section> class="xml-element"></Paper>