Báo cáo khoa học: "SOLUTIONS TO ISSUES DEPEND ON THE KNOWLEDGE REPRESENTATION" ppt

SOLUTIONS TO ISSUES DEPEND ON THE KNOWLEDGE REPRESENTATION Frederick B. Tho~psoH ~li£orni~ Institute o£ Technology Pasadena, C~,li?orni~ In orpQnizing This p~nel, our Ch(tirmon, Bob Moore, expressed the view thor too often discussion o? Hoturra'l l,',nguage occess To dol'o buses has focused on whot p~rticulc~r systems c~*n or cQnnot do, ro'ther than on underlying issues. He Then sd~irr4bly proceeded to orgonize the prJnel nr. ound issues r-qther th~n systems. In responding, I qttempted to ?rr.iMe my ~'emr~rk~, on e,ach o? his five issues in r~ gener~l woy that would not reflec~ ~y ,wn pr4rochiul experience qnd interest, At one point I thought th~.~t I h~d s,cceeded quite well. Howe,.,er~ offer t~king a cleorer eyed view~ it wqs qpparent thor my remarks reflec~c;d ~ssumptions obout knowledge representotion theft were by no Meons univers~ol. This suggests ,a sixth issue which I would like ~o r~omin~t,.,: Are there r'eolly useful generolizNtions ~bou~ comput~Jtionr~l linguistic issues th<~t c~r,e independer~t of r~ssumptions concerning knowledq,.~ r'epre.sentcition? I will come back 'to this sixth issue q?ter discussing t'i~e ?ire cho-~.,n by our Choir~or,. I ~s~,e @i : A.q.ctreq~te_ F.~nc.t i o ns qnd ~uon.!i.S.~. F'irst~ let us csst this issue in o ~omewhot di.fferent way, Irl m~.sny d~tc, b~se ~:ituo.1io~s., there ,',r'e closes of individuals olI of whose Me~bers shcire the ~.:oMs ot'~ributes ~and thus, ÷'rum the point of view o.f the dr~'t~ bose~ ,ir~ .~.ndi~tJ,vuish,~ble. 'Thus there is no need ~o ~dd ,*II of the~.~. , individuals Qs ~':.e. prw, o't(~ entities, To use Bob Moore's ex,a~ple., i? ,:z DEPARTMENT tile h~ ,.~ ~ield Cur NUMPER"OF-EMPL.OYEES~ it strands To r'~~,.~on th~L~' the pczr!icul~:~r [ndividu~.~Is ;~ho ~c~uolly existed in the v,ar'ious dep~r't~tHH~'~ would r~o~ be s~ep,ar,~tely r'~pr'e,>~:~,~ed in the. dc~~obose (for uther'wise there would be o redund,.~ncy whose , :or1~J' :'~.;nc,./ would be h~r'd to police), In ~u,-'h :~i~u,:~io~s we need the. notion of ~'~ ",~olleL'ti,~e," homely ~a single dr~tr~ b,ase object ~'hot ~,.,l(e~ the; pl~.'~c~, of ra number of .ir~divid~.,r, ls end which c,~r~ c,ar'ry their cot~on ,~r!'r.Lbul'e~ together with one ,'~ddition,il item o? in~'ormotion, nomely ~heir r~umber, Thus ~ DEPARTMENT could h~ve ~s o single ~ember su,:h ~ collective ,.,f employees, indeed it could hove severQl such collective MeMbers ond other indiuidu~l MeMbers ~s well. The procedure thor is c~11ed when onswering "how Mony" ~nd "nu(~ber of" questions would know the dif?eren(-e between subcl~s,~es, indiuidurJl (~eMber~ r~nd collective ~embers~ it would know to recurse on subclc~sses, ~.~dd one to its coun'~ for individual MeMbers ond odd the indicoted' number" to its count-for' collec:tJve MeMbers. This ~ppe~rs To be uni?ied ?r,~ework th,,t will h(1~d].e ,~11 of the c:~s,e~ mentioned in Bob Moore's stQtement o? Issue #I. Issue '~2: .T,,iM,e, qnci T_e_p._s_e I should like to split this issue into two, The ?irst sub-issue is the problem of hondling continuusly varying phenoMeno> ~.~uch ra,:, 1he MoveMent of ships~ the chqnging of relotiv,., ~zMount,:~ of ingredients in che~ic~l reQc~ions~ or "the percent completions o9 tnsks, Here it is ~pp(~rent thr~t eoch instonce will require s ~peci~lized procedure to hKLnflle interpol,~tion. Ships cr~nnot s~il Qcross Irjnd~ thus ~n interpolation procedure thor produce~ the position of o ship on The bdJsis of its points o~' dep~u'tur~, ond des'tinotJ, orl will need To know obout the c0cAstlines o? conrinents~ Movements to cheMJ.c~l equilibriums. ~re not line~r~ t~sk coMpletioNs depend on changing personnel ossignf~ents. Just rls we coMputotionol linguists provide to our syste(,~ u~er the (.opobili'¢y to introduce into his dotr~ b~se system ~uch notion.'~ ,a~ loc,=rions of" port.~ end ship~, etc.> we Must ,also provide ~he Met:ins by which he crarl define such '.:ontinuously vr~ryir',~ p,ar(~meters r~s position in such wqys th~.~t ~tpproprJ.ote in~erpo],',tions c~zn be ~de. by ~he generol system in con.junctioN with the p,lrtlculor defini'tJon. For" example, 'the user mr~y de?ine~ "position o? X" in ter~ of ('.r~lcu]:'~%ions, perhops extensive> involvin 9 ~he ,~ctu~.~l 9eo~etr'y of the ~eq~. 169 The second sub-issue on which I would like. to coMmen~ concerns ~hose c~ses where discrete ~ime intervals provide rin r.~dequ(l~(~, r, epresentr.~tion o@ "(he time aspects relevr, nt to the da'tq bqse. In *hese coses~ if` ~he time inf`orf~tion is coMplete~ i,e,, rictu(:l st~Ir~lng ~nd ending rimes of` rill events ripe recorded in ~he d~tQ b,~se, the h~ndling o? time is rrither strriightforwrird, However this c~se of"~en does not ripply, Consider the ?allowing e xQmp], e. : "The Ki~tyhriwk ~rrived in London Monday, The Mriru will soil ?ram London Friday. Will th~ Kitcyhowk ~nd Hrir'u h~ve been in London ~t the s~e time?" One is teP~pted ~o ollow the computer lo give q response: "Possibly," however the introduction of` a three uolued logic is Tr'qught with well known d~ngers of" its own, A More protrricted response gets in the way o~ clnuse [~bedding; how does one hrind].e~ "Will si~ips thrit have been in London together sriil together?" One rins~er would be: "The Ki~i'.vh~.~wk rirrived IriF~t Mondc~y~ the Mor, u will soil next Friday, I~ they will hrive b~.en there qt the same ~iMe~ then not rill ships ~hr~t were in London together will sriil 1'ogether~ bu~ they v:ould be the only exceptions," Choosing ~ relev~mt dic~gnosti,: t,eessoge~ os obove., is o Mrijor r~nd di£?icul't coMputotion,:~l liguis~c issue (:~oing w~],l beyound questions concerning t iMe ~lII,J *,ense, Issue %3 : _~_~r.ULi.f,,.! i.]:Lg ~,p_xo ~u.esx ions This i~ a deep~ philosoph[c,ll ques'~io,~. CoMput~,i. onol linguists hove pr, ogrw:s~ed beyond tl'w. ,:onsiderqtion o? ~:~ing].e ~,.~'ntences, r~nd rire seeking ~o ,?ollow ~che 'focus of ,] dirjlogu~. ~ (And iden'ti.fy 'the theme o? ~ discourse, This is ev~.r}~uqlly ,4n infinite regr'e,~ ultLMritely invo].vi~g cross cul~'ur~l brickgrounds, ~he (perh~/ps Mc, chiclvellion> irlt,~.n~ o~ those ,.,,ha co~,trol ~he u'~e oT o particular ctpplicr:~Lon;. ,:It, Dv~: the eng.i.neering prob].ef~, c~ le,~s'r ~ ~he present eta're o£ • the c~rt,, i~ :~Lapl~_: whr, t response is ~os~ :Jse£u] "~.o ~'he use.r? Consider kwo possible (n~swers ~o ,.'he Following question: "Who ,.~on,.'~ges ec~cl~ deportment?" Ai: "No single person M~nages ~ll o~ the depor~Ments," dept. A ~rin~ger A Unless ~here were ~n undue number o~" dep~rtMents involved~ the second is (.Ice, Ply prei~erred~ ?or it ~uf`f`ices "even i? the f`irst were intended, I. our own experie.,ce, "e~ch" con usefully be in'terpreted ,~s coiling ~or ~, l~b~.led list ,~s onswer in ~l~ost oll coses, The diff"icul¢ies of" being t~ore clever ore great arid will o~°ten result in coMbinqtori(ll explosion, I (~M sur~., for o ]ong tJi~e into the future, we will be seeking simple solutions that (?.~) ore respon,~,ive in Mos~ c(~ses> (b) provide the needed inforMr~eion~ even though redund~nt Jn SOMe cGse~ rind (c) M~ke c:lerir the Misinterpre~rition in the £ew c,~se where this rirjses, even though these solutions May violrite strict linguistic rinqlysis, Issue #4 : ~4e_.,r vi.n~ ~.eM~n~ ic~.!_l y_ CoMp ke X In presenting i~his issue Io the prlnel~ Bob Moore used the ?ollowlng three questions ris ~n example: "Is John Jones (, child oF rln HIT rll U MIIU ~- ? " "Is one o£ John ~ones~s p~rsnts on HIT (ll u~)n u.~ 9" "Did e.i'~her poren~ o~" /ohn ]ones t~ttend HIT?" The appqre.t problem is 'the po.~sibilit~,, of" Multiple des(-rLptions~ o~'ten involving dispor'rite words~ .For, getting ~lt dril:t, in ~he datri h~se, In (.JeBicjrlillg our" systems) we recognize two tru~h~ which ~ppe,~r' to con,flick: (q) the v~lue o.F MiniMizing the reduhdrincy o£ .LnforMf~tion in the dqt(t b,1~e. (b) the necessi-iy o£ non-independent words in the vocobulr~ry, In our' own work~ ,~s Mo.~t o? you know, we hove stressed the use o? definitions c~s u Me~zns of ,'Ich.i.eving o synthe~i~ oF '~-hese *.wo princ:iples. I recoMMe.d it to you u~ ri v~.r~ o~p.i~ul tool in hondlino problems like Bob presents. We illustrate how Bob~s excLMpl~; c,.~n be hr~ndled : "de'fini~ion'child:converse o~" parent ve.rb:John ",;it~end"~ HITmJoh. is ' 'tud~.t,~ o? HIT dei'~inition:~lu~,r-'s'person who hod been ~ student" 170 The ubove three questions then ore ?.~n~1 y zed ~s: "3"ohn )ones is (converse o? parent) o? a person who had been ~ student of HIT?" "One of ~ohn /ones's parents is a person who had been a student o-t HIT?" "W~s ei'ther parent o£ ~ohn ~ones a student of HIT?" I do not wish to slur' over' the fact that ~.= definition Mech~.,nisM ~ust be hifhly :sophis~'~coted in i~s handling of f'ree variables,, bu~ our ~xperience i~dic~*te.~ tha~ ~l'~s can be done quite s~tisfac toril y. Issue #5: Hu~ti-Fil#._~uer'ie_.s This issue has been stated by Bob in terms of G tr'~dixional Multiple file de=to b,~se s'tructure, This issue h~s its coun'ter'p~rt in seM~intic neT data. base structures discussed in pr4per,~ on k~ow].edge representation, Since we use such q semantic net s~ructure For, our data, le't me rephrase the issue in those ~erMs. In Dab's st~tteMerlt of the issue~ he uses tl'~ example of the SHIP file and the PORT .File; wl}ere the SHIP f.ile h,~s fields -For ho~,~¢ port, departure port and destination port. P,~,r'allelLnq his exa~p](:, let us consider ~h~ phrase: "London ship", Suppose ~hr.~t (q) there w~s ship n,~r~ed London, nnd (b) London was a ho~e por~, port of depqrtur~ and des'tir~o'~ion~ not necessarily o~" the same ship, Then "London ship" is four ways ,~Mbifuous~ ~e~ning: (i) the ship London~ (2) London (ho~e port) ships, (3) l.ondon (depr~r~-ur~z par-X) ~hips ~nd (4) London (destinq~ion port) ~hips, In this for~ul~tion of the probleM~ ~II is easy~ insofar ~ the phr~s~; "Londo, ship" is not '.iisc~Mbigu~Ted in con'text~ the user is informed o? the ~lMbiguous M~lrlincjs (Ind the ,~ssoci,:~'ted responses. The difficulty urises when There ar'~. pos.~ibile .i.nterpr,.'<'~'ations ?,~r~her (~field, Fort Collins is n,.~.itl')er ,~ port nor ~ ship~ however ~.he headqunr'ters of the ABC Sllippirt,# CoMpany i~ there un,:l they own ~everol ships. Wh,~'t ?~r'e we ~o ~e~n by "For"t Collins .~hip"? The.~e ~u-e pr'obleM~ tha't wer.e ?irs~ ~1'*:acked by Quillicm, and f qM not ~ur'e ~'t~(~'t unyone I ~.~ c~dded to hi~ !=emii~ol ~r~r~ly':sj.s o£ lhe~, In our own work~ we he*re s~uppecJ at "once re~-1oved" ,.:onnec.'tJons., ,:~ il].u~r~zTed by the four- w~,~y ,~mbiq,)ity ,Ibove. Issue ~6: Solution~ tO. Is~q.es_._D~.~.n As I look back on the abuv~ reMork~ t:oncerning Sob's five issues~ it becomes ~pparent thr~t the u.~efulness of these remarks depends on The degree one is aware of the knowledge representatLon that underlie.s the solution suggested, For ex~Mple~ in the case of the last ,Ls~ue,. il ~ one only knew about traditional file structures~ finding paths theft link fields in More Than one file appears all but unsolvable, Even if one is accustomed to semantic net structures~ the viability of finding connective pnth.~ is highl~ dependccn~ on the existence of back links between attributes and their ,~rgu~en~s and values. Adding a definitional capability~ other thun simple abbreviqtions ~md synonyf4s~ Burns on the way free variables ore handled in 9ener~Jl cmd on the opporo.tus +'or binding theM) for example, in processing the de+'inition: "dei~]nition:are~:length times width" when applied ~o q class> say "areas of ;~hips", how does one ensure ~hat he will ob rain : "lengTh(i) ~k width(i) fop i = I to number of ship~" rather Thonl "lengTh(i) ~ width(j) ?or i~j = i to nut.~ber o? ship =.?" It coMe~ down to how variables qre MainTained in The underlying knowledge represen'tat ion, One is £or'ced to conclude ~hat the basis ~'or the integrcltion of the syntax cu,d ~emonTics o? coMput,~tionr~l linguistic systems i~ -ccoMplished wh n tile d¢ci~4ion~ on knowledge r'epre~en~tiun ~r'e Made, Di~Jcussions 0£ #ur w:~rLous sotut.En.n to ~he J~sues of coMputaTional linguistics can Meaningfully ~uke pl<~ce only in terM~ uf the,:,~ underlying knowledge repr'eser~tot ions. 171 . ro'ther than on underlying issues. He Then sd~irr4bly proceeded to orgonize the prJnel nr. ound issues r-qther th~n systems. In responding, I. ~rrived in London Monday, The Mriru will soil ?ram London Friday. Will th~ Kitcyhowk ~nd Hrir'u h~ve been in London ~t the s~e time?" One is teP~pted