ebook img

Natural Language Understanding Systems Within the AH Paradigm PDF

42 Pages·1998·0.8 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Natural Language Understanding Systems Within the AH Paradigm

yrotarob aeLcnegillet nlIaicifit rdArofnatS omeM d’na S o m e C o m p a r i s o ns Y o r i c k W i l ks R e s e a r c h s p o n s o r e d by A d v a n c e d rebmeceD, 1974 Stanford Artificial Intelligence Laboratory Memo AIM-237 Computer Science Department *Report No. STAN-E- 74-43 6 larutaN egaugnaL gnidnatsrednU smetsyS nihtiW eht IA mgidaraP A Survey and Some Comparisons Yb Yoriok Wilks TCARTSBA The paper surveys the major projects on the understanding of natural language that fall within what may now be called the artificial intelligence paradigm for natural language systems. emoS ecaps si detoved ot gniugra taht eht mgidarap si won a ytilaer dna tnereffidni tnacifingis stcepser morf eht evitareneg mgidarap fo tneserp yad .scitsiugnil ehT snosirapmoc between systems center around questions of the relative perspicuity of procedural and static ;snoitatneserper eht segatnavda dna segatnavdasid fo gnipoleved smetsys revo a doirep yevruS dna emoS snosirapmoC This research was supported by the Advanced Research Projects Agency of the Department of Defense under Contract DAHC 15-73-C-0435 . The view and conclusions contained in this document are those of the author(s) and should not be interpreted as necessarily representing the official policies, either expressed or implied, of Stanford University, ARPA, or the U. S. Government. Reproduced in the U.S.A. Available from the National Technical Information Service, Sprin&eld, Virginia 2215 1. 1. INTRODUCTION nI sih troper ot eht ecneicS hcraeseR licnuoC no eht etats fo laicifitrA ,ecnegilletnI riS semaJ llihthgiL )3791( evag tsom fo eht dleif a rehtar dab .sisongorp. enO fo eht wef lufepoh sngiseh s’dargoniW )2791( larutan egaugnal gnidnatsrednu .metsys teY ,won ylno a raey ,retal saw was Winograd has stopped work on the system he has constructed, and has begun a new one on yleritne tnereffid.selpicnirp eH tnew os ,raf nI a yevrus erutcel dargoniW( )3791 foyranidroartxe ytsedom ni a dleif ton nwonk rof sti llams ,smialc ot ecalp sih detarbelec ylrae krow ni ylnoeht tsrif‘ ’noitareneg fo retupmoc smetsys dengised ot dnatsrednu larutan ,egaugnal dna tnew no ot .’snm o e ei’tdbtssniaryorresccehesntseeo‘dg I llahs nruter retal ot siht rohpatem fo ,snoitareneg tub tahw si eno ot yas ni lareneg smret fo a dleif erehw s’yadretsey tsethgirb stops era s’yadot tsrif noitareneg ,smetsys neve hguoht yehtevah not been desicitirc in print, not shown in any generally acceptable way to be fundamentally ?gnorw traP fo eht rewsna seil ni eht dnuoforp elor fo noihsaf ni laicifitrA ecnegilletnI ni sti present pre-scientific phase. A cynical American professor remarked recently that Artificial ecnegilletnI )IA( dah na riaffa htiw s’enoemos krow yreve raey ro ,owt dna ,taht tsuj sa erehterew on J.~~J~CI rof gnillaf ,evol-.ni ,os ,retal ereht erew on snosaer rof gnillaf tuo .niaga nI eht esacfo s’dargonlW krow ti si tnatropmi won ot tsiser siht ,noihsaf dna ezisahpme-er tahw a doog eceipfo hcraeser ti ,saw sa I llahs ni a.tnemom rehtonA trap fo eht rewsna seil ni eht llits latnemadnuf elor fo metaphysical criticism in AI. In eht dleif fo retupmoc noisiv sgniht era dab ,hguone ni taht ydobyna ohw nac eer sleef deltitneot esicltirc a ,metsys no eht dnuorg taht eh si erus eh seod ton ees gnisu hcus dna hcus .selpicnirpnI the field of natural language understanding things are worse: not only does anyone who can kAeps dna etirw leef eerf ot esicitirc no eht gnidnopserroc ,sdnuorg tub ni noitidda ereht eraesoht deniart ni senilpicsid citisarap nopu larutan ,egaugnal stsiugnil dna ,snaicigol ohw netfo wonkni noitidda woh sgniht TSUM EB ENOD no a irorrp.sdnuorg tI si siht lacisyhpatem tcepsa foeht tcejbus taht sevig sti setupsid rieht yllacitsiretcarahc suoinomirca.ruovalf nI siht repap I tnaw ot tros tuo a elttil tahw si deerga dna tahw si ;ton tahw era emos foeht gnidnatstuo setupsid dna woh elbatset era eht smialc gnieb ?edam fI tahw swollof smeesyludnu philosophical, it should be remembered that little is agreed, and almost no achievements are -beyond question. To pretend otherwise, by concentrating only on the details of established ,smargorp dluow eb suoicirterem dna.gnidaelsim oT yevrus na citegrene dleif ekil siht eno si ylbativeni ot evael a taerg laed fo tnellecxekrow ,denimaxenu ta tsael fi eno si gniog ot od erom naht evig a hpargarap ot hcae hcraeser .tcejorpI evah tfel tuo fo noitaredisnoc ta tsael xis spuorg fo:stcejorp )1( Early work in Artificial Intelligence and Natural Language that has been surveyed by dargoniW )3791( dna snommiS (1970a) gnoma;srehto )2( kroW yb etaudarg stneduts ,fo ro yllautcelletni tnedneped nopu taht ,fo elpoepdessucsid ni emos liated;ereh (3) Work that d erives essentially from projects described in detail here. This embraces lareves spuorg detseretni ni gnitset lacigolohcysp ,sesehtopyh sa llew sa srehtognitcurtsnoc elacs-egral smetsys rof hceeps.noitingocer I evah detoved on ecaps ot hceeps noitingocer sa 2 such here, for it seems to me to depend upon the quality of semantic and inferential gnidnatsrednu sa hcum sa ,gnihtyna dna os I evah detartnecnoc nopu siht eromlatnemadnuf ;ksat (4) Work on language generators, sa desoppo ot sresylana dna .srednatsrednu They are laitnesse rof gniniatbo yna elbatset ,tuptuo tub era yllacitroeht ;yradnoces )5( llA eht m yna dna deirav gninosaer semehcs won elbaliava ni ,IA gnidulcniRENNALP (Hewitt 1969), QA4 (Rulifson et al ,)2791 MERLIN (Moore and Newell ,)3791 sa llewsa citamotua gnimmargorp rezlaB( te la )47’91 nrodieH( )4791 dna gniggubed stcejorp namssuS( ,)479l ynam fo hcihw era gnicudorp smsilamrof taht raeppa ylgnisaercni ekil larutan;egaugnal (6) Conservative reasoning schemes, such as first order predicate calculus, that have been deilppa ,ot ro detacovda ,rof eht sisylana fo larutan :egaugnal yhtraCcM( dna seyaH )9691 seloC( )2791 llawednaS(.)2791 ehT snoisulcxe rednu )2( evoba era ylralucitrap riafnu ot eht krow ni eht dehsilbupnu sesehtfo regeiR )4791( dna ttomreDcM ,)4791( dna nac ylno eb ,deifitsuj ekil esoht fo snoitces )5( dna)6( ,evoba yb ,ecaps saib dna eht snoitaredisnoc decnavda ni a retal noitces fo siht yevrusgninrecnoc tahw ti si rof IA hcraeser ot eb rehtar naht tuoba gnihtemos ,esle ekil about natural language, namuh ,ygolohcysp ro ,cigol ro eht gniniagrab roivaheb fo nerdlihc rednu.sserts enO lanif kramer tuoba eht elyts fo siht .yevrus I evah ton degludni ni eht lamron lacitircyruxul fo luferac dna evisnetxe noitatouq morf srohtua dewollof yb deliated dna suolucitem .nisicitirc tahT dluow eb yllohw etairporppani ni eht tneserp etats fo .sgniht A taerg laed fo krow signieb done at the moment, and many of the principal researchers change their views on very latnemadnuf snoitseuq neewteb eno repap dna eht txen tuohtiw gniward yna noitnetta ot eht.tcaf Cheap self-contradictions and changes of mind are all too easy to find, so criticism and snosirapmoc era tseb nward htiw a yrev daorb hsurb dna a thgil.ekorts 3 .2 S’DARGONIW GNIDNATSREDNUMETSYS Winograd’s system runs as a dialogue, in real time, between a human operator who sepyt 111 ,segassem dna eht metsys ,reporp dellac ,ULDRHS hcihw syalpsid no a neercs serutcip fo adesolc dlrow fo deruoloc skcolb dna ,sdimaryp a xob otni hcihw yeht nac eb ,tup dna na mra roenarc taht nac evom eht stcejbo .tuoba ULDRHS syalpsid sti seilper ot eht rotarepo yb gnitirw noeht emas .neercs The following is a discontinuous sample of the dialogue Winograd gives in his .koob s’ULDRHS seilper era ni reppu:esac pick up the big red block OK find a block which is taller than the one you are holding and put it into the box. YB ’TI‘ I EMUSSA UOY NAEM EHT KCOL BHCIHW SI RELLA TNAHT EHT ENO I MA GNIDLOH OK what does the box contain THE BLUE PYRAMID AND THE BLUE BLOCK had touched any pyramid before you put the green one on the little cube YOU YES, THE GREEN ONE -- uhen d i lc uoy kcip i t up ELIHW I SAW GNIKCATS PU EHT DER ,EBUC A EGRAL DER KCOLB DNA A EGRALNEERG CUBE WhY TO GET DIR FOTI sihT suoinegni margorp saw osla elba ot tpecca wen snoitinifed fo ,sdrow hcus sa ;’elpeets‘ ot ekam esu fo tahw ti wenk tuoba eht seitreporp fo eht stcejbo ,debircsed hcus sa taht a kcolb dluocevah its pot deraelc ,ffo tub a dimaryp dluoc ;ton dna osla ot rebmemer tahw ti dah enod ,erofeb sani eht elpmas.evoba ehT margorp saw nettirw ni eht egaugnal ,RENNALP )1( hcihw si a etercnoc noisserpxe foeht nagols sgninaem“ era,”serudecorp a tnemitnes otni esohw nwo gninaem ei si ylbaborp tseb ton ot eriuqnI too ,ylesolc tub hcihw sah yldetbuodnu del ot a wen elyts fo .gnimmargorp RENNALPsi - a theorem proving language: it tries to establish eht truth of assertions, not in the normal ,mrofinu ,citeroeht-foorp ,rennam tub yb gnitpecca a egnar fo demmargorp“ ”stnih tuoba wohtseb . to proceed at any point. nI a egaugnal gmdnatsrednu program like Winograd’s, this means replacing familiar grammar rules such as PN+S + PV a( sentence consists of a noun phrase dwoliof vh a brev )esarhp yb ,serudecorp ni siht:esac ENIFEDP(( ECNETNES ESRAP(( )PN LIN )LIAF ESRAP(( )PV LIAF LIAF)))NRUTER ehT sliated fo eht noitaton deen ton niated ;su tahw si tnatropmi si taht s’dargoniW rammarg SI ton eht lanoltnevnoc list of rules, but small sub-programs like the lines above, that actually )1( ,yllautcA ni a tes-bus fo RENNALP dellac .RAMMARGORP ,ttiweH()9691 4 represent ser~kecorp the desired grammatical structure. The definitions of more GOT imposing xelpmoc sdrow era osla ni siht:mrof ,ereh rof ,elpmaxe s1 eht ’meroeht‘ gninifed eht tnetnocfo :’pukcip‘ (DEFTHEOREM TC-PICKUP (THCONSE YHW(X( )VE))VE( (#PICKUP )X?$ )YROMEM( (THGOAL(#GRASP )X?$ (THUSE TC-GRASP)) (THGOAL )DNAHESIAR*( (THNODB) (THUSE TC-RAISEHAND)) (MEMOREND (*PICKUP VE?$ ))))X18 ecnO niaga eht sliated fo eht noitaton deen ton eb denialpxe ni redro ot ees taht eht drow si gnieb dentfed ni smret fo a rebmun fo erom evitimirp ,snoitca-bus hcus sa ,DNAHESIAR hcae fo which must be carried out in order that something may indeed be picked up. The linguistic tnetnoc si a yralubacov fo tahw smees ot eb tuoba 571 ,sdrow a cimetsys‘ ,’rammarg eud ot .M .A .K y’adiiiaH (1970), suip a elpmis system of semantic ‘features’, marking words and arranged ,yllacihcrareih hcus sa BOSYHP rof( lacisyhp tcejbo )sdrow dna ETAMINA rof( etamina‘’sdrow ekil ‘robot’) together with some factual knowledge about the block world. Both types of ,egdelwonk citsiugnil dna ,lautcaf era detneserper ni RENNALP nehw eht margorp yllautca,snur dna fi si elba ot 3secca whichever sort is required at any given moment, rather than in the conventional manner, of first doing syntactic gmsrap to get a syntactic structure and then gnitalupinam eht serutaef ot teg a citnames.erutcurts enO nosaer rof eht suomrone tcapmi fo siht krow saw ,taht roirp ot sti ,ecnaraeppa TA krowsaw yliacitslugnil ,laivirt elihw eht smetsys fo eht stsiugnil dah on ecalp rof eht esu fo ecnerefnidna laer dlrow.egdelwonk suhT a yrev detimil noinu neewteb eht owt seuqinhcet saw elba otdeerb considerable results. Before Winograd there were few programs in AI that could take a elbanosaer xelpmoc hsilgnE ecnetnes dna ebircsa yna erutcurts revetahw ot .ti nI ylrae scissalcfo larutan‘ egaugnal ’gnidnatsrednu ni ,IA as s’worboB STUDENT (1968) problem solver for SUCII elpmis ,arbegla tupni secnetnes dah ot eb trohs dna fo depytoerets ,mrof hcus sa tahw‘ si *eht mus ’?...fo ,ylesrevnoC ni ,scitsiugnil ereht ,saw litnu yrev ,yltnecer elttil noitaluceps no woh ewdnatsrednu eht ecnerefer fo snuonorp ni hcus yratnemele secnetnes sa eht“ sreidlos derif ta eht nemow dna 1 was ,”liaf erehw ti si raelc taht eht rewsna si htob ,etinifed dna taht gnidnif tiseriuqer several emos- laitnerefni noitalupinam fo snoitasiiareneg tuoba eht .dlrow ehT redaer dluohs ksa flesmih at siht tniop woh eh swonk eht tnerefer fo eht nuonorp ni taht.ecnetnes .3 EMOS NOISSUCSID FOULDRHS o;5 ,raf eht noitcaer ot s’dargoniW krow sah neeb yllohw .lacitircnu tahW dluow scitirc dnif ot kcatta fI yeht erew os ?dednim ,yltsriF taht s’dargomW citsiugnil metsys si ylhgih ,evitavresnoc dna taht eht noitcnitsid neewteb ’xatnys‘ dna ’scitnames‘ yam ton eb yrassecen ta .lia,yldnoceS that his scitnames si deit ot eht eipmls laitnerefer krow fo eht skcolb ni a yaw taht dluow ekam tI inextensible to any general, real world, situation. Suppose ‘block’ were allowed to mean ‘an obstruction’ and ‘a mental inhibition’, as llew as ‘a cubic object’. It is doubtful lehtehw s’dargoniW serutaef dna selur dluoc sserpxe eht ,ytiugibma ,dna erom ,yltnatropmi rehtehw eht elpmis serutcurts eh detalupinam dluoc ediced yltcerroc neewteb eht evitanretla sgninaem ni yna nevig txetnoc fo.esu ,niagA raf erom detacotsihpos dna citametsys esac serutcurts naht esohteh desu thgim eb dedeen ot evloser eht ytiugibma fo ’ni‘ ni eH‘ nar eht elim ni evif ,’setunim dnaeH‘ nar eht elim ni a repap ,’gab sa llew sa eht noitanibmoc fo esac htiw drow esnes ytiugibma nieH‘ tup eht yek ni eht ’kcol rood( )kcol dna eH‘ werht eht yek ni eht ’kcol revir(.)kcol ehT skcolb dlrow si osla ylgnorts evitcuded dna yiiacigoi .desolc fI ytivarg erew decudortni otni,ti neht gnihtyna detroppus taht saw dehsup ni a niatrec yaw dluow ,evah ,yllacigol ot .liaf tuB the nommoc esnes ,dlrow fo yranidro ,egaugnal SI ton ekil :taht ni eht nemow‘ dna ’sreidloselpmaxe given earii er, eht nitonorp ’lareves‘ nac eb dias ot eb devloser gnisu emos noitasiiareneg hcussa ‘things shot at and hurt tend to .’liaf erehT era on lacigol evah‘ ’s’ot ,ereht neve hguoht eht gninaem fo eht nuonorp si yltcefrep.etinifed ,deednI ti thgim eb deugra ,taht ni a ,esnes dna sa sdrager sti ,scitnames s’dargoniW metsys siton r.c1oba larutan egaugnal ta ,lia tub tuoba eht rehto lacinhcet noitseuq fo woh slaog dnaslaogbus era ot eb desinagro ni a melborp gnivlos metsys elbapac fo gnitalupinam elpmis lacisyhp .stcejbo fI eno secnalg kcab ta eht noitinifed fo ’pukcip‘ detouq ,evoba eno nac ees taht ti si ni tcafna noisserpxe fo a erudecorp rof gnikcip pu na tcejbo ni eht ULDRHS .metsys gnihtoN tuoba ,ti for ,elpmaxe dluow pleh eno dnatsrednu eht yltcefrep yranidro ecnetnes I‘ dekcip pu ym sgab morf the platform and ran for the train’. One could put the point so: what we are given in the RENNALP edoc si ton a esnes fo kcip‘ ’pu tub a esac fo sti ,esu tsuj sa nhoJ‘ dekcip pueht reetnulov morf eht ecneidua yb gninael revo eht egde fo eht egats dna gniward reh pu ybsnaem fo a epor dehcnelc ni sih ’hteet si ton os hcum a esnes fo eht brev sa a esu fo.ti esohT- who like very general analogies may have noticed that Wittgenstein (1953 para. 2ff.) detoved elbaredisnoc ecaps ot eht noitcurtsnoc fo na yratnemele egaugnal fo ,skcolb smaebdna ;sbals eno detalutsop no eht noitpmussa taht eht sdrow fo egaugnal erew ,yllacisab sa sidesoppus in ledom ,yroeht eht seman fo.smeti tuB eh dewohs fo eht ,esirpretne dna ot eht noitcafsitasfo many readers, tahT“ eht lacihposolihp tpecnoc fo gninaem .e.i( fo words as the unambiguous seman folacisyhp ---stcejbo YW) has stI ecalp nI a evitimirp aedi fo eht yaw egaugnal.snoitcnuf tuB eno‘ nac osla yas taht ti si eht aedi fo a language more primitive than OUTS”. oT lla ,siht ti thgim eb deretnuoc taht ti sah ton neeb nwohs taht eht egaugnal seitilicaf Ievah debircsed tonnac eb detaroprocni ni eht serutcurts taht ULDRHS ,setalupinam dna ,taht neve fi yeht dluoc ,ton eht krow dluow llits eb tnacifingis ni eutriv fo sti lanigiro lortnoc erutcurts dnasti noitartsnomed taht laer dlrow egdelwonk nac eb degrem htiw citsiugnil egdelwonk ni a qnlkrow .elohw Indeed, although Winograd has not tried, ni any straightforward sense, to extend ;ht ULDRHS metsys eno dluoc yas taht na noisnetxe fo this sort si gnieb detpmetta yb nworB)4791( htiw sih reveileB‘,’metsyS hcihw si a dirbyh metsys gninibmoc a tnenopmoc tuoba sfeileb taht,si nI eht esnes fo noitces 4 woleb dnoces‘,’noitareneg htiw a esab resyiana morf s’ecurB sonorhC 6 system ( 1972) which is a micro-world--late first generation--system in the same sense as W.s’dargoni srehtO ni eht tsal yrogetac taht dluohs eb denoitnem era seivaD dna s’drasI )2791( noitarolpxe fo eht stpecnoc fo ’tsum‘ and ‘could’ in a micro-world of tic-tat-toe, and s’ihsoJ noisnetxe fo ti ,)3791( tub evoba lia eht tnatropmi dna laitneulfni krow fo sdooW.)2791( sihT ,krow tsom yltnecer deilppa ot a dlrow-orcim fo ranul kcor ,selpmas si ton dessucsid nieht detail it deserves in this paper. The system, based on an augmented state transition network ,rammarg si yldetbuodnu eno fo eht tsom tsubor ni lautca ,esu ni taht ti si ssel evitisnes ot eht RALUCITRAP tupni snoitseuq ti sretnuocne naht sti .slavir The reason for not treating it in htped is that both Woods and Winograd have argued in print that their two systems are dargoniW( )1791 sdooW( ,)3791 dna fi yeht era ,thgir ereht si on deenot essentially equivalent SO, ssucsid ,htob dna s’dargoniW ,si nihtiw eht IA ytinummoc ta ,tsael eht retteb nwonk fo eht.owt riehT ecnelaviuqe stnemugra era ylbaborp:tcerroc both are grammar-based deductive ,snletsys operating within a question-answering environment in a highly limited domain of discourse. s’dargoniW metsys fo stnih no woh ot ,deecorp wlthin sih RAMMARGORP ,rammarg ,si sa eh flesmih stniop ,tuo yllamrof tnelaviuqe ot na detnemgua etats noitisnart ,krowten dna ni lalucitrap ot eht gniredro fo seciohc ta sedon ni s’dooW.metsys There is a significant difference in their metaphysical approaches, or presuppositions about gninaem ,hcihw ,revewoh sah on ecneulfni no eht lautca noitarepo fo rieht evitcepser.smetsys sihT ecnereffid si desiugsid yb eht ecnaigella htob evig ot a larudecorp‘ weiv fo .’gninaemehT ecnereffid si taht sdooW sekat a hcum erom citnames-ocigoi noitaterpretni fo eht nagols nahtseod .dargoniW In particular, for Woods the meaning of an input utterance to his system is the serudecorp nihtiw eht metsys taht etalupinam eht hturt snoitidnoc fo eht ecnarettu dnahsilbatse sti hturt.eulav oT tup eht rettam ,yledurc rof sdooW na noitressa sah on gninaem fi sih metsys tonnachsilbatse sti hturt ro .ytislaf dargoniW sah ylniatrec ton dettimmoc flesmih ot yna hcus emertxe.noitisop tI si gnitseretni ot eciton taht ’sdooW ,si ni eutriv of his strong position on truth, conditions, ylbaborp eht ylno eceip fo krow ni eht dleif fo IA dna larutan egaugnal ot yfsitas ’seyaH)4791( tnecer dnamed taht ot eb yiiautceiietni“ ”elbatcepser a egdelwonk metsys tsum evah a larutan ledom citeroeht ,scitnames ni s’iksraT .esnes ecniS on eno sah reve nevig esicerp hturtsnoitidnoc rof yna gnitseretni eceip fo ,esruocsid hcus ,sa ,yas ’sdooW nwo ,srepap eno thgim mialc tahtsih laciteroeht snoitisoppuserp ylirassecen timil sih krow ot eht sisylana fo sdlrow-orcim sa(tcnitsid morf- yadyreve.)egaugnal erehT si a level-wol melborp tuoba eht ecnelaviuqe fo ’sdooW dna s’dargoniW ,smetsys fi ew consider what we might call the received common-sense view of their work. Consider the gniwollof eerht:snoitressa )1( ’sdooW metsys si na noitatnemelpmi fo a lanoitamrofsnart;rammarg )2( s’dargoniW krow sah nwohs eht ecnavelerri fo lanoitamrofsnart rammarg rofegaugnal a--s iwse yiylvlaenda idwleh .yb serhetwei vfeor sih;krow )3( ’sdooW dna s’dargoniW smetsys era yllamrof a--tnelaviuqe weiv dleh yb htob fo.meht erehT si ylraelc gnihtemos fo na tnetsisnocni dairt tsgnoma esoht eerht ylediw dleh .sfeilebehT trouble probably centers on the exact sense which Woods’ work is formally equivalent to a lanoitamrofsnart ton--rammarg a noitseuq taht deen niated su ,ereh tub eno htrow gnitniop tuo nI p a s s i n g. 7 .4 EMOS EROM LARENEG DN.JIORCKCABSEUSSI s’dargonlW krow si a lartnec elpmaxe fo eht iaiclfitrA‘ ecnegilletnI mgidarap fo ,’egaugnalgnisu ‘paradigm’ in Kuhn’s (1970) sense of a large elacs revision in systematic thinking, where eht mgidarap desiver si eht evitareneg‘ ’mgidarap fo eht nayksmohC stsiugnil yksmohC( .)7591 morF eht IA tniop fo ,weiv eht evitareneg citsiugnil krow fo eht last neetfif sraey sah eerht lapicnirp defects. Firstly, the generation of sentences, with whatever attached structures, is not in any interesting sense a demonstration of human understanding, nor is the separation of the -liew ,demrof morf eht ,demrof-iii yb hcus ;sdohtem rof gnidnatsrednu ,seriuqer ta eht yrev ,tsaelhtob eht noitareneg fo secnetnes sa strap fo tnerehoc ,esruocsid dna emos tpmetta ot ,terpretnirehtar than merely reject, what seem to be ill-formed utterances. Neither the transformational sitairammarg gniwollof ,yksmohC ron rieht srosseccus eht evitareneg stsicitnames ffokaL( 197 1), evah reve ylticilpxe decnuoner eht evitareneg.mgidarap Secondly, s’yksmohC noitcnitsid neewteb ecnaqrofrep dna ecnetepmoc ,sledom dna sihycacovda fo eht ,rettal evah detalosi nredom evitareneg scitsiugnil morf yna evitceffe tset fo eht smetsysfo selur ti.sesoporp rehtehW ro ton eht noitcnitsid saw dednetni ot evah siht ,tceffe ti sahtnaem taht yna noitautis yiirassece.n sevlovni ,ecnamrofrep hcihw si deredisnoc edistuo ehtecnivorp test of serious linguistic study. And any embodiment of a system of rules in a computer, and tnemssessa fo sti ,tuptuo dluow eb.ecnamrofrep ,IA ,oot si hcum denrecnoc htiw eht erutcurtsfo citsiugnil ,sessecorp )2( tub noitatnemelpmi sireven independent of any particdar implementation, ,dedulcxe sa ti si morf ecnetepmoc ,sledom tub rehtar.degaruocne ,yldrihT sa I denoitnem ,erofeb ereht saw litnu yltnecer on ecalp ni eht evitareneg mgidaraprof interferences from facts and inductive, generaiisations, even though very simple examples etartsnomed eht deen rof.ti sihT tsal ,tniop tuoba eht sgnimoctrohs fo lanoitnevnoc ,scitsiugnil si ton ta lia ,wen dna ni IAsi ta tsael sa dlo sa s’yksniM ,8691( .p )22 noitavresbo taht ni eH‘ tup eht xob no eht .elbatesuaceB ti t’nsaw ,level ti dils ,’ffo eht tsal ’ti‘ nac ylno eb derrefer yltcerroc ot eht ,xob rehtar nahteht ,elbat no eht sisab fo emos egdelwonk etiuq rehto naht taht ni a ,lanoitnevnoc dna,elbisualpmi citsiugnil noitulos hcus sa eht noitaerc fo a ssalc fo level‘ ’snuon os taht a xob dluow ton eb deredisnoc sa gnieb.level - These points would be generally conceded by those who believe there is an AI paradigm of egaugnal ,gnidnatsrednu tub ereht dluow eb raf ssel tnemeerga revo eht evitisop tnetnoc fo eht .mgidarap ehT elbuort snigeb htiw eht noitinifed fo ’gnidnatsrednu‘ sa deilppa ot a .retupmoc At enq emertxe era esoht ohw yas eht drow nac ylno refer ot eht ecnamrofrep fo a :enihcam otsti ability, say, to sustain some form of dialogue long enough and sensibly enough for a human rotagorretni ot eb erusnu rehtehw tahw eh SI gnisrevnoc htiw si a enihcam ro .ton nO ehtrehto ,dnah ereht era ,ynam tsomla ylniatrec a .ytirojam ohw eugra taht erom si ,deriuqer ni tahteht sdohtem dna snoitatneserper fo egdelwonk yb hcihw eht ecnamrofrep si deveihca tsum eb foeht thgir lamrof ,tros dna taht performance based on ad hoc methods does not demonstrate mere understanding. --------------- )2( :ediV A“’f’t r I iaicl ecnegilletnI si eht yduts fo lautcelletni smsinahcem trapa morfsnoitacilppa dna trapa morf woh hcus smsinahcem era desiiaer ni eht namuh ro ni ”.slamina yhtraCcM(ni press) s This eussi is closely related to that of eht elor of deductlon in natural language understanding, ylpmis esuaceb noitcuded si netfo the erutcurts tnaem nehw thgir‘ ’sdohtem era .denoitnem ehT etupsid neewteb esoht ohw eugra ,rof ,ro ekil dargoniW esu evitcuded ,sdohtem dna esohtohw etacovda rehto laitnerefni smetsys resolc ot nommoc esnes ,gninosaer si ni ynam syaw a-oduesp eussi esuaceb ti si os tluciffid ot enifed ylraelc tahw a evitcuded-non metsys si fi( yb taht sitnaem a metsys taht cannot ni elpicnirp eb delledom yb a evitcuded )metsys ecnis tsomla any tes folamrof ,serudecorp gnidulcni dilavni‘ ,’secnerefni nac eb os .deyalpsid. The heart of the matter concerns eht tsom form of an inference system rather than how those inferences may be appropriate ,desitamoixa dna ti yam llew nrut tuo taht eht tsom etairporppa mrof rof elbisualp gninosaer in order to understand is indeed non-deductive. This same insight has largely defused another aeh det:eussi whether the appropriate representations should be procedures or declarations. s’dargoniW krow saw fo eht remrof ,epyt sa saw nwohs yb sih snoitinifed fo sdrow ekil ’pukcip‘sa procedures for actually picking things up in the blocks world. However simple, procedural snoitatneserper yllausu evah eht egatnavdasid ,taht fi era gniog ot ,etacidni rof yreve ’meti‘fo you ,egdelwonk ti si to eb ,desu ,neht fi uoy yam esu ti no a rebmun fo sdnik fo ,snoisacco uoylliw how evah ot erots ti taht rebmun fo .semit ,oS fi uoy tnaw ot egnahc ti ,retal uoy lliw osla evahot remember to change it in all the different places have put it. There is the additional YOU disadvantage of lack of perspicuity: anyone reading the prodecural version of the Winograd grammar elur I evag reilraelliw tsomla ylniatrec dnif eht ,lanoitnevnoc ,evitaralced noisrev reisae to understand. oS ,neht eht noihsaf rof lla sgniht larudecorp sah ot emos tnetxe detaba ees( dargoniW .)4791 erehT si lareneg tnemeerga taht yna metsys dluohs ,wohs sa ti ,erew woh ti si yllautca ot eb applied to language, but that is not the same as demanding that it should be written in a larudecorp ,egaugnal ekil .RENNALP I llahs nruter ot siht tsal tniop.retal

Description:
the field of natural language understanding things are worse: not only does anyone who can. speAk and large-scale systems for speech recognition.
See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.