{
    "paper_id": "W89-0122",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T03:44:58.720120Z"
    },
    "title": "Translation of Prepositions by Neural Networks",
    "authors": [],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Translation of prepositions poses a very serious problem to machine translation because prepositions are highly ambigous. In theory prepo sitions can be disambiguated by a filter that excludes already generat ed representational objects with no selection restriction match between preposition and np, but it takes too long time in practice. A neurstl net work makes the disambiguation in fractions o f a second, because it is fast, robust and very powerful.",
    "pdf_parse": {
        "paper_id": "W89-0122",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Translation of prepositions poses a very serious problem to machine translation because prepositions are highly ambigous. In theory prepo sitions can be disambiguated by a filter that excludes already generat ed representational objects with no selection restriction match between preposition and np, but it takes too long time in practice. A neurstl net work makes the disambiguation in fractions o f a second, because it is fast, robust and very powerful.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "In th eo ry , th a t m ea n s fr o m a lin g u istic p o in t o f v ie w , th is c a n b e d o n e in th e fo llo w in g w a y : in th e d ic t io n a r y th e 12 en tries fo r in a re d iffe re n tia te d w rt. th eir ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Linguistic Solution",
                "sec_num": "2"
            },
            {
                "text": "---1---------- 1----------, 1-: 1 State-1 -: 1 temporal-1 I -i 1 1 1 1 non-1 -:-i 1 1 1 State 1----i 1 1 1 individualI -: Inon-1 I-] 1place 1 1 1 1 1 nonindi-1---' 1vidual 1----'",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The Linguistic Solution",
                "sec_num": "2"
            },
            {
                "text": "In T h e c o n c e p t o f s e m a n tic d is ta n c e a n d sem a n tic fitn ess ca n b e o p e ra tio n a liz e d in th e tre e o f s e m a n tic ty p e s . Y o u w a lk in th e tree fr o m th e ty p e w h ich is asked fo r in th e se le c tio n re s tr ic tio n , s te p b y ste p , t o th e ty p e o f th e slotfiller, co u n tin g 1.0 fo r e v e r y s te p t o th e le ft, a n d 0.1 fo r e v e ry ste p t o th e righ t. T h e d ista n ce ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preference Rules",
                "sec_num": "3"
            },
            {
                "text": "fr o m C O N C R E T E (w h ich is se le cte d b y in -1 , P L A C E W H E R E ) t o S C A L E (",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preference Rules",
                "sec_num": "3"
            },
            {
                "text": "-3 . . .X X .. .X ____3 - -2 .. .XX.X............2- -1 .X ..X _X . . . 1 - 0 ..X .X .X ............0 +1 X ___X. .X _____1+ +2 . . . X X . . . X ____",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Preference Rules",
                "sec_num": "3"
            },
            {
                "text": "T h e n eu ra l n e tw o rk is ' tr a in e d ' w ith e x a m p les o f in p u t p a ttern s a n d co rre ct a n sw ers. W h e n th e tra in in g sta rts all th e c o n n e x io n s are ra n ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "...................... 4 - a rg2 : .X X ...X ___ 3 - l o c : .X X .X ...........2 - d i r : .X.. . .X...1- tim e: X.X.X...........0 d ur: . . . X. . X.... 1-)- mea: .XX...X___ 2+ s t a : ...................... 3+ a c t : ...................... 4+ emo: n v p a o i1234567",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "q u a l: c l o t :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "-4 -3 -2 -1 0 +1 +2 +3 +4 .X -1992 . XXX XXX ANSWER xxxxxxxx OUTPUT XX.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "B u t in run nr. 15 th e n e tw ork has ' le a r n e d ' a ru le co m p le te ly , a n d g iv e s th e co r r e c t o u tp u t t o all th e tra in in g sen ten ces. xxxxxxxx It is in terestin g th a t th e e sta b lish ed ru le w ill g iv e th e c o r r e c t an sw er t o n ew sen ten ces t o o , i.e. sen ten ces w h ich h a v e n ev er b e e n g iv en as in p u t p a tte r n b e fo r e .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "In a w a y th e n etw ork has ' le a rn e d ' a lin g u istic ru le in d u c tiv e ly a lth o u g h it has n o t b een fo rm u la te d ex p licitly . It ca n b e seen in th e fo llo w in g th re e e x a m p le s .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "ne w fact SENTENCE: xxxxxxx. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "Mfildet i INPUT PATTERN n v p a o i 1234567 -4 X ___ X ____ X. .4- -3 . .X. X . X ...... 3- -2 X . . .X. .X ..... 2- -1 .X.. X . . . X ____ 1- 0 . .X. X . X ..... 0 +1 ...X . X . . X ____ 1+ +2 X. . .X ....... X2+ +3 ............. 3+ +4 ............. 4+ n v",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Rules in Neural Networks",
                "sec_num": "5"
            },
            {
                "text": "Proceedings of NODALIDA 1989, pages 237-249",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Proceedings of NODALIDA 1989",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "The Neural Network Design S o in th e o r y it c a n b e d o n e , a n d th e h u m an b ra in m u st fo llo w a ru le like th e o n e d e scrib e d w h en it ca lcu la te s th e c o r r e c t rea d in g in fr a c tio n s o f a s e c o n d , b u t it m ust d o it in a sm a rter w a y th an b y c o m p a r is o n in p a irs o f a lre a d y g e n e re te d o b je c t s . T h is sm a rter w a y m u st b e s o m e th in g like w h a t is ca lle d a n eu ra l n e tw ork , w h ich is a s tra te g y fo r p r o g ra m m in g th e p re fe re n ce ru le s o th a t th e m a ch in e ca n c o m p u te th e b e st so lu tio n o f th e p r o b le m in fr a c tio n s o f a s e c o n d , like th e hum an b ra in d o e s.T h e se m a n tic n e tw o rk is d e sig n ed in th e fo llo w in g w a y ; It co n s is ts o f th ree layers, an in p u t la y er w ith 117 n eu ron s, a h id d e n la y er w ith 65 n e u ron s, a n dan o u tp u t la y er w ith 12 n eu ron s. A ll in p u t n e u ro n s a re c o n n e c te d w ith a ll th e h id d e n la yer n eu ron s, a n d all th e h id d e n la y er n eu ro n s a re c o n n e c te d w ith all th e o u tp u t layer n eu ron s. T h a t m ea n s th a t th e re are 767 0 c o n n e x io n s b e tw e e n layer 1 an d 2, an d 792 c o n n e x io n s b etw een layer 2 a n d 3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "t o fo r m u la te th is r e g u la rity as a lin g u istic ru le, n o t e v en as a p refere n ce rule,b e c a u s e o f th e p o s s ib ility o f th e sen ten ce: sh e w ork ed in tw o room s. T h e sem an tic n e tw o r k w ill u tilize th e p r o b a b ilis tic in fo rm a tio n b u t n o t m a k e e rrors in this c r u c ia l e x a m p le , b e c a u s e th e p a tte rn o f c o n n e x io n w eigh ts has lea rn ed th e rule fo r th e c o m b in a t io n o f ca rd in a l n u m b ers a n d m ea su re n ou n s, n o t fo r ca rd in a l n u m b e rs on ly .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Parallel distributed processing. Explorations in the Microstrucure o f Cognition. Vols. 1-3",
                "authors": [
                    {
                        "first": "David",
                        "middle": [
                            "E"
                        ],
                        "last": "Rumelhart",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "James",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mcclelland",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Pd F Sesearch Group",
                        "suffix": ""
                    }
                ],
                "year": 1986,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rumelhart, David E., James L. McClelland and the PD F sesearch Group. 1986. Parallel distributed processing. Explorations in the Microstrucure o f Cognition. Vols. 1-3. MIT Press, Cambridge, Mass.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "type_str": "figure",
                "text": "T h e tra n sla tion o f p r e p o s itio n s p o se s a v e ry seriou s p r o b le m t o m a ch in e tra n s la tio n b e ca u se p r e p o sitio n s are h ig h ly a m b ig io u s -ea ch o f th e m o s t 10 freq u en t p r e p o sitio n s in o n e o f th e 9 E U R O T R A la n g u a g e s is tra n sla te d in to 10 d ifferen t p re p o s itio n s in ea ch o f th e 8 o th e r la n g u a g es-a n d b e c a u s e p r e p o s itio n s a lw ays w ill g e n era te m a n y a tta ch m e n t p a tte rn s. T a k e th e e x a m p le : L en in w r o te th is n o te in his n o t e b o o k in 3 m in u tes in C o p e n h a g e n T h e r e are th e fla t s tru ctu re a n d th e d e e p s tru ctu re a n d 6 a tta ch m e n t p a tte rn s in betw een : l ____ P R E D ________ ARG2 It is o b v io u s in th is e x a m p le th a t in d o e s n o t m ean th e sa m e th in g in th e th ree cla u ses: in his n oteb o ok , in 3 m in u te s a n d in C op en h a g en . T h e first o n e m ean s D I R E C T I O N , th e s e c o n d o n e : T I M E : H O W L O N G , a n d th e th ird o n e : P L A C E W H E R E . In o th e r w o r d s , w e n eed (a t le a st) th ree le x ica l en tries fo r th e w ord in. O t h e r en tries a re n e ed e d fo r in 1897, in a n g er, in d anger, in varioris colou rs, w ith th e m a ilin g s: TI M E W H E N , IN A M O O D O F , D U R I N G A N A C T I V I T Y O F , O F A Q U A L I T Y O F . S oin D a n ish th ere are, in m y o p in io n , a t least 12 d ifferen t m ea n in g s o f th e c o r r e s p o n d in g p r e p o s itio n t (a n d fu rth e rm o re all the fix e d p h ra ses, e.g . in all, to be in f o r ) . T h e d is tin c tio n b e tw e e n D I R E C T I O N a n d P L A C E W H E R E is n o t m a d e in tu itiv ely , b u t w ith th e s o ca lle d ' n o n s e n sica l c o n ju n c tio n re d u c tio n te s t' , w h ich sa y s: i f th e c o n ju n c tio n o f tw o c o n te x ts t o th e sa m e w o r d d o e s n o t m ak e sense, th ere a re tw o re a d in g s o f th e le x ica l ite m , o n e fo r ea ch c o n te x t. S o it d o e s n ot m a k e sen se t o sa y : L e n in w ro te this n o te in his n o te b o o k and three m in u tes, o r L e n in w ro te th is n o te in his n o te b o o k and C op en h a g en , o r L en in w ro te this n o te in th ree m in u te s an d C o p en h a g en o r L e n in w ro te this n o te in th ree m in u tes and 1897. B u t n o te th a t c o n te x ts o f th e sa m e ty p e , i.e. co n te x ts o f th e sa m e rea d in g o f a g iv en w o r d fo r m c a n b e c o o r d in a te d a n d m a k e sense: L en in w rote this n o te in a n g e r and d isa p p o in tm en t.I f w e p a rse th e se n te n ce u sin g a d ic t io n a r y w ith 12 d ifferen t en tries fo r th e w o r d in, a n d 8 d ifferen t a tta ch m e n t p a tte rn s , th e m a ch in e w ill g en era te 8 x 12 x 12 X 12 = 13 .8 2 4 differen t re p re se n ta tio n a l o b je c t s i.e. rea d in g s o f th e sen ten ce. S o th e p r o b le m is h o w t o d is a m b ig u a te th e se n ten ce a n d m a k e th e m a ch in e find th e c o r r e c t r e a d in g a m o n g th e 13 .8 2 4 syn teictica lly p o s s ib le readin gs.",
                "uris": null
            },
            "FIGREF1": {
                "num": null,
                "type_str": "figure",
                "text": "s e l e c t i o n r e s t r i c t i o n s : in _ l (P L A C E W H E R E ) taJres as a rg u m en t 1 n o u n s o f th e ty p e C O N C R E T E , in J (T I M E W H E N ) tak es n o u n s o f th e ty p e Y E A R o r D A T E , in -3 (M O O D ) takes n o u n s o f th e t y p e C O G N I T I O N O R E M O T I O N , in_4 (T I M E H O W L O N G ) takes n o u n s o f th e ty p e S C A L E . T h a t m ea n s th a t all n ou n s in th e d ic t io n a r y a re c o d e d w ith a s e m a n tic ty p e la b e l. I h a v e su g g e ste dth e fo llo w in g set o f s e m a n tic ty p e s fo r n ou n s:",
                "uris": null
            },
            "FIGREF2": {
                "num": null,
                "type_str": "figure",
                "text": "ish ex a m p le s : P A R T I T I V E : s e k to r , sid e, halvdel, S E M I O T I C : afsn it, fo rsla g , aftale, S C A L E : m e te r , decibel, grad, m in u t, dag, T I M E : e fte rk rig stid en , fr e m t i den, Q U A L I T Y : id e n tite t, st\u00f8 r r e ls e , lae n gd e, R E L A T I O N : afh ae n gigh ed , fa k to r , p o sitio n , R E S U L T : p ro d u k tio n .2 , u n d tag else, in v e s te r in g .2 , C O G N /E M O T I O N : in te r e s s e , fr y g t, glae d e, A C T I V I T Y : d atabehandling, a n v en d else, p ro d u k tio n .1 , A C C O M P L I S H M E N T : rev o lu tio n , in v e s t e r i n g .l, u d forsk n in g, P R O P O S I T I O N n ou n s: fo rd el, m ulighed, p rob lem , N O M I N A A G E N T I S : fa b rik a n t af, tilsk u er til, h erre o v er, O R G A N I Z A T I O N : h jem m em a rk ed , in d u stri, reg erin g , C O M M U N I C A T I O N T O O L : p erso n d a ta m a t, v id eob \u00e5 n d op ta g er, radio, P L A C E : E uropa, Computational Linguistics -Reykjavik 1989 K \u00f8 b en h a v n , ild lin jen , M A S S : vand, luft, sand, N A T U R A L K I N D : blom st, tree, s ten , P A R T : kredsl\u00f8b, svin gh ju l, ta ster, W H O L E : d ataanlae g, elektron ik, in fo r m a tion stek n olog i. T h e se le c tio n r e s trictio n c o u ld w o rk as a filter w ith a ' k iller ru le ' , i.e. a rule th a t w o u ld e x c lu d e ( ' k ill' ) all o b je c t s w ith n o m a tch b e tw e en th e ty p e w hich is ask ed fo r in th e s e m a n tic fra m e s p e cifica tio n o f th e h ead , in th is ca se th e p r e p o s itio n , a n d th e ty p e o f th e n o u n th a t fills th e s lo t. T h e r e w ill b e n o m a tch in th e c r e a te d o b je c t w ith in _ l (P L A C E W H E R E ) in th e cla u se in three m in u tes, b e c a u s e m in u te s is a n o u n o f th e t y p e S C A L E , a n d in -1 o n ly selects n ou n s o f th e ty p e C O N C R E T E . T h is ty p e o f ru le w o u ld e x c lu d e m o st o f th e n o t w a n ted o b je c t s a m o n g th e 1 3 .8 2 4 g e n e ra te d re p re se n ta tio n a l o b je c t s . B u t th e ru le is t o o stro n g , b e ca u se it is n o t u n c o m m o n in n a tu ra l te x ts t o fin d m e ta p h o rica l o r slig h tly m e ta p h o rica l sen ten ce s, e .g .: T h e situ a tio n th rea ten s to becom e w o rse . In th is ca se th e selec tio n ru le sa y in g th a t th e v e r b th rea ten o n ly talces n o u n s o f ty p e H U M A N as a r g u m e n tl w ill ' k ill' all th e g e n e ra te d o b je c t s s o th a t n o a n a lysis o r tra n sla tion w ill b e p r o d u c e d a t all.",
                "uris": null
            },
            "FIGREF3": {
                "num": null,
                "type_str": "figure",
                "text": "ste a d it is n ece ssa ry t o u se a p r e f e r e n c e ru le th a t co m p a re s all rep resen ta t io n a l o b je c t s g e n e ra te d fr o m th e sa m e su rfa ce stru ctu re , ran ks th e m w rt. in te r n a l s e m a n tic fitn e s s , a n d selects th e fittest. A s sh ow n in th e first p a ra g ra p h th e s im p le e x a m p le L e n in w ro te this n o te in his n o te b o o k in th ree m in u tes in C o p en h a g en w ill g e n e ra te 8 a tta ch m e n t p a tte rn s w h ich th e n ca n h ave 12 differ en t re a d in g s o f e a ch o f th e th ree p r e p o s itio n s . W h a t is c o m p a r e d b y a preferen ce ru le is n o t tw o cla u ses co n ta in in g th e sa m e tw o o r th ree w ord s, b u t th e sum o f th e s e m a n tic d ista n ce s b e tw e e n all th e p a irs in th e sen ten ce o f 1) a selection r e s tr ic tio n b e a r in g h e a d a n d 2 ) th e c o r r e s p o n d in g slo t filler, a d d e d u p at th e to p n o d e .",
                "uris": null
            },
            "FIGREF4": {
                "num": null,
                "type_str": "figure",
                "text": "th e t y p e o f m in u te s ) is 1 .3, w h ile th e d is ta n c e fr o m S C A L E (w h ich is s elected b y T I M E H O W L O N G ) a n d S C A L E is 0 .0 . C o n s e q u e n tly rea d in g t n^, T I M E H O W L O N G is s e le cte d in th e cla u s e tn th ree m in u tes. T w o rep re sen ta tion a l o b je c t s , tw o tre e stru ctu re s rep resen tin g tw o w h o le sen ten ces, ca n th en b e c o m p 2ired in th e fo llo w in g w a y : 3 + 9 .9 + 0 .1 + 9 .9 B y su ch a p referen ce ru le th e fla t s tr u c tu r e w ith th e D I R E C T I O N re a d in g , th e H O W L O N G rea d in g a n d th e W H E R E rea d in g , re sp ectiv e ly , w ill b e s e le cte dan d th a t is e x a c tly th e c o r r e c t o n e a m o n g th e 13 .8 2 4 p o s s ib le rea d in g s. B u t th is m a ch in ery w ill o n ly w o r k in th eo ry . T h e c o m p a r is o n a m o n g th e o b je c t s w ill b e m a d e in p a irs, s o th ere w ill b e m a d e 1 3 .8 2 4 /2 x 13.825 c o m p a r ison s a n d th a t w ill ta k e a p p r o x im a te ly 6 | h o u rs w ith a fa st m a ch in e a n d a fast p rog ra m . Computational Linguistics -Reykjavik 1989 E a ch o f th e o u tp u t n e u ro n s rep resen ts o n e o f th e p o s s ib le rea d in g s o f the p r e p o s itio n in. T h e 12 re a d in g s o f in M e : A R G l (d e e p s u b je c t ), A R G 2 (d e e p o b je c t ) , L O G (p la c e w h e r e ), D I R (d ir e c t io n ), T I M E , D U R (tim e h ow lo n g ), M E A (m e a s u r e ), S T A (s t a te ), A C T (a c t iv it y ), E M O (e m o t io n /c o g n it io n ), Q U A L (q u a lit y ), C L O T (c lo th e s ). T h e in p u t is a p a tte r n o f th e s y n ta c tic a n d se m a n tic stru ctu re o f a sen ten ce co n ta in in g th e w o r d in. 4 w o rd s t o th e left an d 4 t o th e righ t o f th e p re p o sitio n a re rep rese n ted in th e p a tte r n as s y n ta c tic -s e m a n tic c a te g o rie s. A g iven w ord b e lo n g s t o o n e a n d o n ly o n e o f th e fo llo w in g 56 c a te g o rie s, w h ich in clu d e th e s e m a n tic fe a tu re s d e s c r ib e d in th e first p a ra g ra p h : p r e p o s i t i o n a l ob j e c t s and th e p r e p o s it io n i , t i l , f r a , o v e r , under, f o r , arf, ved PREPOSITIONS: I , PA, T IL , FRA, OM, FOR, AF, MED, UDEN, OVER, n a tu ra l w a y t o rep resen t th e 9 w o r d in p u t p a tte rn w o u ld b e a n array w ith 5 0 4 n e u ro n s o r d e r e d in 9 ro w s a n d 56 c o lu m n s . B u t th a t w o u ld b e a v ery re d u n d a n t re p re se n ta tio n , b e c a u s e o n ly 9 o f th e 504 n eu ro n s w o u ld b e a ctiv a ted in e a ch sen ten ce. T h e in p u t p a tte r n in fo r m a tio n ca n b e rep re se n ted b y o n ly 117 n eu ron s orga^ n ize d in an a rra y w ith 9 ro w s , o n e fo r e a ch w o r d in th e sen te n ce w in d o w , a n d 13 c o lu m n s , in w h ich ea ch o f th e 56 ca te g o r ie s is rep resen ted b y 3 X in a c c o r d a n c e w ith th e fo llo w in g c o d in g k e y (n = n o u n , v = v e r b , p = p r e p o s itio n , a = oth er, o = z e r o , i = 1): __ n o _______n i _____ v o _______ V I ________ p o _____p i ________ a o ________ a i a l e __ p r o p ___t r i -v b __V P O lo c___a f _____f \u00f8 r _______e f t e r ____ a . a d v _ 7 T h a t m ean s th a t th e c a te g o r y I N T R A N S I T I V E V E R B O F T H E S T A T E T Y P E is represen ted b y u o 2. A s a n e x a m p le th e se n te n ce D e t s k e r i 1 9 9 2 ( 'it h a p p e n s in 199 2 ' ) h as th e rep resen ta tion sh ow n b e lo w : SENTENCE: XXX . D e t s k e r -i -1 9 9 2 . XXX XXX INPUT PATTERN n v p a o i 1 2 3 4 5 6 7 -4 ............................... 4 -",
                "uris": null
            },
            "FIGREF5": {
                "num": null,
                "type_str": "figure",
                "text": "2+ +3 ...............................3+ + 4 .............................o u tp u t is represen ted b y 12 ' th e r m o m e te r s ' w h ich sh o w h o w m u ch a g iv en n eu ron , rep resen tin g o n e rea d in g o f th e p r e p o s itio n i, is a ctiv a te d : a r g l : .................. a r g 2 : .................. n a p a tte rn o f in p u t n e u ron s is a c tiv a te d th e n e u ro n s ' fire' , i e. th e y a ctiv a te all th e h id d e n n eu ro n s th e y a re c o n n e c te d w ith , w ith th e w eig h t o r s tre n g th w h ich is a ssign ed to th e s p e cific co n n e x io n . E a ch o f th e h id d en n eu ro n s is n ow a c tiv a te d b y th e su m o f th eir in p u t valu es, w h ich is d e p e n d in g o n b o th th e p a tte rn o f th e firin g in p u t n e u ron s a n d th e w eigh t o f th e th e ir co n n e x io n s . T h e h id d en n e u ron s o n ly fire i f th eir a ctiv a tio n value e x c e e d s a certa in th re sh o ld level, a n d th e o u tp u t n eu ron s are a ctiv a te d in the sa m e way.",
                "uris": null
            },
            "FIGREF6": {
                "num": null,
                "type_str": "figure",
                "text": "d o m iz e d , an d the o u tp u t o f th e n e tw o rk w ill in th e b e g in n in g b e ra th e r in c o r r e c t. T h e n th e co rre ct an sw er is t y p e d as a s e c o n d in p u t, a n d b y a p ro ce s s ca lle d b a ck p ro p a g a tio n allth e c o n n e x io n w eigh ts a c tiv a te d b y th e in p u t se n te n ce are ch a n g ed . T h e c o n n e x io n w e ig h ts y ie ld in g c o r r e c t o u tp u t are in crea sed a n d th e co n n e x io n w eightsy ie ld in g in c o r r e c t o u tp u t a re d e cre a se d w ith a ce rta in rate. B e lo w I m e n tio n s o m e o f th e 100 D a n ish in p u t sen te n ces-o r rath er strings o f 9 w o rd s, th e ce n tra l w o rd i a n d 4 w o rd s to th e left a n d to th e rig h t-an d the c o r r e c t an sw er, i.e. th e b e st re a d in g o f th e p r e p o s itio n i in th e c o n te x t.20. sik re at h v e r d elta g er -i-sa m m e p r o je k t i h ele = A R G 2 21. d elta g er i sa m m e p r o je k t -i-h ele p r o je k te ts l\u00f8betid til = D U R22. d o m in e r e d e tte m arked og -i-stig en d e om fa n g e k sp o rtere fr a = M E A23. n u e r u n d e r o v e r v e je ls e -i-n ogle a f de s t\u00f8 r r e m e d le m s s ta te r = L O C24. til e t s\u00e5 d a n t n y t progra m -i-s t o r m \u00e5 lesto k e r k o m m e t = M E A 25. E s p r it v elk o m m en til m \u00f8 d e t -i-ju n i 1 992 og g od k en d te = T IM E 26. X X X . D e t s k e r -i -199 2 . X X X X X X = T IM E 27. s o m a n v en d es i o p e r a tio n e r -i-m a n g e v e r s io n e r og v a r ia n te r = M E A28. og a fp r\u00f8 v n in g a f V L S I/ s y s te m e r -i-cilisiu m e lle r andre halvledere = L O C29. beslu tn in g en d e n a fg \u00f8relse -i-r\u00e5 d et = L O C30. fu ld t kan s t\u00f8 t te b ru g eren i k o m m u n ik a tio n sp ro cessen , og s o m = A C T 31. ; de vil resu ltere i n y e prod u k ter, p r o c e s s e r = A R G 232. a n v en d else fo r e g \u00e5 r m e g e t L a n g so m m ere i E u rop a en d i Japan = L O C33. a f alle v a r e r fr e m s tille t i fae llessk a b et e r i sm \u00e5 = A R G 2W h e n th e c o n n e x io n stre n g th s h ave b e e n a d ju s te d a n u m b er o f tim es w ith a n u m b e r o f in p u t sen ten ces th e p a tte r n o f th e c o n n e x io n stren g th s w ill represent a ru le w ich w ill y ie ld th e c o r r e c t o u tp u t t o ea ch o f th e in p u t p a tte rn s in th e tra in in g set.It is essen tia l th a t th e in p u t sen ten ces a re a u th e n tic a nd n o t g ra m m a r b o o k se n ten ce s, b e c a u s e a ll re g u la rities in th e in p u t m a teria l, even th e n u m b er o f w o r d s fr o m th e w o r d i t o th e p u n c tu a tio n m a rk , w ill b e m a d e in to a ru le b y th e n e tw ork .It is essen tia l t o o th a t th e n u m b e r o f in p u t sen ten ces is s o h igh th a t all n onlin g u is tic reg u la ritie s o f a n y k in d axe e x c lu d e d . 100 in p u t se n ten ces are certa in ly n o t e n o u g h t o m a k e su re th a t all n o n im p o r ta n t w o r d ty p e s h ave b e e n p la ce d in all 8 p o s itio n s in th e in p u t p ictu re .",
                "uris": null
            },
            "FIGREF7": {
                "num": null,
                "type_str": "figure",
                "text": "I a m n o t su re th a t a w in d o w o f 9 w o rd s is e n o u g h , b u t in th e first 100 a u th e n tic 9 w o r d in p u t sen ten ces th e ru le trig g e rin g w o r d has b een presen t. T h e tra in in g ca n b e seen fro m tw o screen p ictu re s, th e first o n e s h o w in g in p u t p a ttern , asw er an d o u tp u t o f run n o. 2 o f se n te n ce n o . 26. T h e o u tp u t is n o t even slig h tly in th e righ t d ire ctio n .",
                "uris": null
            },
            "FIGREF8": {
                "num": null,
                "type_str": "figure",
                "text": "have n o t y e t-efter 100 in p u t sen ten ces-sta tistics a b o u t h o w m a n y p e rce n t o f c o r r e c t ' gu esses' th e n e tw o rk w ill m a k e a b o u t n ew sen ten ces, b u t it is a lre a d y clea r th a t it is p o s s ib le t o m a k e a n e tw o rk w h ich ca n s o lv e th e p r o b le m o f d isa m b ig u a tio n o f p r e p o sitio n s w ith o u t th e e n o r m o u s o v e rg e n e ra tio n w h ich is m a d e b y filter rules in serial p ro g ra m m in g . It sh ou ld a c c o r d in g t o th e th eorie s b e p o s s ib le t o tra in th e sa m e n e tw o rk t o m a k e th e d isa m b ig u a tio n o f all th e p r e p o s itio n s (o r all th e m o s t freq u e n t a n d a m b ig u o u s p r e p o s itio n s ). T h e n e tw o rk I h a v e d e s c r ib e d is in fa c t d e sig n ed to c o m p u te 15 difieren t p re p o s itio n s . B u t I h a v e n o t y e t tra in e d it w ith o th e r p re p o s itio n s th\u00a3in t. 6 The Power o f Neural Networks I im a g in e th a t th e n eu ra l n etw ork in th e tra n sla tio n p r o c e s s w ill b e p la c e d b e fo r e th e parser. T h e n e tw o rk is fed w ith th e le x ica l w o r d s o f th e in p u t se n te n ce , a n d th e relevan t in fo r m a tio n a b o u t th e s e m a n tic t y p e o f e a ch w o r d ta k en & o m th e d iction a ry . A ll th e p r e p o s itio n s in th e se n ten ce are th en d is a m b ig u a te d b y th e n e tw ork a n d th e re a d in g n u m b er a sig n ed t o th e m b e fo r e th e y a re p a rse d b y th e g ra m m a r parser. T h e p r o d u c t o f th e n e tw o rk w o u ld in th e e x a m p le fr o m th e b eg in n in g o f th is a rticle b e: L en in w rote this n o te in (D I R ) his n o te b o o k in (D U R ) 3 m in u tes in (L O C ) C o p en h a g en in (T I M E ) 1 8 9 7 in (E M O ) a n ger. T h e e n o rm o u s d isa m b ig u a tio n p o w e r o f th e n eu ra l n e tw o rk resu lts & o m th ree fa cto rs : th e p arallel d is tr ib u tio n , w h ich m a k es it fa s t, th e n o n lo c a l re p resen ta tion 248 Computational Linguistics -Reykjavik 1989 o f th e ru le, w h ich m ak es it ro b u s t, a n d th e sta tis tica l a n a lysis, w h ich m ak es it p o w erfu l. T h e m a ch in e d o e s n o t in fa c t c o m p u te th e ru le in pa ra llel, b u t in a seri al m a ch in e th e p r o g r a m sim u la tes th e p arallel p r o ce ssin g , an d th a t is en ou g h t o c o m p u t e th e d is a m b ig u a tio n o f a p r e p o s itio n in fra ctio n s o f a se co n d . 8462 c a lc u la tio n s d o n o t ta k e m o r e th a n a fr a c tio n o f a se co n d . T h e ru les w h ich a re u sed fo r d isa m b ig u a tio n o f th e p r e p o s itio n i, o n e o f w h ich c o u ld b e th a t i fo llo w e d b y a n o u n o f th e sem a n tic ty p e P L A C E w ill n orm a lly b e a i ( L O C ) , axe n o t lo c a t e d in s o m e o f th e c o n n e x io n s , b u t in th e w h o le p a ttern o f c o n n e x io n s b o t h fr o m in p u t la y er t o h id d e n la yer, a n d fr o m h id d en layer to o u t p u t la yer. S o irreg u la rities in th e in p u t, m e ta p h o rs o r s y n ta c tic errors, w ill n o t t o t a lly d is a b le th e ru le, b u t o n ly m a k e m in o r ch a n g es in th e o u tp u t. T h e n e tw o r k w ill a lw a y s fin d th e ' b e s t ' s o lu tio n , i.e. r e co g n iz e th e rea d in g w ith m ost s e m a n tic fitn ess reg a rd less h o w g o o d o r b a d it is-e x a c tly as w e d o even w h en w e rea d th e fa m o u s se n te n ce : C o lo r le s s g reen ideas sleep fu rio u sly. T h e n o n lo c a l re p re se n ta tio n offers a so lu tio n o f th e p r o b le m o f th e so ca lled h e rm e n e u tic cir c le , th e p r o b le m th a t th e w h o le ca n n o t b e u n d e r s to o d b e fo re th e p a rts are u n d e r s t o o d , a n d th e p a rts ca n n o t b e u n d e r s t o o d b e fo r e th e w h o le is u n d e r s t o o d . T h e m e a n in g o f th e sen ten ce con sists o f, b u t is a t th e sa m e tim e m o r e th a n th e su m o f th e sen ses o f th e w o rd s. W it h n o n lo c a l re p re se n ta tio n s th e m e a n in g o f th e w h o le is rep resen ted , n ot as th e su m o f th e m e a n in g o f th e p a rts, b u t as a p a tte rn o r ' m e a n in g ' o f so m e th in g w h ich is s u b s y m b o lic , su b sig n ifica n t o r w ith n o m e a n in g a t all, b u t w ith a d iffe re n tia tin g fu n c tio n , v iz . th e n eu ron s o f th e h id d e n layer. S o th e n etw ork c o m p u te s o r re co g n iz e s th e m e a n in g o f th e w h o le b y c o m p u tin g , n o t th e sum o f th e p a rts , b u t th e p a tte r n o f th e s u b s y m b o lic p a rts (t h e h id d en la yer n eu ron s) o f th e s y m b o lic p a rts (th e w o r d s ) o f th e sen ten ce. T h a t is e x a c t ly th e fu n c tio n o f lette rs o r p h o n e m e s, w h ich h ave n o m ean in g b u t o n ly a d iffe re n tia tio n fu n c tio n , a n d n ev erth eless m a k e it p o s s ib le to tran sm it w o r d sen ses a n d se n te n ce m e a n in g fr o m sen d e r t o re ce iv e r in th e c o m m u n ic a tio n p r o c e s s b e tw e e n h u m an s. B u t m o s t im p o r ta n t, th e n eu ra l n e tw o rk w ill u tilize in fo rm a tio n w h ich can n o t b e u sed in n o r m a l g r a m m a r ru les, v iz . p r o b a b ilis tic in fo m a tio n . It is a lin g u is tic ru le th a t o n ly i n (D U R ) w ill b e fo llo w e d b y a n ou n o f th e ty p e S C A L E : in 3 m in u tes. L e t u s a ssu m e th a t it is a sta tis tica l ru le th a t in (D U R ) is follo w e d b y a c a r d in a l n u m b e r 1 .0 0 0 tim es m o r e o fte n th a n in (L O C ) is. It is n o t p ossib le",
                "uris": null
            },
            "TABREF0": {
                "html": null,
                "text": ".........4-.XX...X....3-.XX.X..... 2-. .X___ X . . .1-X . X.X..... 0 .. .X. .X___ 1+ .XX.. .X---2+ ........... 3+ ........... 4+ nvpaoi1234567",
                "num": null,
                "type_str": "table",
                "content": "<table><tr><td colspan=\"4\">fact no. 26 SENTENCE: XXX . Det INPUT PATTERN nvpaoi1234567 -4 ..run no.15 sker -i argl: arg2: 1992 . XXX XXX ANSWER OUTPUT</td></tr><tr><td>-3</td><td/><td>loc:</td><td/></tr><tr><td>-2</td><td/><td>dir:</td><td/></tr><tr><td>-1</td><td>.X</td><td>time:</td><td>XXXXXXXX</td></tr><tr><td>0</td><td/><td>dur:</td><td/></tr><tr><td>+1</td><td/><td>mea:</td><td/></tr><tr><td>+2</td><td/><td>sta:</td><td/></tr><tr><td>+3</td><td/><td>act:</td><td/></tr><tr><td>+4</td><td/><td>emo:</td><td/></tr><tr><td/><td/><td>qual:</td><td/></tr><tr><td/><td/><td>clot:</td><td/></tr></table>"
            }
        }
    }
}