| { |
| "paper_id": "W95-0201", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T04:37:09.129123Z" |
| }, |
| "title": "Tagging the Teleman Corpus", |
| "authors": [ |
| { |
| "first": "Thorsten", |
| "middle": [], |
| "last": "Brants", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Universit\u00e4t des Saarlandes", |
| "location": { |
| "addrLine": "FR 8.7", |
| "postCode": "151150 D-66041", |
| "settlement": "Computerlinguistik, Postfach, Saarbriicken", |
| "country": "Germany" |
| } |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Christer", |
| "middle": [], |
| "last": "Samuelsson", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Universit\u00e4t des Saarlandes", |
| "location": { |
| "addrLine": "FR 8.7", |
| "postCode": "151150 D-66041", |
| "settlement": "Computerlinguistik, Postfach, Saarbriicken", |
| "country": "Germany" |
| } |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "E x p e r im e n ts w ere ca rried o u t c o m p a r in g th e S w e d ish T e le m a n a n d th e E n g lish S u sa n n e c o r p o r a u sin g an H M M-b a s e d a n d a n o v e l r e d u c tio n is tic s ta tis tica l p a r to f-s p e e c h ta g g e r. T h e y in d ic a te th a t t a g g in g th e T e le m a n co rp u s is th e m o r e d ifficu lt ta sk , a n d th a t th e p e r fo r m a n c e o f th e tw o differen t ta gg ers is c o m p a r a b le .", |
| "pdf_parse": { |
| "paper_id": "W95-0201", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "E x p e r im e n ts w ere ca rried o u t c o m p a r in g th e S w e d ish T e le m a n a n d th e E n g lish S u sa n n e c o r p o r a u sin g an H M M-b a s e d a n d a n o v e l r e d u c tio n is tic s ta tis tica l p a r to f-s p e e c h ta g g e r. T h e y in d ic a te th a t t a g g in g th e T e le m a n co rp u s is th e m o r e d ifficu lt ta sk , a n d th a t th e p e r fo r m a n c e o f th e tw o differen t ta gg ers is c o m p a r a b le .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "T h e e x p e rim e n ts re p o rte d in th e cu rren t a r ticle co n tin u e a lin e o f resea rch in th e fie ld o f p a r t -o f-s p e e c h ta g g in g u sin g s e lf-o rg a n iz in g m o d e ls th a t w a s p re se n te d at th e p r e v io u s (9 t h ) S c a n d in a v ia n C o n fe re n c e o n C o m p u t a t io n a l L in g u is tic s . T h e p a p e r is o r g a n iz e d as fo llo w s ; S e c tio n 2 d iscu sses th e T e le m a n c o r p u s an d th e ta gse ts u sed. S e ctio n 3 d e s crib e s th e H M M -b a s e d ta g g e r a n d S e c tio n 4 th e r e d u c tio n is tic s ta tis tica l o n e . T h e v ita l issue o f h a n d lin g sp a rse d a ta is a d d re sse d in S e ctio n 5 an d the e x p e rim e n ta l resu lts are p rese n ted in S e c tio n 6 . In th e T e le m a n c o rp u s , ea ch w o rd in th e ru n n in g te x t has in a v era g e 2 .3 8 ta g s fo r the s m a ll ta g s e t, a n d 3 .6 9 fo r th e la rg e ta g set. T h e s e n u m b e rs are 2 .0 7 a n d 2.61 fo r th e S u s a n n e c o rp u s , d e s p ite th e fa c t th a t th e ta gse ts fo r th e S u sa n n e c o rp u s are larger th a n th o s e fo r th e T e le m a n c o rp u s . T h u s , th ere is m u ch m o r e w ork fo r th e ta g g e r to T e le m a n a n d S u san n e c o rp u s , s m a ll a n d la rg e ta gsets. \"T h e r e is a w o rd t o th e left th a t is u n a m b ig u o u s ly ta g g e d w ith th e fo llo w in g ta g , a n d th ere are n o in te rv e n in g w o rd s ta g g e d w ith su ch a n d su ch ta g s .\" ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "P ( T | U L i C . ) = n = ^P { T \\ C i ) -P i C i \\ U^^, C i ) 1=1", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The New Approach", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "It is a s \u2022 T e le m a n is a c o n s id e r a b ly to u g h e r n u t t o cra ck th a n S u san n e. R e c a ll ( % ) 1 0 0.00 9 9 .4 6 9 9 .3 5 9 9 .2 3 9 9 .0 3 9 8 .8 2 9 8 .4 3 9 7 .75 T a g s /w o r d 2 .0 7 1.08 1.07 1.06 1.04 1.03 1.02 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "The New Approach", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "P ( E I C ) =^/ N r { E I C ) + P { E I C ) ^/ N + l", |
| "eq_num": "(2)" |
| } |
| ], |
| "section": "Successive Abstraction", |
| "sec_num": "5.1" |
| }, |
| { |
| "text": "L e x ica l s ta tis tics o n ly R e c a ll ( % ) 1 0 0 .0 0 9 9 .3 3 9 9 .2 0 9 8 .9 4 9 8 .6 7 9 8 .1 0 9 7 .4 3 9 5 .28", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Reductionistic Approach", |
| "sec_num": "6.2" |
| }, |
| { |
| "text": "T a g s /w o r d 2 .0 7 1.18 1.16 1.14 1.11 1.08 1.05 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Reductionistic Approach", |
| "sec_num": "6.2" |
| }, |
| { |
| "text": "T r ig r a m a n d le x ic a l s ta tis tics R e c a ll ( % ) 9 9 .2 2 9 8 .4 3 9 8 .2 8 9 8 .11 9 7 .7 8 9 7 .4 3 9 6 .91 9 5 .9 9", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A .B C", |
| "sec_num": null |
| }, |
| { |
| "text": "T a g s /w o r d 2 . L e x ica l s ta tis tics o n ly R e c a ll ( % ) 9 9 .2 2 9 8 .2 7 9 8 .0 3 9 7 .7 8 9 7 .4 5 9 6 .8 0 9 6 .1 5 9 3 .4 2 T r ig r a m a n d le x ic a l s ta tis tics R e c a ll ( % ) 9 9 .2 2 9 8 .4 6 9 8 .2 2 9 7 .9 9 9 7 .5 8 9 7 .1 5 9 6 .4 9 9 5 .5 4 T a g s /w o r d 2 .1 7 1.13 1.10 1.09 1.06 1.05 1.02 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A .B C", |
| "sec_num": null |
| }, |
| { |
| "text": "L e x ica l s ta tis tics o n ly R e c a ll ( % ) 9 9 .2 2 9 8 .2 1 9 7 .8 8 9 7 .6 1 9 7 .3 5 9 6 .4 7 9 5 .4 6 9 2 .8 7 T r ig r a m a n d le x ic a l sta tis tics R e c a ll ( % ) 1 0 0.00 9 9 .2 5 9 9 .1 2 9 8 .9 6 9 8 .7 4 9 8 .4 4 9 8 .0 4 9 6 .8 7", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A .B C", |
| "sec_num": null |
| }, |
| { |
| "text": "T a g s /w o r d 2.61 1.10 1.08 1.07 1.06 1.04 1.03 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A .B C", |
| "sec_num": null |
| }, |
| { |
| "text": "L e x ica l s ta tis tics o n ly R e c a ll ( % ) 1 0 0 .0 0 9 9 .0 5 9 8 .8 8 9 8 .5 9 9 8 .2 0 9 7 .5 8 9 6 .7 2 9 3 .9 8 T a g s /w o r d 2.61 1.23 1.20 1.1 7 1.14 1 .1 0 1.07 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A .B C", |
| "sec_num": null |
| }, |
| { |
| "text": "T r ig r a m a n d le x ic a l s ta tis tics R e c a ll ( % ) 9 8 .3 1 9 6 .9 4 9 6 .5 2 9 6 .1 9 9 5 .6 8 9 5 .0 2 9 4 .2 1 9 2 .7 0 T a g s /w o r d 3 .0 1 1 .2 2 1.18 1.15 1.11 1.08 1.04 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A ,B C", |
| "sec_num": null |
| }, |
| { |
| "text": "L e x ica l s ta tis tics o n ly R e c a ll ( % ) 9 8 .31 9 6 .91 9 6 .4 9 9 5 .9 4 9 5 .5 0 9 4 .4 0 9 3 .4 2 9 0 .2 6 T a g s /w o r d 3 .0 1 1.41 1.35 1.28 1.20 1.14 1.08 1.00", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A ,B C", |
| "sec_num": null |
| }, |
| { |
| "text": "T r ig r a m a n d le x ic a l sta tis tics R e c a ll ( % ) 9 8 .4 9 9 7 .0 3 9 6 .7 2 9 6 .4 1 9 5 .8 8 9 5 .1 6 9 4 . L e x ica l s t a t is t ic s o n ly R e c a ll ( % ) 9 8 .4 9 9 6 .9 5 9 6 .5 5 9 6 .0 5 9 5 .5 7 9 4 .4 4 9 3 .2 6 90.31 [S a m u e ls so n 1994] C h riste r S a m u e ls so n . \" M o r p h o lo g ic a l T a g g in g B a sed E n tire ly o n B a yesia n In fe re n ce \" , in P rocs. 9th Scand inavian C o n fe r e n c e on C o m p u ta tio n a l L in gu istics, p p.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A ,C B", |
| "sec_num": null |
| }, |
| { |
| "text": "2 2 5 -2 3 8 , S t o c k h o lm U n iv ersity , 1994.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A ,C B", |
| "sec_num": null |
| }, |
| { |
| "text": "[T e le m a n 1974] U lf T e le m a n . M a n u a l f \u00f6 r g ra m m a tisk beskrivning av talad och sk riven svenska , ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "A ,C B", |
| "sec_num": null |
| }, |
| { |
| "text": "Proceedings of NODALIDA 1995", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| }, |
| { |
| "text": "21307", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "W e w ish t o th a n k B jo rn G a m b a c k fo r p r o v id in g in fo r m a t io n o n p r e v io u s w ork w ith th e T e le m a n corp u s.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgements", |
| "sec_num": null |
| }, |
| { |
| "text": "9 5 .2 8 % 9 5 .2 8 % -9 1 .4 8 % 9 4 .8 0 % 4 9 .7 2 % 9 1 .2 0 % 9 4 .4 4 % 3 8 .3 7 % a a T r ig r a m T a g g in g9 8 .6 5 % 9 8 .6 5 % -9 5 .7 6 % 9 6 .9 5 % 8 0 .8 1 %9 3 .9 8 % 9 3 .9 8 % -8 6 .9 8 % 9 3 .0 4 % 1 0 .7 8 % 8 8 .1 6 % 9 2 .5 9 % 1 5 .8 1 % a a T r ig r a m T a g g in g 6 g A , B , C A , B , C e \u00ae A , B C ' A , C B 9 9 .8 0 % 9 9 .8 0 % -9 2 .6 1 % 9 5 .6 6 % 5 4 .2 0 % 9 3 .0 7 % 9 5 .4 6 % 5 3 .8 3 % 9 9 .0 2 9 8 .6 6 9 8 .3 5 9 7 .7 8 9 7 .3 7 1.15 1.1 2 1.10 1 .0 7 1.05 T r ig r a m a n d le x ic a l s ta tis tics 9 8 .3 6 9 7 .9 2 9 7 .5 4 9 7 .0 3 9 6 .4 1 T r ig r a m a n d le x ic a l s ta tis tics 9 4 .9 3 9 3 .9 4 9 3 .3 5 9 2 .3 5 9 1 .1 5 L e x ica l s ta tis tics o n ly 9 5 .2 3 9 4 .2 4 9 3 .6 9 9 2 .9 3 9 1 .51 ", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "annex", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "M e rc e r an d P. S. R o o s s in", |
| "authors": [ |
| { |
| "first": "P", |
| "middle": [ |
| "F" |
| ], |
| "last": "Ro W N", |
| "suffix": "" |
| }, |
| { |
| "first": "V", |
| "middle": [ |
| "J" |
| ], |
| "last": ". D E Lla P Ie Tra", |
| "suffix": "" |
| }, |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Lin E K", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [ |
| "D" |
| ], |
| "last": "Afferty", |
| "suffix": "" |
| }, |
| { |
| "first": "R", |
| "middle": [ |
| "L" |
| ], |
| "last": "", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "P. F . B ro w n , V . J. D e lla P ie tra , F . J e lin e k , J. D . L afferty, R . L . M e rc e r an d P. S. R o o s s in . \" C la s s -b a s e d n -g r a m m o d e ls o f n a tu ra l la n g u a g e \" . C o m p u ta tio n a l", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "P rocs. 9th S ca", |
| "authors": [ |
| { |
| "first": "", |
| "middle": [], |
| "last": "D O U G La S S C U Tt In G", |
| "suffix": "" |
| } |
| ], |
| "year": 1994, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "6--11", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "D o u g la s s C u tt in g . \" A P r a c tic a l P a r t-o f-S p e e c h T a g g e r \" , in P rocs. 9th S ca n d in a vian C o n fer en ce on C o m p u ta tio n a l L in gu istics, p p . 6 5 -7 0 , S t o c k h o lm U n iv e rsity , 1994.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "P rocs. 3rd C o n fe r e n c e on A p plied N atural Language P r o c e s s in g", |
| "authors": [ |
| { |
| "first": "", |
| "middle": [], |
| "last": "D O U G La S S R . C U Tt In G , J U Lia N K U P Ie C", |
| "suffix": "" |
| } |
| ], |
| "year": 1992, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "[C u ttin g et al 1992] D o u g la s s R . C u tt in g , J u lia n K u p ie c , J a n P ed e rse n a n d P e n e lo p e S ib u n . \" A P r a c tic a l P a r t-o f-S p e e c h T a g g e r \" , in P rocs. 3rd C o n fe r e n c e on A p plied N atural Language P r o c e s s in g , p p . 1 3 3 -1 4 0 , A C L , 1992.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "E x p e r im e n ts U sin g N e u ra l N e t w o rk s", |
| "authors": [], |
| "year": 1994, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "7--8", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "M a rtin E in e b o rg a n d B jo rn G a m b a c k . \" T a g g in g E x p e r im e n ts U sin g N e u ra l N e t w o rk s\" , in P rocs. 9th Scandinavian C o n fe r e n c e on C o m p u ta tio n a l L in g u istics, p p . 7 1 -8 2 , S t o c k h o lm U n iv ersity , 1994.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "F ran cis a n d H . K u ce ra . F req u en cy A n a ly sis o f E nglish U sage", |
| "authors": [ |
| { |
| "first": "N", |
| "middle": [ |
| "W" |
| ], |
| "last": "", |
| "suffix": "" |
| } |
| ], |
| "year": 1982, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "N. W . F ran cis a n d H . K u ce ra . F req u en cy A n a ly sis o f E nglish U sage, H o u g h to n M ifflin , B o s to n , 1982.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "P o o r E s tim a te s o f C", |
| "authors": [ |
| { |
| "first": "W", |
| "middle": [ |
| "A" |
| ], |
| "last": "Le A N D", |
| "suffix": "" |
| }, |
| { |
| "first": "K", |
| "middle": [ |
| "W" |
| ], |
| "last": "Rch", |
| "suffix": "" |
| } |
| ], |
| "year": 1990, |
| "venue": "Proc. o f the Speech and N atura l La ngu age W o rk sh o p", |
| "volume": "8", |
| "issue": "", |
| "pages": "2--8", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "W . A . G a le a n d K . W . C h u rch . \" P o o r E s tim a te s o f C o n t e x t are W o r s e th a n N o n e \" , in Proc. o f the Speech and N atura l La ngu age W o rk sh o p , p p . 2 8 3 -2 8 7 , M o rg a n K a u fm a n n , 1990.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "type_str": "figure", |
| "text": "T h e n , th e w e ll-e s ta b lis h e d H M M -b a s e d X e r o x ta g g e r, see [C u tt in g 1994 ], w a s c o m p a red w ith s o m e less k n o w n ta g g ers, n a m e ly a n e u ra l-n e tw o rk ta g g e r d e s cr ib e d in [E in e b o rg k. G a m b a c k 1994], a n d a B a y e sia n ta g g e r p re se n te d in [S a m u e ls so n 1994]. T h e X e r o x ta g g e r p e r fo rm s le x ic a l g e n e r a liz a tio n s b y clu s te rin g w o rd s b a se d o n th eir d is tr ib u tio n a l p a tte rn s , w h ile th e la tte r tw o u tiliz e th e m o r p h o lo g ic a l in fo r m a t io n presen t in S w ed ish b y g e n e ra lizin g o v er w o rd su ffixes. T h is tim e , a n o th e r H M M -b a s e d a p p ro a ch is c o m p a r e d w ith a n o v e l r e d u c t io n istic s ta tis tica l ta g g er in sp ired b y th e su cce ssfu l C o n s tra in t G r a m m a r s y s te m , [K a rlsson et al 1995]. T h e p e r fo rm e d e x p e rim e n ts d o n o t o n ly serve t o e v a lu a te th e tw o ta g g e rs , b u t a lso sh ed s o m e n ew lig h t on th e T e le m a n c o rp u s as an e v a lu a tio n d o m a in fo r p a rto f-s p e e c h ta gg ers c o m p a re d t o o th e r, E n g lish , c o r p o r a .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF1": { |
| "type_str": "figure", |
| "text": "h e T e le m a n co rp u s [T elem a n 1974] is a c o rp u s o f c o n t e m p o r a r y S w ed ish , re p resen tin g a m ix tu r e o f d ifferen t te x t gen res like in fo r m a t io n b r o c h u re s o n m ilita r y se rv ice a n d m e d ica l ca re, n ovels, e tc . It c o m p r is e s 8 5 ,4 0 8 w o rd s (to k e n s ; h ere, w o rd s is a c o lle c tiv e d e n o ta tio n o f p r o p e r w o rd s , n u m b e rs, a n d p u n c t u a t io n ). T h e r e are 14,191 differen t w o rd s (t y p e s ); th e m o s t fre q u e n t o n e is w h ich o c c u r s 4 ,6 6 2 tim e s ; th e m o s t fre q u e n t p r o p e r w o rd is \" och \" (a n d ), w h ich o c c u r s 2 ,2 1 7 tim e s . 8 ,4 5 8 o f th e w o rd s o c c u r e x a c tly o n ce , w h ich is 6 0 % o f th e ty p e s b u t o n ly 10% o f th e tok en s. Proceedings of NODALIDA 1995, pages 7-20 T a b le 1: C o m p a r is o n o f T e le m a n an d Tags in the Susanne corp us with indices are cou n ted as separate tags. F or th e e x p e rim e n ts , w e u sed tw o d iffe re n t ta gsets. F irst, w e u sed th e o rig in a l ta g s e t, c o n s is tin g o f 25 8 ta gs. E a ch o f th e 14,191 w o rd ty p e s ca n h a ve b e tw ee n o n e an d 15 o f th e 25 8 ta g s (th e h ig h ly a m b ig u o u s w o rd \"f \u00f6 r \" {fo r, stern , lead, too, . . . ) h a s th e m a x im u m n u m b e r o f ta g s ). W e th en u sed a re d u ce d ta g s e t, c o n s is tin g o f 19 ta g s , w h ich rep resen t c o m m o n s y n ta c tic ca te g o r ie s a n d p u n c tu a t io n . T h is ta gse t is id e n tica l t o th a t used in th e p u b lic a tio n s m e n tio n e d a b o v e . E a ch o f th e w o rd ty p e s th en h a s b etw e e n o n e a n d 7 ta g s ( \"/ o r \" a n d \" i \" h a ve th e m a x im u m n u m b e r o f ta g s ). 2.1 C om p a rison w ith an English C orpus S in ce 1 0 % o f th e w o rd s in th e T e le m a n c o rp u s o c c u r o n ly o n c e , w e e x p e c t fr o m th e G o o d -T u r in g fo r m u la [ G o o d 1953] th a t 1 0 % o f th e w o rd s in n ew te x t b e u n k n o w n , w h ich is a v e ry h ig h p e rce n ta g e . O th e r p u b lic a tio n s ty p ic a lly r e p o r t 5 % . S in ce m o s t o f th e w o rk in th is a rea is o n E n g lish c o r p o r a , w e c o m p a r e d th e T e le m a n c o r p u s w ith an E n g lish c o rp u s , n a m e ly th e S u sa n n e c o rp u s [S a m p s o n 1995], w h ich is a r e -a n n o t a te d p a rt o f th e B ro w n C o r p u s [F ra n cis & K u c e ra 1982], c o m p r is in g d iffe re n t te x t g en res. T h e releva n t fa c ts are s u m m a r iz e d a n d c o m p a r e d in T a b le 1. T h e m a jo r d ifferen ce (a p a r t fr o m c o r p u s size a n d ta g s e ts u sed ) is th e p e r ce n ta g e o f w o rd s th a t o c c u r e x a c tly o n ce : 1 0 % fo r T e le m a n vs. 4 % fo r S u sa n n e. A c c o r d in g to th e G o o d -T u r in g fo r m u la , th is p e r ce n ta g e is id e n tica l t o th e e x p e c t e d p e rce n ta g e o f u n k n o w n w ord s . A c t u a l c o u n ts b y d iv id in g th e c o r p o r a in to tr a in in g a n d test p a rts y ie ld a r o u n d 1 4 % an d 7 % , re sp e ctiv e . T h is in d ic a te s th a t u n seen S w e d ish te x t w ill h a ve s u b s ta n tia lly m o r e u n k n o w n w o rd s th a n u n seen E n g lish , w h ich is m o s t likely d u e t o th e h igh er d eg ree o f m o r p h o lo g ic a l v a r ia tio n in S w ed ish . A fu rth e r d ifficu lty w ith th e S w ed ish c o rp u s is th e h ig h e r d eg ree o f a m b ig u ity .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF2": { |
| "type_str": "figure", |
| "text": "d o in th e T e le m a n co rp u s . S o m e m o r e n u m b e rs : in th e r u n n in g te x t, 5 4 .5 % /6 4 .2 % o f th e w o rd s in th e T e le m a n c o r p u s are a m b ig u o u s , a n d o n ly 4 4 .3 % /4 8 .9 % in the S u s a n n e c o r p u s ( s m a ll/la r g e ta g s e t, resp .; see T a b le 2 fo r fu rth e r d e ta ils ).", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF3": { |
| "type_str": "figure", |
| "text": "The H M M Approach A H id d e n M a r k o v M o d e l (H M M ) co n s ists o f a set o f sta te s, a set o f o u tp u t s y m b o ls a n d a set o f tra n s itio n s . F or ea ch s ta te a n d ea ch s y m b o l, th e p r o b a b ility th a t th is s y m b o l is e m itte d b y th a t s ta te is g iv e n . A ls o , a p r o b a b ility is a s s o c ia te d w ith 8 Proceedings of NODALIDA 1995 T a b le 2: D is tr ib u tio n o f n u m b e r o f c a te g o r ie s p e r w o rd in r u n n in g te x t fo r th e", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF4": { |
| "type_str": "figure", |
| "text": "e a ch tr a n s itio n b etw e e n sta tes (see[R a b in e r 1989] fo r a g o o d in t r o d u c t io n ). T h e tr a n s itio n p r o b a b ility , a n d th u s th e p r o b a b ilit y o f th e fo llo w in g s ta te , d e p e n d s o n ly o n th e p r e v io u s s ta te fo r first o rd e r H M M s , o r o n k p r e v io u s sta te s fo r H M M s o f A:th o r d e r. H M M a p p ro a ch e s t o p a r t -o f-s p e e c h ta g g in g m a k e th e w e ll-k n o w n a s s u m p tio n th a t th e cu rre n t c a te g o r y or p a r t -o f-s p e e c h o f a w o rd d e p e n d s o n ly o n th e p re v io u s (n -1) ca te g o r ie s (M a r k o v a s s u m p t io n ), thus th e y a ssu m e th a t n a tu ra l la n g u a g e is a M a rk o v p r o ce s s o f o r d e r (n -1), w h ich o f c o u rse is n o t tru e, b u t a su ccessfu l a p p r o x im a t io n , n = 3 is ch os en in m o s t o f th e cases, re su ltin g in a trig ra m m o d e l (i.e ., a lw a y s w o rk in g w ith a w in d o w o f size 3 ), s in ce it y ie ld s th e b e s t c o m p r o m is e b e tw e e n size o f c o r p o r a n ee d ed fo r tr a in in g a n d ta g g in g a ccu ra cy . F u rth e rm o re , the cu rre n t w o r d ( s y m b o l) d e p e n d s o n ly o n th e cu rre n t c a te g o r y (s t a t e ). T h u s , in stea d o f c a lc u la tin g a n d m a x im iz in g P { T i .. .Tk \\W\\ . . . VFjt), w ith T{ ta g s a n d W i w ord s, w h ich is im p o s s ib le in all p r a c tic a l cases, o n e c a lc u la te s an d m a x im iz e s d th e b e s t se q u e n ce o f ta g s fo r a g iv e n se q u e n ce o f w o rd s . T h e p a ra m e te r s o f an H M M can b e e s tim a te d d ir e c tly fr o m a p r e ta g g e d co rp u s v ia m a x im u m -lik e lih o o d e s t im a tio n (M L E ). B u t M L E sets a lo t o f th e tra n sitio n p r o b a b ilit ie s t o zero, a n d i f o n e o f th e m u lt ip lie d p r o b a b ilitie s in ( 1) is ze ro , th e p r o d u c t b e c o m e s zero, le a v in g n o m e a n s t o d is tin g u is h b etw e e n d ifferen t p r o d u c ts th a t c o n ta in a ze ro p r o b a b ility . T h is resu lts in p o o r e stim a te s fo r th e p r o b a b ilitie s o f n ew seq u e n ce s o f w o rd s . T h is p r o b le m is a d d re sse d in S e c tio n 5. A n o th e r w ay o f e s tim a tin g th e p a ra m e te rs o f an H M M is to use an u n ta g g e d c o rp u s , a le x ic o n w ith p a rts -o f-s p e e c h lists fo r th e w o rd s a n d th e B a u m -W e lc h a lg o r ith m [B a u m 1972]. T h is a p p ro a ch has th e a d v a n ta g e o f a v o id in g th e te d io u s w ork o f m a n u a lly a n n o ta t in g a co rp u s , b u t it req u ires a s o p h is tic a te d c h o ic e o f in itia l b ia s e s, a n d g e n era lly , th e p e r fo r m a n c e is w orse th a n th a t a ch ie v e d w ith a n n o ta te d c o r p o r a . W h e n u sin g an H M M fo r ta g g in g , th e s y s te m g e ts a strin g o f w o rd s a n d has t o fin d th e m o s t p r o b a b le se q u en ce o f ta g s th a t c o u ld h a ve p r o d u c e d th e s trin g o f w o rd s . T h is is d o n e w ith a d y n a m ic p r o g r a m m in g m e t h o d , th e V it e r b i a lg o r ith m [V ite r b i 1967]. T h e a lg o r ith m fin d s th e m o s t p r o b a b le se q u e n ce o f s ta tes in tim e lin e a r in th e le n g th o f th e in p u t strin g .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF5": { |
| "type_str": "figure", |
| "text": "lt h o u g h n o t y e t fu lly rea liz ed , th e b a s ic p h ilo s o p h y b e h in d th e r e d u c tio n is tic sta tis tica l a p p ro a ch is t o g iv e it th e s a m e e x p res siv e p o w e r as th e C o n s tra in t G ra m m a r s y s te m .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF6": { |
| "type_str": "figure", |
| "text": "h e C o n s t r a in t G r a m m a r sy s te m p e r fo rm s re m a rk a b ly w ell; [V o u tila in e n &; H e ik k ila 1994] r e p o r t 9 9 .7 % reca ll, o r 0 .3 % e rro r ra te, w h ich is ten tim e s sm a lle r th a n th a t o f th e b e s t s t a tis tica l ta gg ers. T h e s e im p re s siv e resu lts are a ch ie v e d b y :1. U tiliz in g a n u m b e r o f d ifferen t in fo r m a t io n so u rce s , an d n o t o n ly th e s te r e o ty p e d le x ic a l s ta tis tics a n d n -g r a m ta g s ta tis tics th a t h ave b e c o m e th e d e fa c t o s ta n d a r d in s ta tis tica l p a r t -o f-s p e e c h ta g g in g . 2. N o t fu lly re s o lv in g all a m b ig u itie s w h en th is w o u ld je o p a r d iz e th e reca ll. 10 Proceedings of NODALIDA 1995 P r o p e r t y 2 m ea n s th a t th e s y s te m tra d e s p re cisio n fo r re ca ll, w h ich m a k e s it id e a l as a p re p ro c e s s o r for n a tu ra l la n g u a g e sy s te m s p e r fo r m in g d e e p e r a n a ly s is. T h e C o n s tra in t G r a m m a r s y s te m w o rk s as fo llo w s : F irs t, th e in p u t s tr in g is a ssign ed all p o s s ib le ta gs fr o m th e le x ic o n , o r ra th e r, fr o m th e m o r p h o lo g ic a l a n a ly ze r. T h e n , ta g s are re m o v e d ite ra tiv e ly b y r e p e a te d ly a p p ly in g a se t o f rules, o r c o n s tra in ts , t o th e ta g g e d strin g . W h e n n o m o r e ta g s are r e m o v e d b y th e la st ite ra tio n , th e p ro ce s s te rm in a te s , a n d m o r p h o lo g ic a l d is a m b ig u a tio n is c o n c lu d e d . T h e n a set o f s y n ta c tic ta g s are a ssign ed t o th e ta g g e d in p u t s trin g a n d a s im ila r p ro ce s s is p e r fo rm e d fo r s y n ta c tic d is a m b ig u a tio n . T h is m e t h o d is o fte n referred to as red u ciion isiic tagging. T h e rules are s o r t -o f fo r m u la t e d as fin ite s ta te a u t o m a t a [T a p a n a in e n , p e r s o n a l c o m m u n ic a tio n ], w h ich a llo w s v ery fa s t p ro ce s s in g . E a ch ru le a p p lie s to a cu rren t w o rd w ith a set o f c a n d id a te ta g s . T h e s tru c tu re o f a rule is ty p ic a lly : \" In th e fo llo w in g c o n t e x t , d is ca rd th e fo llo w in g ta g s .\" \" In th e fo llo w in g c o n t e x t , c o m m it t o th e fo llo w in g t a g .\" W e w ill ca ll d is ca rd in g o r c o m m it t in g to ta g s th e rule a ction . A ty p ic a l rule co n tex t is:", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF7": { |
| "type_str": "figure", |
| "text": "T h e stru ctu re o f th e C o n s tra in t G r a m m a r rules r e a d ily a llo w s th e ir c o n t e x t s t o b e v ie w e d as th e c o n d itio n in g s o f c o n d itio n a l p r o b a b ilitie s , a n d th e a c tio n s h a v e an o b v io u s in te r p r e ta tio n as th e c o rr e s p o n d in g p r o b a b ilitie s . E a ch c o n t e x t ty p e can b e seen as a se p a ra te in fo r m a t io n s o u r ce , a n d w e w ill c o m b in e in fo r m a t io n so u rces S i , . . . ,S n b y m u lt ip ly in g th e s c a le d p r o b a b ilitie s : P { T \\ S i , . . . , S \" ) P { T ) n i = l P j T I S i) P { T ) T h is fo r m u la can b e e s ta b lis h e d b y B a y esia n in v e rsio n , th en p e r fo r m in g th e in d e p e n d e n ce a s s u m p tio n s , an d ren ew ed B a yesia n in v ersion : P{T\\Si,...,S\") = P{T)-P{Sl,...,Sr.\\T) P { S i , . . . , S n ) sta n d a rd s t a tis tica l p a r t -o f-s p e e c h ta g g in g th ere are o n ly tw o in fo r m a t io n so u rce s -th e le x ic a l p r o b a b ilitie s an d th e ta g s a ssign e d to n e ig h b o u r in g w o rd s . 11 Proceedings of NODALIDA 1995 P ( T a g I L e x ic o n a n d n -g r a m s ) = P ( T a g I L e x ic o n ) \u2022 P ( T a g | N -g ra m s ) T h e c o n t e x t w ill in gen eral n o t b e fu lly d is a m b ig u a te d . R a th e r th a n e m p lo y in g d y n a m ic p r o g r a m m in g o v e r th e la ttic e o f r e m a in in g c a n d id a te ta g s , th e n ew a p p r o a c h uses th e w e ig h te d a v era g e ov e r th e r e m a in in g c a n d id a te ta g s t o e s tim a te th e p r o b a b ilitie s :", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF8": { |
| "type_str": "figure", |
| "text": "su m e d th a t { C i : t = 1 , . . . , n } co n s titu te s a p a r t itio n o f th e c o n t e x t C , i.e ., th a t C -a n d th a t CiC\\Cj = 0 fo r i j . In p a rticu la r, tr ig ra m p r o b a b ilitie s are c o m b in e d as fo llo w s : P ( T \\ C ) = = P i T \\ T , , T r ) P ( { T , , T r ) \\ C ) (T ,,T ,)e c H ere T d e n o te s a c a n d id a te ta g o f th e cu rre n t w o r d , Tj d e n o te s a c a n d id a te ta g o f th e im m e d ia te left n e ig h b o u r , a n d Tr d e n o te s a c a n d id a te ta g o f th e im m e d ia te righ t n e ig h b o u r . C is th e set o f o r d e re d p a irs (T j, T r ) d ra w n fr o m th e set o f ca n d id a te ta gs o f th e im m e d ia t e n e ig h b o u r s . P { T \\ T ; ,T r ) is th e s y m m e t r ic tr ig ra m p r o b a b ility . T h e ta g g e r is r e d u c tio n is tic sin ce it re p e a te d ly re m o v e s lo w -p r o b a b ilit y ca n d id a te ta g s. T h e p r o b a b ilit ie s are th en r e ca lcu la te d , a n d th e p r o ce s s te rm in a te s w h en th e p r o b a b ilit ie s h a ve s ta b iliz e d a n d n o m o r e ta g s ca n b e r e m o v e d w ith o u t je o p a r d iz in g th e re ca ll; c a n d id a te ta g s are o n ly r e m o v e d i f th eir p r o b a b ilitie s are b e lo w so m e th re s h o ld valu e. d lin g sp a rse d a ta c o n s ists o f tw o d iffere n t tasks: 1. E s tim a t in g th e p r o b a b ilitie s o f ev en ts th a t d o n o t o c c u r in th e tra in in g d a ta . 2. I m p r o v in g th e e stim a te s o f c o n d it io n a l p r o b a b ilitie s w h ere th e n u m b e r o f o b s e r v a tio n s u n d er th is c o n d itio n in g is s m a ll. C o p in g w ith u n k n o w n w o rd s , i.e ., w ord s n o t e n co u n te re d in th e tra in in g set, is an a r c h e ty p ic a l e x a m p le o f th e fo r m e r ta sk . E s tim a t in g p r o b a b ility d is tr ib u tio n s c o n d it io n a l o n s m a ll c o n te x ts is an e x a m p le o f th e la tte r ta sk . W e w ill e x a m in e se v e ra l a p p ro a ch e s t o th ese tasks. F or th e H M M , it is n ecessa ry t o a v o id ze ro p r o b a b ilitie s . T h e m o s t s tr a ig h t fo r w a rd s tra te g y is e m p lo y in g th e e x p e c t e d -lik e lih o o d e s tim a te (E L E ), w h ich s im p ly a d d s 0 .5 t o e a ch fre q u e n c y c o u n t a n d th e n c o n s tr u c ts a m a x im u m -lik e lih o o d esti m a te ( M L E ) , (see e.g . [G a le & C h u rch 1 9 9 0 ]). T h e M L E o f th e p r o b a b ility is th e re la tiv e fr e q u e n c y r . A n o th e r p o s s ib ilit y is th e G o o d -T u r in g m e t h o d [G o o d 1953], w h ere ea ch fr e q u e n c y / is r e p la ce d b y / * = ( / -f-\\ ) N j^i / N f , w here N j d e n o te s th e fr e q u e n c y o f fre q u e n c y / . A lte r n a tiv e ly , o n e ca n use lin ea r in te r p o la tio n o f th e p r o b a b ilit ie s o b t a in e d b y M L E , P { c \\ a ,b ) = A i r ( c ) + A2t ( c | b) -f A3t ( c ] a , 6). 12 Proceedings of NODALIDA 1995 [B ro w n et al 1992] let th e A valu es d e p e n d e n t o n th e c o n t e x t , w h ich im p r o v e s th e ta g g in g a ccu ra cy . T h is is rela ted t o th e id e a o f s u cce s siv e a b s t r a c t io n p re se n te d in S e c tio n 5 .1 . T o a ch ie v e im p r o v e d e s tim a te s o f le x ic a l p r o b a b ilitie s , w o rd s ca n b e clu s te re d to g e th e r, see [C u tt in g et al 1992]. T h e r e are severa l w a y s t o h a n d le u n k n o w n w o rd s . T h e s e in clu d e ; 1. M a k in g e v ery ta g a p o s s ib le ta g fo r th a t w o rd w ith e q u a l p r o b a b ilit y a n d fin d in g th e m o s t p r o b a b le ta g so le ly b a se d o n c o n t e x t p r o b a b ilitie s . T h e re su lts ca n b e slig h tly im p r o v e d b y tr y in g o n ly o p e n -c la s s ta g s fo r u n k n o w n w ord s. 2. A s an e x te n s io n t o ca se 1, c h o o s in g d iffe re n t b u t a g a in c o n s ta n t p r o b a b ilitie s fo r ea ch p o s s ib le ta g. T h is co n s titu te s an a p r io r i d is tr ib u tio n fo r u n k n o w n w o rd s , re fle ctin g fo r e x a m p le th a t m o s t o f th e u n k n o w n w o rd s are n o u n s. T h e p r o b a b ilitie s c o u ld b e o b t a in e d fr o m a se p a ra te tr a in in g p a rt, o r fr o m th e d is tr ib u tio n o f w o rd s th a t o c c u r o n ly o n c e in th e tr a in in g c o rp u s . T h e s e w o rd s reflect th e d is tr ib u tio n o f u n k n o w n w o rd s a c c o r d in g t o th e fo r m u la p resen ted in [G o o d 1953]. 3. E x p lo itin g w o r d -fo r m in fo r m a t io n as p r o p o s e d in [S a m u e ls so n 1994]. H ere, th e p r o b a b ility d is tr ib u tio n s are d e te rm in e d fr o m th e la st n c h a ra cte rs o f th e w o rd , a n d th e r e m a in in g n u m b e r o f s y lla b le s . T h is m e t h o d h as b e e n p ro v e n su cce ssfu l fo r S w ed ish te x t. 4. U tiliz in g o r th o g r a p h ic a l cu es su ch as c a p it a liz a t io n .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF9": { |
| "type_str": "figure", |
| "text": "s s u m e th a t w e w a n t t o e s tim a te th e p r o b a b ilit y P { E j C ) o f th e e v e n t E g iv e n a c o n t e x t C fr o m th e n u m b e r o f tim e s N e it o c c u r s in N = jC j tria ls, b u t th a t th is d a ta is sp arse. A s s u m e fu rth e r th a t th ere is a b u n d a n t d a ta in a m o r e g en era l c o n t e x t C Z) C th a t w e w a n t t o use t o g e t a b e t te r e s tim a te o f P [ E j C ) . I f th ere is an o b v io u s lin ea r o r d e r C = Cm C C m -i C \u2022 \u2022 \u2022 C C i = C ' o f th e v a rio u s g e n e ra liz a tio n s Ck o f C , w e ca n b u ild th e e s tim a te s o f P { E j C * ) o n th e re la tiv e fre q u e n cy r { E j C k ) o f ev en t E in c o n t e x t Ck a n d th e p r e v io u s e s t im a te s o f P ( E I C k -i ) . W e ca ll th is m e t h o d lin ear su ccessiv e abstra ction . A s im p le e x a m p le is e s tim a tin g th e p r o b a b ilit y P { T j l \" , . . I n -j ) o f a ta g T g iv e n I n -j , \u25a0 \u25a0 .,l n , th e last J -I-1 letters o f th e w o rd . In th is case, th e e s tim a te w ill b e b a s e d o n th e re la tiv e fre q u e n cies r ( T j r ( T j . . . , r ( T j l \" ) , r ( T ) . P r e v io u s e x p e rim e n ts [S a m u elsson 1994] in d ic a t e th a t th e fo llo w in g is a s u ita b le fo rm u la ;", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF10": { |
| "type_str": "figure", |
| "text": "h is fo r m u la s im p ly u p -w e ig h ts th e re la tiv e fre q u e n c y r b y a fa c t o r y/ N , th e sq u a re r o o t o f th e size o f c o n t e x t C , w h ich is th e a c tiv e in g r e d ie n t o f th e s ta n d a rd d e v ia t io n o f r . I f th ere is o n ly a p a rtia l o r d e r o f th e v a rio u s g e n e r a liz a tio n s , th e s c h e m e is still v ia b le . F or e x a m p le , c o n s id e r g e n e ra liz in g s y m m e t r ic tr ig r a m s ta tis tic s , i.e ., s ta tis tics o f th e fo r m P { T j T i,T r ). H ere, b o t h T| a n d Tr are o n e -s te p g e n e r a liz a tio n s o f th e c o n t e x t T i,T r , a n d b o t h h a v e in tu rn th e c o m m o n g e n e r a liz a tio n f i . W e m o d ify E q u a tio n 2 a c c o r d in g ly ; P ( T \\ T ,,T r ) = y/\\T,,T r\\r(T \\T ,,T r) + P{T \\ T ,) + P{T \\ T r) ll th is p a rtial su ccessiv e abstraction . 6 Experiments F or th e e x p e rim e n ts , b o t h c o r p o r a w ere d iv id e d in to th ree sets, o n e la rg e set an d tw o s m a ll sets. W e used th ree d ifferen t d iv is io n s in to tr a in in g a n d te s tin g sets. F irst, all th ree sets w ere u sed fo r b o t h tr a in in g a n d te s tin g . In th e s e c o n d a n d th ird case, tr a in in g a n d test sets w ere d is jo in t, th e la rg e set a n d o n e o f th e s m a ll sets w ere used fo r tr a in in g , th e r e m a in in g s m a ll set w as u sed fo r te s tin g . A s a b a se lin e t o in d ic a te w h a t is g a in e d b y ta k in g th e c o n t e x t in to a c c o u n t, w e p e r fo r m e d an a d d itio n a l set o f e x p e rim e n ts th a t u sed le x ic a l p r o b a b ilitie s o n ly , a n d ig n o r e d th e c o n te x t.", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF11": { |
| "type_str": "figure", |
| "text": "h e e x p e rim e n ts o f th is s e c tio n w ere p e r fo r m e d w ith a tr ig ra m ta g g e r as d e s crib e d in S e c tio n 3. Z e ro fre q u e n cie s w ere a v o id e d b y u sin g e x p e c t e d -lik e lih o o d e s tim a tio n . U n k n o w n w o rd s w ere h a n d le d by a m ix t u r e o f m e t h o d s 2 a n d 3 lis te d in S e ctio n 5: If th e su ffix o f 4 c h a ra cte rs (3 ch a ra cte rs fo r th e S u sa n n e c o r p u s ) o f th e u n k n o w n w o rd s w as fo u n d in th e le x ic o n , th e ta g d is tr ib u tio n fo r th a t su ffix w as u sed. O th e rw is e w e u sed th e d is tr ib u tio n o f ta g s fo r w o rd s th a t o c c u r r e d o n ly o n c e in th e tra in in g co rp u s . A s o p p o s e d t o tr ig ra m ta g g in g , le x ic a l ta g g in g ig n o re s c o n t e x t p r o b a b ilitie s an d is b a se d s o le ly o n le x ic a l p r o b a b ilitie s . E a ch w o rd is a ssign ed its m o s t fre q u e n t ta g f r o m th e tr a in in g co rp u s . U n k n o w n w o rd s w ere a ssign e d th e m o s t fre q u e n t ta g o f w o rd s th a t o c c u r r e d e x a c tly o n c e in th e tr a in in g c o rp u s . T h e m o s t fre q u e n t ta gs fo r s in g le o c c u r r e n c e w o rd s are fo r th e T e le m a n c o rp u s NIMSS (in d e fin ite n o u n -n o u n c o m p o u n d ) an d noun (la rg e a n d s m a ll ta g s e t, r e s p .), fo r th e S u sa n n e co rp u s NN2 (p lu r a l c o m m o n n o u n ) a n d NN ( c o m m o n n o u n ; a g a in la rg e a n d s m a ll ta gse t re sp .). T a g g in g sp eed weis g e n era lly b etw e e n 1000 an d 20 00 w o rd s p er s e c o n d o n a S p a rcS e rv e r 1000; m o s t o f th is v a ria tio n w as d u e t o v a ria tio n s in th e n u m b e r o f u n k n o w n w o rd s . T h e resu lts fo r th e T e le m a n c o rp u s are sh o w n in T a b le 5 an d th e resu lts fo r th e S u s a n n e c o rp u s in T a b le 6 . W h a t im m e d ia te ly a t tr a c ts a t te n tio n is th e re m a r k a b ly lo w p e r fo r m a n c e o f th e tr ig r a m a p p r o a c h fo r th e T e le m a n c o rp u s . A lr e a d y th e b a selin e o b t a in e d b y le x ica l t a g g in g is b e lo w 8 0 % fo r n ew te x t, u su al resu lts are a r o u n d 9 0 % . N o r m a l resu lts can b e o b t a in e d o n ly fo r k n o w n w o rd s o r w h e n u sin g th e s m a ll ta g s e t, th e la tte r b e in g in fa c t a v ery s im p le ta sk , s in ce th e a lg o r ith m h as t o ch o o s e fr o m o n ly 19 ta g s . For th e la rg e ta g s e t, tr ig ra m ta g g in g a ch ie v e s o n ly 8 3 % a cc u r a c y . T h is low figu re is d u e t o th e u n u su a lly h ig h n u m b e r o f u n k n o w n w ord s a n d th e la rg er degree o f a m b ig u ity c o m p a r e d to E n g lish c o r p o r a , as is d iscu sse d in S e c tio n 2. U sin g a la rg e S w ed ish le x ic o n o r m o r p h o lo g ic a l a n a ly ze r s h o u ld im p r o v e th e resu lts sign ifica n tly. A n o th e r in te re stin g resu lt is th a t a c c u r a c y in crea se s w h en th e size o f th e ta gse t in cre a se s fo r th e Ccises w h ere k n o w n te x t is ta g g e d a n d c o n te x t p r o b a b ilitie s are ta k en in t o a c c o u n t. T h is m e a n s th a t th e a d d itio n a l in fo r m a t io n a b o u t th e c o n t e x t in th e la rg e r ta g s e t is very h e lp fu l fo r d is a m b ig u a tio n , b u t o n ly w h en d is a m b ig u a tin g 14 Proceedings of NODALIDA 1995", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF12": { |
| "type_str": "figure", |
| "text": "w n te x t. T h is c o u ld arise fr o m th e fa c t th a t a la rg e n u m b e r ( > 5 0 % ) o f th e tr ig ra m s th a t o c c u r in th e tr a in in g te x t o c c u r e x a c tly o n c e . A n d m o s t o f th e p o s sib le tr ig ra m s d o n o t o c c u r a t all (g e n e ra lly m o r e th a n 9 0 % , d e p e n d in g o n th e size o f th e ta g s e t). N o w , th e tr ig ra m a p p ro a ch h as a d is tin c t b ia s t o th o se tr ig ra m s th a t o c c u r r e d o n c e ov e r th o se th a t n ever o c c u r r e d . T h e s e h a p p e n t o b e th e righ t on es fo r k n o w n te x t b u t n o t n ecessa rily fo r n ew t e x t, th u s th e p o s itiv e effect o f a larger ta g s e t va n ish es fo r fresh te x t. T h e re su lts fo r th e S u sa n n e c o r p u s are sim ila r t o th o se r e p o rte d in o th e r p u b li c a tio n s fo r (o t h e r ) E n g lish c o r p o r a .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF13": { |
| "type_str": "figure", |
| "text": "h e r e d u c tio n is tic s t a tis tica l ta g g e r d e s cr ib e d in S e c tio n 4 w as te s ted o n th e sa m e d a t a as th e H M M ta g g e r. T h e in fo r m a t io n so u r ce s e m p lo y e d in th e e x p e rim e n ts w ere le x ic a l s ta tis tics a n d c o n t e x t u a l in fo r m a t io n , w h ich co n s is te d o f s y m m e tr ic tr ig ra m s ta tis tics . U n k n o w n w o rd s w ere h a n d le d b y c re a tin g a d e cis io n tree o f th e fo u r la st le tte rs fr o m w o rd s w ith th ree or less o c cu rr e n ce s . E a ch n o d e in th e tree w as a s s o c ia te d w ith a p r o b a b ility d is tr ib u tio n (o v e r th e ta g s e t) e x tr a c te d fr o m these w o rd s , a n d th e p r o b a b ilitie s w ere s m o o th e n e d th r o u g h lin ea r su cce ssiv e a b s tra ctio n , see S e c tio n 5.1 . T h e r e w ere tw o c u t -o ff v a lu es fo r c o n te x ts : F irstly , an y c o n t e x t w ith less th a n 10 o b s e r v a t io n s w as d is ca rd e d . S e c o n d ly , an y c o n t e x t w h ere th e p r o b a b ilit y d is tr ib u ti o n s d id n o t d iffer s u b s ta n tia lly fr o m th e u n c o n d itio n a l o n e w as a lso d is ca rd e d . O n ly th e r e m a in in g on e s w ere u sed fo r d is a m b ig u a tio n . D u e t o th e c o m p u t a t io n a l m o d e l e m p lo y e d , o m it t e d c o n te x ts are e q u iv a le n t t o b a c k in g o f f t o w h a te v e r th e cu rren t p r o b a b ilit y d is tr ib u tio n is. T h e d is tr ib u tio n s c o n d itio n a l o n c o n t e x t s are h ow ev er s u s c e p t ib le t o th e p r o b le m o f sp a rse d a ta . T h is w as h a n d le d u sin g p a rtia l su ccessiv e a b s t r a c t io n as d e s cr ib e d in S e c tio n 5.1 .T h e re su lts are sh o w n in T a b le s 7 a n d 8 . T h e y cle a rly in d ic a te th a t:\u2022 T h e e m p lo y e d tr e a tm e n t o f u n k n o w n w o rd s is q u ite e ffe ctiv e .\u2022 U sin g c o n t e x t u a l in fo r m a t io n , i.e ., trig ra m s , im p ro v e s t a g g in g a ccu ra cy .\u2022 T h e p e r fo r m a n c e is o n p a ir w ith th e H M M ta g g e r an d c o m p a r a b le to sta teo f-t h e -a r t s ta tis tica l p a r t -o f-s p e e c h ta gg ers.", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF14": { |
| "type_str": "figure", |
| "text": "su lts u sin g th e S u sa n n e c o r p u s are s im ila r t o th o s e r e p o rte d fo r th e L a n ca s te r-O s lo -B e r g e n ( L O B ) c o r p u s in [de M a rck e n 1990], w h ere a s t a tis tica l n -b e s t -p a t h a p p r o a c h w as e m p lo y e d t o tra d e p r e cis io n fo r reca ll. T h e ta g g in g sp ee d w as ty p ic a lly a c o u p le o f h u n d re d w o rd s p er s e c o n d o n a S p a rcS e rv e r 1000, b u t va ried w ith th e size o f th e ta g s e t a n d th e a m o u n t o f re m a in in g a m b ig u ity .", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF15": { |
| "type_str": "figure", |
| "text": "e x p e rim e n ts w ith th e H M M a p p r o a c h sh o w th a t it is m u ch h a rd er t o p r o ce s s th e S w e d ish th a n th e E n g lish co rp u s . A lt h o u g h th e tw o c o r p o r a are n o t fu lly c o m p a r a b le b e c a u s e o f th e d ifferen ce s in size a n d ta g s e ts u sed , th e y revea l a s tro n g te n d en cy . T h e d ifficu lty in p r o ce s s in g is m o s t ly d u e t o th e ra th e r la rg e n u m b e r o f u n k n o w n w o rd s in th e S w ed ish c o rp u s a n d th e h ig h e r d e g ree o f a m b ig u ity d e s p ite h a v in g s m a lle r ta g s e ts . T h e s e e ffects m a in ly arise fr o m th e h ig h er m o r p h o lo g ic a l v a r ia tio n 16 Proceedings of NODALIDA 1995", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF16": { |
| "type_str": "figure", |
| "text": "T r a in in g T e s tin g | T h r e s h o ld : 0 Proceedings of NODALIDA 1995", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF17": { |
| "type_str": "figure", |
| "text": "r a m a n d le x ica l s ta tis tics", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF18": { |
| "type_str": "figure", |
| "text": "w edish w h ich ca lls fo r a d d itio n a l stra te g ie s t o b e a p p lie d . T h e s e c o u ld b e th e use o f a la rge co rp u s -in d e p e n d e n t le x ic o n a n d a se p a ra te m o r p h o lo g ic a l a n a ly s is.It is rea ssu rin g t o see th a t th e r e d u c tio n is tic ta g g e r p e r fo r m s as w ell as th e H M M ta g g e r, in d ic a t in g th a t th e n ew fr a m e w o r k is as p o w e rfu l as th e c o n v e n t io n a l o n e w hen u sin g s trictly c o n v e n tio n a l in fo r m a t io n so u rce s .T h e n ew fr a m e w o r k also en a b le s u sin g th e sa m e s o rt o f in fo r m a t io n as th e h ig h ly su cce s sfu l C o n s tra in t G r a m m a r a p p ro a ch , an d th e h o p e is th a t th e a d d itio n o f fu rth e r in fo r m a t io n so u rce s can a d v a n ce s t a t e -o f-t h e -a r t p e r fo r m a n c e o f s ta tis tica l ta gg ers. V ie w ed as an e x te n s io n o f th e C o n s tra in t G r a m m a r a p p r o a c h , th e n ew s ch e m e a llo w s m a k in g d e cis io n s on th e ba sis o f n o t fu lly d is a m b ig u a te d p o r t io n s o f th e in p u t s trin g . T h e a b s o lu te v a lu e o f th e p r o b a b ilit y o f e a ch ta g ca n b e u sed as a q u a n tita tiv e m ecisure o f w h en to r e m o v e a p a rticu la r c a n d id a te ta g a n d w h e n to lea ve in th e a m b ig u ity . T h is p r o v id e s a t o o l t o c o n t r o l th e t r a d e o ff b e tw e e n reca ll (a c c u r a c y ) a n d p re cisio n (r e m a in in g a m b ig u ity ). Proceedings of NODALIDA 1995 [ G o o d 1953] I. J. G o o d . \" T h e p o p u la t io n fre q u e n cie s o f sp e cie s a n d th e e s t im a tio n o f p o p u la tio n p a ra m e te r s \" , B io m e tr ik a 4 0 , p p . 2 3 7 -2 6 4 , 1953. [K a rlss o n ei al 1995] F red K a rls s o n , A t r o V o u tila in e n , J u h a H eik k il\u00e4 a n d A r t o A n t t ila (e d s ). C o n s traint G ra m m a r. A L a n gu a g e-In d ep en d en t S y s tem f o r P a rsin g U n restricted T ext, M o u t o n d e G r u y t e r , B erlin / N ew Y o r k , 1995. [de M a rek en 1990] C a rl G . d e M a reken . \" P a rsin g th e L O B C o r p u s \" , in P rocs. 28th A n n u a l M eetin g o f the A s s o c ia tio n f o r C o m p u ta tio n a l L in gu istics, p p . 2 4 3 -2 5 1 , A C L 1990. [R a b in e r 1989] L . R . R a b in e r . \" A tu to r ia l o n h id d e n M a rk o v m o d e ls a n d s e le cte d a p p lic a t io n s in s p e e ch r e c o g n it io n \" , in Proceed in gs o f the I E E E 7 7 (2 ), p p . 2 5 7 -2 8 5 , 1989. [S a m p s o n 1995] G e o ffre y S a m p s o n . English f o r the C o m p u te r , O x fo r d U n iv e rsity P ress, O x fo r d , 1995.", |
| "uris": null, |
| "num": null |
| }, |
| "FIGREF19": { |
| "type_str": "figure", |
| "text": "in S w e d is h ), S tu d e n tlitte ra tu r, L u n d , S w ed en 1974. [V ite r b i 1967] A . V it e r b i. \" E rro r b o u n d s fo r c o n v o lu t io n a l c o d e s a n d an a s y m p t o t ic a lly o p t im u m d e c o d in g a lg o r it h m \" , in I E E E T ra n sa ction s on In fo r m a tio n T h eory, p p . 2 u tila in e n h H e ik k il\u00e4 1994] A t r o V o u tila in e n a n d J u h a H eik k il\u00e4 . \" A n E n g lish co n s tr a in t g r a m m a r ( E N G C G ) : a s u r fa c e -s y n t a c tic pa rser o f E n g lis h \" , in P rocs. 14th In tern a tio n a l C o n fe r e n c e on English La ngu age R esea rch on C o m p u terize d C o rp o ra , p p . 1 8 9 -1 9 9 , Z u rich , 1994.", |
| "uris": null, |
| "num": null |
| }, |
| "TABREF1": { |
| "type_str": "table", |
| "html": null, |
| "content": "<table/>", |
| "text": "Results of the HMM experiments with the Teleman corpus", |
| "num": null |
| }, |
| "TABREF2": { |
| "type_str": "table", |
| "html": null, |
| "content": "<table/>", |
| "text": "Results of the reductionistic experiments with the Teleman corpus", |
| "num": null |
| }, |
| "TABREF3": { |
| "type_str": "table", |
| "html": null, |
| "content": "<table/>", |
| "text": "Results of the reductionistic experiments with the Susanne corpus", |
| "num": null |
| } |
| } |
| } |
| } |
