{"id":1976,"date":"2021-11-14T15:00:47","date_gmt":"2021-11-14T15:00:47","guid":{"rendered":"https:\/\/cog-ist.com\/?post_type=blog_content&p=1976"},"modified":"2025-09-07T20:16:08","modified_gmt":"2025-09-07T20:16:08","slug":"esek-yuku-ve-nukteli-sozler-makine-cevirisinin-eksikleri-uzerine-arthur-goldhammer","status":"publish","type":"blog_content","link":"https:\/\/cog-ist.com\/en\/blog_content\/esek-yuku-ve-nukteli-sozler-makine-cevirisinin-eksikleri-uzerine-arthur-goldhammer\/","title":{"rendered":"\u201cE\u015fek y\u00fck\u00fc\u201d ve \u201cN\u00fckteli s\u00f6zler\u201d: Makine \u00c7evirisinin Eksikleri \u00dczerine \u2014 Arthur Goldhammer"},"content":{"rendered":"

\u00d6zg\u00fcn Ad\u0131:\u00a0Shitloads and zingers: on the perils of machine translation<\/a><\/p>\n\n\n\n

Arthur Goldhammer<\/em> yazar, \u00e7evirmen, e\u011fitmen ve Frans\u0131z politikas\u0131 hakk\u0131nda bir blog yazar\u0131d\u0131r. Frans\u0131zca\u2019dan Tocqueville\u2019in Amerika\u2019da Demokrasi<\/em>\u2019si ve Thomas Piketty\u2019nin 21\u2019inci Y\u00fczy\u0131lda Kapital<\/em> kitaplar\u0131 da dahil 120\u2019den fazla kitap \u00e7evirmi\u015ftir. Harvard\u2019s Center for European Studies \u00fcyesidir ve di\u011fer yaz\u0131lar\u0131n\u0131n yan\u0131 s\u0131ra The American Prospect\u2019te yer alan yaz\u0131lar\u0131 da bulunmaktad\u0131r. Cambridge, Massachusetts\u2019te ya\u015famaktad\u0131r.<\/p>\n\n\n\n

Y\u0131llar \u00f6nce, Amsterdam\u2019dan Boston\u2019a giden bir u\u00e7akta, sa\u011f taraf\u0131mda oturan iki Amerikal\u0131 rahibe, Amerika Birle\u015fik Devletleri\u2019ne do\u011fru yola \u00e7\u0131kan heyecanl\u0131 ve gen\u00e7 bir Hollandal\u0131y\u0131 dinliyordu. Rahibelere nereli olduklar\u0131n\u0131 sordu. Ne yaz\u0131k ki, Framingham, Massachusetts seyahat program\u0131nda yoktu, ancak \u201ce\u015fek y\u00fck\u00fcyle (shitload) zaman\u0131 oldu\u011funu ve daha bir\u00e7ok yeri ziyaret edebilece\u011fini\u201d belirtti.

Ne\u015feli gen\u00e7 Hollandal\u0131, g\u00f6r\u00fcn\u00fc\u015fe g\u00f6re, \u201ce\u015fek y\u00fck\u00fc\u201dn\u00fcn, \u201c\u00e7ok fazla\u201d anlam\u0131na geldi\u011finin fark\u0131ndayd\u0131. \u0130ngilizcenin sentaks\u0131na hakimdi ve geni\u015f bir kelime hazinesi vard\u0131. Fakat kelimelerin sosyal ba\u011flamda nas\u0131l kullan\u0131ld\u0131\u011f\u0131na dair deneyimi eksikti.

Yak\u0131n zamanda Google \u00c7eviri motorunun ifade-temelli (phrase-based) bir sistemden n\u00f6ral a\u011f sistemine ge\u00e7ece\u011fi haberi akl\u0131ma aniden bu an\u0131y\u0131 getirdi. (Teknik farkl\u0131l\u0131klar burada<\/a> a\u00e7\u0131klanm\u0131\u015ft\u0131r.) Her iki y\u00f6ntem de makineyi c\u00fcmle \u00e7iftlerinden olu\u015fan bir \u201ckorpus\u201d ile e\u011fitmeye (train) dayan\u0131r: kaynak metin ve \u00e7eviri. Bilgisayar daha sonra, kaynak metindeki kelime s\u0131ras\u0131na g\u00f6re, hedef dilden en olas\u0131 kelime s\u0131ras\u0131n\u0131 \u00e7\u0131karmak (infer) i\u00e7in kurallar \u00fcretir.

Y\u00f6ntem olarak \u00f6r\u00fcnt\u00fc e\u015fle\u015ftirme (pattern matching) kullan\u0131l\u0131r. Ak\u0131ll\u0131 telefonunuzdan sizi \u201cBrookline\u2019e g\u00f6t\u00fcrmesini\u201d istedi\u011finizde s\u00f6yledi\u011finiz heceleri yorumlamak i\u00e7in veya bir foto\u011fraf uygulamas\u0131 arkada\u015f\u0131n\u0131z\u0131n y\u00fcz\u00fcn\u00fc etiketledi\u011finde benzer \u00f6r\u00fcnt\u00fc e\u015fle\u015ftirme algoritmalar\u0131 kullan\u0131l\u0131r. Makine, y\u00fczleri veya hedefleri \u201canlamaz\u201d, onlar\u0131 say\u0131 vekt\u00f6rlerine indirger ve i\u015fler.

Frans\u0131zcadan 125 kitap \u00e7evirmi\u015f profesyonel bir \u00e7evirmen olarak, Google\u2019\u0131n yeni \u00e7eviri motorunun neredeyse bir insan \u00e7evirmen kadar iyi oldu\u011fu, 0 ve 6 aras\u0131nda bir puanlama yapt\u0131\u011f\u0131m\u0131zda insanlar\u0131n 5.1, Google \u00e7eviri motorunun da 5.0 puan alabilece\u011fi iddialar\u0131na sert \u00e7\u0131kmam beklenebilir. Fakat ben ayn\u0131 zamanda doktoras\u0131n\u0131 matematik alan\u0131nda yapm\u0131\u015f, Avrupa gazetelerini d\u00f6rt dilde \u201cokuyan\u201d ve sonu\u00e7lar\u0131 konuya g\u00f6re kategorize eden bir yaz\u0131l\u0131m geli\u015ftirmi\u015f biriyim de ayn\u0131 zamanda. Dolay\u0131s\u0131yla bir makine \u00e7evirmenin yerimi alabilece\u011fi ihtimalinden korkmaktan ziyade, hangi makinelerin nas\u0131l ola\u011fan\u00fcst\u00fc ba\u015far\u0131lara sahip oldu\u011funun fark\u0131nday\u0131m ve Google\u2019\u0131n \u00e7al\u0131\u015fmalar\u0131n\u0131n teknik karma\u015f\u0131kl\u0131\u011f\u0131na ve ustal\u0131\u011f\u0131na hayranl\u0131k beslemekteyim.

Ancak hayranl\u0131\u011f\u0131m, makine \u00e7evirisinin eksikliklerine kar\u015f\u0131 beni k\u00f6r etmiyor. \u201cE\u015fek y\u00fck\u00fcyle\u201d \u0130ngilizce bilen Hollandal\u0131 gen\u00e7 gezgini d\u00fc\u015f\u00fcn\u00fcn. Gen\u00e7 adam\u0131n ak\u0131c\u0131l\u0131\u011f\u0131, \u201cbeyninin\u201d -canl\u0131 n\u00f6ral a\u011f\u0131n\u0131n da diyebilirsiniz- dili do\u011fal k\u0131lan ince kurallar\u0131 (ve istisnalar\u0131) yakalayacak kadar iyi e\u011fitildi\u011fini g\u00f6steriyordu. Bilgisayar dilleri ise ba\u011flam-ba\u011f\u0131ms\u0131z gramerlere <\/a>sahiptir. Ancak gen\u00e7 Hollandal\u0131, anadili \u0130ngilizce olan ki\u015finin diksiyonunu, tonunu ve yap\u0131s\u0131n\u0131 \u015fekillendiren daha ince kurallar\u0131 kavramak i\u00e7in \u0130ngilizce sosyal deneyiminden yoksundu. Anadili \u0130ngilizce olan ki\u015fi de belirli etkileri elde etmek i\u00e7in bu kurallar\u0131 \u00e7i\u011fnemeyi se\u00e7ebilir. Bir \u00e7ift rahibeye \u201cbir s\u00fcr\u00fc yer\u201d yerine \u201ce\u015fek y\u00fck\u00fcyle yer\u201d deseydim, bununla bir anlam ifade ederdim. Hollandal\u0131 gen\u00e7 yanl\u0131\u015fl\u0131kla kendini bir komedide buldu.

Google\u2019\u0131n \u00e7eviri motoru, haber kaynaklar\u0131ndan <\/a>Wikipedia\u2019ya kadar \u00e7e\u015fitli korporalar \u00fczerinde \u201ce\u011fitilmi\u015ftir\u201d. Her korpusun yal\u0131n tan\u0131mlamas\u0131, ortaya \u00e7\u0131kt\u0131\u011f\u0131 ba\u011flam\u0131n tek g\u00f6stergesidir. Bu kadar az bir bilgiden, \u201ce\u015fek y\u00fck\u00fc\u201d gibi bir kelimenin uygunlu\u011funu veya uygunsuzlu\u011funu \u00e7\u0131karmak zor olurdu. Frans\u0131zcaya \u00e7eviri yap\u0131l\u0131yorsa, makine beaucoup (\u00e7ok fazla) veya plusieurs (\u00e7ok, \u00e7e\u015fitli) iyi bir e\u015fle\u015fme sunabilir. Bu, n\u00f6tr plusieurs\u2019in <\/em>aksine toplumsal olarak belirlenmi\u015f<\/em> \u201ce\u015fek y\u00fck\u00fcne\u201d dayanan komediyi<\/em> de\u011fil, ifadenin anlam\u0131n\u0131 verirdi. Algoritma ne kadar karma\u015f\u0131k olursa olsun, verilen enformasyona dayanmal\u0131d\u0131r ve ba\u011flama, \u00f6zellikle de sosyal ba\u011flama ili\u015fkin ipu\u00e7lar\u0131n\u0131 kodda iletmek korkutucu \u015fekilde zordur.

Frans\u0131zca \u201cpetite phrase<\/em>\u201di ele alal\u0131m. \u201cPhrase\u201d<\/em>, \u0130ngilizce\u2019de \u201cc\u00fcmle\u201d veya \u201cpar\u00e7a\u201d anlam\u0131na gelebilir. Marcel Proust bunu \u00c0 la recherche du temps perdu<\/em> (1913\u201327) adl\u0131 roman\u0131nda, \u201cla petite phrase de Vinteuil<\/em>\u201d sat\u0131r\u0131nda m\u00fczikal bir ba\u011flamda kulland\u0131\u011f\u0131nda(phrase<\/em>), \u201cpar\u00e7a\u201d anlam\u0131na gelmelidir; \u00e7\u00fcnk\u00fc \u201cc\u00fcmle\u201dnin burada hi\u00e7bir anlam\u0131 yoktur. Google \u00c7eviri (eski ifade-tabanl\u0131 sistem; yeni tarafs\u0131z a\u011f hen\u00fcz yaln\u0131zca Mandarin \u00c7incesi i\u00e7in kullan\u0131labilmektedir) bu konuda olduk\u00e7a ba\u015far\u0131l\u0131d\u0131r. Yaln\u0131zca petite phrase<\/em>\u2019i \u00e7evirdi\u011finizde \u201ck\u0131sa c\u00fcmle\u201d (kar\u015f\u0131l\u0131\u011f\u0131n\u0131) verir. La petite de Vinteuil<\/em>\u2019i (Vinteuil, bestekar olan bir karakterin ad\u0131d\u0131r) \u00e7evirirseniz, yay\u0131nlanm\u0131\u015f Proust \u00e7evirilerini yank\u0131layarak \u201cVinteuil\u2019in k\u0131sa par\u00e7as\u0131\u201d kar\u015f\u0131l\u0131\u011f\u0131n\u0131 sunar. \u201cVinteuil\u201d ad\u0131n\u0131n nadirli\u011fi gerekli ba\u011flam\u0131 sa\u011flar \u2013ki bunu da istatistiksel algoritma se\u00e7er. Fakat \u201cla petitephrasede Sarkozy<\/em>\u201dyi \u00e7evirirseniz, do\u011fru olan \u201cSarkozy\u2019nin n\u00fckteli s\u00f6z\u00fc\u201d yerine \u201ck\u00fc\u00e7\u00fck par\u00e7a Sarkozy\u201d kar\u015f\u0131l\u0131\u011f\u0131n\u0131 verir. \u00c7\u00fcnk\u00fc, eski cumhurba\u015fkan\u0131n\u0131n ad\u0131n\u0131n da g\u00f6sterdi\u011fi \u00fczere une petite phrase siyasi<\/em> ba\u011flamdad\u0131r ve m\u00fczikal bir par\u00e7adan ziyade siyasi rakibinize kar\u015f\u0131 s\u00f6ylenen n\u00fckteli bir s\u00f6zd\u00fcr. Ancak Sarkozy ad\u0131 o kadar \u00e7e\u015fitli c\u00fcmlelerde ge\u00e7er ki, istatistik motoru onu d\u00fczg\u00fcn bir \u015fekilde kaydedemez- ve ard\u0131ndan bu hatay\u0131 da talihsiz bir gramer hatas\u0131yla birle\u015ftirir.<\/p>\n\n\n\n

MIT\u2019deki \u00f6\u011frencilik g\u00fcnlerime kadar uzanan \u00f6nceki t\u00fcm yapay zeka (AI) yaratma giri\u015fimlerinde oldu\u011fu gibi sorun, zekan\u0131n inan\u0131lmaz derecede karma\u015f\u0131k olmas\u0131d\u0131r. Zeki olmak, yaln\u0131zca kurallardan mant\u0131ksal olarak veya d\u00fczenliliklerden istatistiksel olarak \u00e7\u0131kar\u0131mlarda bulunabilmek de\u011fildir. Bundan \u00f6nce hangi kurallar\u0131n uygulanabilece\u011fini bilmek, duruma duyarl\u0131l\u0131k bilinci gerektiren bir sanat gerekir. Programc\u0131lar \u00e7ok zekidir, ancak anlam\u0131n ortaya \u00e7\u0131kt\u0131\u011f\u0131 \u00e7ok \u00e7e\u015fitli ba\u011flamlar\u0131 tahmin edecek kadar zeki de\u011fillerdir. Bu nedenle, en iyi algoritmalar bile bir \u015feyleri ka\u00e7\u0131racakt\u0131r- ve Henry James\u2019in dedi\u011fi gibi, ideal \u00e7evirmen \u201c\u00fczerinde hi\u00e7bir \u015feyin kaybolmad\u0131\u011f\u0131\u201d bir ki\u015fi olmal\u0131d\u0131r.

Bu, mekanik \u00e7evirinin yararl\u0131 olmad\u0131\u011f\u0131 anlam\u0131na gelmez. \u00c7o\u011fu \u00e7eviri i\u015fi rutindir. Zaman zaman, makineler uygun bir i\u015f yapabilir. Bununla birlikte, mucizeler, ba\u015far\u0131l\u0131 edebi \u00e7eviriler veya uygun \u015fekilde terc\u00fcme edilmi\u015f politik i\u011fneli s\u00f6zler beklemeyin. Ortaya at\u0131lan a\u015f\u0131r\u0131 \u00f6zg\u00fcvenli iddialar, YZ ara\u015ft\u0131rmalar\u0131n\u0131 daha en eski zamanlar\u0131ndan beri besliyor. Bunu i\u015fimden olurum korkusuyla s\u00f6ylemiyorum: \u00c7evirmenlikten emekli oldum ve bug\u00fcnlerde zaman\u0131m\u0131n bir k\u0131sm\u0131n\u0131 \u2026 kod yazmaya ay\u0131r\u0131yorum.<\/p>","protected":false},"featured_media":1977,"template":"","meta":{"_acf_changed":false},"event_publishing_tags":[426,84,280,286,97,94,93,342,953,691,64,92,323,281,285,87,955,62,348,88,954,341,89,848,951,849,950,100,952,425,61,349],"kategori":[305],"class_list":["post-1976","blog_content","type-blog_content","status-publish","has-post-thumbnail","hentry","event_publishing_tags-artificial","event_publishing_tags-artificial-intelligence","event_publishing_tags-bilgisayar","event_publishing_tags-bilgisayar-bilimi","event_publishing_tags-bilim","event_publishing_tags-bilis","event_publishing_tags-bilissel-bilim","event_publishing_tags-buyuk-dil-modelleri","event_publishing_tags-ceviri","event_publishing_tags-cogist","event_publishing_tags-cognition","event_publishing_tags-cognitive-science","event_publishing_tags-cogsci","event_publishing_tags-computer","event_publishing_tags-computer-science","event_publishing_tags-dil","event_publishing_tags-dil-modelleri","event_publishing_tags-dilbilim","event_publishing_tags-intelligence","event_publishing_tags-language","event_publishing_tags-language-models","event_publishing_tags-large-language-models","event_publishing_tags-linguistics","event_publishing_tags-machine","event_publishing_tags-machine-translation","event_publishing_tags-makine","event_publishing_tags-makine-cevirisi","event_publishing_tags-science","event_publishing_tags-translation","event_publishing_tags-yapay","event_publishing_tags-yapay-zeka","event_publishing_tags-zeka","kategori-ceviri"],"acf":[],"_links":{"self":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content\/1976","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content"}],"about":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/types\/blog_content"}],"version-history":[{"count":0,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content\/1976\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media\/1977"}],"wp:attachment":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media?parent=1976"}],"wp:term":[{"taxonomy":"event_publishing_tags","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/event_publishing_tags?post=1976"},{"taxonomy":"kategori","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/kategori?post=1976"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}