{"id":26873,"date":"2019-08-12T12:42:18","date_gmt":"2019-08-12T09:42:18","guid":{"rendered":"http:\/\/www.revolvia.com\/?post_type=mmg_articles&p=26873"},"modified":"2019-08-12T12:42:22","modified_gmt":"2019-08-12T09:42:22","slug":"yapay-zekanin-yazdigi-sahte-metinleri-ayirt-edebilen-yapay-zeka","status":"publish","type":"mmg_articles","link":"https:\/\/www.revolvia.com\/en\/mmg-article\/yapay-zekanin-yazdigi-sahte-metinleri-ayirt-edebilen-yapay-zeka\/","title":{"rendered":"Yapay zekan\u0131n yazd\u0131\u011f\u0131 sahte metinleri ay\u0131rt edebilen yapay zeka"},"content":{"rendered":"\n
\n

Yapay zekaya ili\u015fkin karanl\u0131k senaryolar, birka\u00e7 y\u0131l \u00f6ncesine kadar yaln\u0131zca bilim-kurgu edebiyat\u0131n\u0131n ve sineman\u0131n konusuydu. Oysa bug\u00fcn bilim ve teknoloji otoriteleri taraf\u0131ndan ciddi olarak masaya yat\u0131r\u0131lan meseleler aras\u0131nda. Yapay zekan\u0131n\u00a0insanl\u0131\u011f\u0131 yok edip etmeyece\u011fi<\/a>, yok ederse\u00a0neler olaca\u011f\u0131<\/a>\u00a0s\u0131k\u00e7a tart\u0131\u015f\u0131l\u0131yor. \u00c7\u00fcnk\u00fc ak\u0131ll\u0131 telefonlar arac\u0131l\u0131\u011f\u0131yla yapay zeka art\u0131k hayat\u0131m\u0131z\u0131n tam i\u00e7inde.<\/p>\n

Metin ve g\u00f6rsel i\u00e7erik \u00fcretimi de yapay zekayla d\u00f6n\u00fc\u015fmeye ba\u015flayan alanlardan. Hat\u0131rlayanlar\u0131n\u0131z olacakt\u0131r, ge\u00e7ti\u011fimiz y\u0131l Mynet\u2019in T\u00fcrk\u00e7e i\u00e7erik \u00fcretebilecek bir yapay zeka algoritmas\u0131 \u00fczerinde \u00e7al\u0131\u015ft\u0131\u011f\u0131na\u00a0yer vermi\u015ftik<\/a>. Habercilikten reklamc\u0131l\u0131\u011fa;\u00a0Adobe<\/a>ve\u00a0Nvidia \u00fcr\u00fcnleri<\/a>,\u00a0Remove.bg<\/a>\u00a0ve\u00a0AI Portraits<\/a>\u00a0gibi g\u00f6rsel d\u00fczenleme ara\u00e7lar\u0131na kadar yarat\u0131c\u0131 sekt\u00f6r ve ileti\u015fim sekt\u00f6r\u00fcnden pek \u00e7ok alan y\u00f6n\u00fcn\u00fc yapay zekaya \u00e7evirdi. Peki d\u00fcnyan\u0131n ba\u015f\u0131 sahte haberlerle bu denli beladayken yapay zekan\u0131n \u00fcretebilece\u011fi sahte i\u00e7eriklerle nas\u0131l m\u00fccadele edilebilir? Harvard \u00dcniversitesi ve MIT-IBM Watson Lab\u2019dan ara\u015ft\u0131rmac\u0131lar bu soruya odakland\u0131lar ve b\u00f6ylece yapay zeka taraf\u0131ndan \u00fcretilen metinleri tespit edebilen yapay zekal\u0131 bir ara\u00e7\u00a0geli\u015ftirdiler<\/a>. Bu arac\u0131n ad\u0131\u00a0GLTR<\/span>\u00a0(Giant Language Model Test Room\/Dev Dil Modeli Test Odas\u0131).<\/p>\n

GLTR, belirli bir metin par\u00e7as\u0131n\u0131n bir dil modeli algoritmas\u0131 taraf\u0131ndan \u00fcretilip \u00fcretilmedi\u011fini tespit etmek \u00fczere yarat\u0131lm\u0131\u015f bir sistem. Bir metin girip arac\u0131 denemek isterseniz GLTR\u2019a\u00a0buradan<\/a>\u00a0eri\u015febilirsiniz.<\/p>\n

G\u00fcn\u00fcm\u00fczde sahte haberler \u00fcretme ve yanl\u0131\u015f bilgi yayma amac\u0131yla \u00e7e\u015fitli yapay zeka ve do\u011fal dil olu\u015fturma modelleri kullan\u0131l\u0131yor. Uzmanlar\u0131n ve deneyimli ki\u015filerin bu metinleri ay\u0131rt etmesi m\u00fcmk\u00fcn olabilir. GLTR ise b\u00f6yle bir bilgi ve deneyime sahip olmayan okuyucu i\u00e7in fayda sa\u011flayabilecek bir ara\u00e7.<\/p>\n

\"GLTR\"<\/p>\n

GLTR, testlerde %72 ba\u015far\u0131 oran\u0131na sahip oldu\u011funu g\u00f6sterdi<\/h2>\n

Ara\u015ft\u0131rmac\u0131lar taraf\u0131ndan payla\u015f\u0131lan sonu\u00e7lara g\u00f6re, GLTR insanlar\u0131n sahte metin tespit oran\u0131n\u0131 y\u00fczde 54\u2019ten y\u00fczde 72\u2019ye \u00e7\u0131kard\u0131. Algoritma, farkl\u0131l\u0131klar\u0131 tan\u0131mlayabilmek i\u00e7in metindeki istatistiksel s\u00f6zc\u00fck da\u011f\u0131l\u0131mlar\u0131n\u0131 kullan\u0131\u0131yor. Temel \u00f6nermesi \u015f\u00f6yle; e\u011fer metin bir dil modeli kullan\u0131larak olu\u015fturulmu\u015fsa, o zaman bir insan taraf\u0131ndan yaz\u0131landan daha tahmin edilebilir bir kelime dizisine sahip olur. Ayr\u0131ca yapay zekal\u0131 \u00fcreticiler taraf\u0131ndan olu\u015fturulan c\u00fcmleler, dilbilgisi a\u00e7\u0131s\u0131ndan do\u011fru olsa da ger\u00e7ekte anlams\u0131z metinlerden olu\u015fabilir. Bunu daha basit bir \u015fekilde \u00f6zetlersek, GLTR yapay zekayla metin yaz\u0131l\u0131rken kelimelerin ya da c\u00fcmlelerin ger\u00e7ek anlam\u0131na de\u011fil metindeki istatistiksel \u00f6r\u00fcnt\u00fclere dayand\u0131\u011f\u0131 ger\u00e7e\u011fini temel al\u0131yor. B\u00f6ylece metnin bir insan taraf\u0131ndan yaz\u0131lamayacak kadar tahmin edilebilir oldu\u011funu s\u00f6yleyebiliyor. Sistem en tahmin edilebilir kelimeleri ye\u015fille, daha az tahmin edilebilir kelimeleri k\u0131rm\u0131z\u0131yla ve sar\u0131yla, en az tahmin edilebilir kelimeleri ise morla vurguluyor. Vurgulanan metin \u00e7o\u011funlukla ye\u015fil ve sar\u0131yla vurgulanm\u0131\u015f s\u00f6zc\u00fcklerden olu\u015fuyorsa bu, yapay zeka taraf\u0131ndan olu\u015fturuldu\u011funa dair g\u00fc\u00e7l\u00fc bir g\u00f6sterge say\u0131l\u0131yor. GLTR, yaln\u0131zca sahte metinleri tespit etmekle kalm\u0131yor. Ayn\u0131 zamanda \u00fclkelerin se\u00e7im ve demokratik s\u00fcre\u00e7lerini dahi etkileyebilen Twitter botlar\u0131n\u0131n tespit edilmesi i\u00e7in de potansiyel ta\u015f\u0131yor.<\/p>\n

\"\"
G\u00f6rsel:\u00a0YouTube<\/a><\/em><\/p>\n

\u00a0<\/div>\n<\/section>\n","protected":false},"excerpt":{"rendered":"

Harvard \u00dcniversitesi ve MIT-IBM Watson Lab ara\u015ft\u0131rmac\u0131lar\u0131, yapay zeka taraf\u0131ndan \u00fcretilen metinleri tespit edebilen GLTR (Giant Language Model Test Room) ad\u0131nda yapay zekal\u0131 bir ara\u00e7 geli\u015ftirdiler.<\/p>\n","protected":false},"author":13,"featured_media":26874,"template":"","meta":{"_acf_changed":false,"footnotes":""},"tags":[],"class_list":["post-26873","mmg_articles","type-mmg_articles","status-publish","has-post-thumbnail","hentry"],"translation":{"provider":"WPGlobus","version":"3.0.0","language":"en","enabled_languages":["tr","en"],"languages":{"tr":{"title":true,"content":true,"excerpt":true},"en":{"title":false,"content":false,"excerpt":false}}},"acf":[],"_links":{"self":[{"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/mmg_articles\/26873","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/mmg_articles"}],"about":[{"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/types\/mmg_articles"}],"author":[{"embeddable":true,"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/users\/13"}],"version-history":[{"count":0,"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/mmg_articles\/26873\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/media\/26874"}],"wp:attachment":[{"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/media?parent=26873"}],"wp:term":[{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.revolvia.com\/en\/wp-json\/wp\/v2\/tags?post=26873"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}