{"id":2097,"date":"2023-06-07T15:00:40","date_gmt":"2023-06-07T15:00:40","guid":{"rendered":"https:\/\/cog-ist.com\/?post_type=blog_content&p=2097"},"modified":"2025-09-19T19:50:18","modified_gmt":"2025-09-19T19:50:18","slug":"kara-kutuyu-yorumlamak-buyuk-dil-modelleri-ve-dil-bilgisi-giris-deniz-ekin-yavas","status":"publish","type":"blog_content","link":"https:\/\/cog-ist.com\/en\/blog_content\/kara-kutuyu-yorumlamak-buyuk-dil-modelleri-ve-dil-bilgisi-giris-deniz-ekin-yavas\/","title":{"rendered":"Kara Kutuyu Yorumlamak: B\u00fcy\u00fck Dil Modelleri ve Dil Bilgisi \u2014 Giri\u015f \u2014 Deniz Ekin Yava\u015f"},"content":{"rendered":"

\u201cKara Kutuyu Yorumlamak\u201d serisinin t\u00fcm yaz\u0131lar\u0131na buradan<\/a> eri\u015febilirsiniz.<\/p>\n\n\n\n

Deniz Ekin Yava\u015f, Heinrich-Heine \u00dcniversitesi\u2019nde Hesaplamal\u0131 Dilbilim alan\u0131nda doktora \u00f6\u011frencisi ve ara\u015ft\u0131rma g\u00f6revlisi. \u00d6ne\u011fitimli dil modellerini kullanarak s\u00f6zl\u00fcksel anlambilim ve anlambilim-s\u00f6zdizim kesi\u015fimini ara\u015ft\u0131r\u0131yor.<\/em><\/p>\n\n\n\n

Son zamanlarda GPT, BERT gibi b\u00fcy\u00fck dil modelleri (large language models) hem bilimsel, hem de bilimsel olmayan camiada b\u00fcy\u00fck bir yank\u0131 uyand\u0131rd\u0131. Bunun nedenlerinin ba\u015f\u0131nda bu modellerin bir\u00e7ok farkl\u0131 do\u011fal dil i\u015flemleme g\u00f6revinde \u015fu ana kadarki model performanslar\u0131n\u0131n \u00e7ok \u00fczerinde performans g\u00f6stermesi geliyor.<\/p>\n\n\n\n

Bu modellerin \u00fcst\u00fcn ba\u015far\u0131s\u0131yla beraber akademik alanda yorumlanabilirlik (interpretability)<\/em> \u00e7al\u0131\u015fmalar\u0131 da \u00e7ok b\u00fcy\u00fck bir ilgi g\u00f6rmeye ba\u015flad\u0131. Bu \u00e7al\u0131\u015fmalar, modellerin davran\u0131\u015flar\u0131n\u0131n alt\u0131nda yatan nedenleri ortaya \u00e7\u0131kararak bu modellerin ger\u00e7ekten neler bildi\u011fini ortaya koymay\u0131 ama\u00e7lar. Bu yaz\u0131 serisinde farkl\u0131 konulardaki yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na de\u011finerek bu \u00e7al\u0131\u015fmalar\u0131n bize bu modellerin sahip oldu\u011fu bilgiler hakk\u0131nda neler s\u00f6yledi\u011fini g\u00f6rece\u011fiz. Bu modeller ger\u00e7ekten d\u00fc\u015f\u00fcnd\u00fc\u011f\u00fcm\u00fcz kadar \u201czeki\u201d mi? \u0130nsana benzer \u015fekilde \u201cd\u00fc\u015f\u00fcnme\u201d, kavramlar\u0131 temsil etme ve kavramlar aras\u0131nda ili\u015fki kurma \u00f6zelliklerine sahip mi? Bu yaz\u0131 dizisinde bu sorular\u0131, alanda yap\u0131lan akademik \u00e7al\u0131\u015fmalardan \u00f6rnekler vererek yan\u0131tlamaya \u00e7al\u0131\u015faca\u011f\u0131z. Bu serinin sonunda amac\u0131m\u0131z ise, bu dil modellerinin bili\u015fsel olarak insan zihnine benzerli\u011fine, yani bili\u015fsel ger\u00e7ek\u00e7ili\u011fine<\/em> (cognitive plausibility)<\/em> ili\u015fkin fikir sahibi olmak.<\/p>\n\n\n\n

Bu yaz\u0131da \u00f6ncelikli olarak b\u00fcy\u00fck dil modelleri hakk\u0131nda temel bilgilere de\u011finece\u011fiz ve \u201cyorumlanabilirlik\u201d terimine ve yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131n\u0131n \u00f6nemine a\u00e7\u0131kl\u0131k getirece\u011fiz.<\/p>\n\n\n\n

\u00d6ne\u011fitimli B\u00fcy\u00fck Dil Modelleri<\/strong><\/h2>\n\n\n\n

Son y\u0131llarda do\u011fal dil i\u015flemleme alan\u0131ndaki geli\u015fmeler alanda b\u00fcy\u00fck bir \u00e7\u0131\u011f\u0131r a\u00e7t\u0131. Bu geli\u015fmelerin ba\u015f\u0131nda yapay sinir a\u011flar\u0131 (artificial neural networks) modellerinin bir devam\u0131 niteli\u011findeki BERT (Devlin vd. 2018), GPT (Floridi ve Chiriatti, 2020) gibi, \u00f6ne\u011fitimli (pre-trained) transformers modellerinin<\/em> geli\u015ftirilmesi gelmektedir. \u00d6ne\u011fitimli transformers modelleri<\/em>, dil modelleme amac\u0131yla e\u011fitilirler ve bu nedenle b\u00fcy\u00fck dil modelleri<\/em> olarak da adland\u0131r\u0131l\u0131rlar. Dil modelleme amac\u0131yla e\u011fitilme s\u00fcre\u00e7leri \u00f6ne\u011fitim s\u00fcreci<\/em> olarak adland\u0131r\u0131l\u0131r. Modeller bu s\u00fcre\u00e7te \u00e7ok b\u00fcy\u00fck say\u0131da ham metne maruz b\u0131rak\u0131l\u0131r ve bu s\u00fcre\u00e7 sonunda bu metinler \u00fczerinden dildeki s\u00f6zc\u00fcklerin da\u011f\u0131l\u0131m\u0131, yani dile ili\u015fkin temel istatiksel bilgiyi edinmi\u015f olurlar.<\/p>\n\n\n\n

Bu modellerin farkl\u0131 do\u011fal dil i\u015flemleme g\u00f6revlerine uyarlanmas\u0131 i\u00e7in ise ek bir e\u011fitim s\u00fcrecinden ge\u00e7meleri gerekmektedir; bu s\u00fcre\u00e7 ince ayar<\/em> (fine tuning)<\/em> s\u00fcrecidir. B\u00fcy\u00fck dil modellerinin bu noktadaki avantaj\u0131 ise g\u00f6revlere y\u00f6nelik e\u011fitimlerinin \u00e7ok h\u0131zl\u0131 ve kolay olmas\u0131d\u0131r. Bu modeller, g\u00f6reve ili\u015fkin az say\u0131da veri ile ve k\u0131sa bir ince ayar s\u00fcrecinden sonra g\u00f6rev i\u00e7in kullan\u0131labilir hale getirilirler. Son zamanlarda bu modellerin bir\u00e7ok farkl\u0131 do\u011fal dil i\u015flemleme g\u00f6revi i\u00e7in kullan\u0131ld\u0131\u011f\u0131n\u0131 ve bu g\u00f6revlerde rekor puanlar\u0131 rahatl\u0131kla ge\u00e7tiklerini g\u00f6r\u00fcyoruz. Bunun temel nedeni ise modellerin \u00f6n\u00f6\u011frenme s\u00fcrecinde edindi\u011fi dile ili\u015fkin genel bilginin ince ayar s\u00fcreciyle g\u00f6reve aktar\u0131labilmesi ve bu bilginin do\u011fal dil i\u015flemleme g\u00f6revleri i\u00e7in olduk\u00e7a kullan\u0131\u015fl\u0131 olmas\u0131d\u0131r.[1]<\/a><\/p>\n\n\n\n

Yorumlanabilirlik ve B\u00fcy\u00fck Dil Modelleri<\/strong><\/h2>\n\n\n\n

Yapay sinir a\u011f\u0131 modelleri do\u011falar\u0131 gere\u011fi bir kara kutudur. Modele verilen bir girdi sonucunda bir \u00e7\u0131kt\u0131 al\u0131n\u0131r, ancak modelin bu s\u00fcre\u00e7teki davran\u0131\u015flar\u0131 ve kararlar\u0131 hakk\u0131nda bir fikir sahibi olunmaz. Bir ba\u015fka \u015fekilde ifade etmek gerekirse, bu modellerin davran\u0131\u015flar\u0131n\u0131n ve model \u00f6ng\u00f6r\u00fclerinin alt\u0131nda yatan nedenler ara\u015ft\u0131rmac\u0131lar i\u00e7in eri\u015filebilir ve do\u011frudan yorumlanabilir de\u011fildir. Yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131 ise bu nedenleri ortaya \u00e7\u0131karmay\u0131 ama\u00e7lar. Bu, hem modelin \u00e7al\u0131\u015fma prensipleri hakk\u0131nda daha \u00e7ok bilgi verirken hem de modelin performans\u0131n\u0131n iyile\u015ftirilmesi i\u00e7in yap\u0131labilecekler hakk\u0131nda ipucu verir.<\/p>\n\n\n\n

\"\"\/
Kaynak: https:\/\/xkcd.com\/1838\/<\/a><\/figcaption><\/figure>\n\n\n\n

Peki, b\u00fcy\u00fck dil modelleri bir\u00e7ok farkl\u0131 do\u011fal dil i\u015flemleme g\u00f6revinde \u00e7ok ba\u015far\u0131l\u0131 sonu\u00e7lar elde ederken modellerin performanslar\u0131ndaki ba\u015far\u0131lar\u0131n\u0131 tam olarak neye bor\u00e7luyuz? Bu modeller \u00f6ne\u011fitim s\u00fcre\u00e7lerinde dilin istatiksel bilgisinin d\u0131\u015f\u0131nda dile, dilin yap\u0131s\u0131na, kavramlara ili\u015fkin bilgileri \u00f6\u011freniyor mu? Yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131 b\u00fcy\u00fck dil modelleri \u00f6zelinde bu sorular\u0131 yan\u0131tlamaya \u00e7al\u0131\u015f\u0131r. Bu nedenle de modellerin sahip oldu\u011fu dil bilgisini s\u00f6zdizim, anlambilim gibi farkl\u0131 alanlarda test etmeyi ama\u00e7lar; \u00f6rne\u011fin, dilbilgisellik yarg\u0131s\u0131, \u00f6zne-y\u00fcklem uyumu, s\u00f6zc\u00fck t\u00fcr\u00fc bilgisi, s\u00f6zdizimsel ili\u015fkiler, kavramlar aras\u0131 olgusal ili\u015fkiler, s\u00f6zc\u00fckler aras\u0131 anlamsal ili\u015fkiler, \u00e7okanlaml\u0131l\u0131k, vb.<\/p>\n\n\n\n

\"\"\/
\u015eekil 2: Soru Yan\u0131tlama G\u00f6revi<\/em><\/figcaption><\/figure>\n\n\n\n

Do\u011fal dil i\u015flemleme g\u00f6revlerindeki performans ve dil bilgisi ili\u015fkisini birka\u00e7 \u00f6rnek \u00fczerinden a\u00e7\u0131klamak bunu daha anla\u015f\u0131l\u0131r k\u0131lacakt\u0131r. \u00d6rne\u011fin, soru yan\u0131tlama<\/em> g\u00f6revini ele alal\u0131m. Soru yan\u0131tlama g\u00f6revinde modelden, verilen bir metin \u00fczerinden sorulan soruyu yan\u0131tlamas\u0131 beklenir.[2]<\/a> B\u00f6yle bir g\u00f6revin model taraf\u0131ndan ba\u015far\u0131l\u0131 bir \u015fekilde ger\u00e7ekle\u015ftirilmesi i\u00e7in \u00f6\u011feler aras\u0131ndaki artg\u00f6nderimsel (anaphoric) ili\u015fkinin \u00e7\u00f6z\u00fclebilmesi<\/em> olduk\u00e7a \u00f6nemlidir[3]<\/a>. \u00d6rne\u011fin; \u015eekil 2\u2019de verilen \u00f6rnekte \u201c\u0130stanbul\u2019da ya\u015f\u0131yorum.\u201d t\u00fcmcesinin \u00f6znesinin ba\u015far\u0131l\u0131 bir \u015fekilde Minno\u015f olarak belirlenebilmesi i\u00e7in (bo\u015f) artg\u00f6nderimin do\u011fru \u015fekilde \u00e7\u00f6z\u00fcmlenebilmesi gerekmektedir. Bunun yan\u0131 s\u0131ra ba\u015fka dilsel bilgiler de bu g\u00f6revde etkili olabilir. \u00d6rne\u011fin; t\u00fcmcenin \u00f6\u011feleri, bu \u00f6\u011feler aras\u0131ndaki hiyerar\u015fik ili\u015fkiler, eylemlerin \u00fcye yap\u0131lar\u0131 ve \u00fcyelerin anlambilimsel rolleri, vb. \u00d6rne\u011fin \u015eekil 2\u2019deki sorunun do\u011fru yan\u0131tlanmas\u0131nda, \u201cya\u015fa-\u201d eylemi ve eylemin \u00fcyeleri aras\u0131ndaki ili\u015fkinin do\u011fru tan\u0131mlanmas\u0131 etkili olabilir. Bunun i\u00e7in de bu dilsel bilgiler gereklidir.<\/p>\n\n\n\n

\"\"\/
\u015eekil 3: Duygu Analizi G\u00f6revi<\/em><\/figcaption><\/figure>\n\n\n\n

Ba\u015fka bir \u00f6rnek olarak duygu analizi g\u00f6revi <\/em>verilebilir. Bu g\u00f6revdeki ama\u00e7 t\u00fcmceleri olumlu veya olumsuz duygu ta\u015f\u0131malar\u0131na g\u00f6re s\u0131n\u0131fland\u0131rmakt\u0131r. Bu g\u00f6rev i\u00e7in en \u00f6nemli dilsel bilgi, olumsuzlama, olumsuzlama t\u00fcrleri (\u00f6rne\u011fin, t\u00fcmce veya s\u00f6zc\u00fck d\u00fczeyinde) ve olumsuzlaman\u0131n kapsam alan\u0131na ili\u015fkin bilgidir. \u00d6rne\u011fin, \u201cFilm k\u00f6t\u00fc de\u011fildi.\u201d gibi olumsuz duygu ta\u015f\u0131yan bir s\u00f6zc\u00fckle beraber olumsuzlaman\u0131n kullan\u0131ld\u0131\u011f\u0131 bir t\u00fcmce, model i\u00e7in zorlay\u0131c\u0131 olacakt\u0131r. \u00c7\u00fcnk\u00fc bir t\u00fcmcenin olumsuz bir \u00f6\u011fe i\u00e7ermesi, o t\u00fcmcenin olumsuz duygu ta\u015f\u0131d\u0131\u011f\u0131 anlam\u0131na gelmemektedir ama bu ayr\u0131m model i\u00e7in \u00e7ok net de\u011fildir.[4]<\/a><\/p>\n\n\n\n

Yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131yla ara\u015ft\u0131rmac\u0131lar, b\u00fcy\u00fck dil modellerinin farkl\u0131 konulardaki bilgilerini \u00f6l\u00e7mek i\u00e7in tan\u0131sal testler geli\u015ftirmeyi ve bu testlerle model davran\u0131\u015flar\u0131n\u0131 veya g\u00f6sterimlerini \u201cyorumlanabilir\u201d hale getirmeyi ama\u00e7lamaktad\u0131r. B\u00f6ylelikle, modellerin sahip oldu\u011fu dilsel bilgi hakk\u0131nda bir sonuca ula\u015fmak m\u00fcmk\u00fcn olmaktad\u0131r. Ancak bu testlerin dilin istatisti\u011fine ili\u015fkin bilgiyle de\u011fil, dilsel bilginin kendisiyle ilgili sonu\u00e7lara ula\u015ft\u0131rd\u0131\u011f\u0131ndan emin olmak gerekmektedir.<\/p>\n\n\n\n

Sonu\u00e7<\/strong><\/h2>\n\n\n\n

Bu makalede b\u00fcy\u00fck dil modelleri ve yorumlanabilirlik kavram\u0131na ili\u015fkin genel bilgilere de\u011findik. Serinin bir sonraki yaz\u0131lar\u0131nda farkl\u0131 konular \u00f6zelinde akademideki yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na yak\u0131ndan bakaca\u011f\u0131z ve bu \u00e7al\u0131\u015fmalar\u0131n ba\u015far\u0131l\u0131 b\u00fcy\u00fck dil modellerinin bili\u015fsel ger\u00e7ek\u00e7ili\u011fi konusunda hangi sonu\u00e7lara vard\u0131\u011f\u0131n\u0131 g\u00f6rece\u011fiz.<\/p>\n\n\n\n

Notlar<\/strong><\/h2>\n\n\n\n

[1]<\/a> Farkl\u0131 \u00f6ne\u011fitimli b\u00fcy\u00fck dil modelleri, \u00f6ne\u011fitim y\u00f6ntemleri, ince ayar y\u00f6ntemleri ve bu modellerin kullan\u0131ld\u0131\u011f\u0131 farkl\u0131 g\u00f6revler ile ilgili daha fazla bilgi i\u00e7in bkz: Qiu vd. 2020, Han vd. 2021.<\/p>\n\n\n\n

[2]<\/a> Bu g\u00f6rev metin odakl\u0131 olmadan da yap\u0131labilir. Ba\u011flam verilerek yap\u0131lan soru yan\u0131tlama g\u00f6revi a\u00e7\u0131k soru yan\u0131tlama,<\/em> verilmeden yap\u0131lan ise kapal\u0131 soru yan\u0131tlama<\/em> g\u00f6revi olarak adland\u0131r\u0131lmaktad\u0131r.<\/p>\n\n\n\n

[3]<\/a> Artg\u00f6nderim \u00e7\u00f6z\u00fcmlemesinin soru yan\u0131tlama g\u00f6revleri a\u00e7\u0131s\u0131ndan \u00f6nemi i\u00e7in bkz: Vicedo ve Ferr\u00e1ndez 2000.<\/p>\n\n\n\n

[4]<\/a> Duygu analizi ve olumsuzlama ili\u015fkisi i\u00e7in bkz: Wiegand vd. 2010.<\/p>\n\n\n\n

Kaynak\u00e7a<\/h2>\n\n\n\n

Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.<\/p>\n\n\n\n

Floridi, L., & Chiriatti, M. (2020). GPT-3: Its nature, scope, limits, and consequences. Minds and Machines, 30, 681\u2013694.<\/p>\n\n\n\n

Han, X., Zhang, Z., Ding, N., Gu, Y., Liu, X., Huo, Y., \u2026 & Zhu, J. (2021). Pre-trained models: Past, present and future. AI Open, 2, 225\u2013250.<\/p>\n\n\n\n

Qiu, X., Sun, T., Xu, Y., Shao, Y., Dai, N., & Huang, X. (2020). Pre-trained models for natural language processing: A survey. Science China Technological Sciences, 63(10), 1872\u20131897.<\/p>\n\n\n\n

Vicedo, J. L., ve Ferr\u00e1ndez, A. (2000). Importance of pronominal anaphora resolution in question answering systems. In Proceedings of the 38th Annual Meeting of the Association for Computational Linguistics<\/em> (pp. 555\u2013562).<\/p>\n\n\n\n

Wiegand, M., Balahur, A., Roth, B., Klakow, D., & Montoyo, A. (2010). A survey on the role of negation in sentiment analysis. In Proceedings of the workshop on negation and speculation in natural language processing<\/em> (pp. 60\u201368).<\/p>","protected":false},"featured_media":2098,"template":"","meta":{"_acf_changed":false},"event_publishing_tags":[479,426,84,280,286,94,93,924,627,342,416,345,691,64,923,92,323,281,285,87,1057,62,1056,348,88,1058,1059,1053,341,89,343,344,76,1054,1055,346,425,61,349,96],"kategori":[725],"class_list":["post-2097","blog_content","type-blog_content","status-publish","has-post-thumbnail","hentry","event_publishing_tags-ai","event_publishing_tags-artificial","event_publishing_tags-artificial-intelligence","event_publishing_tags-bilgisayar","event_publishing_tags-bilgisayar-bilimi","event_publishing_tags-bilis","event_publishing_tags-bilissel-bilim","event_publishing_tags-bilissel-dilbilim","event_publishing_tags-buyuk-dil-modeli","event_publishing_tags-buyuk-dil-modelleri","event_publishing_tags-chatbot","event_publishing_tags-chatgpt","event_publishing_tags-cogist","event_publishing_tags-cognition","event_publishing_tags-cognitive-linguistics","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-modeli","event_publishing_tags-dilbilim","event_publishing_tags-dogal-dil-isleme","event_publishing_tags-intelligence","event_publishing_tags-language","event_publishing_tags-language-model","event_publishing_tags-language-modeli","event_publishing_tags-large-language-model","event_publishing_tags-large-language-models","event_publishing_tags-linguistics","event_publishing_tags-llm","event_publishing_tags-llms","event_publishing_tags-mind","event_publishing_tags-natural-language-processing","event_publishing_tags-nlp","event_publishing_tags-openai","event_publishing_tags-yapay","event_publishing_tags-yapay-zeka","event_publishing_tags-zeka","event_publishing_tags-zihin","kategori-yazi"],"acf":[],"_links":{"self":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/blog_content\/2097","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\/2097\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media\/2098"}],"wp:attachment":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media?parent=2097"}],"wp:term":[{"taxonomy":"event_publishing_tags","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/event_publishing_tags?post=2097"},{"taxonomy":"kategori","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/kategori?post=2097"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}