{"id":2152,"date":"2024-01-10T15:00:50","date_gmt":"2024-01-10T15:00:50","guid":{"rendered":"https:\/\/cog-ist.com\/?post_type=blog_content&p=2152"},"modified":"2025-09-20T14:09:33","modified_gmt":"2025-09-20T14:09:33","slug":"kara-kutuyu-yorumlamak-buyuk-dil-modelleri-ve-dil-bilgisi-sozdizimi-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-sozdizimi-deniz-ekin-yavas\/","title":{"rendered":"Kara Kutuyu Yorumlamak: B\u00fcy\u00fck Dil Modelleri ve Dil Bilgisi- S\u00f6zdizimi\u00a0\u2014 Deniz Ekin Yava\u015f"},"content":{"rendered":"

Son y\u0131llarda b\u00fcy\u00fck dil modelleri bir\u00e7ok farkl\u0131 g\u00f6revde elde ettikleri ba\u015far\u0131larla b\u00fcy\u00fck ilgi \u00e7ekti. Bu ba\u015far\u0131lar, bizlere b\u00fcy\u00fck dil modellerinin \u00f6ne\u011fitim (pre-training)<\/em> s\u00fcre\u00e7lerinde dilin istatistiksel bilgisinin \u00f6tesinde dile ili\u015fkin \u00e7ok daha fazlas\u0131n\u0131 \u00f6\u011frenip \u00f6\u011frenmedi\u011fini sormam\u0131za neden oldu. Bunun sonucu olarak bu soruyu yan\u0131tlamay\u0131 ama\u00e7layan yorumlanabilirlik (interpretability)<\/em> \u00e7al\u0131\u015fmalar\u0131 bilimsel alanda b\u00fcy\u00fck bir \u00f6nem kazand\u0131. Bu serinin amac\u0131, farkl\u0131 konulardaki yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na de\u011finerek modellerin bili\u015fsel olarak ger\u00e7ek\u00e7ili\u011fini (cognitive plausibility)<\/em> sorgulamak.<\/p>\n\n\n\n

\u00d6nceki yaz\u0131lar\u0131m\u0131zda modellerin s\u00f6zc\u00fck anlam\u0131 ile kavram ve kategorilere ili\u015fkin bilgilerine odakland\u0131k. Bu yaz\u0131yla beraber serimizde oda\u011f\u0131m\u0131z\u0131 dilbilgisine \u00e7ekece\u011fiz ve s\u00f6zdizimi (syntax)<\/em> ile ba\u015flayaca\u011f\u0131z. S\u00f6zdizimi, bir dilde s\u00f6zc\u00fcklerin bir araya gelme kurallar\u0131n\u0131 ve t\u00fcmce olu\u015fum kurallar\u0131n\u0131 \u00e7al\u0131\u015fan dilbilimin bir alt dal\u0131d\u0131r. S\u00f6zdizimsel bilgi, s\u00f6zc\u00fck t\u00fcrleri (part of speech)<\/em>, s\u00f6zc\u00fcklerin t\u00fcmcelerdeki dilbilgisel i\u015flevleri (grammatical function)<\/em>, s\u00f6zc\u00fck s\u0131ras\u0131 (word order) <\/em>gibi bir\u00e7ok bilgiyi i\u00e7erir.<\/p>\n\n\n\n

Bu yaz\u0131da b\u00fcy\u00fck dil modellerinin s\u00f6zdizimine ili\u015fkin bilgi edinip edinmedi\u011fini sorgulayaca\u011f\u0131z ve bu konuda yap\u0131lm\u0131\u015f yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na yer verece\u011fiz. Bu \u00e7al\u0131\u015fmalara ge\u00e7meden \u00f6nce s\u00f6zdizimine ve s\u00f6zdiziminin temel kavramlar\u0131na de\u011finerek ba\u015flayal\u0131m.<\/p>\n\n\n\n

S\u00f6zdizimi ve Temel Kavramlar\u0131<\/strong><\/p>\n\n\n\n

S\u00f6zdizimi, bir dilde s\u00f6zc\u00fcklerin bir araya gelme kurallar\u0131n\u0131 ve t\u00fcmce (s\u00f6zdizimi alan\u0131nda \u201cc\u00fcmle\u201d yerine \u201ct\u00fcmce\u201d terimi tercih edilmektedir) olu\u015fum kurallar\u0131n\u0131 \u00e7al\u0131\u015fan dilbilimin alt dal\u0131d\u0131r. S\u00f6zc\u00fcklerin nas\u0131l bir araya gelerek dilbilgisel t\u00fcmceler olu\u015fturuldu\u011funu, yani t\u00fcmcelerin dilbilgiselli\u011fini (grammaticality) <\/em>a\u00e7\u0131klamay\u0131 ama\u00e7lar.<\/p>\n\n\n\n

Dilde s\u00f6zc\u00fckler, \u201cad\u201d, \u201cs\u0131fat\u201d, \u201czarf\u201d, \u201ceylem\u201d ve birka\u00e7 t\u00fcr daha olmak \u00fczere i\u015flev ve anlamlar\u0131na g\u00f6re farkl\u0131 t\u00fcrlere ayr\u0131l\u0131r; s\u00f6zc\u00fck t\u00fcrleri (part of speech)<\/em> dildeki kurallara g\u00f6re bir araya gelir ve \u00f6bekleri (phrase) <\/em>olu\u015fturur. \u00d6rne\u011fin, \u201ckedi\u201d s\u00f6zc\u00fc\u011f\u00fc bir add\u0131r ve bir s\u0131fat olan \u201ctatl\u0131\u201d s\u00f6zc\u00fc\u011f\u00fc ile birle\u015fmesi ile \u201ctatl\u0131 kedi\u201d \u00f6be\u011fini olu\u015fturur. Ancak bu iki s\u00f6zc\u00fck t\u00fcr\u00fcn\u00fcn nas\u0131l bir araya getirilebilece\u011fine ili\u015fkin kurallar vard\u0131r. Bu \u00f6be\u011fin dilbilgisel bir \u00f6bek olmas\u0131 i\u00e7in ad, s\u0131fattan sonra gelmelidir:<\/p>\n\n\n\n

tatl\u0131 <\/strong>+ kedi \u2192 tatl\u0131 kedi<\/strong><\/p>\n\n\n\n

tatl\u0131 <\/strong>+ kedi \u2947<\/strong> kedi tatl\u0131<\/strong><\/p>\n\n\n\n

T\u00fcmceler ise \u00f6beklerin belirli kurallara g\u00f6re bir araya gelmesiyle olu\u015fur; \u201cTatl\u0131 kedi piyano \u00e7al\u0131yor.\u201d. T\u00fcmcedeki her \u00f6bek di\u011fer \u00f6beklerle kurdu\u011fu ili\u015fkiye g\u00f6re belirli bir dilbilgisel i\u015fleve (grammatical function) <\/em>sahiptir. \u00d6rne\u011fin, \u201ctatl\u0131 kedi\u201d t\u00fcmcenin \u00f6znesi, \u201cpiyano\u201d nesnesi, \u201c\u00e7al\u0131yor\u201d ise eylemidir.<\/p>\n\n\n\n

Tatl\u0131 kedi piyano \u00e7al\u0131yor<\/strong>.<\/p>\n\n\n\n

\u00d6zne Nesne Eylem<\/p>\n\n\n\n

B\u00fct\u00fcn diller i\u00e7in ortak olan kavramlar olsa da baz\u0131 \u00f6zellikler dilden dile farkl\u0131l\u0131k g\u00f6sterebilir. \u00d6rne\u011fin bir\u00e7ok dilde Ad, Eylem gibi s\u00f6zc\u00fck t\u00fcrlerinden ve \u00d6zne, Nesne gibi dilbilgisel i\u015flevlerden bahsedebiliriz. Ancak s\u00f6zc\u00fcklerin ve \u00f6beklerin bir araya gelme kurallar\u0131 dilden dile farkl\u0131l\u0131k g\u00f6sterir.<\/p>\n\n\n\n

Dilden dile farkl\u0131l\u0131k g\u00f6steren \u00f6zelliklerden biri s\u00f6zc\u00fck s\u0131ras\u0131d\u0131r (word order)<\/em>. T\u00fcrk\u00e7ede t\u00fcmcede \u00f6nce \u00d6zne<\/strong> sonra Nesne<\/strong> en son da Eylem<\/strong> gelir, yani T\u00fcrk\u00e7e \u00d6zne-Nesne-Eylem s\u0131ralamas\u0131na (SVO order)<\/em> sahip bir dildir. \u0130ngilizcede ise t\u00fcmcede \u00f6nce \u00d6zne sonra Eylem en son ise Nesne gelir, yani \u0130ngilizce \u00d6zne-Eylem-Nesne s\u0131ralamas\u0131na (SOV order)<\/em> sahip bir dildir. Ancak T\u00fcrk\u00e7e s\u00f6z s\u0131ras\u0131 a\u00e7\u0131s\u0131ndan olduk\u00e7a esnektir ve s\u00f6zc\u00fcklerin farkl\u0131 s\u0131ralanmalar\u0131 \u00e7o\u011funlukla dilbilgisel t\u00fcmceler olu\u015fturur. \u0130ngilizce i\u00e7in ise bu ge\u00e7erli de\u011fildir ve sadece \u00d6zne-Eylem-Nesne s\u0131ralamas\u0131 dilbilgisel t\u00fcmceler olu\u015fturur. A\u015fa\u011f\u0131da \u00f6rneklerde bu fark\u0131 g\u00f6rebilirsiniz. S\u00f6zdiziminde bir t\u00fcmcenin dilbilgisel olmad\u0131\u011f\u0131 t\u00fcmcenin ba\u015f\u0131na konulan \u201c*\u201d i\u015fareti ile belli edilir.<\/p>\n\n\n\n

T\u00fcrk\u00e7e:<\/em><\/p>\n\n\n\n

1. Tatl\u0131 kedi piyano \u00e7al\u0131yor.<\/p>\n\n\n\n

2. Tatl\u0131 kedi \u00e7al\u0131yor piyano.<\/p>\n\n\n\n

3. Piyano \u00e7al\u0131yor tatl\u0131 kedi.<\/p>\n\n\n\n

4. *Piyano tatl\u0131 kedi \u00e7al\u0131yor.<\/p>\n\n\n\n

5. \u00c7al\u0131yor piyano tatl\u0131 kedi.<\/p>\n\n\n\n

6. \u00c7al\u0131yor tatl\u0131 kedi piyano.<\/p>\n\n\n\n

\u0130ngilizce:<\/em><\/p>\n\n\n\n

1. The cute cat is playing the piano.<\/p>\n\n\n\n

2. *The cute cat the piano is playing.<\/p>\n\n\n\n

3. *The piano is playing the cute cat.<\/p>\n\n\n\n

4. *The piano the cute cat is playing.<\/p>\n\n\n\n

5. *Is playing the cute cat the piano.<\/p>\n\n\n\n

6. *Is playing the piano the cute cat.<\/p>\n\n\n\n

S\u00f6zdizimi \u00d6zelinde Sorgulama \u00c7al\u0131\u015fmalar\u0131<\/strong><\/p>\n\n\n\n

Serinin bir \u00f6nceki yaz\u0131s\u0131nda do\u011fal dil i\u015flemlemede s\u00f6zc\u00fcklerin vekt\u00f6rlerle temsil edildiklerini ve bu vekt\u00f6rlere s\u00f6zc\u00fck g\u00f6mmesi (word embedding)<\/em> ad\u0131 verildi\u011fini g\u00f6rd\u00fck. Bu g\u00f6mmelerin s\u00f6zc\u00fcklerin anlamsal \u00f6zelliklerine ili\u015fkin bilgi ta\u015f\u0131d\u0131klar\u0131na ve bunun sonucu olarak ise, anlamsal olarak benzer olan s\u00f6zc\u00fcklerin vekt\u00f6rlerinin birbirlerine benzerlik g\u00f6sterdi\u011fine de\u011findik. Ancak s\u00f6zc\u00fck g\u00f6mmeleri sadece s\u00f6zc\u00fcklerin anlamsal \u00f6zelliklerine ili\u015fkin bilgiyi kodlamaz.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 1: \u201ckedi\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn vekt\u00f6r temsili.<\/figcaption><\/figure>\n\n\n\n

B\u00fcy\u00fck dil modelleri de ayn\u0131 insanlar gibi dili, kullan\u0131m\u0131 \u00fczerinden \u00f6\u011frenir; yani s\u00f6zc\u00fcklerle, t\u00fcmcelerde ve metinlerde kar\u015f\u0131la\u015f\u0131r. Peki \u00e7ocuklarda oldu\u011fu gibi, bunun sonucu olarak dil modelleri de dilin i\u00e7sel kurallar\u0131n\u0131, yani dilbilgisini, bu maruz kalman\u0131n sonucu olarak \u00f6\u011frenir mi? S\u00f6zdizimi oda\u011f\u0131ndaki yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131 bu soruyu yan\u0131tlamay\u0131 ama\u00e7lar. Bu konudaki \u00f6nemli bir \u00e7al\u0131\u015fma alt alan\u0131 ise sorgulama \u00e7al\u0131\u015fmalar\u0131d\u0131r (probing studies)<\/em>. S\u00f6zdizimi oda\u011f\u0131ndaki sorgulama \u00e7al\u0131\u015fmalar\u0131, bu modellerin s\u00f6zdizimine ili\u015fkin ne kadar bilgi \u00f6\u011frendi\u011fini test etmeyi ama\u00e7lamaktad\u0131r. Bu \u00e7al\u0131\u015fmalar dil modellerinin s\u00f6zc\u00fck s\u0131ras\u0131, t\u00fcmcelerin dilbilgiselli\u011fi, s\u00f6zc\u00fck t\u00fcrleri, s\u00f6zc\u00fckler aras\u0131ndaki dilbilgisel ili\u015fkileri \u00f6\u011frenip \u00f6\u011frenmedi\u011fini ortaya \u00e7\u0131karmay\u0131 ama\u00e7lar.<\/p>\n\n\n\n

Sorgulama \u00e7al\u0131\u015fmalar\u0131nda sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131lar\u0131 (probing classifiers)<\/em> kullan\u0131l\u0131r. Sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131larla b\u00fcy\u00fck dil modellerinin g\u00f6mmelerinin ta\u015f\u0131d\u0131\u011f\u0131 bilgiyi ortaya \u00e7\u0131karmak ama\u00e7lan\u0131r. Bu s\u0131n\u0131fland\u0131r\u0131c\u0131lar kullan\u0131larak s\u00f6zc\u00fck g\u00f6mmeleri, dilbilgisel \u00f6zelliklerine g\u00f6re s\u0131n\u0131fland\u0131r\u0131l\u0131r. Bunun i\u00e7in s\u0131n\u0131fland\u0131r\u0131c\u0131ya girdi olarak bir s\u00f6zc\u00fck g\u00f6mmesi verilir. S\u0131n\u0131fland\u0131r\u0131c\u0131 g\u00f6mmeden yola \u00e7\u0131karak bu g\u00f6mmenin temsil etti\u011fi s\u00f6zc\u00fc\u011f\u00fcn dilbilgisel \u00f6zelliklerini tahmin etmeye \u00e7al\u0131\u015farak g\u00f6mmeyi s\u0131n\u0131fland\u0131r\u0131r (bkz. G\u00f6rsel 2). Sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131lar farkl\u0131 dilbilgisel \u00f6zelliklere odaklanabilir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 2: \u201ctatil\u201d ve \u201cyapmak\u201d s\u00f6zc\u00fcklerinin s\u00f6zc\u00fck t\u00fcr\u00fcne g\u00f6re s\u0131n\u0131fland\u0131r\u0131lmas\u0131.<\/figcaption><\/figure>\n\n\n\n

Sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131lar\u0131 ile yap\u0131lan \u00e7al\u0131\u015fmalardaki temel fikir \u015fudur: E\u011fer aranan bilgi s\u00f6zc\u00fck g\u00f6mmelerinde kodlanm\u0131\u015fsa \u00e7ok basit bir s\u0131n\u0131fland\u0131r\u0131c\u0131 bile do\u011fru etiketi \u00fcretebilecektir. Bunun sonucu olarak da s\u0131n\u0131fland\u0131r\u0131c\u0131 bir\u00e7ok \u00f6rnekle ba\u015far\u0131l\u0131 performans g\u00f6sterecektir. Burada ama\u00e7 sadece s\u00f6zc\u00fck g\u00f6mmesinin kendisinden yola \u00e7\u0131karak bir s\u00f6zc\u00fc\u011fe ili\u015fkin dilbilgisel bilgilerin tahmin edilmesidir. Bu nedenle de s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n performans\u0131 bize bu bilginin ne kadar ba\u015far\u0131l\u0131 bir \u015fekilde s\u00f6zc\u00fck g\u00f6mmesinden \u00e7\u0131kar\u0131labildi\u011fini g\u00f6sterir.<\/p>\n\n\n\n

Farkl\u0131 s\u00f6zdizimi bilgileri i\u00e7in sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131lar nas\u0131l e\u011fitilir ve bu s\u0131n\u0131fland\u0131r\u0131c\u0131larla bu bilgiler nas\u0131l ortaya \u00e7\u0131kar\u0131l\u0131r yak\u0131ndan bakal\u0131m.<\/p>\n\n\n\n

S\u00f6zc\u00fck T\u00fcr\u00fc Bilgisi: <\/strong>S\u00f6zc\u00fck t\u00fcr\u00fc bilgisini test etmek i\u00e7in e\u011fitilen bir sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131s\u0131n\u0131n amac\u0131, verilen bir s\u00f6zc\u00fc\u011fe kar\u015f\u0131l\u0131k o s\u00f6zc\u00fc\u011f\u00fcn s\u00f6zc\u00fck t\u00fcr\u00fcn\u00fc (Ad, Eylem, S\u0131fat, vb.) tahmin etmektir. \u00d6rne\u011fin G\u00f6rsel 2\u2019de oldu\u011fu gibi, bu s\u0131n\u0131fland\u0131r\u0131c\u0131ya \u201ctatil\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn s\u00f6zc\u00fck g\u00f6mmesini verdi\u011fimizde s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n \u201cAd\u201d etiketini \u00fcretmesini bekleriz. Benzer \u015fekilde bu s\u0131n\u0131fland\u0131r\u0131c\u0131ya \u201cyapmak\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn s\u00f6zc\u00fck g\u00f6mmesini verdi\u011fimizde s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n \u201cEylem\u201d etiketini \u00fcretmesini bekleriz.<\/p>\n\n\n\n

Dilbilgisel \u0130\u015flev Bilgisi: <\/strong>S\u00f6zc\u00fcklerin dilbilgisel i\u015flevleri i\u00e7in e\u011fitilen bir s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n amac\u0131 benzer bir \u015fekilde, verilen bir s\u00f6zc\u00fc\u011fe kar\u015f\u0131l\u0131k o s\u00f6zc\u00fc\u011f\u00fcn belirli bir t\u00fcmcedeki dilbilgisel i\u015flevini (\u00d6zne, Eylem, Nesne, vb.) tahmin etmektir. \u00d6rne\u011fin bu s\u0131n\u0131fland\u0131r\u0131c\u0131ya G\u00f6rsel 3\u2019deki gibi \u201ctatil\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn iki farkl\u0131 t\u00fcmcedeki g\u00f6mmesini verdi\u011fimizde s\u0131n\u0131fland\u0131r\u0131c\u0131dan farkl\u0131 tahminler bekleriz. \u00c7\u00fcnk\u00fc ilk t\u00fcmcede \u201ctatil\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn dilbilgisel i\u015flevi \u00d6znedir, ikincide ise Nesnedir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 3: \u201ctatil\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fcn farkl\u0131 t\u00fcmcelerdeki dilbgilsel i\u015flevine g\u00f6re s\u0131n\u0131fland\u0131r\u0131lmas\u0131.<\/figcaption><\/figure>\n\n\n\n

\u00d6bek Yap\u0131 Bilgisi: <\/strong>S\u00f6zc\u00fck g\u00f6mmelerinde \u00f6bek yap\u0131 bilgisinin kodlan\u0131p kodlanmad\u0131\u011f\u0131n\u0131 test edebiliriz. \u00d6rne\u011fin \u201cGe\u00e7en yaz Barcelona\u2019da g\u00fczel bir tatil yapt\u0131k.\u201d t\u00fcmcesini d\u00fc\u015f\u00fcnelim. Bu t\u00fcmcenin \u00f6bekleri \u015funlard\u0131r: \u201cge\u00e7en yaz\u201d, \u201cBarcelona\u2019da\u201d \u201cg\u00fczel bir tatil\u201d ve \u201cyapt\u0131k\u201d. G\u00f6rsel 4\u2019te oldu\u011fu gibi, e\u011fitti\u011fimiz s\u0131n\u0131fland\u0131r\u0131c\u0131ya bu t\u00fcmceyi verdi\u011fimizde s\u0131n\u0131fland\u0131r\u0131c\u0131dan t\u00fcmcedeki s\u00f6zc\u00fckleri ait olduklar\u0131 \u00f6beklere g\u00f6re etiketlemesini bekleriz. Yani bu t\u00fcmce i\u00e7in s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n \u201c1, 1, 2, 3, 3, 3, 4\u201d etiketlerini \u00fcretmesi beklenir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 4: \u201cGe\u00e7en yaz Barcelona\u2019da g\u00fczel bir tatil yapt\u0131k.\u201d t\u00fcmcesinin \u00f6bek s\u0131n\u0131fland\u0131r\u0131lmas\u0131.<\/figcaption><\/figure>\n\n\n\n

S\u00f6zc\u00fck S\u0131ras\u0131: <\/strong>S\u00f6zc\u00fck g\u00f6mmelerinde s\u00f6zc\u00fck s\u0131ras\u0131 bilgisinin kodlan\u0131p kodlanmad\u0131\u011f\u0131n\u0131 da test edebiliriz. Daha \u00f6nceki \u00f6rnekten \u201cTatl\u0131 kedi piyano \u00e7al\u0131yor.\u201d t\u00fcmcesini ve onun s\u00f6zc\u00fck s\u0131ras\u0131 bozulmu\u015f hali \u201cPiyano tatl\u0131 kedi \u00e7al\u0131yor.\u201d ele alal\u0131m. Bu g\u00f6revde s\u0131n\u0131fland\u0131r\u0131c\u0131n\u0131n amac\u0131 hangi t\u00fcmcenin bozuk, hangisinin bozuk olmad\u0131\u011f\u0131n\u0131 tahmin etmektir. BERT gibi baz\u0131 modeller s\u00f6zc\u00fck g\u00f6mmelerinin \u00f6tesinde t\u00fcmceler i\u00e7in de g\u00f6mmeler \u00fcretir. B\u00f6yle bir modele bir t\u00fcmce verdi\u011finizde o t\u00fcmceyi temsil eden t\u00fcmce g\u00f6mmesini (sentence embedding) <\/em>modelden alabilirsiniz. Bu g\u00f6rev i\u00e7in t\u00fcmce g\u00f6mmesini kullanabiliriz. Bunun i\u00e7in, G\u00f6rsel 5\u2019deki gibi, bu t\u00fcmceleri temsil eden g\u00f6mmeler s\u0131n\u0131fland\u0131r\u0131c\u0131ya verilir ve s\u0131n\u0131fland\u0131r\u0131c\u0131dan \u201cTatl\u0131 kedi piyano \u00e7al\u0131yor.\u201d t\u00fcmcesi i\u00e7in bozuk de\u011fil; \u201cPiyano tatl\u0131 kedi \u00e7al\u0131yor.\u201d t\u00fcmcesi i\u00e7in ise bozuk etiketi vermesi beklenir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 5: T\u00fcmcelerdeki s\u00f6zc\u00fck s\u0131ras\u0131 ihlalinin s\u0131n\u0131fland\u0131r\u0131lmas\u0131.<\/figcaption><\/figure>\n\n\n\n

Bahsetti\u011fimiz bilgi t\u00fcrleri \u00fczerine yap\u0131lan sorgulay\u0131c\u0131 \u00e7al\u0131\u015fmalar, b\u00fcy\u00fck dil modellerinin bu bilgileri kodlad\u0131\u011f\u0131n\u0131 g\u00f6stermektedir. Liu vd. (2019) ve Tenney vd. (2019) \u00e7al\u0131\u015fmalar\u0131nda \u0130ngilizce BERT ve GPT\u2019nin s\u00f6zc\u00fck t\u00fcr\u00fcn\u00fc, s\u00f6zc\u00fcklerin dilbilgisel i\u015flevlerini ve \u00f6bek yap\u0131s\u0131na ili\u015fkin bilgilerini test etmek i\u00e7in sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131lar e\u011fitilmi\u015ftir. Bu iki \u00e7al\u0131\u015fman\u0131n da sonu\u00e7lar\u0131 bu s\u0131n\u0131fland\u0131r\u0131c\u0131lar\u0131n iki modelin g\u00f6mmeleriyle \u00e7ok ba\u015far\u0131l\u0131 performans g\u00f6sterdi\u011fini (%90 civarlar\u0131nda) ortaya koymaktad\u0131r. Bu da bize bu bilgilerin modellerin g\u00f6mmelerinde kodland\u0131\u011f\u0131n\u0131 g\u00f6stermektedir.<\/p>\n\n\n\n

Otmakhova vd. (2022) ise s\u00f6zc\u00fck s\u0131ras\u0131 bilgisinin kodlanmas\u0131n\u0131n Rus\u00e7a ve \u0130ngilizce BERT\u2019te farkl\u0131 olup olmad\u0131\u011f\u0131na odaklan\u0131r. Bu iki dilin se\u00e7ilme nedeni Rus\u00e7an\u0131n t\u0131pk\u0131 T\u00fcrk\u00e7e gibi s\u00f6zc\u00fck s\u0131ras\u0131 a\u00e7\u0131s\u0131ndan esneklik g\u00f6stermesidir. Bu \u00e7al\u0131\u015fmada beklenen, s\u00f6zc\u00fck s\u0131ras\u0131 esnek olan dillerin modellerinin g\u00f6mmelerinde bu ayr\u0131m\u0131n daha az kodlanmas\u0131d\u0131r. \u00c7\u00fcnk\u00fc bu dillerde s\u00f6z s\u0131ras\u0131 esnektir ve hangi yap\u0131n\u0131n bozuk oldu\u011funu tespit etmek daha zor olabilir. Ancak \u00e7al\u0131\u015fman\u0131n sonu\u00e7lar\u0131, iki dil i\u00e7in de s\u00f6zc\u00fck s\u0131ras\u0131 bilgisinin modellerin g\u00f6mmelerinden kolayl\u0131kla tespit edilebildi\u011fini g\u00f6stermektedir. \u0130lgin\u00e7 \u015fekilde Rus\u00e7a i\u00e7in performans, \u0130ngilizceye k\u0131yasla daha bile y\u00fcksektir.<\/p>\n\n\n\n

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

Serinin bu yaz\u0131s\u0131nda b\u00fcy\u00fck dil modellerinin s\u00f6zdizimsel bilgilerini sorgulayan yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na de\u011findik. Bu yaz\u0131da de\u011findi\u011fimiz \u00e7al\u0131\u015fmalar, sorgulay\u0131c\u0131 s\u0131n\u0131fland\u0131r\u0131c\u0131larla bu modellerin bilgilerini ortaya \u00e7\u0131karmay\u0131 ama\u00e7lamaktad\u0131r. Bu \u00e7al\u0131\u015fmalar\u0131n sonucunda s\u00f6zc\u00fck t\u00fcr\u00fc, s\u00f6zc\u00fcklerin dilbilgisel i\u015flevleri, s\u00f6zc\u00fck s\u0131ras\u0131 ve \u00f6bek yap\u0131 ile ilgili s\u00f6zdizimsel bilgilerin b\u00fcy\u00fck modellerin g\u00f6mmelerinde kodland\u0131\u011f\u0131 sonucuna vard\u0131k.<\/p>\n\n\n\n

Kaynak\u00e7a<\/strong><\/p>\n\n\n\n

Liu, N. F., Gardner, M., Belinkov, Y., Peters, M. E., & Smith, N. A. (2019). Linguistic knowledge and transferability of contextual representations. arXiv preprint arXiv:1903.08855<\/em>.<\/p>\n\n\n\n

Otmakhova, J., Verspoor, K., & Lau, J. H. (2022, Temmuz). Cross-linguistic comparison of linguistic feature encoding in BERT models for typologically different languages. In Proceedings of the 4th Workshop on Research in Computational Linguistic Typology and Multilingual NLP<\/em> (ss. 27\u201335).<\/p>\n\n\n\n

Tenney, I., Xia, P., Chen, B., Wang, A., Poliak, A., McCoy, R. T., \u2026 & Pavlick, E. (2019). What do you learn from context? probing for sentence structure in contextualized word representations. arXiv preprint arXiv:1905.06316<\/em>.<\/p>","protected":false},"featured_media":2153,"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-2152","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\/2152","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\/2152\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media\/2153"}],"wp:attachment":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media?parent=2152"}],"wp:term":[{"taxonomy":"event_publishing_tags","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/event_publishing_tags?post=2152"},{"taxonomy":"kategori","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/kategori?post=2152"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}