{"id":2114,"date":"2023-08-02T15:00:28","date_gmt":"2023-08-02T15:00:28","guid":{"rendered":"https:\/\/cog-ist.com\/?post_type=blog_content&p=2114"},"modified":"2025-09-19T19:53:40","modified_gmt":"2025-09-19T19:53:40","slug":"kara-kutuyu-yorumlamak-buyuk-dil-modelleri-ve-dil-bilgisi-kavramlar-ve-kategoriler","status":"publish","type":"blog_content","link":"https:\/\/cog-ist.com\/en\/blog_content\/kara-kutuyu-yorumlamak-buyuk-dil-modelleri-ve-dil-bilgisi-kavramlar-ve-kategoriler\/","title":{"rendered":"Kara Kutuyu Yorumlamak: B\u00fcy\u00fck Dil Modelleri ve Dil Bilgisi \u2014 Kavramlar ve Kategoriler"},"content":{"rendered":"\n

\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 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 istatiksel bilgisinin d\u0131\u015f\u0131nda 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

Serinin bu yaz\u0131s\u0131nda s\u00f6zc\u00fck anlam\u0131na (word meaning<\/em>) yak\u0131ndan bakaca\u011f\u0131z. S\u00f6zc\u00fck anlam\u0131, dilbilimde farkl\u0131 yakla\u015f\u0131mlar taraf\u0131ndan farkl\u0131 \u015fekillerde a\u00e7\u0131klansa da bili\u015fsel bilim temelli yakla\u015f\u0131mlar, s\u00f6zc\u00fcklerin anlam boyutunu zihindeki kavramlarla ili\u015fkilendirerek a\u00e7\u0131klar[1]<\/a>. Bu yaz\u0131da b\u00fcy\u00fck dil modellerinin kavramlara ili\u015fkin sahip olduklar\u0131 bilgiyi sorgulayaca\u011f\u0131z ve bu konuda yap\u0131lm\u0131\u015f yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na yer verece\u011fiz.<\/p>\n\n\n\n

Bu \u00e7al\u0131\u015fmalara ge\u00e7meden \u00f6nce \u201ckavramlara\u201d ili\u015fkin birka\u00e7 temel bilgiye de\u011finerek ba\u015flayal\u0131m.<\/p>\n\n\n\n

Kavramlar ve Kategoriler<\/strong><\/h2>\n\n\n\n

\u0130nsanlar, kategoriler arac\u0131l\u0131\u011f\u0131yla d\u00fcnya deneyimlerini ve bu deneyimlerde rol alan varl\u0131klar\u0131 tan\u0131r, anlamland\u0131r\u0131r, birbirinden ay\u0131rt eder ve birbiriyle ili\u015fkilendirir. Kategoriler belirli say\u0131da ortak \u00f6zelli\u011fe sahip \u00f6\u011fenin bir araya gelmesiyle olu\u015fur. Kavramlar ise kategorilerin insan zihnindeki temsilleridir. \u00d6rne\u011fin, ger\u00e7ek d\u00fcnyadaki bir g\u00fcvercin \u201cku\u015f\u201d kategorisine aittir ve bu kategori zihnimizde \u201cku\u015f\u201d kavram\u0131na denk gelir.<\/p>\n\n\n\n

Bunun yan\u0131 s\u0131ra \u00f6\u011frendi\u011fimiz yeni bilgiler de kavramlar ve kategoriler arac\u0131l\u0131\u011f\u0131yla anlaml\u0131 hale gelir. \u00d6rne\u011fin, yeni bir \u00f6\u011feyle kar\u015f\u0131la\u015ft\u0131\u011f\u0131m\u0131zda, e\u011fer bu \u00f6\u011fenin bir kategoriye ait oldu\u011funu biliyorsak genel \u00f6zellikleri hakk\u0131nda da bilgi sahibiyiz demektir. Benzer bir \u015fekilde bir kategoriye ili\u015fkin yeni bilgi edindi\u011fimizde bu bilgi art\u0131k bizim i\u00e7in kategori \u00fcyeleri i\u00e7in de ge\u00e7erlidir.<\/p>\n\n\n\n

Ancak kategorilerin tan\u0131mlar\u0131 ve kategori \u00fcyeli\u011fini belirleyen \u00f6zellikler d\u00fc\u015f\u00fcn\u00fcld\u00fc\u011f\u00fc kadar kesin \u00e7izgilerle belirlenmi\u015f de\u011fildir. \u00d6rne\u011fin, \u201cku\u015f\u201d kategorisi \u201cu\u00e7ar\u201d, \u201ckanatl\u0131d\u0131r\u201d gibi \u00f6zellikler ile tan\u0131mlanabilir, ancak u\u00e7mayan bir ku\u015f (penguen) veya kanats\u0131z bir ku\u015f (deformasyonla do\u011fmu\u015f bir ku\u015f) yine de kategori \u00fcyeli\u011fini koruyacakt\u0131r. \u00c7al\u0131\u015fmalar, kategorilerin bulan\u0131k (fuzzy) <\/em>oldu\u011funu, yani <\/em>s\u0131n\u0131rlar\u0131n\u0131n belirsiz oldu\u011funu g\u00f6stermektedir (Hampton, 1979). Herhangi bir \u00f6\u011fenin bir kategoriye ait olup olmad\u0131\u011f\u0131na ili\u015fkin yarg\u0131 hem ki\u015fiden ki\u015fiye (Hampton, 1979), hem de ayn\u0131 ki\u015fi i\u00e7in farkl\u0131 zamanlarda farkl\u0131l\u0131k g\u00f6sterir (McCloskey ve Glucksberg 1978). Bu nedenle ara\u015ft\u0131rmac\u0131lar \u00e7al\u0131\u015fmalar\u0131nda \u00f6\u011felerin bir kategoriye ait olup olmad\u0131\u011f\u0131ndan \u00e7ok \u00fcyelerin tipikli\u011fine (typicality) <\/em>odaklanmaktad\u0131r. Rosch (1973) \u00e7al\u0131\u015fmas\u0131nda kategori \u00fcyelerinden baz\u0131lar\u0131n\u0131n insanlar taraf\u0131ndan daha \u201ciyi\u201d \u00f6rnekler olarak alg\u0131land\u0131\u011f\u0131n\u0131 ortaya koyar. \u00d6rne\u011fin, \u201cmeyve\u201d kategorisini d\u00fc\u015f\u00fcnelim. \u201czeytin\u201d, \u201cmeyve\u201d kategorisine ait bir \u00f6\u011fe olsa da \u201cportakal\u201d, \u201celma\u201d gibi \u00f6\u011feler \u201czeytin\u201d \u00f6\u011fesine k\u0131yasla bu kategorinin daha iyi \u00f6rnekleridir.<\/p>\n\n\n\n

Tipik \u00f6\u011feler<\/em>, bir kategoriyi en iyi temsil eden \u00fcyelerdir.[2]<\/a> Buna kar\u015f\u0131l\u0131k atipik \u00f6\u011feler<\/em>, \u00e7o\u011funlukla kategori \u00fcyeli\u011fi konusunda farkl\u0131 yarg\u0131lara vard\u0131\u011f\u0131m\u0131z, kategori \u00f6zelliklerinin b\u00fct\u00fcn hepsine ya da ay\u0131r\u0131c\u0131 olanlar\u0131na[3]<\/a> sahip olmayan \u00fcyelerdir. \u00d6\u011felerin tipikli\u011finin yeni bilginin \u00f6\u011frenilmesinden dil \u00fcretimine bir\u00e7ok konuda etkisi oldu\u011fu \u00e7e\u015fitli \u00e7al\u0131\u015fmalarda ortaya konulmu\u015ftur (Rosch ve Mervis 1975, Rips vd., 1976, Garrod ve Sanford 1977, Mervis ve Pani 1980, Osherson vd. 1990, Onishi vd. 2008). \u00d6rne\u011fin, Rosch ve Mervis (1975) \u00e7al\u0131\u015fmas\u0131nda kategorilerin tipik \u00fcyelerinin atipik \u00fcyelerden daha kolay \u00f6\u011frenildikleri g\u00f6stermektedir.<\/p>\n\n\n\n

Kavramlar ve Kategoriler \u00d6zelinde Yorumlanabilirlik \u00c7al\u0131\u015fmalar\u0131<\/strong><\/h2>\n\n\n\n

Dil modellerinin bilgilerini sorgulayan yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131nda modellerin bilgilerini ortaya \u00e7\u0131karmak i\u00e7in bir\u00e7ok farkl\u0131 y\u00f6ntem kullan\u0131lmaktad\u0131r. Bunlar\u0131n en pop\u00fclerlerinden biri ise istem <\/em>(prompt<\/em>) temelli tan\u0131 testleridir. Burada ele alaca\u011f\u0131m\u0131z 4 \u00e7al\u0131\u015fmada da (Weir vd. (2020), Ettinger (2020), Ravichander vd. (2020), Misra vd. (2021)), farkl\u0131 ara\u015ft\u0131rma sorular\u0131 i\u00e7in istem temelli testler geli\u015ftirilmi\u015ftir. Bu nedenle, bu \u00e7al\u0131\u015fmalar\u0131 daha iyi anlamak i\u00e7in \u00f6ncelikle istem temelli testleri tan\u0131yal\u0131m ve bu testlerin neleri ama\u00e7lad\u0131klar\u0131na a\u00e7\u0131kl\u0131k getirelim.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 1:<\/strong> Model s\u00f6zc\u00fck tahmini.<\/figcaption><\/figure>\n\n\n\n

\u0130stem temelli testlerde, modele bir istem sunulur ve modelden bu istem do\u011frultusunda s\u00f6zc\u00fck tahmini yapmas\u0131 beklenir. Modelden isteme kar\u015f\u0131l\u0131k olarak s\u0131ral\u0131 bir s\u00f6zc\u00fck listesi al\u0131n\u0131r ve bu liste, o istem i\u00e7in model taraf\u0131ndan en uygun bulunan s\u00f6zc\u00fckleri ve model taraf\u0131ndan bu s\u00f6zc\u00fcklere atanan olas\u0131l\u0131k puanlar\u0131na ili\u015fkin bilgiyi i\u00e7erir. G\u00f6rsel 1\u2019de modele verilen \u201cku\u015flar\u201d istemine kar\u015f\u0131l\u0131k olarak modelin \u00fcretti\u011fi s\u00f6zc\u00fckleri ve ayn\u0131 \u015fekilde s\u00f6zc\u00fcklerin s\u0131ras\u0131n\u0131 ve bu s\u00f6zc\u00fcklere model taraf\u0131ndan atanan olas\u0131l\u0131k puan\u0131n\u0131 g\u00f6rebilirsiniz.<\/p>\n\n\n\n

Ara\u015ft\u0131rmac\u0131lar, istem temelli testler arac\u0131l\u0131\u011f\u0131yla modellerin belirli dilsel ba\u011flamlardaki davran\u0131\u015flar\u0131n\u0131 incelerler. B\u00f6ylelikle bu istemler i\u00e7in \u00fcretilen s\u00f6zc\u00fckler, bu s\u00f6zc\u00fcklerin olas\u0131l\u0131k puanlar\u0131 ve istemler aras\u0131ndaki farkl\u0131l\u0131klar\u0131n model \u00fcretimine etkisi gibi farkl\u0131 bilgilerden yola \u00e7\u0131karak modellerin sahip olduklar\u0131 bilgiler hakk\u0131nda genellemelere varmaya \u00e7al\u0131\u015f\u0131rlar.<\/p>\n\n\n\n

Kavramlar ve \u00d6zellikleri<\/strong><\/h2>\n\n\n\n

Weir vd. (2020) b\u00fcy\u00fck dil modellerinin kavramlar ve kavramlar\u0131n \u00f6zellikleri aras\u0131ndaki ili\u015fkiye ili\u015fkin bilgiyi \u00f6ne\u011fitim s\u00fcre\u00e7lerinde edinip edinmediklerini ele al\u0131r. Bunun i\u00e7in, modellerin bilgisinde kavramlara ili\u015fkin basmakal\u0131p \u00f6rt\u00fck varsay\u0131mlar\u0131n (stereotypic tacit assumptions)<\/em> varl\u0131\u011f\u0131n\u0131 test eder.<\/p>\n\n\n\n

Basmakal\u0131p \u00f6rt\u00fck varsay\u0131mlar (B\u00d6V),<\/em> varl\u0131klar hakk\u0131nda yayg\u0131n olarak kabul edilen inan\u00e7lard\u0131r (Prince 1978) ve varl\u0131klar\u0131n ili\u015fkilendirildi\u011fi kategoriler ve kavramlarla ilgilidirler. B\u00d6V\u2019ler asl\u0131nda kavram \u00f6zellikleridir. B\u00d6V\u2019leri di\u011fer kavram \u00f6zelliklerinden ay\u0131ran, kavramlarla ili\u015fkilendirilen \u00f6zelliklerin ki\u015fiden ki\u015fiye farkl\u0131l\u0131k g\u00f6sterebilmesi, B\u00d6V\u2019lerin ise en s\u0131k ve bir\u00e7ok insan taraf\u0131ndan kavramla ili\u015fkilendirilen \u00f6zellikler olmas\u0131d\u0131r.<\/p>\n\n\n\n

Bu ama\u00e7la Weir vd. (2020), McRae vd. (2005)\u2019de insanlar i\u00e7in tasarlanan testlerden yola \u00e7\u0131karak b\u00fcy\u00fck dil modelleri i\u00e7in iki farkl\u0131 test geli\u015ftirir. \u0130lk testte kavramlardan ba\u015flan\u0131r ve modellerin kavramlara ait \u00f6zellikleri s\u0131ralay\u0131p s\u0131ralayamad\u0131\u011f\u0131na odaklan\u0131l\u0131r. \u0130kinci testte ise B\u00d6V\u2019ler verilerek modellerin bu \u00f6zelliklerden yola \u00e7\u0131karak do\u011fru kavramlara ula\u015f\u0131p ula\u015famad\u0131\u011f\u0131na odaklan\u0131l\u0131r.<\/p>\n\n\n\n

Ku\u015flar\u0131n ___ vard\u0131r.<\/strong><\/p>\n\n\n\n

McRae vd. (2005) \u00e7al\u0131\u015fmas\u0131nda, insanlardan kavramlara ili\u015fkin farkl\u0131 \u00f6zellikleri listelemelerini ister, \u00f6rne\u011fin; ku\u015flar canl\u0131d\u0131r, ku\u015flar\u0131n gagalar\u0131 vard\u0131r, ku\u015flar u\u00e7ar, vb. Weir vd. (2020) ise bu deneyi BERT ve RoBERTa modellerinin bilgisini test etmek i\u00e7in istem temelli testler tasarlayarak bu modeller i\u00e7in uyarlar. Bunun i\u00e7in modellere G\u00f6rsel 2\u2019deki gibi, kavramlar\u0131n farkl\u0131 \u00f6zelliklerini \u00fcretebilecekleri istemler sunar. Bu istemler i\u00e7in \u00fcretilen s\u00f6zc\u00fckler, McRae vd. (2005)\u2019de insanlar taraf\u0131ndan \u00fcretilen \u00f6zelliklerle kar\u015f\u0131la\u015ft\u0131r\u0131l\u0131r ve bu iki liste aras\u0131nda \u00f6rt\u00fc\u015fme olup olmad\u0131\u011f\u0131 de\u011ferlendirilir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 2: Kavram \u00d6zellikleri i\u00e7in Sunulan Farkl\u0131 \u0130stemler.<\/figcaption><\/figure>\n\n\n\n

Bu kar\u015f\u0131la\u015ft\u0131rman\u0131n sonucunda model ve insan \u00fcretimlerinde \u00f6rt\u00fc\u015fmeler oldu\u011fu ortaya \u00e7\u0131kmaktad\u0131r. Ancak bunun yan\u0131 s\u0131ra modellerin sunulan kavramlardan ba\u011f\u0131ms\u0131z olarak s\u0131kl\u0131kla tekrar etti\u011fi \u00fcretimler vard\u0131r. Modellerden bir s\u0131fat \u00fcretilmesi istedi\u011finde modeller \u201cdangerous\u201d (\u2018tehlikeli\u2019) s\u0131fat\u0131n\u0131 addan ba\u011f\u0131ms\u0131z olarak \u00e7ok s\u0131k \u00fcretir. Benzer bir \u015fekilde, \u201cmade of __\u201d (\u2018_dan yap\u0131lm\u0131\u015ft\u0131r\u2019) istemi verildi\u011finde modeller s\u0131kl\u0131kla \u201cwood\u201d (\u2018tahta\u2019) s\u00f6zc\u00fc\u011f\u00fcn\u00fc \u00fcretmektedir. \u00d6rne\u011fin, modele \u201cHamsterlar\u0131n ___ oldu\u011funu herkes bilir.\u201d istemi verildi\u011finde bir\u00e7ok ad i\u00e7in oldu\u011fu gibi bo\u015fluk i\u00e7in \u201ctehlikeli\u201d s\u00f6zc\u00fc\u011f\u00fcn\u00fc \u00fcretir.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 3: \u201cHamsterlar\u0131n tehlikeli oldu\u011funu herkes bilir.\u201d (DALL-E ile \u00fcretilmi\u015ftir. \u0130stem: \u201cA photograph of a dangerous hamster.\u201d)<\/figcaption><\/figure>\n\n\n\n

Peki insanlar ve modellerden al\u0131nan yan\u0131tlar\u0131n kar\u015f\u0131la\u015ft\u0131r\u0131lmas\u0131 bize ger\u00e7ekten anlaml\u0131 sonu\u00e7lar verebilir mi? Bu konuda g\u00f6z \u00f6n\u00fcnde bulundurulmas\u0131 gereken birka\u00e7 \u00f6nemli detay var. \u0130lk olarak modellerin ve insanlar\u0131n bu deneylerdeki s\u00f6zc\u00fck \u00fcretim s\u00fcre\u00e7leri farkl\u0131d\u0131r. Bir istem verildi\u011finde modeller, o istem i\u00e7in dilbilgisel ve hatta anlamsal olarak en uygun s\u00f6zc\u00fckleri s\u0131ralayabileceklerdir. Ancak s\u0131ralad\u0131klar\u0131 s\u00f6zc\u00fckler, insanlardan al\u0131nan yan\u0131tlar gibi deneyin amac\u0131na uygun olmak zorunda de\u011fildir. Weir vd. (2020)\u2019nin de belirtti\u011fi gibi bunun nedeni, insanlardan deney i\u00e7in a\u00e7\u0131k bir \u015fekilde kavramlara ait \u00f6zelliklerin s\u0131ralamas\u0131n\u0131 istememizdir, buna kar\u015f\u0131n modeller b\u00f6yle bir \u00f6nhaz\u0131rlama (priming)<\/em> s\u00fcrecinden ge\u00e7mezler[4]<\/a>.<\/p>\n\n\n\n

Buna ek olarak b\u00f6yle bir g\u00f6revde insanlar, kavrama ait ay\u0131r\u0131c\u0131 \u00f6zellikleri s\u0131ralayacakken[5]<\/a> modelden al\u0131nan yan\u0131t\u0131n bir \u00f6zellik olmas\u0131 durumunda bile, bu \u00f6zelli\u011fin kavram i\u00e7in ay\u0131r\u0131c\u0131 bir \u00f6zellik olmas\u0131n\u0131 beklemek \u00e7ok ger\u00e7ek\u00e7i de\u011fildir. \u00d6rne\u011fin, G\u00f6rsel 1\u2019de \u201cku\u015flar\u201d istemi i\u00e7in model taraf\u0131ndan \u00fcretilen s\u00f6zc\u00fcklerden baz\u0131lar\u0131 \u00f6zellik de\u011fildir (\u201c\u00f6ld\u00fc\u201d, \u201cburada\u201d), baz\u0131lar\u0131 ise \u00f6zellik olsa bile (\u201cg\u00fczeldir\u201d) ay\u0131r\u0131c\u0131 de\u011fildir. Bunlar\u0131 \u00f6nlemek i\u00e7in bu \u00e7al\u0131\u015fmada istemler, \u201cEveryone knows that\u201d (\u2019herkes bilir ki\u2019) yap\u0131s\u0131 ile ba\u015flat\u0131l\u0131r. Bu \u015fekilde modelin, kavramlar\u0131n genel \u00f6zellikleri hakk\u0131nda \u00fcretim yapmas\u0131n\u0131n sa\u011flanmas\u0131 ama\u00e7lan\u0131r.<\/p>\n\n\n\n

___ kanatl\u0131d\u0131r, yumurtlar ve gagalar\u0131 vard\u0131r.<\/strong><\/p>\n\n\n\n

Weir vd. (2020) ikinci deneyde modellerin, kavramlar\u0131n B\u00d6V\u2019lerinden yola \u00e7\u0131karak kavram ad\u0131na ula\u015f\u0131p ula\u015famad\u0131klar\u0131na odaklan\u0131r. Bunun i\u00e7in McRae vd. (2005)\u2019de kavramlar i\u00e7in \u00fcretilen \u00f6zellikleri kullanarak istemler olu\u015fturulur. Her kavram i\u00e7in farkl\u0131 say\u0131da \u00f6zellikler i\u00e7eren (1\u201310) istemler yarat\u0131l\u0131r ve bu istemler i\u00e7in modelden kavram ad\u0131n\u0131 tahmin etmesi istenir. G\u00f6rsel 4\u2019te, Weir vd. (2020)\u2019da \u201cay\u0131\u201d kavram\u0131 i\u00e7in farkl\u0131 say\u0131da B\u00d6V i\u00e7eren istemleri ve buna kar\u015f\u0131l\u0131k modelin s\u00f6zc\u00fck tahminlerini g\u00f6rmekteyiz.<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 4: B\u00d6V\u2019lere g\u00f6re Model Tahmini.<\/figcaption><\/figure>\n\n\n\n

Deneyin sonu\u00e7lar\u0131nda, sadece bir \u00f6zellik verilen istemler i\u00e7in bile modelden al\u0131nan s\u00f6zc\u00fck listelerinde do\u011fru yan\u0131t\u0131n oldu\u011funu g\u00f6rmekteyiz. Verilen \u00f6zellik say\u0131s\u0131 artt\u0131k\u00e7a da do\u011fru tahminlerin say\u0131s\u0131 ve do\u011fru tahmine model taraf\u0131ndan atanan olas\u0131l\u0131k puan\u0131 artmaktad\u0131r (bkz: G\u00f6rsel 4).<\/p>\n\n\n\n

\u0130stemler i\u00e7in \u00fcretilen s\u00f6zc\u00fck listelerine yak\u0131ndan bakt\u0131\u011f\u0131m\u0131zda, do\u011fru yan\u0131t d\u0131\u015f\u0131ndaki di\u011fer y\u00fcksek puanl\u0131 tahminlerin genellikle ayn\u0131 \u00f6zelliklere sahip ba\u015fka kavramlar oldu\u011funu g\u00f6rmekteyiz. Verilen \u00f6zellik say\u0131s\u0131n\u0131n artt\u0131r\u0131lmas\u0131 ile do\u011fru yan\u0131t daha \u00e7ok netle\u015fir ve bu da modelin tahminlerine olumlu olarak yans\u0131r. \u00d6rne\u011fin, \u201c___ b\u00fcy\u00fckt\u00fcr ve t\u00fcyleri ve pen\u00e7eleri vard\u0131r.\u201d t\u00fcmcesi, \u201ckaplan\u201d ve \u201caslan\u201d tahminlerinin ikisi i\u00e7in de do\u011frudur. \u201cormanda ya\u015far\u201d gibi bir \u00f6zelli\u011fin bu \u00f6zelliklere eklenmesi ile hangi kavramdan bahsedildi\u011fi daha \u00e7ok netle\u015fir. Bunun nedeni temel olarak anlam belirsizli\u011finin giderilmesidir (disambiguation)<\/em>. \u0130nsanlara benzer bir deney d\u00fczene\u011fi sunuldu\u011funda insanlar\u0131n da modellere benzer \u015fekilde davranaca\u011f\u0131 tahmin edilebilir.<\/p>\n\n\n\n

Kategori \u00dcyeli\u011fi<\/strong><\/h2>\n\n\n\n
\"\"\/
G\u00f6rsel 5: Kategori \u00dcyeli\u011finin Genellenebilirli\u011fi.<\/figcaption><\/figure>\n\n\n\n

G\u00fcvercin bir ku\u015ftur ve g\u00fcvercin bir ku\u015f de\u011fildir.<\/strong><\/h2>\n\n\n\n

Ettinger (2020) \u00e7al\u0131\u015fmas\u0131nda olumsuzlaman\u0131n (negation)<\/em>, BERT modelinin kategori ad\u0131 \u00fcretimindeki etkisini ele al\u0131r. \u0130lk olarak, modele \u00fcye adlar\u0131n\u0131n oldu\u011fu istemler verilir ve modelin bu istemler i\u00e7in do\u011fru kategori ad\u0131n\u0131 \u00fcretip \u00fcretmedi\u011fi incelenir. Deneyin sonu\u00e7lar\u0131, modelin her istem i\u00e7in do\u011fru kategori ad\u0131n\u0131 \u00fcretebildi\u011fini g\u00f6stermektedir.<\/p>\n\n\n\n

Ancak olumsuzlaman\u0131n dahil edilmesiyle i\u015fler ilgin\u00e7le\u015fir. Ettinger (2020) ayn\u0131 istemleri olumsuz formda modele tekrar verir (\u201cG\u00fcvercin bir ___.\u201d ve \u201cG\u00fcvercin bir ___ de\u011fildir.\u201d). Bunun sonucunda modelin s\u00f6zc\u00fck tahminlerinde herhangi bir de\u011fi\u015fiklik tespit etmez. Model hem olumlu hem de olumsuzluk i\u00e7eren istemler i\u00e7in ayn\u0131 kategori adlar\u0131n\u0131 \u00fcretir. Bu da modelin insanlardan farkl\u0131 anlam \u00fcretme s\u00fcre\u00e7lerine sahip oldu\u011fu ve modelin \u00f6\u011frenmesinin s\u00f6zc\u00fcklerin birlikte bulunma olas\u0131l\u0131klar\u0131n\u0131 ezberlemekten ibaret oldu\u011fu d\u00fc\u015f\u00fcncelerini peki\u015ftirmektedir.<\/p>\n\n\n\n

G\u00fcvercin bir ku\u015ftur. Peki, g\u00fcvercinler ku\u015f mudur?<\/strong><\/h2>\n\n\n\n

Ravichander vd. (2020), Ettinger (2020)\u2019in \u00e7al\u0131\u015fmas\u0131ndan yola \u00e7\u0131karak b\u00fcy\u00fck dil modellerinin kategori \u00fcyeli\u011fine ili\u015fkin bilgilerinin genellenebilirli\u011fini test eder. Bunun i\u00e7in, modelin farkl\u0131 yap\u0131lar i\u00e7in tutarl\u0131 bir \u015fekilde kategori \u00fcyeli\u011fine uygun \u015fekilde s\u00f6zc\u00fck \u00fcretip \u00fcretemedi\u011fine odaklan\u0131r. Bunun i\u00e7in BERT modeline \u00fcye adlar\u0131n\u0131 i\u00e7eren istemleri hem tekil hem de \u00e7o\u011ful olmak \u00fczere iki formda sunar ve modelin iki form i\u00e7in de do\u011fru kategori ad\u0131n\u0131 \u00fcretip \u00fcretmedi\u011fi de\u011ferlendirir. \u00c7al\u0131\u015fman\u0131n sonu\u00e7lar\u0131, Ettinger (2020)\u2019in \u00e7al\u0131\u015fmas\u0131ndan farkl\u0131 olarak tekil t\u00fcmcelerde de modelin her zaman do\u011fru yan\u0131t\u0131 \u00fcretemedi\u011fi g\u00f6stermektedir. Buna ek olarak, modelin \u00e7o\u011ful t\u00fcmcelerde kategori ad\u0131n\u0131 \u00fcretmede daha ba\u015far\u0131s\u0131z oldu\u011fu ortaya \u00e7\u0131kmaktad\u0131r. Bu da bu bilginin modeller i\u00e7in genellenebilir olmad\u0131\u011f\u0131n\u0131 g\u00f6stermektedir.<\/p>\n\n\n\n

Bu bilgi e\u011fer genellenebilir de\u011filse, o zaman bu testlerde model, s\u00f6zc\u00fcklerin s\u0131kl\u0131\u011f\u0131 ve s\u00f6zc\u00fcklerin birlikte bulunma olas\u0131l\u0131klar\u0131 gibi istatiksel bilgilerden mi yararlan\u0131r? Ravichander vd. (2020) ek olarak bu soruyu ayd\u0131nlatmay\u0131 hedefler ve model davran\u0131\u015flar\u0131n\u0131 ve model e\u011fitim verilerini daha yak\u0131ndan inceler. Model davran\u0131\u015flar\u0131nda, do\u011fru yan\u0131t verilen \u00fcye adlar\u0131n\u0131n s\u0131kl\u0131\u011f\u0131 y\u00fcksek s\u00f6zc\u00fckler oldu\u011fu g\u00f6r\u00fclmektedir. Model e\u011fitim verilerinde ise, testlerde ele al\u0131nan \u00fcye ve kategori ad\u0131 \u00e7iftlerinin %78’inin s\u0131n\u0131fland\u0131r\u0131c\u0131 t\u00fcmcelerinin (taxonomic sentences) (\u201cX bir Y\u2019dir.\u201d) bulundu\u011fu belirlenir. Bu y\u00fczde model performans\u0131ndan bile y\u00fcksek bir orand\u0131r[6]<\/a>. Bu da modelin performans\u0131n\u0131n e\u011fitim verilerinde bulunan yap\u0131lar\u0131n ezberlenmesinden kaynakland\u0131\u011f\u0131n\u0131 g\u00f6sterir.<\/p>\n\n\n\n

Tipik \u00d6\u011feler<\/strong><\/h2>\n\n\n\n

Misra vd. (2021) \u00f6\u011felerin tipikli\u011fine ili\u015fkin insanlarda g\u00f6zlemlenen e\u011filimlerin b\u00fcy\u00fck dil modellerinde de g\u00f6r\u00fcl\u00fcp g\u00f6r\u00fclmedi\u011fini ara\u015ft\u0131r\u0131r. Bunun i\u00e7in \u00f6\u011felerin tipikli\u011finin yap\u0131lan deneysel \u00e7al\u0131\u015fmalarda g\u00f6zlemlenen iki \u00f6nemli etkisine odaklan\u0131r. Birincisi, tipik \u00f6\u011felerin kategoriler ile ili\u015fkilendirilmesinin atipik \u00f6\u011felere k\u0131yasla daha kolay olmas\u0131d\u0131r (Rips vd., 1973; Rosch, 1973). \u0130kincisi ise, daha \u00f6nceden bilinmeyen \u00f6zelliklerin tipik \u00f6\u011felerle sunumunun kategorilere ili\u015fkin t\u00fcmevar\u0131msal \u00e7\u0131kar\u0131m\u0131 kolayla\u015ft\u0131rmas\u0131d\u0131r (Rips 1975, Osherson vd. 1990).<\/p>\n\n\n\n

\"\"\/
G\u00f6rsel 6: Tipiklik Testi.<\/figcaption><\/figure>\n\n\n\n

Penguen bir ku\u015ftur ama g\u00fcvercin kadar de\u011fil.<\/strong><\/h2>\n\n\n\n

\u0130nsanlarla yap\u0131lan deneysel \u00e7al\u0131\u015fmalar, \u00f6\u011felerin tipikli\u011finin taksonomik t\u00fcmce do\u011frulama g\u00f6revlerinde (taxonomic sentence verification tasks)<\/em> yan\u0131t s\u00fcreleri \u00fczerinde etkisinin oldu\u011funu g\u00f6stermektedir (Rips vd., 1973; Rosch, 1973). Deneklere verilen \u201cX bir Y\u2019dir.\u201d gibi, X\u2019in bir \u00f6\u011fe ve Y\u2019nin ise kategori ad\u0131 oldu\u011fu s\u0131n\u0131fland\u0131r\u0131c\u0131 bir t\u00fcmcede, X\u2019in tipik bir \u00f6\u011fe oldu\u011fu durumda t\u00fcmcenin do\u011frulu\u011funun onaylanma h\u0131z\u0131, atipik oldu\u011fu durumdan \u00e7ok daha fazlad\u0131r.<\/p>\n\n\n\n

Misra vd. (2021) ayn\u0131 etkiyi b\u00fcy\u00fck dil modellerinin dil \u00fcretiminde g\u00f6rmeyi ama\u00e7lamaktad\u0131r. Bunun i\u00e7in GPT, GPT2, ELECTRA, ALBERT, BERT ve RoBERTa modelleri ile \u00e7al\u0131\u015f\u0131rlar. Bu modellere G\u00f6rsel 6\u2019da g\u00f6sterildi\u011fi gibi \u201cG\u00fcvercin bir\u201d veya \u201cPenguen bir\u201d gibi tipiklikleri farkl\u0131 \u00f6\u011feler i\u00e7eren istemler vererek \u201cku\u015ftur\u201d s\u00f6zc\u00fc\u011f\u00fcne atanan olas\u0131l\u0131k puan\u0131n\u0131 kar\u015f\u0131la\u015ft\u0131r\u0131r.[7]<\/a> Burada beklenen sonu\u00e7 tipikli\u011fi y\u00fcksek \u00f6\u011felerle olas\u0131l\u0131k puan\u0131n\u0131n artmas\u0131d\u0131r, b\u00f6ylelikle modellerin tipiklik konusunda insanlara benzer e\u011filimler g\u00f6sterdi\u011fi sonucuna var\u0131lacakt\u0131r. Deneyin sonu\u00e7lar\u0131 da bunu g\u00f6stermektedir, tipikli\u011fi y\u00fcksek \u00f6\u011felerle olas\u0131l\u0131k puan\u0131 artmaktad\u0131r.<\/p>\n\n\n\n

Bu \u00e7al\u0131\u015fman\u0131n \u00f6nemli yanlar\u0131ndan biri, b\u00f6yle bir davran\u0131\u015f\u0131n temelinde istatiksel bilginin \u00f6\u011frenilmesinin \u00f6tesinde ba\u015fka bir \u00f6\u011freniminin yat\u0131p yatmad\u0131\u011f\u0131n\u0131 incelemesidir. Bunun i\u00e7in Misra vd. (2021), sadece istatiksel bilgiyi \u00f6\u011frenen bir modelle b\u00fcy\u00fck dil modellerinin bu testteki performanslar\u0131n\u0131 kar\u015f\u0131la\u015ft\u0131r\u0131r. Bu kar\u015f\u0131la\u015ft\u0131rma ise b\u00fcy\u00fck dil modellerinin, istatiksel modelle benzer performansa sahip oldu\u011funu g\u00f6stermektedir. Bu da model ve insanlar\u0131n tipiklik yarg\u0131lar\u0131 aras\u0131ndaki benzerliklerin basit istatiksel \u00f6\u011frenmeden kaynakland\u0131\u011f\u0131n\u0131 g\u00f6stermektedir.<\/p>\n\n\n\n

G\u00fcvercinler so\u011fan sever. Peki, ku\u015flar so\u011fan sever mi?<\/strong><\/h2>\n\n\n\n

\u0130nsanlarla yap\u0131lan \u00e7al\u0131\u015fmalar, \u00f6\u011felerin tipikli\u011finin t\u00fcmevar\u0131msal \u00e7\u0131kar\u0131m\u0131 (inductive inference)<\/em> etkiledi\u011fini ortaya koymaktad\u0131r (Rips 1975, Osherson vd. 1990). Osherson vd. (1990) \u00e7al\u0131\u015fmas\u0131nda, kategorilerle ilgili verilen yeni bilgilerin, bilgiler tipik \u00f6\u011felerle sunuldu\u011funda daha inand\u0131r\u0131c\u0131 oldu\u011funu g\u00f6stermektedir. Osherson vd. (1990), bu \u00e7al\u0131\u015fma i\u00e7in bo\u015f y\u00fcklemlemelerden (blank predicates)<\/em> yararlan\u0131r. Bo\u015f y\u00fcklemlemeler, kategoriye ait al\u0131\u015f\u0131lmad\u0131k \u00f6zellikleri ifade eder. \u00d6rne\u011fin, ku\u015flar i\u00e7in \u201cso\u011fan sevmek\u201d b\u00f6yle bir \u00f6zelliktir, \u00e7\u00fcnk\u00fc insan zihninde bu \u00f6zellik ve kategori aras\u0131nda halihaz\u0131rda bir ili\u015fki yoktur. Bu y\u00fcklemlemelerin kullan\u0131m\u0131yla deneyde, zihinde halihaz\u0131rda var olan bilgilerin t\u00fcmevar\u0131msal \u00e7\u0131kar\u0131m s\u00fcrecine etkisi kontrol edilmi\u015f olunur. Osherson vd. (1990)\u2019nin \u00e7al\u0131\u015fmas\u0131 bize \u201cG\u00fcvercinler so\u011fan sever. Ku\u015flar so\u011fan sever.\u201d ve \u201cPenguenler so\u011fan sever. Ku\u015flar so\u011fan sever.\u201d gibi iki \u00e7\u0131kar\u0131m sunuldu\u011funda ilk \u00e7\u0131kar\u0131m\u0131n bizim i\u00e7in daha kolay ve inand\u0131r\u0131c\u0131 oldu\u011funu g\u00f6stermektedir. Bunun nedeni ise \u201cg\u00fcvercin\u201din daha tipik bir \u00f6\u011fe olmas\u0131d\u0131r.<\/p>\n\n\n\n

Misra vd. (2021) bu etkiyi b\u00fcy\u00fck dil modelleri \u00f6zelinde ortaya \u00e7\u0131karmay\u0131 ama\u00e7lar. Bunun i\u00e7in Osherson vd. (1990)\u2019da ele al\u0131nan \u00f6\u011feler ve kategorilere yer verilir ancak bo\u015f y\u00fcklemlemeler yerine bu deneyde anl\u0131k s\u00f6zc\u00fckler (nonce words)<\/em>[8]<\/em><\/strong><\/a> kullan\u0131l\u0131r. Bunun nedeni ise, benzer bir \u015fekilde, var olan s\u00f6zc\u00fcklere ili\u015fkin bilginin model karar\u0131n\u0131 etkilemesini engellemektir. Bu nedenle de modelin s\u00f6z varl\u0131\u011f\u0131nda bu \u00f6zelliklerin bulunmad\u0131\u011f\u0131na emin olmak gerekmektedir.<\/p>\n\n\n\n

Misra vd. (2021), bu deneyde de ilk deneye benzer bir \u015fekilde istemdeki \u00f6\u011fenin tipikli\u011finin \u00fcretilen s\u00f6zc\u00fc\u011f\u00fcn olas\u0131l\u0131k puan\u0131n\u0131 etkileyip etkilemedi\u011fine odaklan\u0131r. Bu deneyde modele sunulan istem ilk t\u00fcmce \u201cG\u00fcvercinler tular.\u201d, model taraf\u0131ndan \u00fcretilmesi beklenen ise ikinci t\u00fcmce \u201cKu\u015flar tular.\u201dd\u0131r (bkz: G\u00f6rsel 6). Bu deneyin sonu\u00e7lar\u0131nda da istemdeki \u00f6\u011fenin tipikli\u011finin \u00fcretimde etkisinin oldu\u011funu g\u00f6rmekteyiz. Ancak bu \u00e7al\u0131\u015fmada \u00f6\u011felerin s\u00f6zc\u00fck s\u0131kl\u0131\u011f\u0131n\u0131n etkileri g\u00f6z \u00f6n\u00fcnde bulundurulmam\u0131\u015ft\u0131r Tipikli\u011fin \u00f6tesinde s\u00f6zc\u00fck s\u0131kl\u0131\u011f\u0131n\u0131n da bu testlerde benzer davran\u0131\u015f \u00f6r\u00fcnt\u00fclerine neden olabilece\u011fi dikkate al\u0131nmal\u0131d\u0131r.<\/p>\n\n\n\n

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

Serinin bu yaz\u0131s\u0131nda b\u00fcy\u00fck dil modellerinin kavramlara ili\u015fkin bilgilerini sorgulayan yorumlanabilirlik \u00e7al\u0131\u015fmalar\u0131na de\u011findik. De\u011finilen \u00e7al\u0131\u015fmalarda kavramlar ve \u00f6zellikleri aras\u0131ndaki ili\u015fkiler, kategori \u00fcyeli\u011fi, tipiklik gibi farkl\u0131 konular oda\u011f\u0131nda b\u00fcy\u00fck dil modellerinin bilgilerinin test edildi\u011fini g\u00f6rd\u00fck. Bu \u00e7al\u0131\u015fmalar\u0131n sonucunda modellerin kavramlarla ilgili bilgilere sahip olduklar\u0131n\u0131 ve model davran\u0131\u015flar\u0131 ile insan davran\u0131\u015flar\u0131 aras\u0131nda bir \u00f6rt\u00fc\u015fme oldu\u011funu tespit etsek de bu bilginin ve \u00f6rt\u00fc\u015fmelerin \u00e7o\u011fu zaman dilin istatistiki bilgisinin \u00f6\u011frenilmesiyle ili\u015fkili oldu\u011fu sonucuna vard\u0131k.<\/p>\n\n\n\n

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

[1]<\/a>S\u00f6zc\u00fck anlam\u0131na ili\u015fkin farkl\u0131 bili\u015fsel bilim temelli yakla\u015f\u0131mlar vard\u0131r. Bu yakla\u015f\u0131mlar, s\u00f6zc\u00fck anlam\u0131 ve kavramlar aras\u0131nda ili\u015fki, \u00f6nce kavramlar\u0131n m\u0131 yoksa s\u00f6zc\u00fcklerin mi geldi\u011fi, kavramlar\u0131n zihinsel olu\u015fumunda dilin etkisi gibi konulardaki a\u00e7\u0131klamalar\u0131 a\u00e7\u0131s\u0131ndan farkl\u0131l\u0131k g\u00f6stermektedir. Genel bilgi i\u00e7in \u015fu ansiklopedi maddesine bakabilirsiniz: https:\/\/plato.stanford.edu\/entries\/word-meaning\/.<\/a><\/p>\n\n\n\n

[2]<\/a> Bir kategori i\u00e7in en tipik \u00fcye prototiptir (prototype)<\/em>.<\/p>\n\n\n\n

[3]<\/a> \u00d6rne\u011fin, \u201cku\u015f\u201d kategorisi i\u00e7in \u201ckanatl\u0131d\u0131r\u201d ay\u0131r\u0131c\u0131 bir \u00f6zelliktir, yani \u201cku\u015f\u201d kategorisini di\u011fer kategorilerden farkl\u0131 ve belirgin k\u0131lan bir \u00f6zelliktir.<\/p>\n\n\n\n

[4]<\/a> Bu \u00e7al\u0131\u015fman\u0131n eskice oldu\u011fu g\u00f6z \u00f6n\u00fcnde bulundurulursa, ChatGPT\u2019yle yap\u0131lacak bir \u00e7al\u0131\u015fmada bu sorun ortadan kald\u0131r\u0131labilir.<\/p>\n\n\n\n

[5]<\/a> McRae vd. (2005), insanlar\u0131n bu \u00f6zelli\u011fin do\u011fru oldu\u011funu bilseler de bir k\u00f6pe\u011fi tan\u0131mlarken \u201ckalemden daha b\u00fcy\u00fck\u201d \u00f6zelli\u011fini kullanmayaca\u011f\u0131n\u0131 belirtir.<\/p>\n\n\n\n

[6]<\/a> Modelin tekil istemlerdeki kategori adlar\u0131n\u0131n \u00fcretimindeki do\u011fruluk oran\u0131 %77\u2019dir.<\/p>\n\n\n\n

[7]<\/a>\u00d6\u011felerin tipikli\u011finin belirlenmesinde, Rosch\u2019ta (1975) insanlardan al\u0131nan tipiklik yarg\u0131lar\u0131 sonu\u00e7lar\u0131 temel al\u0131nm\u0131\u015ft\u0131r.<\/p>\n\n\n\n

[8]<\/a> Anl\u0131k s\u00f6zc\u00fckler, <\/em>dilde var olmayan ve anl\u0131k kullan\u0131m i\u00e7in \u00fcretilen s\u00f6zc\u00fcklerdir.<\/p>\n\n\n\n

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

Ettinger, A. (2020). What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models. Transactions of the Association for Computational Linguistics<\/em>, 8, 34\u201348.<\/p>\n\n\n\n

Garrod, S., & Sanford, A. (1977). Interpreting anaphoric relations: The integration of semantic information while reading. Journal of Verbal Learning and Verbal Behavior<\/em>, 16.77\u201390.<\/p>\n\n\n\n

Hampton, J. A. (1979). Polymorphous concepts in semantic memory. Journal of Verbal Learning and Verbal Behavior<\/em>, 18, 441\u2013461.<\/p>\n\n\n\n

McCloskey, M. E., & Glucksberg, S. (1978). Natural categories: Well defined or fuzzy sets? Memory & Cognition<\/em>, 6, 462\u2013472.<\/p>\n\n\n\n

McRae, K., Cree, G. S., Seidenberg, M. S., & McNorgan, C. (2005). Semantic feature production norms for a large set of living and nonliving things. Behavior Research Methods<\/em>, 37(4), 547\u2013559.<\/p>\n\n\n\n

Mervis, C. B. & Pani J. R. (1980) Acquisition of basic object categories. Cognitive Psychology<\/em>. October; 12(4): 496\u2013522.<\/p>\n\n\n\n

Misra, K., Ettinger, A., & Rayz, J. T. (2021). Do language models learn typicality judgments from text?. In Proceedings of the 43rd Annual Conference of the Cognitive Science Society<\/em>.<\/p>\n\n\n\n

Onishi, K. H., Murphy, G. L., & Bock, K. (2008). Prototypicality in sentence production. Cognitive Psychology<\/em>, 56(2), 103\u2013141.<\/p>\n\n\n\n

Osherson, D. N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based induction. Psychological review<\/em>, 97(2), 185.<\/p>\n\n\n\n

Ravichander, A., Hovy, E., Suleman, K., Trischler, A., & Cheung, J. C. K. (2020). On the systematicity of probing contextualized word representations: The case of hypernymy in BERT. In Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics<\/em> (pp. 88\u2013102).<\/p>\n\n\n\n

Rips, L. J., Shoben, E. J., & Smith, E. E. (1973). Semantic distance and the verification of semantic relations. Journal of verbal learning and verbal behavior<\/em>, 12(1), 1\u201320.<\/p>\n\n\n\n

Rips, L. J. (1975). Inductive judgments about natural categories. Journal of verbal learning and verbal behavior,<\/em> 14(6), 665\u2013681.<\/p>\n\n\n\n

Rosch, E. H. (1973). On the internal structure of perceptual and semantic categories. In T. E. Moore (Ed.), Cognitive development and the acquisition of language<\/em> (pp. 111\u2013144). New York, NY: Academic Press.<\/p>\n\n\n\n

Rosch, E., & Mervis, C. B. (1975). Family resemblance: Studies in the internal structure of categories. Cognitive Psychology<\/em>, 7, 573\u2013605.<\/p>\n\n\n\n

Weir, N., Poliak, A., & Van Durme, B. (2020). Probing neural language models for human tacit assumptions. arXiv preprint arXiv:2004.04877.<\/p>\n","protected":false},"featured_media":2115,"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-2114","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\/2114","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\/2114\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media\/2115"}],"wp:attachment":[{"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/media?parent=2114"}],"wp:term":[{"taxonomy":"event_publishing_tags","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/event_publishing_tags?post=2114"},{"taxonomy":"kategori","embeddable":true,"href":"https:\/\/cog-ist.com\/en\/wp-json\/wp\/v2\/kategori?post=2114"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}