Tokenisation sous-baatu
Tokenization subword dafay xaaj mbind ci pàcc yu gëna ndaw baat yi waaye gëna mag ci araf yi, lu melni 'token' boole ci 'ization'.
Résumé
It is the standard way modern language models turn text into the discrete IDs they actually process, balancing vocabulary size against meaning.
Plongeur bu xóot
Baatu yi bari nañu lool te kenn mënu leen lim (vocabulaire yi dina ñu bari lool te duñu am baat yu bari), waaye araf yu benn-benn yi duñu am lu bari luñuy tekki, te dañuy tax ñu toppalante lu gudd lool. Subword tokenization mooy kompromis bi: dafay tëye baat yu bari yi ñu mat waaye di xaaj baat yu bariwul wala yu jafee xam ci ay pàcc yu am solo. 'Tiis' mën na nekk 'un', 'happi', 'ness'. Algorithm yi gëna mag ñooy Kodaasu Ñaari Byte (GPT moo koy jëfandikoo), WordPiece (BERT moo koy jëfandikoo), ak Unigram/SentencePiece (T5 ak model yu bari yu làkk yu bari dañu koy jëfandikoo). Xeetu jëfandikoo gii dafay jëfandikoo kàddu yuñu gisul ci anam wu rafet, di séddoo ay piyees ci kàddu yu jëm ci ('fo', 'fo', 'fo'), te jàppale làkk wumu mëna doon. Fragment bu nekk dafay méngoo ak ID integer, te ID yooyu ñooy li couche embedding model bi di soppi def ko ay vecteur.
Gis-gis xarala
Algorithm yu bari dañuy tànnee subwords ci anam wu wuute: BPE dafay boole ñaari mbir yu bari ci suuf, WordPiece dafay tànn boole yi gëna yokk corpus, Unigram dafay tàmbali ak vocabulaire bu yaatu ba noppi dagg ay token yu gëna néew luñu mëna gaañ. WordPiece dafay màrki piyeesu biir baat yi ak prefix '##', waaye SentencePiece dafay jëfandikoo espace yi ni màndarga bu amul fenn suko defee mu liggéey ci bind bu ñor te du njëkka xaaj espace bu weex, lu baax la ci làkk yi amul espace.
njeextalu pexe
Gaawaay ak yaatuwaay
Liggéeyukaay yi ci làkk yi mën nañu gëna gaaw te duñu yàq deggoo gi.
Dugg ak yegg
Dafay yaatal jëfandikoo gi ci làkk yi ak ci anam yi ñuy jokkoo.
dogal yu gëna leer
Ekip yi mën nañu gëna yàgg ci àtte ci jamono ji otomatisation di liggéey ci baamtu.
Ëlëgu Tokenization Subword
Tokenization subword dina wéy di am doole ndax gaaw na te dëgër, waaye ay ñakk kattan, xaaj bu jaxasoo ci math, kode, ak script yu bari, boole ci njëgu token yu tolloowul ci làkk yi, ñu ngi yóbbe gëstu ci niveau byte ak xeetu token-free. Xaarandil tokenizers yu gëna am xel, amaana jàng wala méngoo ak yoon bu gëna baax ci làkk yu bari suko defee ñu baña daan mbind yi nekkul ci Àngle ak ay token yu bari ci frase bu nekk.
Doxal ci àdduna dëgg
BERT dafay jëfandikoo tokenization WordPiece, di màrke piyeesu wéyal yu melni '##ing' ngir tabaxaat kàddu yu njëkk ya.
T5 ak xeetu làkk yu bari dañuy jëfandikoo SentencePiece, mooy jëfandikoo làkk yu amul bayaal yu melni Japone ci saasi.
Modèle chat yi dañuy xaaj ab terme technique bu bariwul ay fragment yuñ xamee ci plaasu ñu lajj ci ab baat buñu xamul.
Tokenizers yi dañuy séddoo ay baat yu ndaw ci 'daw', 'dawal', ak 'dawkat', loolu mooy tax model bi mëna yamale morphologie ci anam wu jaar yoon.
Risk yi ak balustrade yi
Lépp lu jaarul yoon mën na dugg ci rapoor yi, jàppale ci liggéey bi, wala ci njariñu gëstu bi.
Sensibilite bu gaaw mën na jur njariñ yu wuute ci laajte yu noonu mel.
Done yu am solo mën nañu feeñ sudee seytu jëfandikoo gi néew doole.
Roadmap ngir samp gi
Mandargal formaa génne gi, melokaan bi, ak standard kalite yi laata ngay dugal ko.
Tontu yu am solo ak balluwaay yu wóor saa yu dëggu bi di am solo.
Fexeel am barabu xool nit ñi ngir am njariñ yu am solo.
Toppal anami gacce yi ak di faral di tàggataat ay laaj wala def-liggéey.
Weyal di banneexu
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Subword Tokenization quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Gis bi ci topp
Tokenisasioŋ
Laaj yi ñuy faral di laaj
What is Subword Tokenization?
Tokenization subword dafay xaaj mbind ci pàcc yu gëna ndaw baat yi waaye gëna mag ci araf yi, lu melni 'token' boole ci 'ization'. Mooy anam wi ñuy jëfandikoo ngir soppi làkk yi ci jamono jii, ñu soppi mbind yi ci ID yu ñuy jëfandikoo, ñuy méngale dayo vocabulaire bi ak li ñuy tekki.
Ban jafe-jafe la subword tokenization di saafara suñu ko méngale ak jëfandikoo baat yu mat?
Baatu-baatu lëmm yi dañu yaatu te ba leegi amul baatu-baatu yu bari; subwords yi deñuy yombal vocabulaire yi te dañuy xaaj baat yiñ xamul ci anam wu rafet.
Ci yii, yan ci ñoom mooy algorithm tokenisation bu BERT di jëfandikoo?
BERT dafay jëfandikoo WordPiece, mooy tànnee boole yi gëna yokk mën mënu corpus tàggat.
naka la WordPiece di màrkee piyees buy wéyal baat bi?
WordPiece dafay njëkka '##' ci biir baat bi, suko defee 'playing' nekk 'playing' boole ci '##ing'.
Lan moo tax SentencePiece méngoo bu baax ak làkk yu melni Japone?
SentencePiece dafay dox ci bind bu ñor, ba noppi enkode espace yi ni màndarga bu amul fenn, suko defee mu dox doonte amul espace ci diggante kàddu yi.
Lan la fragment subword bu nekk di mujjee ci karte balaa muy yegg ci model bi?
Baatu-baatu bu nekk dañu koy jox ID bu mat sëkk, bi couche bi koy samp daal di koy soppi def ko vecteur bi model bi mëna jëfandikoo.