Ilabhorethri yobuhlakani bokwenziwa ephethwe yi-Google imemezele ngesikhathi sendawo ngomhlaka-12 ukuthi yethule imodeli enkulu yezilimi eziningi yokuguqula ulimi lwezandla lube umbhalo (SL2T).
Ngenxa yale modeli, ubuhlakani bokwenziwa bolimi lwezandla bungena okokuqala emikhiqizweni yabathengi: izinhlelo ze-Gboard ne-Live Transcribe kumafoni ahlakaniphile ochungechunge lwe-Google Pixel 11 zisekela ukufakwa kombhalo ngolimi lwezandla (ekuqaleni kusekelwa uLimi Lwezandla LwaseMelika kuphela, kanti kamuva luzonwetshwa kwezinye izilimi zezandla nakwamanye amadivayisi).
Abantu abangezwa kahle abangaba yizigidi ezingu-70 emhlabeni wonke basebenzisa izilimi zezandla ezingaphezu kuka-200. Lo msebenzi omusha ubenza bajabulele ulwazi oluguquguqukayo “lokuthayipha esikhundleni sokukhuluma”, olufana nokufakwa kombhalo ngezwi. Ukuhlolwa kubonise ukuthi ukusebenzisa uLimi Lwezandla LwaseMelika kushesha, kungokwemvelo futhi kujabulisa kakhulu kunokuthayipha ngesiNgisi.
I-DeepMind ithe ulimi lwezandla luwulimi lwemvelo oluzimele, olunohlelo lolimi nesilulumagama salo esihlukile. Ngakho-ke, ukuhumusha ulimi lwezandla kudinga ukuhumusha komshini kwangempela, hhayi ukuguqula ulimi lwezandla lube umbhalo ngokulandelana. Imodeli ekwazi ukwenza umsebenzi we-SL2T kumele ikwazi “ukuqonda” ukunyakaza okuhambisana ngesikhathi kwezingxenye eziningi zomzimba zomsebenzisi wolimi lwezandla, okudinga amandla amakhulu okucubungula okubona ngekhompyutha.
Imodeli ye-SL2T ye-Google DeepMind yaqeqeshwa ngedatha yamahora angu-100 000 yezilimi zezandla ezingaphezu kuka-50. Iguqula ulimi lwezandla lube ulwazi lwezindawo zamaphuzu ezimpawu zokuma ukuze ivikele ubumfihlo nokuphepha kwabasebenzisi, ngaphandle kokuthumela ukusakazwa kwevidiyo kwasekuqaleni kuseva. Ngesikhathi esifanayo, iyayeka ukusebenzisa “izichasiselo eziphakathi” ebezisetshenziswa kabanzi ngaphambilini, futhi ihumushe ngokuqondile ulwazi lwendawo.