IBM yana buɗe CodeNet don tsarin koyo na inji wanda ke fassara da tabbatar da lamba

Kamfanin IBM ya kaddamar da shirinsa na CodeNet, wanda ke da nufin samarwa masu bincike bayanan da ke ba su damar yin gwaji tare da amfani da dabarun koyon na'ura don ƙirƙirar masu fassarar harshe na shirye-shirye, masu samar da code da masu tantancewa. CodeNet ya ƙunshi tarin misalan lambobi miliyan 14 waɗanda ke magance matsalolin shirye-shirye na gama gari 4053. Gabaɗaya, tarin ya ƙunshi kusan layukan lamba miliyan 500 kuma ya ƙunshi yarukan shirye-shirye 55, duka harsunan zamani kamar C++, Java, Python da Go, da kuma yarukan gado waɗanda suka haɗa da COBOL, Pascal da FORTRAN. Ana rarraba ci gaban aikin a ƙarƙashin lasisin Apache 2.0, kuma ana shirin rarraba bayanan bayanan a cikin hanyar jama'a.

Misalan an ƙirƙira su kuma aiwatar da algorithms iri ɗaya a cikin harsunan shirye-shirye daban-daban. Ana sa ran tsarin da aka tsara zai taimaka wajen horar da tsarin koyon injin tare da haɓaka sabbin abubuwa a fagen fassarar da tantance lambobin na'ura, kwatankwacin yadda bayanan bayanan hoto na ImageNet ya taimaka wajen haɓaka ƙirar ƙira da tsarin hangen nesa na kwamfuta. An bayyana gasar shirye-shirye daban-daban a matsayin daya daga cikin manyan hanyoyin samar da tarin.

Ba kamar masu fassara na gargajiya ba, waɗanda ake aiwatar da su bisa ka'idojin fassara, tsarin koyon injin na iya ɗauka tare da la'akari da yanayin amfani da lambar. Lokacin juyawa daga wannan yaren shirye-shirye zuwa wani, mahallin yana da mahimmanci kamar yadda lokacin fassara daga wannan harshe na ɗan adam zuwa wani. Wannan rashin la'akarin mahallin shine abin da ke hana canza lambar daga harsunan gado kamar COBOL.

Samun babban bayanan aiwatar da algorithm a cikin yaruka daban-daban zai taimaka ƙirƙirar tsarin koyo na injin na duniya wanda, maimakon fassarar kai tsaye tsakanin takamaiman harsuna, sarrafa mafi ƙarancin wakilcin lambar, mai zaman kanta daga takamaiman yarukan shirye-shirye. Irin wannan tsarin za a iya amfani da shi azaman mai fassara, fassara lambar da aka watsa a cikin kowane harsunan da aka goyan baya zuwa cikin ƙayyadaddun wakilci na ciki, wanda daga nan za a iya samar da lambar a cikin harsuna da yawa.

Hakanan tsarin zai sami damar yin sauye-sauye na biyu. Misali, bankuna da hukumomin gwamnati suna ci gaba da yin amfani da ayyuka a cikin yaren COBOL da suka tsufa. Mai fassara na tushen ilmantarwa na inji zai iya canza lambar COBOL zuwa wakilcin Java, kuma, idan ya cancanta, juya guntun Java zuwa lambar COBOL.

Baya ga fassarar tsakanin harsuna, an ambaci irin wuraren aikace-aikacen CodeNet kamar ƙirƙirar tsarin bincike na lambar waya da sarrafa kansa na gano clone, gami da haɓaka haɓakawa da tsarin gyaran lambar atomatik. Musamman, misalan da aka gabatar a CodeNet suna sanye take da metadata da ke kwatanta sakamakon gwajin aiki, sakamakon girman shirin, yawan amfani da ƙwaƙwalwar ajiya da jihar, wanda ke ba mu damar bambance lambar daidai daga lambar tare da kurakurai (don bambanta lambar daidai daga lambar da ba ta dace ba, tarin musamman ya haɗa da misalai tare da kurakurai, wanda rabon su shine 29.5%). Tsarin koyan na'ura na iya ɗaukar wannan metadata don samar da mafi kyawun lambar ko don gano koma baya a cikin lambar da aka bincika (tsarin zai iya fahimtar cewa algorithm a cikin lambar da aka ƙaddamar ba a aiwatar da shi da kyau ko ya ƙunshi kurakurai).

source: budenet.ru

Add a comment