Facebook (yakarambidzwa muRussian Federation) yakaburitsa zviri kuitika zveNLLB (Hapana Mutauro Wakasara Kuseri) chirongwa, chakanangana nekugadzira modhi yemuchina wepasi rose wekushandura zvakananga zvinyorwa kubva kune mumwe mutauro kuenda kune mumwe, nekupfuura shanduro yepakati muChirungu. Iyo yakatsanangurwa modhi inovhara mitauro inodarika mazana maviri, kusanganisira isingawanzo mitauro yevanhu vemuAfrica neAustralia. Chinangwa chikuru chepurojekiti ndechekupa nzira yekukurukurirana kune chero vanhu, zvisinei nemutauro wavanotaura.
Iyo modhi inopihwa rezenisi pasi perezinesi reCreative Commons BY-NC 4.0, iyo inobvumidza kukopa, kugovera patsva, kugadzirisa, uye mabasa anobva kunze, chero iwe uchipa kupihwa, kuchengetedza rezinesi, uye kurishandisa kune zvisiri zvekutengesa chete. Zvishandiso zvekushanda nemamodheru zvinopihwa pasi peMIT rezenisi. Kukurudzira budiriro uchishandisa modhi yeNLLB, zvakasarudzwa kugovera zviuru mazana maviri emadhora kuti vape rubatsiro kune vanoongorora.
Kurerutsa kusikwa kwemapurojekiti uchishandisa iyo yakarongwa modhi, kodhi yemashandisirwo anoshandiswa kuyedza uye kuongorora mhando yemamodhi (FLORES-200, NLLB-MD, Toxicity-200), kodhi yemhando yekudzidzira uye encoder yakavakirwa paLASER3 raibhurari. Mutauro-Agnostic mutsara) zvakare akavhurika sosi. Iyo yekupedzisira modhi inopiwa mune mbiri shanduro - yakazara uye yakapfupikiswa. Iyo yakapfupikiswa vhezheni inoda zviwanikwa zvishoma uye inokodzera kuongororwa uye kushandiswa mumapurojekiti ekutsvaga.
Kusiyana nemamwe masisitimu eshanduro anoenderana nemasisitimu ekudzidza muchina, mhinduro yeFacebook yakakosha pakuti inopa mhando imwechete yemitauro yese mazana maviri, inovhara mitauro yese uye isingade kushandiswa kwemhando dzakasiyana pamutauro wega wega. Kushandura kunoitwa kubva mumutauro wemabviro kuenda kumutauro wauri kududzira, pasina shanduro yepakati muChirungu. Kugadzira masisitimu eshanduro yepasi rose, modhi yeLID (Kuzivikanwa Kwemitauro) inokurudzirwa zvakare, izvo zvinoita kuti zvikwanise kuona mutauro unoshandiswa. Avo. iyo sisitimu inogona kungoziva kuti ndeupi mutauro ruzivo rwunopihwa uye rwuchiturikira mumutauro wemushandisi.
Kuturikira kunotsigirwa munzira ipi neipi, pakati pemitauro 200 inotsigirwa. Kusimbisa kunaka kweshanduro pakati pemitauro ipi neipi, FLORES-200 referensi test set yakagadzirwa, izvo zvakaratidza kuti iyo NLLB-200 modhi maererano nemhando yeshanduro iri paavhareji 44% yepamusoro kune yaimbotaurwa muchina-based tsvakurudzo masisitimu kana uchishandisa. Mametriki eBLEU achienzanisa shandurudzo yemuchina neshanduro yakajairwa nevanhu. Kune isingawanzo mitauro yeAfrica nemitauro yeIndia, hukuru hwemhando hunosvika 70%. Zvinokwanisika kuongorora nemaziso kunaka kweshanduro pane yakanyatsogadzirirwa demo site.
Source: opennet.ru