Ukwahlulahlula okubhaliweyo kusetyenziswa incokoGPT

Uhlalutyo lokubhaliweyo, okanye ukugrumba isicatshulwa, bubuchule obubalulekileyo bokufumana ulwazi oluxabisekileyo kwixabiso elikhulu ledatha yokubhaliweyo engacwangciswanga. 

Ibandakanya ukusetyenzwa kunye nokuhlalutya isicatshulwa ukufumana iipateni, iintsingiselo kunye nobudlelwane.

Ivumela iinkampani, abaphandi kunye nemibutho ukuba benze izigqibo ezisekelwe kulwazi olufunyenwe kwiitekisi. 

Njengoko umthamo wedatha engacwangciswanga uqhubeka ukhula ngokukhawuleza, imfuno yezixhobo ezichanekileyo nezisebenzayo zokuhlalutya itekisi iye yabaluleka kakhulu kuwo wonke amashishini ahlukeneyo njengentengiso, imali, ukhathalelo lwempilo, kunye nenzululwazi yezentlalo.

Ngokwesiko, uhlalutyo lombhalo lwenziwe kusetyenziswa iindlela ezisekelwe kumgaqo kunye nobuchule bokufunda ngomatshini njenge-SpaCY kunye ne-transformer technique. Ngelixa ezi ndlela zingqineke zisebenza, zifuna umgudu omkhulu kunye nobuchule bokugqibelela.

Ngokufika kwemifuziselo yolwimi olukhulu (LLM) njenge NcokolaGPT di I-OpenAI. Ibonise ubuchule obumangalisayo ekuveliseni isicatshulwa esifana nomntu kunye nomxholo wokuqonda, okwenza ukuba ibe sisixhobo esithembisayo kwimisebenzi yohlalutyo lwesicatshulwa njenge entity recognition, sentiment analysis, kwaye topic modeling.

Makhe sibone ngoku ukuba singakwenza njani ukwahlula okubhaliweyo sisebenzisa iChatGPT.

Indlela yesiNtu (imifuziselo enye) vs. LLM

Ngaphambili, besisoloko sisebenzisa iimodeli ezahlukeneyo kwimisebenzi eyahlukeneyo yokufunda koomatshini. Umzekelo, ukuba ndifuna ukukhupha ulwazi kwisicatshulwa, kuya kufuneka ndisebenzise imodeli yokuqaphela iqumrhu (NER - Named Entity Recognition), ukuba ndifuna ukwahlula isicatshulwa sam kwiiklasi ezahlukeneyo, ndiza kufuna imodeli yohlelo. Umsebenzi ngamnye owahlukileyo ubufuna ukuba iimodeli ziqeqeshwe ngokwahlukeneyo kumsebenzi ngamnye, nokuba kungokutshintshelwa ukufunda okanye ngoqeqesho.

Ngokusungulwa kwe Large Language Models (LLM), imodeli ye-LLM iya kukwazi ukwenza imisebenzi emininzi ye-NLP kunye okanye ngaphandle koqeqesho. Wonke umsebenzi unokuba defiigqitywe ngokulula ngokuguqula imiyalelo kwimiyalelo.

Ngoku makhe sibone indlela yokwenza umsebenzi we-NLP wesintu NcokolaGPT kwaye uyithelekise nendlela yemveli. Imisebenzi ye-NLP eya kwenziwa ngu NcokolaGPT kweli nqaku zezi:

• Ukutsalwa kolwazi (NER)
• Ukuhlelwa kwesicatshulwa
• Sentiment analysis
• Isishwankathelo

Ukutsalwa kolwazi (NER)

I-Entity Entity Recognition (NER) ibhekisela kumsebenzi wokuchonga ngokuzenzekelayo amagama kwiibhloko ezahlukeneyo zedatha yombhalo. Isetyenziswa ikakhulu ukukhupha iindidi zamaqumrhu abalulekileyo afana namagama amachiza kumanqaku ekliniki, imigaqo enxulumene nengozi kumabango e-inshurensi, kunye neminye imimiselo ethe ngqo kwisizinda kwiirekhodi.

Qaphela ukuba lo msebenzi ukhethekileyo kwi-domain yezonyango. Yayisoloko ifuna ukuba sichaze kwaye siqeqeshe ngaphezu kwe-10.000 imiqolo yedatha yemodeli enye ukwazi iklasi ethile kunye nekota kwisicatshulwa. I-ChatGPT inokuchonga ngokuchanekileyo igama ngaphandle kwesicatshulwa esiqeqeshwe kwangaphambili okanye ukulungiswa kakuhle, okusisiphumo esihle kakhulu!

Ukuhlelwa kwesicatshulwa

Ukuhlelwa kombhalo kubhekiselele kwinkqubo ezenzekelayo yokufumana kunye nokuhlelwa kwesicatshulwa kwiindidi ezivela kwidatha enkulu, idlala indima ebalulekileyo ekufumaneni idatha yombhalo kunye nokutsalwa. Imizekelo yezicelo zokuhlelwa kombhalo ziquka izilumkiso zeklinikhi okanye ukuhlelwa kwezinto zomngcipheko, ukuhlelwa kokuxilongwa ngokuzenzekelayo, kunye nokufunyanwa kwe-spam.

Sentiment analysis

Sentiment analysis kubandakanya ukugqiba imvakalelo okanye uvakalelo oluvakaliswa kwisiqwengana sesicatshulwa. Ijolise ekuhleleni okubhaliweyo ngokweendidi zangaphambilidefiNite, ezifana ne-positive, negative okanye ingathathi hlangothi, ngokusekelwe kwimvakalelo esisiseko edluliswa ngumbhali. 

Izicelo zohlalutyo lweemvakalelo ziquka:

• uhlalutyo lokuphononongwa kwabathengi kunye nengxelo,
• ukulandelela iimvakalelo kwimidiya yoluntu,
• esweni iintsingiselo zemarike e
• umlinganiselo weemvakalelo zezopolitiko ngexesha lamaphulo onyulo.

Isishwankathelo

Izishwankathelo ezizenzekelayo zibhekisela kwinkqubo apho izihloko eziphambili zephepha elinye okanye ngaphezulu zichongiwe kwaye zinikezelwe ngendlela emfutshane nechanekileyo. Oku kuvumela umsebenzisi ukuba ajonge kwiinqununu ezinkulu zedatha ngexesha elifutshane. Imizekelo yezicelo ibandakanya inkqubo yesishwankathelo evumela ukuveliswa ngokuzenzekelayo kwezinto ezingabonakaliyo kumanqaku eendaba kunye nokushwankathelwa kolwazi ngokukhupha izivakalisi kwi-abstracts yamaphepha ophando.

I-ChatGPT sisixhobo esisisishwankathelo esigqwesileyo, ngakumbi kumanqaku amade kunye nophononongo olunzima. Ngokuncamathisela uphononongo kwi-ChatGPT, sinokusazi ngokulula isishwankathelo sophononongo lwemveliso ngokujonga nje.

Umda weeLLMs

Kuba injongo yeli nqaku kukuphonononga amandla e-LLM ukwenza imisebenzi yocazululo lwesicatshulwa, kubalulekile ukuqaphela imida yazo. Eminye yemida ephambili yeeLLMs ibandakanya:

1. Ukusetyenziswa kobutyebi : Ukusebenzisa ii-LLM kufuna izibonelelo ezibalulekileyo zokubala kunye nezemali, ezinokuba ngumngeni kwimibutho emincinci okanye kubaphandi ngabanye abanezibonelelo ezinqongopheleyo. Ukusukela namhlanje, i-ChatGPT yamkela kuphela malunga namathokheni angama-8.000 wokufaka kunye nemveliso, ukuhlaziya inani elikhulu ledatha, kufuna ukuba umsebenzisi aphule umbhalo kwiinqununu ezininzi zedatha, kwaye unokufuna iifowuni ezininzi ze-API kwimisebenzi.
2. Uvakalelo olukhawulezileyo lwebinzana : Ukusebenza kwee-LLM kunokuchatshazelwa yindlela i-prompts ebhalwa ngayo. Utshintsho oluncinci kumagama akhawulezayo lunokuvelisa iziphumo ezahlukeneyo, ezinokuba yimbangela yenkxalabo xa ukhangela imveliso ehambelanayo nethembekileyo.
3. Ukunqongophala kobuchule obuthile besizinda : Ngelixa ii-LLM zinokuqonda ngokubanzi kwemimandla eyahlukeneyo, ayinakuba nenqanaba elifanayo lobuchwephesha njengeemodeli ezikhethekileyo eziqeqeshwe kwidatha ye-domain-specific. Ngenxa yoko, ukusebenza kwabo kusenokungabi yegqwesileyo kwezinye iimeko kwaye kunokufuna ukulungiswa kakuhle okanye ulwazi lwangaphandle, ngakumbi xa besebenza ngolwazi olukhethekileyo okanye lobugcisa.

Ercole Palmeri