Ukuhlaziya umbhalo kusetshenziswa i-chatGPT

Izibalo zombhalo, noma ukumba umbhalo, kuyindlela ebalulekile yokuthola imininingwane ebalulekile enanini elikhulu ledatha yombhalo engahlelekile. 

Kubandakanya ukucubungula nokuhlaziya umbhalo ukuze uthole amaphethini, amathrendi kanye nobudlelwano.

Ivumela izinkampani, abacwaningi nezinhlangano ukuthi zenze izinqumo ezisekelwe olwazini olutholwe emibhalweni. 

Njengoba umthamo wedatha engahlelekile uqhubeka ukhula ngokuphawulekayo, isidingo samathuluzi okuhlaziya umbhalo anembile nasebenza kahle siye saba bucayi kakhulu ezimbonini ezihlukahlukene, njengokumaketha, ezezimali, ukunakekelwa kwezempilo, kanye nesayensi yezenhlalo.

Ngokwesiko, ukuhlaziywa kombhalo kuye kwenziwa kusetshenziswa izindlela ezisekelwe emithethweni nezindlela zokufunda zomshini ezifana ne-SpaCY kanye nenqubo ye-transformer. Nakuba lezi zindlela seziphumelele, zidinga umzamo omkhulu kanye nobuchwepheshe ukuze ziphelele.

Ngokufika kwezinhlobo zezilimi ezinkulu (LLM) ezifana I-ChatGPT di I-OpenAI. Ibonise amakhono amangalisayo ekukhiqizeni umbhalo ofana nomuntu kanye nesimo sokuqonda, okuwenza ube yithuluzi elithembisayo lemisebenzi yokuhlaziya umbhalo njenge entity recognition, sentiment analysis, futhi topic modeling.

Ake sibone manje ukuthi singakwenza kanjani ukuhlukanisa umbhalo sisebenzisa i-ChatGPT.

Indlela yendabuko (amamodeli awodwa) vs. I-LLM

Esikhathini esidlule, besihlala sisebenzisa amamodeli ahlukene emisebenzini ehlukene ekufundeni komshini. Isibonelo, uma ngifuna ukukhipha ulwazi embhalweni, ngizodinga ukusebenzisa imodeli yokuqashelwa kwebhizinisi (NER - Named Entity Recognition), uma ngidinga ukuhlukanisa umbhalo wami ube amakilasi ahlukene, ngizodinga imodeli yokuhlukanisa. Umsebenzi ngamunye ohlukene ubudinga ukuthi amamodeli aqeqeshwe ngendlela ehlukene ngomsebenzi ngamunye, ngokudlulisela ukufunda noma ngokuqeqeshwa.

Ngokwethulwa kwe Large Language Models (LLM), imodeli ye-LLM izokwazi ukwenza imisebenzi eminingi ye-NLP ngokuqeqeshwa noma ngaphandle kokuqeqeshwa. Noma yimuphi umsebenzi kungaba defikuqedwe kalula ngokushintsha imiyalo kumiyalo.

Manje ake sibone ukuthi ungawenza kanjani umsebenzi we-NLP wendabuko I-ChatGPT futhi uyiqhathanise nendlela yesintu. Imisebenzi ye-NLP ezokwenziwa ngu I-ChatGPT kulesi sihloko kukhona:

• Ukukhishwa kolwazi (NER)
• Ukuhlukaniswa kombhalo
• Sentiment analysis
• Isifinyezo

Ukukhishwa kolwazi (NER)

I-Entity Entity Recognition (NER) ibhekisela emsebenzini wokuhlonza ngokuzenzakalelayo amabhulokhi edatha yombhalo. Isetshenziselwa kakhulu ukukhipha izigaba zebhizinisi ezibalulekile njengamagama ezidakamizwa kumanothi omtholampilo, imigomo ehlobene nengozi evela ezimangalweni zomshwalense, neminye imigomo eqondene nesizinda kumarekhodi.

Qaphela ukuthi lo msebenzi uqondene ngqo nesizinda sezokwelapha. Bekuvame ukudinga ukuthi sichasise futhi siqeqeshe imigqa yedatha engaphezu kuka-10.000 kumodeli eyodwa ukuze yazi isigaba esithile kanye nethemu embhalweni. I-ChatGPT ingakwazi ukuhlonza kahle igama ngaphandle kwanoma yimuphi umbhalo oqeqeshwe kusengaphambili noma ukucushwa kahle, okuwumphumela omuhle uma kuqhathaniswa!

Ukuhlukaniswa kombhalo

Ukuhlukaniswa kombhalo kubhekisela enqubweni ezenzakalelayo yokuthola nokuhlukanisa umbhalo ngezigaba kusuka kudatha enkulu, idlala indima ebalulekile ekubuyiseni nasekukhishweni kwedatha yombhalo. Izibonelo zezinhlelo zokusebenza zokuhlukanisa umbhalo zifaka izexwayiso zomtholampilo noma ukuhlukaniswa kwezinto eziyingozi, ukuhlukaniswa ngezigaba okuzenzakalelayo kokuxilonga, nokutholwa kogaxekile.

Sentiment analysis

Sentiment analysis kubandakanya ukunquma imizwa noma imizwa evezwa esiqeshini sombhalo. Ihlose ukuhlukanisa umbhalo ngokwezigaba zangaphambilidefinite, njengephozithivu, engemihle, noma engathathi hlangothi, ngokusekelwe emuzweni owumsuka ovezwe umlobi. 

Izicelo zokuhlaziya imizwa zifaka:

• ukuhlaziya ukubuyekezwa kwamakhasimende kanye nempendulo,
• ukulandelela imizwa yenkundla yezokuxhumana,
• ukuqapha izitayela zezimakethe e
• ukukalwa kwemizwelo yezepolitiki phakathi nemikhankaso yokhetho.

Isifinyezo

Izifinyezo ezizenzakalelayo zibhekisela enqubweni lapho izihloko eziyinhloko zombhalo owodwa noma ngaphezulu zibonwa futhi zethulwe ngendlela emfushane nenembile. Lokhu kuvumela umsebenzisi ukuthi abheke izingcezu ezinkulu zedatha ngesikhathi esifushane. Isibonelo sezinhlelo zokusebenza zifaka isistimu yesifinyezo evumela ukukhiqizwa okuzenzakalelayo kwezifinyezo ezivela kuma-athikili ezindaba kanye nokufingqwa kolwazi ngokukhipha imisho ezifushaneni zephepha locwaningo.

I-ChatGPT iyithuluzi elihle kakhulu lokufingqa, ikakhulukazi lezindatshana ezinde nezibuyekezo eziyinkimbinkimbi. Ngokunamathisela ukubuyekezwa ku-ChatGPT, singakwazi kalula ukwazi isifinyezo sokubuyekezwa komkhiqizo shazi.

Umkhawulo wama-LLM

Njengoba inhloso yalesi sihloko iwukuhlola ikhono lama-LLM okwenza imisebenzi yokuhlaziya umbhalo, kubalulekile ukuqaphela ukulinganiselwa kwawo. Eminye yemikhawulo eyinhloko yama-LLM ihlanganisa:

1. Ukusetshenziswa kwezinsiza : Ukusebenzisa ama-LLM kudinga izinsiza ezibalulekile zokubala kanye nezezimali, okungaba inselele ezinhlanganweni ezincane noma abacwaningi ngabanye abanezinsiza ezilinganiselwe. Kusukela namuhla, i-ChatGPT yamukela kuphela amathokheni acishe abe ngu-8.000 okokufaka nokukhiphayo, ukuhlaziya inani elikhulu ledatha, kudinga ukuthi umsebenzisi ahlukanise umbhalo ube izingcezu eziningi zedatha, futhi ingadinga amakholi amaningi e-API ukuze enze imisebenzi.
2. Ukuzwela kokwazisa umusho : Ukusebenza kwama-LLM kungathinteka ngendlela okushiwo ngayo amagama. Ushintsho oluncane ekusetshenzisweni kwamagama ngokushesha lungaveza imiphumela ehlukene, okungaba yimbangela yokukhathazeka lapho ufuna okukhiphayo okungaguquki nokuthembekile.
3. Ukushoda kobuchwepheshe obuthile besizinda : Nakuba ama-LLM enokuqonda okujwayelekile kwezizinda ezihlukahlukene, angase angabi nalo izinga lobuchwepheshe elifanayo njengamamodeli akhethekile aqeqeshwe kudatha eqondene nesizinda. Ngenxa yalokho, ukusebenza kwabo kungase kungabi kuhle kwezinye izimo futhi kungase kudinge ukucushwa kahle noma ulwazi lwangaphandle, ikakhulukazi lapho usebenza ngolwazi olukhethekile kakhulu noma lobuchwepheshe.

Ercole Palmeri