Riba:
- Ƙarfin bambance ƙarfi da raunin NMT da fassarar tushen LLM da zaɓar injin da ya dace daidai da nau'in kasuwanci.
- Ikon ƙaddamar da cancantar rubutu don na'ura da kuma tantance shi don ainihin lokaci da farashi
- Ƙarfin don auna ingancin fitarwar injin da sauri cikin girma biyar da haɗarin tuta ba tare da rikitar da santsi tare da daidaito ba
Mafi ƙarfin hanzarin da kuke da shi a matsayin mai fassara shine fassarar inji (MT) - amma amfani da kayan aiki da hankali yana nufin amfani da shi zuwa aikin da ya dace, a cikin yaren da ya dace, kuma tare da kyakkyawan fata. A cikin wannan rukunin, zaku san manyan iyalai guda biyu na MT (NMT da fassarar LLM), koya wanne ne ya fi dacewa da wane aiki, yadda ake saurin auna ɗanyen ingancin fassarar kafin bayarwa, da yadda ake zaɓar injin daidai da aikin. Manufar ita ce kawar da ra'ayin "duk iri ɗaya ne, manna-fassara" kuma ya zama ƙwararren wanda zai iya hango ko wane rubutu ya dace da na'ura kuma wanda ke buƙatar aikin ɗan adam.
Iyalai biyu: NMT da fassarar tushen LLM
NMT (Fassarar Injin Neural) tsarin ne waɗanda suka koya daga miliyoyin jimlolin harsuna biyu kuma suna fassara jumla gaba ɗaya; Google Translate, DeepL, Microsoft Translator su ne misalan waɗannan. Ƙarfi: mai saurin gaske, daidaito, ƙware a cikin gajerun jimloli. Rauni: sau da yawa baya ganin mahallin bayan iyakar jimla (ma'ana daga dukkan rubutun); Yana iya zama da wuya a yanke shawarar ko kalmar “banki” tana nufin banki ko bakin kogi ta hanyar duba sakin layi kuma bai dace da jerin sharuddan da aka bayar ba.
Fassarar tushen LLM (Babban Samfurin Harshe) shine amfani da samfura da aka kora kamar su ChatGPT, Claude, Gemini don fassara. Ƙarfi: yana ɗaukar umarni - yana fahimtar umarni kamar "amfani da sauti na yau da kullun," "bi wannan jerin sharuddan," "masu sauraro na yara yara"; yana ganin mafi girman mahallin; yana ɗaukar karin magana da sautin mafi kyau. Rauni: wani lokacin yana iya zama “mafi ƙirƙira” kuma yana ƙara zuwa tushen (haɗarin hasashe), rasa daidaituwa cikin dogon rubutu, kuma farashi/gudu ya bambanta dangane da aikin.
Dokokin babban yatsan hannu: a madaidaiciya, maimaitawa, rubutun fasaha NMT yawanci yana da sauri kuma yana isa; LLM ya fi sassauƙa tare da rubutun da ke buƙatar horo a cikin sautin, mahallin, ƙamus, da koyarwa. Yawancin ƙwararru a yau suna amfani da duka biyu tare: mai sauri daftarin aiki daga NMT, sautin / gyaran lokaci daga LLM.
Tukwici: Na'urar ingancin harshe biyu (misali Baturke → Turanci) na iya bambanta da kishiyar alkibla (Turanci →Turanci). Harsuna tare da agglutinative, m syntax, kamar Baturke, gabaɗaya sun fi ƙalubale ga MT. Yi wa kanku hukunci wane injin ya fi kyau a cikin naku biyu tare da ƴan rubutun samfuri.
Dacewar na'ura na rubutu: tambayi wannan da farko
Ba kowane rubutu ya dace daidai da na'ura ba. Kafin fara fassarar, wuce rubutun ta hanyar tace "dace":
- Ya dace da na'ura: jagorar fasaha, bayanin samfur, rubutu mai maimaitawa, daidaitaccen wasiku, tebur/jeri. Harshen a bayyane yake, ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun kalmomi ba su da yawa.
- Matsakaicin dacewa: labarai, abun ciki na kamfani, kayan ilimi. Injin yana samar da fayyace masu kyau amma yana buƙatar babban gyara bayan sauti da kwarara.
- Bai dace da na'ura ba (haɗari mai girma): kwangilar doka, rubutun likitanci, talla / taken, rubutun wallafe-wallafe, ban dariya, shayari. A nan injin yana ba da ra'ayoyi kawai; Aiki ya fi fadi ga mutane.
Idan kun yi wannan bambance-bambance daga farkon, zaku ba abokin ciniki lokaci/farashin da ya dace kuma ku kare kanku daga makauniyar amincewa da albarkatun ƙasa.
Da sauri auna ingancin fitarwa mai ɗanɗano
Lokacin da kuka ga fitowar MT kuna mamakin "yana da kyau ko mara kyau?" Amsa tambayar tare da jerin abubuwan dubawa da sauri, ba hankali ba. Dubi waɗannan ma'auni guda biyar: daidaito (ma'anar aminci ga tushen?), cikawa (shin an cire jumlar ko ƙarawa?), Kalmomi (shin daidai ne kuma daidai ne?), ƙwarewa (shin halitta ne a cikin harshen da ake nufi?), Tsarin (sunan tags, lambobi, tsarawa?). Za mu zurfafa waɗannan girma a cikin naúrar inganci na gaba; A yanzu, bari ya zama tunaninka kafin isarwa.
Tsanaki: Kurakurai mafi haɗari na fitowar MT sune "masu kyau amma ba daidai ba". Jumlar na iya kasancewa cikin cikakken Turkanci, amma "rangwamen 15%" a cikin tushen ƙila an canza shi zuwa "rangwamen 50%". Kada ku ji daɗin magana; Koyaushe duba lamba, korau, da madaidaicin suna daban.
uku mini lokuta
Case 1 - Injin da ya dace ya yanke lokacin cikin rabi. Wani mai fassara ya zaɓi ya fassara ƙirar software mai kalmomi 12,000 tare da NMT da ɗan littafin tallace-tallace mai girma iri ɗaya tare da LLM. Matsakaicin NMT a wurin sadarwa ya sa aikin ya yi sauri; Sassauci tonal na LLM a cikin ɗan littafin ya rage nauyin sake rubutawa da kashi 40%. Ba da ayyukan biyu ga injin guda zai bata lokaci.
Case 2 - Pre-kimantawa ya ceci farashin. Wani ofishi yana shirin bayar da rubutun doka saboda "ana iya yin shi da injina, zai yi arha". A cikin kimantawa na farko, an fassara samfurin kalmomi 500; Na'urar ta dawo da sharuddan doka guda biyu tare da tsarin da ba daidai ba daidai yake. Ofishin ya sanya farashin aikin a matsayin "nauyi bayan gyara" kuma ya ba da tabbataccen lokacin ƙarshe; Ya guje wa asarar kuɗi saboda tayin da ba daidai ba.
Case 3 — Bambancin nau'in harshe. Tawagar ta yi tunanin cewa injin da ya yi aiki sosai a Turanci →Jamus yana da inganci iri ɗaya a cikin Baturke →Larabci. Gwajin kalmomi 300 ya nuna cewa fitarwar Larabci yana buƙatar ƙarin gyara. Tawagar ta ware injuna daban-daban da kuma kasafin gyara daban-daban dangane da nau'in harshe.
Samfura huɗu masu kwafi
1) Ƙimar dacewa da rubutu:
Matsayinku: babban ƙwararren MTPE. Ƙimar rubutu mai zuwa don dacewa da fassarar inji: na zahiri/na fasaha ko ƙirƙira/na yau da kullun? Rarraba shi a matsayin (1) mai son na'ura sosai, (2) matsakaici, (3) babban haɗari kuma rubuta hujjar ku. Yi ƙididdige nauyin da ake tsammani bayan gyara kamar haske/matsakaici/ nauyi. Misalin rubutu: [manna kalmomi 200-300]
2) Fassarar LLM da aka ba da umarni (tare da kalmomi da sarrafa sauti):
Fassara wannan rubutu [source] →[manufa].Masu sauraro: [wanda]. Sautin: [misali. hukuma]. Filin: [misali. Finance].Jerin sharuddan (tabbatar amfani): [A→a, B→b].Dokoki: kar a ƙara abin da ba a cikin tushen ba, kiyaye lambobi/kwanaki/suna, maye gurbin kowace jumla da jumla. Yi alama a inda ba ku da tabbas da [?]. Rubutu: [...]
3) Kwatankwacin injin guda biyu:
A ƙasa akwai fassarori guda biyu daban-daban na jimlar tushe ɗaya (A da B). Kwatanta kowane ta fuskar daidaito, kalmomi da dabi'a, kuma ku gaya mani wanda zai buƙaci ƙarancin gyara, tare da hujja. Source: [...] | Fassarar A: [...] | Fassarar B: [...]
4) Binciken saurin fitar da albarkatun kasa:
A ƙasa akwai fassarar tushe da inji. Kawai yi alama kamar haka: (1) bayanan da aka cire/karin, (2) lamba/kwanaki/suna ba daidai ba, (3) kuskuren rashin amincewa, (4) rashin daidaituwa. Kada ku rubuta wani gyara, kawai jera wuraren haɗari. Source: [...] | Fassara: [...]
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rauni: "Fassara wannan rubutun kuma zai yi kyau." (Don injin, "mai kyau" ba a bayyana shi ba; babu sauti, babu lokaci, babu taro.)
Güçlü: "Fassara wannan bayanin samfurin daga Ingilishi zuwa Baturke. Yanki: kasuwancin e-commerce. Masu sauraro: matasa mabukaci
Bambance-bambance: hanzari mai ƙarfi yana ba injin abin da ake iya aunawa; fitowar kai tsaye tana kusantar daftarin da aka gyara.
Taswirar kwatanta NMT vs LLM
Girman
NMT (Google, DeepL)
LLM (ChatGPT, Claude)
gudun
da sauri sosai
matsakaici
Karɓi umarni/sauti
mai rauni
mai karfi
Yi biyayya da jerin wa'adin
iyakance
Da kyau (tare da sauri)
Faɗin mahallin
mai rauni
mai kyau
Hadarin hallucinations
ƙananan
Matsakaici (ana buƙatar sarrafawa)
mafi dacewa aiki
Madaidaici / fasaha / maimaituwa
Sautin / yanayi / halitta
Kuskuren gama gari
- Yin amfani da injin iri ɗaya don kowane aiki. Rashin zaɓar injin bisa ga nau'in aikin yana haifar da asarar lokaci da inganci.
- Bayar da farashi/lokaci ba tare da tantancewa ba. Gwajin kalma 200-300 yana bayyana abubuwan mamaki daga farkon.
- Kuskure iyawa don daidaito. Kyakyawar jimla baya nufin jimla daidai.
- Yin watsi da nau'in harshe da bambancin shugabanci. Kyakkyawan injin a cikin ɗayan biyu yana iya zama mai rauni a cikin wani.
- Ba a lura da ƙarin LLM ba. LLM wani lokaci yana ƙara jimloli ba a cikin tushen "don taimako"; duba wannan.
Tagar yanayi da daidaito
Lokacin fassara dogon rubutu tare da LLM, wajibi ne a san iyakar fasaha guda ɗaya: taga mahallin - adadin rubutun da samfurin zai iya "gani" kuma yana riƙe da hankali a lokaci guda. Idan rubutun ya fi wannan taga girma, dole ne a raba shi cikin chunks, kuma samfurin na iya "manta" a cikin ɓangarorin lokaci na gaba ko yanke shawara da kuka yi a cikin ɓangarorin da suka gabata. Don haka, maimakon fassarar dogon littafi ko daftari a cikin yanki ɗaya, tunatar da kowane yanki na ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun kalmomi da salon su yana kiyaye daidaito. Wannan matsalar ba ta faruwa a cikin gajerun rubutu; Koyaya, a cikin dogayen ayyuka kamar litattafai da litattafai, ya zama dole a bugu da žari duba daidaito a cikin sassa. Magani mai amfani: don shirya "ƙwaƙwalwar aikin" (sharuɗɗan + haruffa + yanke shawara) kuma ƙara shi zuwa farkon kowane yanki kuma bincika daidaito tsakanin sassan a ƙarshen fassarar. A cikin NMT, ana fuskantar wannan matsalar ta wata hanya dabam: Tun da NMT ke fassara jimla da jimla, daidaiton sakin layi ya riga ya yi rauni, don haka termbase yana ɗaukar kalmar horo.
a takaice
Akwai iyalai guda biyu na fassarar injin: NMT, mai sauri da daidaito, da LLM, wanda ke ɗaukar koyarwa kuma yana ganin mahallin da sauti mafi kyau. Babban mai fassara yana ba da ƙimar dacewar rubutu don na'ura a gaba, yana zaɓar injin daidai da aikin, da sauri ya auna ingancin kayan da aka fitar kafin bayarwa, kuma baya rikitar da kaifin basira da daidaito. Wannan ra'ayin kima kafin kima yana ba ku damar ba da lokaci/farashi na gaske kuma ku kare kanku daga amana makaho.
Aikin aikace-aikace
Fassara rubutu guda 200 iri ɗaya tare da kayan aikin NMT da LLM. Kwatanta abubuwan da aka fitar guda biyu akan ma'auni biyar (sahihanci, cikawa, kalmomi, iyawa, tsari) kuma rubuta dalilan da suke buƙatar ƙarancin gyara bayan wannan rubutu. Sannan batun tuta/matsala/haɗari akan duka biyun tare da samfurin “tabbataccen fitarwa mai sauri”.
jerin abubuwan dubawa
- [ ] Na rarraba dacewa da rubutun don na'ura (ƙananan / matsakaici / babban haɗari).
- [ ] Dangane da nau'in aikin, na yanke shawarar ko zan yi amfani da NMT ko LLM.
- [ ] Na yi ƙima na farko tare da ƙaramin samfurin kuma na ƙayyade tsawon lokaci / farashin daidai.
- [ ] Na yi sauri na duba kayan da aka fitar a cikin girma biyar.
- [ ] Na duba lamba, kwanan wata, suna da abubuwan da ba su dace ba daban ba tare da iyawa na yaudare ni ba.