Naúrar 4 / 11

Gudanar da Kalmomin Kalmomi da Rubuce-rubuce: Daidaituwa da Harshen Kamfanin

Riba:

  • Ikon tabbatar da daidaiton kalmomi ta hanyar kafa ma'auni da bayyana ma'anar da aka amince da su da kuma haramtattun makamantansu.
  • Yin amfani da hankali na wucin gadi don fitar da ƴan takara da sauri da yanke shawara na ƙarshe tare da gwani da abokin ciniki
  • Ikon haɗa ma'auni zuwa kayan aiki/sauri yayin fassara da gano rashin jituwa tare da yawon shakatawa na QA na bayan fassarori.

A cikin fassarar fasaha, idan kun fassara kalmar "Asusun mai amfani" a matsayin "asusun mai amfani" a wuri ɗaya, "account member" a wani wuri, "profile mai amfani" a wani wuri, za a rasa mai amfani, alamar za ta bayyana rashin daidaituwa, kuma bincike / takardun za a karya, ko da rubutun "daidai ne". Shiru amma mafi mahimmancin ingancin fassarar shine daidaiton kalmomi. A cikin wannan rukunin, zaku koyi sarrafa kalmomi, ƙirƙira ma'auni (ma'ajin bayanai na lokaci), cirewar lokaci tare da AI, da kuma yadda ake kiyaye horon ƙamus a cikin fassarar. Manufar shine ya zama ƙwararren wanda ya dace da ra'ayi iri ɗaya a kowane aiki a cikin hanya ɗaya a ko'ina kuma yana da aminci ga harshen kamfani.

Ka'idoji na asali

Kalmomi sune kalmomi da maganganu waɗanda suka dace da takamaiman ra'ayi na filin ko cibiyoyi ("Checkout", "manufofin dawowa", "sem cell"). Termbase (TB — bayanan bayanan lokaci) ƙamus ne da aka tsara wanda ke kiyaye daidaitattun maƙasudin tushen, ma'anar, bayanin amfani da "haramtattun makamancin" waɗannan sharuɗɗan; Yana haɗi zuwa kayan aikin CAT kuma yana ba da shawarar fassara ta atomatik. Ciro wa'adi shine tsarin tattara kalmomin ta atomatik waɗanda ke buƙatar fassara/daidaita su daga rubutu. Kamus shine mafi sauƙi, jerin sharuɗɗan ginshiƙai biyu; Ana amfani da shi maimakon termbase a cikin ƙananan ayyuka.

Me yasa yake da mahimmanci haka? Saboda fassarar inji da LLM da kansu ba su bada garantin daidaiton kalmomi ba. NMT na iya fassara kowace jumla da kanta; LLM na iya canza shawararsa a cikin dogon rubutu. Abin da ke ba da ladabtarwa na lokaci shine ƙayyadaddun bayanan da kuka shirya kuma ku haɗa da abin hawa.

Tukwici: Ku ciyar da mintuna 30 don saita lokaci akan aikinku na farko tare da abokin ciniki. Wannan jarin yana haɓaka duka sauri da inganci a cikin duk ayyukan da ke gaba kuma yana ba da suna "samuwa daidai a cikin fassarorin ku".

Haɓakawa na lokaci da ginin termbase tare da AI

AI yana haɓaka ɓangaren mafi wahala na kafa ƙa'ida: nemo 'yan takara. Kuna iya ba da rubutun tushe ga LLM kuma ku ce, "jera takamaiman ƙa'idodin yanki a cikin wannan rubutu, waɗanda suke maimaituwa, da waɗanda ke buƙatar daidaito cikin fassarar." Bayan haka kun amince da wannan jerin 'yan takara, ƙayyade madaidaicin amsawar manufa kuma tabbatar da abokin ciniki idan ya cancanta. Mahimmin mahimmanci: AI yana ba da shawarar 'yan takara na lokaci, gwani da abokin ciniki sun yanke shawara. Ko alama ta kira "alƙawari" ko "yi ajiyar kuɗi" zaɓin abokin ciniki ne, ba na AI ba.

Ga kowane lokaci, sanya mai zuwa cikin Termbase: kalmar tushe, takwaransa da aka amince da shi, taƙaitaccen bayanin/ma'anar, "kada ku yi amfani" (an haramta) daidai, da yanki/tambarin aikin idan an zartar. Rubuta haramtattun makamantan ita ce hanya mafi inganci don guje wa rashin daidaituwa (don "account", KAR a yi amfani da "account" kawai; "profile", "membership").

Kula da horon kalmomi a duk tsawon fassarar

Shigar da Termbase bai isa ba; Yana buƙatar aiwatarwa. Layer uku:

  1. Lokacin fassara: Lokacin fassara zuwa LLM, ƙara maƙasudin kalma zuwa ga faɗakarwa ("DOLE a yi amfani da waɗannan sharuɗɗan tare da waɗannan kwatankwacin"). Idan kana amfani da kayan aikin CAT, haɗa ma'auni; Kayan aiki yana nuna wasan nan take.
  2. Bayan Fassara QA: Bincika rubutun da aka gama a kan ma'auni. Na tambayi LLM "akwai wurare a cikin wannan rubutun da ba su bi ka'ida ba?" ko gudanar da aikin QA na kayan aikin CAT.
  3. Sabuntawar ƙarshen aikin: Ƙara sabbin sharuɗɗa da ra'ayoyin abokin ciniki zuwa ma'auni lokacin da aikin ya ƙare; termbase takarda ce mai rai.
Tsanaki: Lokacin da kuka ba LLM dogon lokaci yana iya "manta" wasu sharuɗɗan ko canza su dangane da mahallin. Don haka kar a taɓa tsallake zagayen QA bayan fassarar; Ƙaddamarwa ba shine madadin dubawa ba.

uku mini lokuta

Case 1 - Termbase ya share rashin daidaituwa. A cikin takaddun kalmomi 60,000 da wani kamfani na software ya yi tare da masu fassara masu zaman kansu 5 daban-daban, an fassara kalmomi 40 iri ɗaya ta hanyoyi daban-daban. Da zarar an kafa tushe gama gari kuma duk fassarorin sun daidaita da shi, korafe-korafe kamar "Ba zan iya samun wannan allon ba" don tallafin mai amfani ya ragu sosai.

Shari'a 2 - AI ta rage ƙima daga kwanaki 2 zuwa awanni 2. Maimakon fitar da jerin sharuɗɗan da hannu daga takaddar na'ura mai shafi 200, ƙungiyar fassarar likita ta tsara LLM: an jera sharuɗɗan ɗan takara 380 a cikin sa'o'i 2. Masanin ya amince da waɗannan a cikin kwana 1 kuma ya fayyace martanin su; Zai ɗauki makonni idan an yi shi da hannu. Hukunce-hukuncen na mutane kuma.

Shari'a ta 3 - Dokar haramtawa ta kawar da rikici. A cikin ma'auni na alamar, "abokin ciniki" an amince da shi don "abokin ciniki" kuma an hana "mai amfani" (saboda sanya alamar). Lokacin da sabon mai fassara ya buga "mai amfani", kalmar QA ta kama shi kuma an gyara murguɗin harshe kafin ya ci gaba da rayuwa.

Samfura huɗu masu kwafi

1) Cire wa'adin ɗan takara:

Matsayinku: Masanin kimiyya. Daga waɗannan sharuɗɗan cire rubutu [yanki] waɗanda ke buƙatar daidaito a cikin fassarar: ƙayyadaddun ra'ayoyi na yanki, jimlar jimloli masu maimaitawa, samfuran samfura/fasali, gajarta. Jera kowane kalma a cikin yaren tushe, ƙara taƙaitaccen bayanin kula. Hukuncin nawa ne; Kawai ku ba da shawarar ɗan takara. Ƙara kalmomin gama gari (da, amma).Text: [...]

2) Ƙirƙirar layin Termbase:

Shirya layin tushe don sharuɗɗan masu zuwa: kalmar tushe | mayar da martani mai niyya | taƙaitaccen bayanin | ba daidai ba amsoshi don kauce wa. Idan ba ku da tabbas, yi alamar amsar da [?]; Zan yanke shawara.Alan: [...] | Sharuɗɗan: [...]

3) Fassarar da ta dace ta Termbase:

Fassara wannan rubutu [source]→[manufa]. DOLE a yi amfani da ƙa'idar da ke ƙasa; KADA KA YI AMFANI da abubuwan da aka haramta:- [source1] → [target1] (amfani: [an haramta1]) - [source2] → [target2] Idan kun ci karo da kalmar da ba a cikin Termbase ba, yi masa alama da [?].Text: [...]

4) QA bayan fassarori:

Bincika fassarar da ke ƙasa akan ka'idar da na bayar. Kawai jera RASHIN DADI: wace kalma aka yi amfani da ita, inda, tare da wace ma'anar kuskure, menene yakamata ya zama daidai. Bayar da rahotannin wurare masu jituwa.Tambayoyi: [...] | Fassara: [...]

Rauni mai ƙarfi / Ƙarfi mai ƙarfi

Rauni: "Ku yi hankali da sharuɗɗan." (Wane lokaci, wanda yayi daidai, wanne haramcin - ba a sani ba; injin ya zaɓi kansa.)

Mai ƙarfi: "Lokacin da ake fassara wannan rubutu, yi amfani da ƙa'idar da ke gaba: 'dashboard'→'dashboard', 'login'→'login'. Idan kun ci karo da kalmar da ba ta dace ba, sanya [?] kuma ku tsara shi da kanku."

Bambance-bambance: hanzari mai ƙarfi yana ƙayyade takwaransa da aka yarda, da takwaransa da aka haramta, da kuma halin da ba a sani ba; an tabbatar da daidaito.

Table of the terminology tools/ approach

kusanci

Me yake yi?

iyaka

Kalmomi (ginshiƙai biyu)

Ƙananan kasuwanci, farawa da sauri

Babu ma'anar / hani

Termbase (an haɗa zuwa CAT)

Shawarwari ta atomatik, QA

Shigarwa yana buƙatar aiki

Cire wa'adi tare da LLM

Yana hanzarta jerin 'yan takara

Mutane sun yanke shawara

Tsawon lokaci QA

Yana kama dissonance

Dole ne Termbase ya kasance na zamani

Jerin abubuwan da aka haramta

Yana kiyaye yaren alamar

Yana buƙatar kulawa akai-akai

Kuskuren gama gari

  • Fara babban aiki ba tare da shigar da Termbase ba. Rashin daidaituwa daga baya yana girma sosai.
  • Karɓar martanin da AI ya ba da shawara ba tare da tambaya ba. Shawarar ƙa'idar ta dogara ga abokin ciniki / gwani.
  • Ba rubuta amsoshi da aka haramta ba. Kawai rubuta "dama" baya hana "ba daidai ba".
  • Kalmar bayan fassarar tana tsallake QA. Sanya termbase a cikin hanzari baya bada garantin aiwatarwa.
  • Ba ana sabunta Termbase ba. Mataccen mataccen lokaci zai ɓatar da lokaci.

Kalmomin harsuna da yawa da amincewar abokin ciniki

A cikin manyan ayyuka, ma'anar kalmar sau da yawa yana da yaruka da yawa: ra'ayi yana da takwaransa a cikin yare fiye da ɗaya (source → Turanci, Jamusanci, Baturke, Larabci). Wannan yana tabbatar da daidaiton fassarar ra'ayi iri ɗaya a cikin duk harsuna kuma yana haɓaka aiki lokacin da aka ƙara sabon harshe. Abin lura a cikin ma’anar harsuna da yawa shi ne cewa kowane harshe yana da nasa al’adu da na nahawu; Ba daidai ba ne a makance kwafi daidai da yare zuwa wani.

Kalmomin kalmomi kuma yana da sarkar yarda. A cikin sharuddan mahimmanci, ƙwararrun batun abokin ciniki ko ƙungiyar tambarin abokin ciniki yakamata su yanke shawarar ƙarshe, ba mai fassara ba; Domin sun fi sanin jami'in/sha'a/alama daidai da wani lokaci. Shi ya sa a kan manyan ayyuka yana da kyau a yi wa lakabin lakabin "an yarda", "dan takara" da "masu jayayya": kuna tambayar abokin ciniki don neman ɗan takara da sharuɗɗan jayayya gaba ɗaya, samun amincewa, sannan ku kulle su. Wannan horo yana hana tsakiyar ɗawainiya "wannan kalmar ya kamata a zahiri" abin mamaki da sake fassarar taro.

A takaice

Daidaituwar ƙa'idodi shine ganuwa amma mafi mahimmancin ingancin fassarar. Babban mai fassara ya kafa ma'auni akan aikin; Yana amfani da AI don fitar da 'yan takara da sauri, amma ya bar yanke shawara tare da abokin ciniki da gwani; a sarari ya rubuta amsa da aka yarda da kuma haramta; Yana haɗa termbase zuwa faɗakarwa/kayan aiki yayin fassarar kuma baya tsallake kalmar bayan fassarar QA. Termbase takarda ce mai rai; yana girma a kowane kasuwanci kuma yana kare alamar.

Aikin aikace-aikace

Zaɓi rubutun filin (na fasaha, likita ko tallace-tallace). Ciro aƙalla ƴan takara na wa'adi 15 tare da samfuri na "ƙaddamar da ɗan takara", tabbatar da su kuma rubuta daidaitattun ma'anoni da ma'anoni da aka haramta tare da samfurin "layin termbase". Sa'an nan kuma fassara ɗan gajeren rubutu tare da wannan ma'auni, bincika rashin daidaituwa tare da samfurin "term QA" kuma gyara abin da kuka samu.

jerin abubuwan dubawa

  • [ ] Na kafa kalma / ƙamus a wurin aiki.
  • [ ] Na amince da ƴan takara na AI kuma na yanke shawara.
  • [ ] Baya ga amsoshin da aka amince, na kuma rubuta abubuwan da aka haramta.
  • [ ] Na haɗa ma'auni zuwa faɗakarwa/kayan aiki yayin fassarar.
  • [ ] Na yi binciken QA na bayan fassarar kuma na gyara rashin jituwa.