paint-brush
AI mubyukuri nibyiza gusobanukirwa ibyangombwana@uri
Amateka mashya

AI mubyukuri nibyiza gusobanukirwa ibyangombwa

na Uri Merhav9m2024/09/09
Read on Terminal Reader

Birebire cyane; Gusoma

Ntabwo ari byiza ariko mubyukuri ni ingirakamaro. LLMs ni abasazi neza gusobanukirwa ibyanditse no kubikuramo amakuru. Bakeneye gusa urukundo ruke hamwe na OCR nziza no gusobanukirwa kumeza.
featured image - AI mubyukuri nibyiza gusobanukirwa ibyangombwa
Uri Merhav HackerNoon profile picture
0-item

Wifuzaga ko robot ikora ikawa yawe, ariko ukabona ibisubizo bya JSON bivuye mubyangombwa aho.

Igihe cyose tekinoroji nshya igaragaye, irarohama muri hyperbole. Twitter yanjye yuzuyemo "influencers" bavuga ko bubatse urubuga rwuzuye hamwe nigitekerezo kimwe, ariko umuntu wese uri hanze ugerageza kubaka imbuga azi ko kuri ubu aribyiza bihagije kugirango ashyire mubikorwa bito hanyuma akava kumurongo wimbitse kuri buriwese intera ndende hamwe nububiko bwuzuye kode nkimiterere.


Wibuke igihe twasezeranijwe gutwara imodoka ejo hashize imyaka icumi? Gutwara ibinyabiziga ni ikibazo cyakemutse, nk'uko byatangajwe na Elon Musk, umuvugizi wanyuma, hashize imyaka 8 .


Mugihe twari dutegereje ko Teslas itangira gukora amafranga yonyine, imbaraga nke zishimishije zari zikomeje. Mobileye yubatse sensor igenda beep mugihe ugiye kwiruka mubintu. Bakijije ubuzima butabarika, kandi bagabanya ubwishingizi hafi 90%. Bubatse isosiyete $ 17B.


Nizera ko gusobanukirwa inyandiko ari tekinoroji ya Mobileye kuri LLMs. Gusobanukirwa nimbonerahamwe yimari, gutanga ibisobanuro byubwishingizi no gufata ibyemezo byubuvuzi bivuye kwa muganga bisa nkibyoroheje ugereranije ninzozi ndende. Ariko niba ukanze inshuro ebyiri kuri iki kibazo, uzasanga mbere kitarakemutse kandi ko gifungura agaciro kenshi.

Inyuma

Imyaka icumi ishize, nakoreye itsinda ryiza rya LinkedIn. Twagerageje gukemura ikibazo kimwe cyoroshye uburiganya: nigute ushobora kumvikanisha résumé, aho yaba ituruka hose, hanyuma ugashushanya imitwe yacyo kumurongo muto wamazina yemewe?


Wagira ngo ibi byoroshye. Ndashaka kuvuga, "software injeniyeri" ni umutwe wuzuye, sibyo? Ariko tuvuge iki mugihe umuntu yanditse "mugenzi"? Bashobora kuba babitse ububiko cyangwa bakuramo umushahara wimibare itandatu mubigo byamategeko. Ikiganza cya Sitasiyo (Aussie Cowboy) niki, umujyanama ni iki (gishobora gusobanura umujyanama / wigenga, ariko birashobora gusobanura Muganga niba uri umwongereza kandi ufite amateka akwiye kuri yo)? Niba ugerageza guhuza imitwe yakazi murutonde rwibintu bizwi kugirango ubashe gutondekanya gushakisha, kugurisha, nibindi - nigute ushobora kubaka icyitegererezo kizi imiterere yindimi zose n’umuco, kandi ntukibeshye "Assistant Assistant" kuri ube umuyobozi, mugihe Assistant Regional Manager mubyukuri yungirije umuyobozi wakarere?


Nibyiza, ibi rero nibyiza, ariko niba nkorera LinkedIn , nzakenera ubwoko bwamakuru yihariye. Ndashaka JSON .


Imirimo myinshi irakenewe gushushanya amakarita yakazi muri tagisi isanzwe - urutonde rwanyuma rwakazi rwemewe mbere. Ariko urashobora kubona uburyo ikintu cyari gikomeye cyane mubihe byashize kiba gito.

Ibikorwa byo mu biro bihinduka ikibuga cya AI

Gusoma résumés nikibazo cyiza cyo gukoresha, ariko ndatekereza ko atari impinduramatwara. LinkedIn nisosiyete yikoranabuhanga kandi yamye ikoresha urwembe rukarishye kubibazo. Irashobora kuba nziza, ariko turimo gusimbuza kode imwe yo gutangiza iyindi.


Ibintu birashimishije cyane mugihe usimbuye imirimo irambiwe. Igice kinini cy'ubukungu gishingiye ku bantu bakora imirimo y'impuguke itangira “gusoma inyandiko, kumenya icyo ivuga, no gusubiramo iyo nzira ad isesemi.”


Reka nkubite ingero zimwe:

  • Gucunga amafaranga: Ufite fagitire, kandi umuntu akeneye kuyihindura kurutonde rwimibare - icyishyuwe, uwo, nifaranga. Byumvikane neza? Ntabwo iyo yashyinguwe mu kajagari kamakuru yinyongera, ameza atuzuye, cyangwa PDF zisa nkumuntu wabinyujije muri blender.


  • Gutunganya ibirego byubuzima: Iyi ninzozi mbi, ikemurwa ningabo zubuvuzi busaba ubuvuzi. Bashungura imisozi ya fagitire, inoti z’abaganga, na fagitire zose zigomba guhurira hamwe mu kajagari kangiritse hamwe na kopi, kandi zigomba kuyihuza na politiki y’ubwishingizi bw’ubuzima iriho hanyuma bakamenya niba amafaranga yishyurwa, mu cyiciro, na kugeza ku mubare. Ariko iyo umanutse kuri yo usanga ahanini ari ugusoma, gutondeka, no kuranga. Ibyemezo ntabwo bigoye; ni ugukuramo amakuru nicyo kibazo.


  • Kwandika inguzanyo: Gusubiramo ibyemezo bya banki no gutondekanya amafaranga yinjira. Na none, nibyinshi muburyo bwo gutunganya amakuru atubatswe kuruta siyanse ya roketi.


Birashimishije? Oya. Ni ingirakamaro? Ndatekereza ko.

Gukuramo Inyandiko nigikorwa gifatika

Kugeza ubu, LLM izwi cyane kuri salusiyo - aka gukora shit. Ariko ikigaragara ni cyiza: salusiyo iteganijwe mugihe ubajije ubumenyi bwisi ariko bikurwaho mubikorwa bifite ishingiro .


LLM ntabwo ari nziza cyane mugusuzuma ibyo "bazi" - nibyinshi mubyiza byamahirwe ko bashobora kubikora rwose kuko batabitojwe kubwibyo. Amahugurwa yabo y'ibanze ni uguhitamo no kuzuza inyandiko zikurikirana. Ariko, mugihe LLM ihawe inshingano zifatika - imwe aho ibyinjijwe byonyine bisabwa kugirango habeho guhanura, igipimo cya salusiyo gishobora kumanuka kugeza kuri zeru. Kurugero, niba wanditse iyi blog muri ChatGPT ukabaza niba isobanura uburyo bwo kwita kumatungo yawe ferret, icyitegererezo kizatanga igisubizo cyukuri 100%. Igikorwa kiba giteganijwe. LLMs ifite ubuhanga bwo gutunganya igice kinini cyinyandiko no guhanura uburyo umusesenguzi ubishoboye yuzuza ibibuze, kimwe muri byo gishobora kuba {“care ferret care”: ibinyoma}.


Nkuwahoze ari umujyanama wa AI, twakoze imishinga yibanda ku gukura amakuru mu nyandiko, cyane cyane mu nganda nk’ubwishingizi n’imari. Ubwoba rusange bwari "LLMs hallucinate," ariko mubikorwa, ibibazo bikomeye byakunze guterwa namakosa yo gukuramo ameza cyangwa ibindi bitavuguruzanya. LLMs birananirana gusa iyo tunaniwe kubereka ibyinjijwe bisukuye, bidasobanutse. Hariho ibintu bibiri byingenzi kuri gutangiza neza gutunganya inyandiko:


  1. Gukuramo inyandiko neza - Ibi bikubiyemo guhindura inyandiko mumyandiko isukuye, imashini isomeka, harimo gukoresha ameza, inyandiko zandikishijwe intoki, cyangwa imiterere itandukanye. LLM ikeneye inyandiko isobanutse, yumvikana gukorana nayo.


  2. Igishushanyo gikomeye - Izi gahunda zigomba gusobanura ibisubizo urimo gushaka, uburyo bwo gukemura ibibazo byimbere, hamwe nimiterere yamakuru, kwemeza ko sisitemu izi neza icyo gukuramo muri buri bwoko bwinyandiko.


Ikinyuranyo hagati yingaruka zishobora guterwa na salusiyo nimbogamizi za tekiniki zishobora kuba nini, ariko hamwe nifatizo zifatika, urashobora gukoresha LLMs neza mugutunganya inyandiko.


Gukuramo inyandiko biroroshye kuruta kubanza guhura nijisho

Dore igitera LLMs guhanuka no gutwikwa, no kubona ibisubizo bibi bisekeje:

  1. Iyinjiza ifite imiterere igoye nkuburyo bubiri-inkingi, hanyuma ukoporora & wandike inyandiko kuva eg PDF kuva ibumoso ugana iburyo, ufata interuro rwose bivuye mumirongo.
  2. Iyinjiza ifite agasanduku, ibimenyetso, ibisobanuro byanditseho intoki, kandi wabibuze rwose muguhindura inyandiko
  3. Ndetse birushijeho kuba bibi: watekereje ko ushobora kuzenguruka ugahindura inyandiko, kandi wizeye ko uzashyiraho gusa ifoto yinyandiko kandi ufite GPT kubwibyo wenyine. IYI ikujyana mumujyi wa salusiyo. Gusa saba GPT kwandukura ishusho yimeza hamwe na selile zimwe zirimo ubusa uzabona byishimye bigenda apeshit no gukora ibintu kubushake.


Burigihe bifasha kwibuka ibyo akajagari kabisa kagenda mubyangombwa-byukuri. Dore urupapuro rwimisoro rudasanzwe:

inkomoko: umusoreshwa winshuti muri leta ya Amerika


Birumvikana ko impapuro zimisoro nyazo zujujwe imirima yose, akenshi mukwandika


Cyangwa dore resumè yanjye

Inkomoko: njye


Cyangwa urugero ruboneka kumugaragaro raporo ya laboratoire (iyi ni ibisubizo byurupapuro rwambere kuva Google)



Inkomoko: irembo ryubushakashatsi, ishusho rusange


Ikintu kibi cyane ushobora gukora, nukuvuga, ni ukubaza ubushobozi bwa GPT bwo kwandukura imbonerahamwe. Gerageza niba utinyutse - birasa neza ukirebye neza, rwose ukora ibintu bitunguranye hejuru ya selile zimwe zameza, bikuramo ibintu rwose bivuye mumirongo, nibindi.

Niba hari ibitagenda neza kwisi, iyubake sosiyete ya SaaS kugirango ikosore

Igihe twashinzwe gusobanukirwa n'ubwoko bw'inyandiko, njye na mugenzi wanjye Nitai Dean twatangaye twatangajwe ko nta gisubizo kiboneye cyo kumvikanisha aya masomo.


Abantu bamwe bavuga ko babikemuye, nka AWS Textract. Ariko bakora amakosa menshi kumyandiko iyo ari yo yose igoye twagerageje. Noneho ufite umurizo muremure wibintu bito nkenerwa, nko kumenya ibimenyetso, buto ya radio, inyandiko zambutse, inyandiko zandikishijwe intoki kurupapuro, nibindi nibindi.


Noneho, twubatse Docupanda.io - ibanza gutanga inyandiko isukuye yerekana page yose uyijugunye. Ku ruhande rw'ibumoso uzabona inyandiko y'umwimerere, naho iburyo, urashobora kubona ibyasohotse.


isoko: docupanda.io


Imbonerahamwe nayo ikoreshwa. Munsi ya hood, duhindura imbonerahamwe muburyo bwa LLM-busomeka neza:

Inkomoko: docupanda.io


Igice cya nyuma cyo kumvikanisha amakuru hamwe na LLMs kibyara kandi kigakurikiza imiterere ikomye. Nibyiza cyane ko dushobora gukora AI ibumba ibisohoka muri JSON, ariko kugirango dushyire mubikorwa amategeko, gutekereza, kubaza, nibindi kumakuru - dukeneye kubikora buri gihe. Amakuru akeneye guhuza nu murongo wateganijwe mbere yuko tuzuzuza ibirimo. Mwisi yamakuru, twita iyo Schema .

Kubaka Schemas nikigeragezo nikosa… Ibyo LLM ishobora gukora

Impamvu dukeneye igishushanyo nuko amakuru ntacyo amaze adahoraho. Niba turi gutunganya inyandiko zabarwayi, hanyuma bagashushanya kuri "gabo" "Umugabo" "m" na "M" - dukora akazi kabi.


Nigute ushobora kubaka igishushanyo? Mu gitabo, urashobora kubaka igishushanyo wicaye kirekire kandi gikomeye ukareba urukuta, ukanasobanura icyo ushaka gukuramo. Wicaye aho, witondere ibikorwa byubuzima bwawe, hanyuma ugende "Ndashaka gukuramo izina ryumurwayi, itariki, igitsina nizina rya muganga. Yego, kandi uburinganire bugomba kuba M / F / Ibindi. ”


Mubuzima busanzwe, kugirango usobanure icyo wakura mubyangombwa, uhindagurika ureba inyandiko zawe… byinshi. Utangirana nibintu nkibi byavuzwe haruguru, ariko rero ureba inyandiko ukabona ko umwe muribo afite URUTONDE rwabaganga aho kuba umwe. Kandi bamwe muribo banashyiraho adresse kubaganga. Aderesi zimwe zifite numero yumubare numero yinyubako, birashoboka rero ko ukeneye ikibanza kubyo. Komeza kandi birakomeza.


Icyo twaje kubona nuko kuba ushobora gusobanura neza ibintu byose ushaka gukuramo, byombi bidafite akamaro, biragoye kandi bikemurwa cyane na AI.


Nicyo gice cyingenzi cya DocuPanda. Aho gusaba gusa LLM kunoza ibisohoka kuri buri nyandiko, twubatse uburyo bukwemerera:


  1. Kugaragaza ibintu ukeneye kubona mu nyandiko mu rurimi rwubuntu
  2. Gira ikarita yacu ya AI hejuru yinyandiko nyinshi hanyuma umenye igishushanyo gisubiza ibibazo byose kandi cyakira kinks nibitagenda neza bigaragara mubyangombwa.
  3. Hindura igishushanyo hamwe nibitekerezo kugirango uhindure ibyo ukeneye mubucuruzi


Ibyo urangije ni gahunda ikomeye ya JSON - inyandikorugero ivuga neza neza icyo ushaka gukuramo muri buri nyandiko, hamwe namakarita arenga ibihumbi magana yabyo, ukuramo ibisubizo kuri bose, mugihe ukurikiza amategeko nko guhora ukuramo amatariki muri imiterere imwe, yubaha urutonde rwibyiciro byateganijwe mbere, nibindi

isoko: docupanda.io

Byinshi!

Kimwe numwobo wose wurukwavu, burigihe hariho ibintu byinshi kuruta kubanza guhura nijisho. Igihe cyashize, twabonye ko hakenewe ibintu byinshi:

  • Akenshi amashyirahamwe agomba guhangana nuruzinduko rwinyandiko zitazwi, nuko duhita tubashyira mubyiciro hanyuma tugahitamo gahunda yo kubakoresha

  • Inyandiko rimwe na rimwe zihuza inyandiko nyinshi, kandi ukeneye igisubizo cyubwenge kugirango utandukane inyandiko ndende cyane mubice bya atome, seperate

  • Kubaza inyandiko zukuri ukoresheje ibisubizo byatanzwe nibyiza cyane


Niba hari ikintu kimwe cyakuwe kuriyi nyandiko, ni uko ugomba kureba mugukoresha LLMs kugirango wumve inyandiko muburyo busanzwe. Niba hari ibintu bibiri byafashwe, ni uko ugomba no kugerageza Docupanda.io . Impamvu niyubaka nuko ndayizera. Birashoboka ko iyo ari impamvu nziza ihagije yo kubireka?


Uzaba umukozi wo mu biro (Source: unsplash.com)