Kwenzekeni
Iphrinta ye-arXiv yethula i-HIRA, uhlelo lwe-AI olungaqeqeshelwanga, lwasendaweni ukuze luhlukanise imibhalo ezimbonini ezilawulwayo. Ihlanganisa ukubuyiswa kwegama elingukhiye elisuka kumbhalo we-OCR, ukushumeka kombhalo ominyene nokumelwa kweleveli yesithombe, bese isebenzisa imodeli yolimi esingethwe endaweni ukuze kuqinisekiswe izimo ezingaqinisekile ngaphambi kokuzidlulisela kubabuyekezi abangabantu. Ababhali babika imiphumela ku-corpus ye-trade-finance yeklasi yangasese ye-80-class kanye ne-benchmark ye-Tobacco-3482 elungisiwe.
Leli phepha lichaza i-HIRA njengenkundla engaqeqeshwanga, yokubuyiselwa emagcekeni-yokungathandwa kwabathelisi esikubona ukuze kuhlukaniswe imibhalo. Isigaba sayo sokuqala sihlanganisa amasignali amathathu: ukusesha kwe-BM25 phezu kombhalo we-OCR, ukushumeka kombhalo ominyene kanye nokumelwa kwezinga lesithombe. Lawa masiginali ahlanganiswa kusetshenziswa ukuhlanganisa kwezinga elilinganiselwe lokuqinisekisa. Uhlelo luhlukanisa amadokhumenti ngokuqondile uma ukuzethemba kokubuyisa kuphezulu. Amadokhumenti acatshangelwa ukuthi awaqinisekile noma adideke ngokubonakalayo adluliselwa kusiqinisekisi semodeli yolimi olukhulu esiphethwe endaweni, esithola umbhalo we-OCR, izibonelo ezibuyisiwe, izincazelo zamalebula namagama ahlotshaniswa nokudideka okungenzeka. Uma isiqinisekisi sihlala singaqinisekile, idokhumenti ithunyelwa kumbuyekezi ongumuntu.
Ababhali bathi ukulungiswa kwabantu kugcinwa njengezibonelo zokubuyisa okunesisindo somugqa kunokuba kusetshenziselwe ukubuyekeza amapharamitha emodeli. Lezo zilungiso zibuye zibuyekeze igrafu yokudideka ene-Dirichlet-smoothed, ehloselwe ukusiza i-cascade isingathe ukungaqondakali kwezigaba okuphindaphindayo. Ngokwezinto ezibonakalayo, idizayini ebikiwe ishintsha ukuzivumelanisa nokusuka ekuqeqesheni kabusha imodeli kuye ekubuyekezeni inkumbulo yokubuyisa yesistimu kanye nokumelwa kwayo kokudideka okuvamile. Umthombo wethula lokhu njengendlela yokubhekana namalebula okuqalisa abandayo kanye nomthamo wokubuyekeza oyivela kancane ekusetshenzisweni okulawulwayo, kodwa i-abstract ayinikezi imininingwane eyanele yokuhlola umthwalo wokuqalisa noma inqubo yokuphatha.
Ekhophasini yangasese yesigaba sama-80 yezohwebo, iphepha libika ukucubungula ukusakazwa kokukhiqizwa kwamadokhumenti angu-30,233 ngenkathi licela ukulungiswa komuntu kwemibhalo engu-1,945, noma u-6.4% wokusakaza. Ibika ukuthi i-Macro-F1 inyuke isuka ku-0.6218 yaya ku-0.8548. Kubhentshimakhi ye-Tobacco-3482 elungisiwe, i-HIRA kubikwa ukuthi izuze i-Macro-F1 ka-0.9423 ngenkathi isebenzisa isiqinisekisi esisingathwe endaweni se-DeepSeek-R1-Distill-Qwen-32B. Leli phepha lithi lokhu bekungamaphesenti angu-17.4 ngaphezu kwesisekelo solimi olungasho lutho, kuyilapho isiqinisekisi sicelelwe cishe u-40% wemibhalo futhi ingqikithi yezingcingo eziyimodeli yolimi zehle cishe ngama-60%. Umthombo ukhomba umsebenzi njengenguqulo ye-arXiv ethunyelwe ngo-Aug. 22, 2026 futhi ubala i-CIKM 2026 kumazwana ayo; ayichazi isimo sokubuyekezwa kontanga noma ukuqinisekiswa kwangaphandle.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Leli phepha likhuluma ngenkinga yokusatshalaliswa okungokoqobo: izinhlangano ezilawulwayo zingase zibe namalebula anomkhawulo, ukuhanjiswa kwedatha okunomkhawulo, ababuyekezi abayivela kancane kanye nezinqubo ezibizayo zokuphatha amamodeli. Imiphumela yayo ebikiwe iphakamisa ukuthi ukubuyisa inkumbulo kanye nempendulo yomuntu ekhethiwe kungathuthukisa ukuhlukaniswa ngaphandle kokushintsha izisindo zemodeli ngokuphindaphindiwe. Okutholiwe kusasele kuyizimangalo ezivela ku-arXiv ephrintiwe ngaphambili futhi akuqinisekisi ukuthi i-HIRA izokwenza okufanayo kuzo zonke ezinye izinhlangano, izinhlobo zamadokhumenti noma izimo zokusebenza.
Umthelela obalulekile osebenzayo ukuthi izinhlangano zingase zingadingi ukuthumela wonke amadokhumenti ngemodeli yolimi olukhulu noma ziphinde ziqeqeshe ngokuphindaphindiwe isihlungi ukuze sithuthukise ukusebenza kumakesi anomsila omude. I-cascade ebikiwe ye-HIRA igcine izinyathelo ezibiza kakhulu noma ezidinga ukubuyekezwa kumadokhumenti abonakala anzima. Uma imiphumela iba yinto ejwayelekile, lokho kungase kunciphise inani lezinto ezibucayi ezicutshungulwa imodeli yolimi, kukhawulele imigqa yombuyekezi futhi kwenze ukuthuthukiswa okwandayo kube lula ukubhala. Leyo imithelela engaba khona yedizayini ebikiwe, hhayi imiphumela esungulwe ngokuzimela.
Isakhiwo se-human-in-the-loop sibalulekile futhi ekuphenduleni. Uhlelo aluphathi imodeli yolimi njengegunya lokugcina kuzo zonke izimo ezingaqinisekile: izimo ezingaxazululiwe zidluliselwa ekubuyekezweni komuntu, futhi lokho kulungiswa kuba yingxenye yenkumbulo yokubuyisa. Echibini likagwayi-3482, leli phepha libika ukuthi ukulungiswa kwabantu okungu-518, okumele u-24.8% wechibi, kwanele ukuthi i-HIRA ifane ne-oracle yechibi elibhalwe ngokugcwele lapho yonke imibhalo engu-2,086 ibhalwe khona amalebula eqiniso eliphansi. Lo mphumela uphakamisa ukuthi impendulo ekhethiwe ingase inikeze iningi levelu yokulebula okuphelele kulokho kuhlolwa, nakuba ukuqhathanisa kuncike ku-oracle ekhethiwe yephepha kanye nokusetha kokuhlola.
Umsebenzi uhambisana nezilungiselelo ezilawulwayo ngoba ababhali baklama ngokucacile mayelana nokuba khona kwedatha nokusingathwa kwendawo. Isiqinisekisi sendawo singenza ezinye izinhlelo zokusebenza zibe lula kunokuthumela amadokhumenti kusevisi yangaphandle, kodwa umthombo awuqinisekisi ukuthobelana nomthetho, ukuphepha, ukusebenza kahle kokulawula ukufinyelela noma ukumelana nokuvuza kwedatha. Futhi ayibonisi ukuthi isistimu iqeda ukuchema noma amaphutha okuhlukanisa. I-Macro-F1 ifingqa ukusebenza kuzo zonke izigaba, ngakho-ke ayivezi ngokwayo ukuthi yiziphi izigaba zamadokhumenti ezihlala zibuthakathaka, noma izigaba ezingavamile ziyathuthuka, noma ukuthi amaphutha anemiphumela engalingani yini. Leyo mikhawulo ibalulekile lapho ukuhlukaniswa kuthinta ukucutshungulwa kwezimali, ukubuyekezwa kokuthobelana noma okunye ukugeleza komsebenzi okulandelanayo.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukulandelela okubaluleke kakhulu ukuhlola okuzimele ekuqoqweni okwengeziwe kwedokhumenti elawulwayo, ngokubika okucacile ngamaphutha, ukubambezeleka, izindleko zokusebenza, izilawuli zobumfihlo kanye nomsebenzi wombuyekezi. Umthombo awuwuvezi umklamo ogcwele wesistimu, ukusatshalaliswa kwamalebula namaphutha, noma ukuthi ukusebenza kushintsha kanjani lapho amadokhumenti ehluka ezibonelweni ezigcinwe kumemori yokubuyisa. Futhi ayiqinisekisi ukuthi indlela yokusebenza isetshenzisiwe yini endaweni yokukhiqiza elawulwayo ngale komkhakha wokukhiqiza okubikiwe.
Ukuphindaphinda okuzimele kufanele kuhlole ukuthi izinzuzo ezibikiwe zisekhona yini ngaphandle kokusakaza kwezimali ezizimele zababhali kanye nebhentshimakhi elungisiwe ye-Tobacco-3482. Okuguquguqukayo okubalulekile kuhlanganisa ikhwalithi yombhalo, amaphutha e-OCR, ukungalingani kwezigaba, izinguquko ezifanekisweni zedokhumenti, izinto ezisetshenziswa ngezilimi eziningi kanye nezigaba ebezingabonwa ngaphambilini. Umthombo awusho ukuthi zingaki izibonelo ezikhona ngaphambi kokuthi impendulo iqale, ukuthi idatha yokuqinisekisa ihlukaniswe kanjani nephuli yokubuyisa, noma ukuthi noma yikuphi ukuvuza kolwazi bekungase kuthinte imiphumela yebhentshimakhi. Leyo mininingwane izodingeka ukwahlulela ukuthi izinombolo zisebenza kangakanani.
Ukubika ngokusebenza kungomunye umbuzo ovulekile. I-abstract inikeza ingxenye yamadokhumenti athunyelwe ukuze alungiswe abantu kanye nokwehliswa kwezingcingo zemodeli yolimi, kodwa ayinikezi ukubambezeleka kokuphela ukuya ekupheleni, izidingo zokubala, ukukhula kwesitoreji, isikhathi sombuyekezi esimweni ngasinye noma izindleko eziphelele. Futhi ayiqhathanisi i-cascade nezinhlelo ezilula zokubuyisa, izihlukanisi ezivamile ezigadiwe noma ezinye izinhlobo zezilimi zendawo. Ngenxa yokuthi i-HIRA igcina izilungiso njengezibonelo zokubuyisa, ukuhlola okuzayo kufanele kuhlole ukuthi amaphutha angaphambi kwesikhathi noma amalebula achemile angavezwa ngokuphindaphindiwe futhi aqiniswe kanjani, nokuthi abalawuli bazisusa kanjani izibonelo eziphelelwe yisikhathi noma ezingalungile.
Izicelo zokubusa kwephepha nazo zidinga ukuhlolwa okusebenzayo. Ukuthunyelwa okulawulwayo kuzodinga amarekhodi ubufakazi balokho obusekela ukuhlukaniswa ngakunye, kungani idokhumenti yenyuka, ukuthi umqinisekisi uphendule ini nokuthi ukulungiswa komuntu kwashintsha kanjani izinqumo zakamuva. Umthombo uthi i-HIRA isebenzisa ukuzethemba, amagama aqondene nokudideka kanye nokubuyela emuva kwesibuyekezo somuntu, kodwa ayisho ukuthi kukhethwa imingcele yokuzethemba, ukuthi ababuyekezi babhala kangaki phezu kwesistimu, noma kwenzekani lapho ubufakazi bokubuyisa bungqubuzana nesiqinisekisi. Kuze kube yilapho leyo mibuzo iphendulwa futhi nemiphumela iphindwaphindwa, isiphetho esiqine kakhulu esisekelwayo sinomkhawulo: lesi siphetho sibika ukwakheka okukhethiwe kokubuyekeza okuthembisayo kanye nemiphumela emikhulu kuzilungiselelo ezimbili zokuhlola ezishiwo, hhayi isisombululo esivamile esifakazelwe sokuhlukaniswa kwedokhumenti elawulwayo.