Ababhali:  (1) Clemencia Siro, University of Amsterdam, Amsterdam, The Netherlands;  (2) Mohammad Aliannejadi, University of Amsterdam, Amsterdam, The Netherlands;  (3) Maarten de Rijke, University of Amsterdam, Amsterdam, The Netherlands.  Ithebula Lezixhumanisi   Abstract kanye 1 Isingeniso   2 Indlela yokusebenza kanye 2.1 Idatha yokuhlola nemisebenzi   2.2 Ukuzenzakalela kwezimo zezingxoxo ezehlukene   2.3 Ukuhlolwa kwe-Crowdsource   2.4 Izimo zokuhlola   2.5 Abahlanganyeli   3 Imiphumela kanye nokuhlaziywa kanye 3.1 nezibalo zedatha   3.2 I-RQ1: Umthelela yenani elihlukahlukene lomongo wengxoxo   3.3 RQ2: Umthelela womongo wengxoxo okhiqizwa ngokuzenzakalelayo   4 Ingxoxo Nemiphumela   5 Umsebenzi Ohlobene   6 Isiphetho, Ukulinganiselwa, Nokucatshangelwa Kwezimiso Zokuziphatha   7 Ukubonga kanye Nezikhombo   A. Isithasiselo  Abstract  Amalebula e-Crowdsourced adlala indima ebalulekile ekuhloleni ama-task-oriented dialogue systems (TDSs). Ukuthola amalebula eqiniso eliyisisekelo ekhwalithi ephezulu navumelanayo avela ezichasiselweni kuletha izinselele. Lapho kuhlolwa i-TDS, izichasiselo kufanele ziqonde ngokugcwele ingxoxo ngaphambi kokunikeza izahlulelo. Ucwaningo lwangaphambilini luphakamisa ukuthi kusetshenziswe ingxenye kuphela yomongo wengxoxo kunqubo yezichasiselo. Nokho, umthelela walo mkhawulo kukhwalithi yelebula awukahlolisiswa. Lolu cwaningo luphenya umthelela womongo wengxoxo kwikhwalithi yesichasiselo, kucatshangelwa umongo onqanyuliwe wokufaneleka nokulebula okusebenzisekayo. Siphinde futhi siphakamise ukusebenzisa amamodeli wolimi amakhulu (LLMs) ukuze kufinyezwe umongo wengxoxo ukuze kunikezwe incazelo ecebile nefushane yomongo wengxoxo futhi kufundwe umthelela wokwenza lokho ekusebenzeni kwesichasiselo. Ukunciphisa umongo kuholela ezilinganisweni ezinhle kakhulu. Ngokuphambene, ukuhlinzeka ngawo wonke umongo wengxoxo kuveza izilinganiso zokuhambisana kwekhwalithi ephezulu kodwa kwethula ukungaqondakali ezilinganisweni zokusebenziseka. Ukusebenzisa inkulumo yomsebenzisi yokuqala njengomongo kuholela ezilinganisweni ezingaguquki, ngokufana nalezo ezitholwe kusetshenziswa yonke ingxoxo, ngomzamo wezichasiselo onciphe kakhulu. Esikutholile kukhombisa ukuthi ukuklama komsebenzi, ikakhulukazi ukutholakala komongo wengxoxo, kuyithinta kanjani ikhwalithi nokuvumelana kwamalebula okuhlola anemithombo eminingi.[1]  1 Isingeniso  Ngokuthuthuka kwakamuva kumamodeli olimi aqeqeshwe ngaphambilini namamodeli ezilimi ezinkulu (LLMs), amasistimu ezingxoxo ezigxile kumsebenzi (TDSs) achaze kabusha indlela abantu abafuna ngayo ulwazi, bethula indlela engokwemvelo yabasebenzisi yokusebenzelana nemithombo yolwazi (Budzianowski and Vulic′, 2019; Wu et al., 2020). Njengoba ama-TDS aya ngokuya eba yingxenye yezinqubo zokufuna ulwazi, umbuzo wokuthi kufanele kuhlolwe kanjani ngokunembe nangempumelelo ukusebenza kwawo uba bucayi. Ngenxa yokungahlangani kahle kwamamethrikhi azenzakalelayo anamalebula akhiqizwe abantu (Deriu et al., 2021), ukuhlolwa kwama-TDS kushintshele ekuthembeleni kuzilinganiso zabasebenzisi noma amalebula agcwele abantu njengezinyathelo zeqiniso eliphansi (Li et al., 2019).  Kusetshenziswe izindlela ezahlukahlukene zokuqoqa abantu ukuze kuqoqwe amalebula eqiniso eliyisisekelo, njengokulebula okulandelanayo (Sun et al., 2021), lapho izichasiselo zidlula enkulumweni ngayinye futhi ziwachasise ngamunye ngamunye. Le ndlela yethula ubungozi obuthile enqubweni yezichasiselo, njengokukhathala kwezichasiselo kanye nomthwalo ophakeme wokuqonda ezingxoxweni ezinde kakhulu, ezidinga ukuthi zikhumbule futhi zilandelele isimo sengxoxo njengoba zichasisa izinkulumo (Siro et al., 2022). Nakuba ukulandela nokuqonda umongo wengxoxo kubalulekile futhi kungaba nomthelela ezilinganisweni zabachasi, ukufunda nokuqonda izingxoxo ezinde kakhulu kungaholela ekusebenzeni okonakele.  Ukuze kubhekwane nalolu daba, omunye umugqa wocwaningo uhlongoza ukuthatha ngokungahleliwe nje amazwi ambalwa engxoxweni ngayinye azochazelwa (Mehri no-Eskenazi, 2020; Siro et al., 2022, 2023). Ngenkathi kubhekwana nomthwalo ophezulu wokuqonda nokukhathala, ukukhawulela ukuqonda kwezichasiseli inkhulumomphendvulwano kubangela ubungozi obusobala, njengamalebula angathembekile nachemile (Schmitt and Ultes, 2015; Siro et al., 2022). Ikakhulukazi, inani lomongo wengxoxo lingaholela ekuchemani. Isibonelo, izichasiselo ezingenawo umongo ocebile zingase zincike ngokungenhloso ezilinganisweni ezinhle noma ezimbi, zinganaki ikhwalithi ebanzi yempendulo. Ngakho, ukunikeza izichasiselo umongo omncane kakhulu kubeka engcupheni izahlulelo ezidukisayo, ezingase ziholele kumalebula anganembile noma angahambisani. Ngokuphambene, izichasiselo ezigcwalayo ezinolwazi oluningi zingazikhukhumeza, okungaholela emfundweni ephansi ngokwekhwalithi yelebula.  Umsebenzi wangaphambili uphenye izici ezithinta ikhwalithi nokuvumelana kwamalebula okuhlola axutshwe nesixuku, okuhlanganisa izici zesichasiseli, arXiv:2404.09980v1 [cs.CL] 15 Apr 2024 idizayini yomsebenzi, umthwalo wengqondo, kanye nezinqubo zokuhlola (bona, isb, Parteroi202 et al., Rom. 2020; Santhanam et al., 2020). Kodwa-ke, awukho umsebenzi wangaphambilini ofunda umthelela wesampula okungahleliwe kanye nenani lamazwi ayisampula kukhwalithi yesichasiselo.  Kulolu cwaningo, sihlose ukubhekana naleli gebe locwaningo ngokuphenya ukuthi amanani ahlukene olwazi olungokomongo ayithinta kanjani ikhwalithi nokuvumelana kwamalebula agcwele imithombo ye-TDS, okunikela ekuqondeni umthelela walezo zinketho zokuklama. Sihlola amalebula e-crowdsourcing ezicini ezimbili ezinkulu zokuhlola, okungukuthi, ukuhlobana nokuba usizo ngaphansi kwezimo ezihlukene, lapho siqhathanisa ikhwalithi yesichasiselo ngaphansi kwamasu ahlukene okunqamula umongo wengxoxo.  Ukubhekana nenselelo yokuqukethwe okunganele ezingeni lokujika, siphakamisa ukusebenzisa izindlela ze-heuristic kanye nama-LLM ukuze sikhiqize isidingo solwazi lomsebenzisi nesifinyezo sengxoxo. Ama-LLM angadlala indima yabasizi bezichasiselo (u-Faggioli et al., 2023) ngokufingqa umlando wenkhulumomphendvulwano, ukusiza ukuqonda okusebenza kahle nangempumelelo komongo wengxoxo ngaphambi kokuchasisa inkulumo. Kulokhu, sisebenzisa i-GPT-4 ekufinyezeni umongo wengxoxo futhi siqhathanise ukusebenza kwezichasiselo ngaphansi kwezimo ezihlukene, kanye nosayizi abahlukene bomongo. Ngalokhu kuhlolwa, siphendula imibuzo emibili eyinhloko: (RQ1) Ingabe ukuhluka kwenani lomongo wengxoxo kuthinta kanjani ukuhlolwa okugcwele kwe-TDSs? (RQ2) Ingabe ukuvumelana kwamalebula e-crowdsourced kungathuthukiswa ngomongo ongeziwe okhiqizwa ngokuzenzakalelayo?  Okutholakele kwethu kuveza ukuthi ukutholakala komongo wengxoxo yangaphambilini kunomthelela kakhulu ezilinganisweni zabachasi, okunomthelela obonakalayo kukhwalithi yabo. Ngaphandle komongo wangaphambili, izichasiselo zivame ukunikeza izilinganiso ezinhle kakhulu ezimpendulweni zesistimu, okungenzeka ngenxa yobufakazi obanele bokujeziswa, okwethulwa ukuchema okuqondile. Ngokuphambene, ukwethula wonke umongo wengxoxo kuveza izilinganiso eziphezulu zokuhambisana. Ngokuqondene nokuba usizo, ukwethula yonke indikimba yengxoxo yethula ukungaqondakali futhi kwehlisa kancane isivumelwano sesichasi. Lokhu kugqamisa ibhalansi ethambile olwazini lwesimo olunikeziwe ukuze luhlolwe. Ukufakwa komongo wengxoxo ekhiqizwe ngokuzenzakalelayo kuthuthukisa isivumelwano sesichasiselo esimweni sokungabi bikho komongo (C0) kuyilapho kunciphisa isikhathi sesichasiselo uma siqhathaniswa nesimo somongo ogcwele (C7), kwethula ibhalansi ekahle phakathi komzamo wesichasiselo nokusebenza.  Okutholakele kwethu kudlulela kweminye imisebenzi yengxoxo egxile emsebenzini efana nosesho lwengxoxo nokunxenxa izintandokazi, kokubili kuncike ekuhloleni okuxubile ukuze kuhlolwe ukusebenza kwesistimu.  Leli phepha   ngaphansi kwelayisensi ye-CC BY 4.0 DEED. litholakala ku-arxiv  [1] Ukukhuthaza ucwaningo kule ndawo, sikhipha idatha yethu esidlangalaleni ku-https://github.com/Clemenciah/ Effects-of-Dialogue-Context

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Model Tuning Technology

Transferring the essence of optimal performance, and saving the model from the abyss of underfitting.

Ukulinganiswa kwe-AI Chatbots, I-Context I-Double-Edged Sword

About Author

IMIBONO

HANG TAGS

LESI SIHLOKO SETHULWE NGAPHAKATHI

Related Stories

The Billionaires F*cked Up

The Billionaires F*cked Up

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps