Chii chaitika
arXiv preprint inosuma MoPLEx, tarisiro-yekuwedzera yekudzidza musanganiswa wePlackett-Luce modhi kubva kune dzakawanda-nzira nhanho mhinduro. Iyo nzira yakagadzirirwa kurongeka kweAI uye kukwidziridzwa kwekuda, uko vatsananguri vangave vasingagovane imwe inopindirana yaunofarira chimiro.
Iro bepa rinodzidza maitiro ekudzidza musanganiswa wek Plackett-Luce modhi kubva kune akawanda-nzira masanji anopihwa nevatsananguri. Nenzira inoshanda, nzira iyi inofungidzira kuti mhinduro dzezvimiro dzinogona kubva kune akati wandei ari pasi pemapoka ekuda kwete kubva kune imwechete ine kuodha kwakagovaniswa. Vanyori vanogadzira izvi zvinoenderana neAI kurongeka uye kukwidziridzwa kwekuda, uko kutonga kwevanhu kunowanzoshandiswa kutungamira maitiro emitauro. Uku kuumbwa kunochengeta tarisiro yekuti mazinga akacherechedzwa akarongwa nekududzirwa sei, pane kufunga kuti kusawirirana kunongoerekana kwakanganisa mudhata.
Vanyori vanocherekedza fungidziro yekumisikidza mubasa rekutanga pamisanganiswa yeBradley-Terry modhi zvichibva pakuenzanisa paviri. Vanotaura kuti mamodheru emusanganiswa anova nedzidziso isingazivikanwe kana k yakakura kupfuura m/2, uko m ndihwo hurefu hwenzvimbo. Iyo sosi hairatidze kuti kudzikiswa uku kunobata ese aripo ekuda-optimization system; inopa mamiriro acho sedambudziko rinokurudzira nzira yakarongwa. Musiyano iwoyo une basa nekuti iyo yakataurwa inomisikidza inosimudzira nzira yekuenzanisira pasina, pachayo, inotsanangura maitiro emamwe ese masisitimu kana dataset.
MoPLEx inoshandisa matanho maviri makuru. Chekutanga, inokwidziridza zvimiro zviripo kusvika kuhukuru hukuru nekugadzira mhinduro nyowani kubva kumhando yemutauro wekutanga. Chechipiri, inoshandisa gradient-based fungidziro munzvimbo yekumisikidza yekumisikidza kudzikisa mutengo wekufungidzira. Iwo fungidziro anokonzerwa anobatanidzwa mune tarisiro-yekuwedzera maitiro ekukodzera musanganiswa wePlackett-Luce modhi. Vanyori vanoshuma kuti iyi gradient-based approximation inofungidzira mukana wechokwadi neisingasviki 5% kukanganisa pamamodhi anosvika 34 bhiriyoni paramita, asi kwakabva hakuna kutsanangudza kuzere kwekuyedza kuseta kana huwandu hwemabasa anoshandiswa bvunzo iyoyo. Pamwe chete, matanho aya anotsanangura kufambiswa kwebasa kubva kune yakawedzera chinzvimbo data kuburikidza nefungidziro yemukana uye musanganiswa wakakodzera.
Nei zvichikosha
Iro basa rinogadzirisa ganhuriro mukuenzanisira kwekuda: nzira dzinofunga imwe yekuda maitiro dzinogona kuvharidzira kusawirirana kunoreva pakati pevanotsanangura. Kana mhedzisiro yakashumwa ikabata kupfuura zviyedzo zvebepa, kuenzanisira kwakasiyana-siyana zvido zvinogona kubatsira vanogadzira kunzwisisa data rekurongeka nenzira kwayo uye kuderedza njodzi yekubata kusawirirana seruzha.
Chinonyanya kukosha ndechekuti data rekuda rinogona kunge riine kusawirirana kwakarongwa. Imwe chete yekumisikidza modhi inogona kudzvanya mitongo yakasiyana muavhareji chiratidzo, nepo musanganiswa modhi uchiedza kuparadzanisa mapatani. Pabasa rekuenzanisa, musiyano iwoyo unogona kukosha kana vatauri vakasiyana nekuti vanokoshesa zvinangwa zvakasiyana, kududzira mirairo zvakasiyana, kana kumiririra vanhu vakasiyana. Iro bepa rinopa MoPLEx senzira yekuyera iwo misiyano kuburikidza neakawanda-nzira masanji. Musiyano wakakosha nekuti chimiro chekusawirirana chinogona kukanganisa manzwisisiro anoita chiratidzo chekugadzirisa.
Zvinoenderana nezviri kutaurwa, zviedzo pamaseti ekuda-optimization zvakaona kuti MoPLEx yakavandudza kuunganidzwa kweavhareji neavhareji ye43.7% uye kurongeka kweavhareji neavhareji ye15.2% pamusoro pekutanga uchishandisa imwe chete yekumisikidza modhi uye musanganiswa weBradley-Terry modhi. Izvo zvichemo zvakaitwa nevanyori ve preprint, kwete yakazvimirira yakagadziriswa mhedzisiro. Kwanobva hakupi mazita ekutanga, saizi yedataset, nguva dzekuvimbwa, kana-per-dataset inodiwa kutonga kuenderana kana kusimba kwenhamba maavhareji. Mavhareji akashumwa saka anoratidza gwara uye saizi yezvakawanikwa nevanyori, asi achisiya zvakakosha kuti zviongororwe.
Iko kukosha kunoshanda kwaizoenderana nekuti nani kudzoreredza kwemapoka ekuda kunotungamira kune nani maitiro emuenzaniso mukutumirwa. Kubatana kwepamusoro kana kurongeka kwenzvimbo hausi humbowo hwekuti AI sisitimu yakachengeteka, yakanaka, kana kuti inowirirana nevashandisi. Kunobva zvakare hakutauri zvakabuda muvanhu, kutumirwa kwekugadzira, kudzikiswa kwehunhu hunokuvadza, kana kuenzanisa nedzimwe nzira dzechizvino-zvino dzekudzidza. Mupiro waro saka unonyanya kuita methodological: inokurudzira nzira inonzwisisika yekuongorora zviratidzo zvekugadzirisa uye inoshuma kubudirira kwekuedza kwekutanga. Izvo zvipingamupinyi zvinochengeta kukosha kwebasa rakasungirirwa kuongororo yezvavanoda pane kuriwedzera kusvika pakuendeswa kwakakura kana kuchengetedza mhedziso.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
Zvekutarisa zvinotevera
Mibvunzo mikuru ndeyokuti MoPLEx inowanda kupfuura iyo dataseti yakashumwa, ingani computational pamusoro penzvimbo yekuwedzera kwayo inosuma, uye kana kufanana kwegradient ipuroksi inovimbika kune anonota zvaanoda. Iro bepa ndere preprint yakarongwa kuoneka paEMNLP 2026, saka zvarinoda zvinoramba zviri pasi pekuongororwa kwakawanda uye kudzokororwa.
Kudzokorodza kunofanira kuratidza kana kuvandudzwa kwakataurwa kuchienderera mberi muhurefu hwakasiyana, nhamba dzemapoka ezvavanoda, huwandu hwevatsananguri, uye mhuri dzemhando dzemitauro. Iyo abstract mishumo miyedzo pane yekuda-optimization datasets asi haivazivise kana kutsanangura kuti vanomiririra sei. Mibairo inogona kusiyana zvakanyanya zvichienderana nekuti izvo zvaunofarira zvakasiyana sei uye nemazvo sei mhinduro dzakagadzirwa dzinoratidza basa rekutanga. Kudzokorora kwakadaro kwaizobatsira kuona kana iyo yakataurwa pateni yakagadzikana pamamiriro ese anoenderana neyakarongwa nzira.
Iyo nhanho-yekuwedzera nhanho inofanirwa kuongororwa. MoPLEx inogadzira dzimwe mhinduro kubva kumhando yemutauro wekutanga isati yafungidzira musanganiswa. Izvo zvinogadzira kutsamira kunokwanisika pane iyo base modhi kugona, kusarura, uye kugovera mhinduro. Kwabva hakutauri kuti nzira iyi ine hunyambiri sei kune idzo sarudzo, ingave mhinduro dzakagadzirwa dzichigona kukanganisa mapoka anofungidzirwa, kana kuti yakawanda sei fungidziro inodiwa muzvigadziriso zvinogoneka. Izvi zvinoita kuti hukama pakati pedzimwe nzira dzakagadzirwa uye mapoka akafungidzirwa chikamu chakakosha chekuongorora maitiro anoshanda eiyo nzira.
Rimwe basa rinofanira kujekesa mutengo wekombuta uye miganho yekuongorora. Iro bepa rinoshuma kukanganisa kusingasviki 5% yekufungidzira-kwakavakirwa gradient pane modhi inosvika 34 bhiriyoni paramita, asi mhedzisiro yacho haigadzirise kuita kwakaenzana kwemamodheru makuru kana ekupedzisira-kusvika-kumagumo kudzidziswa. Kwakabvawo hakutsananguri nyaya dzekukundikana, fungidziro yekusavimbika, kuchengetedzwa kwekuvanzika kwedata reanotinha, kana maitirwo anofanirwa kuitwa nevagadziri kana mapoka evanoda achikakavadzana. Bepa rakaendeswa kuarXiv musi waNyamavhuvhu 25, 2026, uye rakanyorwa serichauya paEMNLP 2026; kuongorora kwevezera uye kubereka kwakazvimirira kunoramba kuchikosha zvisingazivikanwe. Iyi mibvunzo yakavhurika inotsanangura mukaha wasara pakati penzira yakashumwa uye chivimbo nezvekushandiswa kwayo kwakakura.