Iya enkundleni yedijithali: Kusetshenziswa i-ai ekhiqizayo ukuthola nokunciphisa ukucwazimula ekuxhumaneni nomphakathi

I-Artificial Intelligence (AI) ibusa izihloko zanamuhla – idunyiswe njengendlela yokuphumelela ngolunye usuku, yaxwayisa ngokuthi yingozi ngokulandelayo. Kodwa-ke iningi le mpikiswano lenzeka nge-bubble, ligxile kumathemba angaqondakali kanye nokwesaba kunokuba izixazululo zokhonkolo. Ngaleso sikhathi, enye inselelo ephuthumayo yayivame ukunganakwa ukuthi ukunyuka kwezinkinga zempilo yezengqondo emiphakathini eku-inthanethi, lapho ukushintshana okubandlululayo noma okunobutha kushukumisa ukwethembana nokuphepha kwengqondo.

Le ndatshana yethula ukusetshenziswa okusebenzayo kwe-AI okuhloswe ngayo kuleyo nkinga: Ipayipi lokufunda lomshini elenzelwe ukuthola nokunciphisa ukucwazimula kokuqukethwe okukhiqizwe umsebenzisi. Uhlelo luhlanganisa amamodeli wokufunda ajulile wokuhlukaniswa ngamamodeli amakhulu wezilimi (LLMS) ngokuqamba izimpendulo ezibucayi zomongo. Uqeqeshelwe amazwana angaphezu kwezigidi ezimbili kanye nama-twitter, kufinyelelwe ukunemba okuphezulu (F1 = 0.99) futhi kukhiqizwe imiyalezo yokulinganisa ehambisanayo nge-Virtual Moderator Persona.

Ngokungafani neningi le-hype ezungeze i-AI, lo msebenzi ukhombisa ithuluzi elibonakalayo, elisebenzisekayo elisekela inhlala-kahle yedijithali. Kubonisa ukuthi i-AI ingamkhonza kanjani nje ukusebenza kahle kwebhizinisi noma inzuzo, kepha ukudalwa kwe-fairer, izikhala ezibandakanya kakhulu lapho abantu bexhuma khona online. Kulokho okulandelayo, ngichaza ipayipi, ukusebenza kwalo, kanye nemiphumela yaso ebanzi yemiphakathi eku-inthanethi ne-Digital Well. Kubafundi abanesifiso sokuhlola ucwaningo ngokujula okwengeziwe, kufaka phakathi ividiyo yesethulo poster echaza izindawo zekhodi kanye nombiko wokucwaninga ogcwele, izinsiza zitholakala ku-GitHub. [1]

Ipayipi lokufunda lomshini elisebenzisa ubuhlakani bokufakelwa obulalayo ukuze babhekane namanethiwekhi okuxhumana nomphakathi kubaluleke kakhulu emphakathini wezengqondo. Ukuningi ngokwengeziwe, ukuxhumana komuntu namakhompyutha ukuthembela izimpendulo ezinikezwa amamodeli amakhulu wezilimi ezikhuluma ngengxoxo yokubonisana.

Isu

Uhlelo lwenzelwe njengepayipi lesigaba ezintathu: ukuqoqa, ukuthola nokunciphisa. Isigaba ngasinye esihlanganisiwe semvelo solimi lwe-Natural Proceding (NLP) ngamamodeli wesimanje we-Transformer ukuthwebula womabili lesikali nobuqili bolimi olucwasiwe online.

Isinyathelo 1. Ukuqoqwa kwedatha nokulungiselela

Ngithole okuthunyelwe okungu-1 million twitter kusuka ku-datent140 dataset [2] kanye ne-1 million reddit imibono kusuka ku-pushshift Corpus (2007-2014) [3]. Amazwana ahlanzwa, angaziwa, futhi anikezele. Ukuqanjwa kokulungiselela okubandakanya i-tokazizatization, i-lemmatization, ukususwa kwamagama, kanye nokufana kwamagama ukusebenzisa i-NLTK ne-Spacy.

Ukuqeqesha amamodeli ngempumelelo, izimpawu ze-metadata ezingezinyosi – ezinjenge-bias_terms, has_bias, kanye ne-bias_type-evumele ukuhlukaniswa kwe-subsset okubandlululayo. Ithebula le-1 lifingqa lezi zici, ngenkathi ama-1 libonisa imvamisa yamagama we-bias kuwo wonke ama-datasets.

Ukubhekana nokuvuza kwedatha kanye nezinkinga ze-model ngokweqile zibalulekile ezigabeni zokuqala zedatha.

Amasu wokufunda agadiwe asetshenziselwa ukulebula imigomo ye-bias futhi azihlukanise njengamafomu afakiwe noma acashile.

Isinyathelo 2. Isichasiselo se-bias kanye nelebula

I-Bias yachazwa ngama-axes amabili: ukuba khona (okubandlululwayo vs. ukungakhethi) kanye nefomu (eliveziwe, elicacile, noma alikho). Ukuqashwa okuphelele kuchazwe njengolimi olucashile noma olunamakhodi (isib Isibonelo, “umkhulu uBiden wawela izitebhisi” wawunamakhodi njengo-Age Age, kuyilapho “uBiden ungumkhulu othanda umndeni wakhe” kwakungenjalo. Lokhu kufakwa amakhodi okuqukethwe kuncishise ama-positives angamanga.

Isinyathelo 3. Umbono kanye namamodeli wokuhlela

Amamodeli amabili we-transformer anika amandla isigaba sokutholwa:

– Roberta [4] kwahlelwa kahle ukuhlukaniswa kwemizwa. Imiphumela yayo (enhle, engathathi hlathi, engemihle) isize ukwehlisa izwi lokuphawula okubandlululayo.

– Distilbert [5] waqeqeshwa kudathasethi ecebisiwe enamalebula angenamkhawulo / acacile, enika amandla ukuhlukaniswa okuqondile kwezindawo ezicashile.

Ngemodeli yokutholwa oqeqeshiwe ngokunemba okuphezulu kakhulu, amazwana ahlolwe ngemodeli enkulu yolimi futhi kukhiqizwa impendulo.

Isinyathelo 4. Isu lokunciphisa

Ukutholwa kwe-Bias kwalandelwa ukunciphisa isikhathi sangempela. Lapho sekukhonjwa ukuphawula okubandlululayo, uhlelo lwakha impendulo elungiselelwe uhlobo lwe-BAS:

– Ukucacisa okucacile: Ukulungiswa okuqondile, okuqinisekayo.
– Ukuqasha okuphelele: Ukuphindaphinda okuthambile noma iziphakamiso zemfundo.

Izimpendulo zakhiqizwa nge-chatgpt [6]ekhethelwe ukuguquguquka kwawo nokuzwela komongo. Zonke izimpendulo zakhiwa nge-fant egxile eModerati, iJenai-Moderator ™, eyayilondoloze izwi nethoni (isithombe 3).

Isinyathelo 5. Ukwakhiwa kwesistimu

Ipayipi eligcwele liboniswa kuMfanekiso 4. Ihlanganisa ukuvela kwangaphambili, ukutholwa kwe-bias, kanye nokunciphiswa okuzayo. Imiphumela yedatha kanye nemodeli igcinwe kwi-postgresql ehlobene ne-schema, ivumela ukungena, ukucwaningwa kwamabhuku, nokuhlanganiswa okuzayo nezinhlelo zabantu-in-loop.

Umphumela

Lolu hlelo luhlolwe kwidathabhethi yemibono engaphezu kwezigidi ezimbili kanye namazwana we-twitter, egxile ngokunemba, i-nuance, kanye nokusebenza kwangempela komhlaba.

Isici esikhizayo

Njengoba kukhonjisiwe kuMfanekiso 1, amagama ahlobene nomjaho, ubulili, kanye nobudala avele ngokungafani nokuphawula kwabasebenzisi. Ekudlule kokuqala kokuhlola idatha, yonke imininingwane yedatha yahlolwa, futhi kwakukhona ukwenzeka okungamaphesenti amane okuthi abakwa-Bias bakhonjwa kumazwana. Ukuhlukaniswa kwasetshenziselwa ukubhekana nokungalingani kokungabikho kokuvela kokubandlulula. Amagama we-Bias afana nomkhiqizo nokuxhashazwa kwavela kaningi, kanti abakwa-Seconto be-BIORTION babonisa ngokuvelele njengabanye ukulingana okuhlobene nokulingana.

Ukusebenza kwemodeli

– URoberta uzuze ukunemba okuqinisekisiwe kwe-98.6% ngu-Epoch wesibili. Ama-curves awo alahlekelwe (Umdwebo 5) ahlangana ngokushesha, nge-matrix yokudideka (Umdwebo 6) ekhombisa ukuhlukaniswa okuqinile kwesigaba.

– UDidilbert, oqeqeshelwe amalebula angenasisekelo / acacile, afinyelela amamaki we-99% F1 (Umdwebo 7). Ngokungafani nokunemba okuluhlaza, i-F1 engcono ikhombisa ukulinganisela kokunemba nokukhumbula emininingwaneni emibalabala[7].

Ukusatshalaliswa kohlobo lwe-Bias

Umdwebo 8 ukhombisa ama-boxplots ohlobo lwe-bias asatshalaliswa ngaphezulu kwamarekhodi okuqokwa amarekhodi. Ubude bebhokisi lebhokisi lamazwana amabi lapho cishe amarekhodi angama-20,000 e-database ehlanganisiwe ehlanganisa imibono emibi kakhulu futhi engemihle ehlanganisiwe. Ukuze uthole imibono emihle, okungukuthi, ukuphawula okubonisa umuzwa wothando noma ongewona ama-bias, ibhokisi leziza iziza ezingaba ngu-10,000 amarekhodi. Ukuphawula okungathathi hlangothi kwakungamarekhodi acishe abe ngu-10,000. I-bias kanye nokubikezela imizwa okubikezelwe kuqinisekisa logic okunolwazi lokuhlelwa kwemizwa.

Ukusebenza kahle

Izimpendulo ezikhiqizwe ezivela eJenai-Moderator eziboniswe kuMfanekiso 3 zahlolwa ngababuyekezi babantu. Izimpendulo zahlulelwa zinembe ezilimini nangokwesimo esifanele, ikakhulukazi zokukhetha okugcwele. Ithebula 2 linikeza izibonelo zokubikezelwa kohlelo ngokuphawula kwangempela, okubonisa ukuzwela kwamacala acashile.

Ukukhuluma

Ukulinganisa kuvame ukufakwa inkinga njengenkinga yokuhlunga kwezobuchwepheshe: Thola igama elivinjelwe, susa amazwana, bese uqhubekela phambili. Kepha ukumodelana nakho kungukuxhumana phakathi kwabasebenzisi nezinhlelo. Ku-HCI ucwaningo, ukulunga akuyena kuphela ezobuchwepheshe kepha kukodwa [8]. Lolu hlelo luhlanganisa lo mbono, ukunciphisa ukunciphisa inkhulumomphendvulwano ngomongameli oqhutshwa ngumuntu: UJenai-Moderator.

Ukulinganisela njengokuxhumana

Ukuqagela okucacile kuvame ukudinga ukulungiswa okuqinile, kuyilapho kufakwa izinzuzo ze-lias bias ezivela kwimpendulo eyakhayo. Ngokuvusekisa kunokususa, uhlelo lukhuthaza ukuboniswa nokufunda [9].

Ukulunga, ithoni, nokwakhiwa

Tone izindaba. Ukulungiswa okunokhahlo ngokweqile kubasebenzisi abahlukanisayo; ngokweqile izexwayiso ezinengozi ngokunganakwa. Lolu hlelo luyehluka ngethoni: Ziqiniseke nge-exprevelit bias, efundisayo yokuba yi-lias bias (Umdwebo 4, Ithebula 2). Lokhu kuqondanisa nocwaningo okubonisa ukungakhethi kuncike kumongo [10].

Ukukala nokuhlanganiswa

Idizayini evumelanayo isekela ukuhlanganiswa okususelwa ku-API namapulatifomu. Ukungena ngemvume okwakhelwe ngaphakathi kunika amandla ukuvela obala nokubuyekezwa, ngenkathi izinketho zabantu ezikulo

Ukucatshangelwa okuhle nokuhlangana komphakathi

Ukutholwa kwe-Bias kubeka engcupheni ama-positive angamanga noma amaqembu angenamali angaphezulu. Indlela yethu igqugquzela lokhu ngokukhipha idatha yolwazi lomuntu siqu, ukugwema amalebula abunguophonoji, nokugcina izingodo ezihlaziya. Noma kunjalo, ukwengamela kubalulekile. Njengoba uMehrabi et al. [7] phikisana, uBias akakaze aqedwe ngokuphelele kodwa kufanele ahlatshwe.

Ukugcina

Le phrojekthi ikhombisa ukuthi i-AI ingathunyelwa ekwakheni emiphakathini eku-inthanethi – hhayi nje ukuthola ukukhetha, kepha ukunciphisa ngezindlela ezigcina isithunzi somsebenzisi futhi kuqhakambisa inhlala-mbuthano yedijithali.

Ukwamukela

Le phrojekthi igcwalise izidingo ze-Milestone II ne-Capstone Yezidingo ze-Master of Appled Idatha Science Science Science Science Science Science (Mads) e-University of Michigan School of Information (UMSI). Iphosta yephrojekthi ithole umklomelo we-Mads e-USSI Expososition 2025 Seshini. UDkt Laura Stagnaro ukhonze njengomeluleki wephrojekthi ye-capstone, noDkt Jinseok Kim wakhonza njengomeluleki wephrojekthi weMilestone II.

Mayelana nomlobi

UCelia B. Banks ungosayensi wezenhlalo kanye ne-data sosayensi okhiqizwa ngumsebenzi wabantu izinhlelo zabantu futhi wasebenzisa isayensi yedatha. Ucwaningo lwakhe odokotela ezinhlelweni zabantu nenhlekelele luhlolisisa ukuthi izinhlangano zivela kanjani ezindaweni ezibonakalayo, zibonisa intshisekelo yakhe ebanzi ekuxhumaneni kwabantu, ubuchwepheshe kanye nezakhiwo. UDkt Banks ungumfundi wempilo yonke, futhi ukugxila kwakhe kwamanje kwakha kulesi sisekelo ngocwaningo olusetshenzisiwe ku-Data Science kanye ne-Analytics.

Ukunqubekela phambili

[1] C. Amabhange, uCelia Banks Portfolio Reposito: I-University of Michigan School of Information Poster Session (2025) [Online]. Kuyatholakala: [Accessed 10 May 2025]

[2] A. Go, Ukuhlaziywa kwemizwa ye-Twitter (2009), I-Entropy, p. 252

[3] Ukuqapha1, Ukubekwa kwamazwana okubomvu okungu-1 billion kusuka ngo-2005-2019 [Data set] (2019), PushShift ngeso. Kuyatholakala: [Accessed 1 September 2024]

[4] Y. Liu, Roberta: Indlela elungiselelwe kahle ye-Bert Predraining Akess (2019), I-Arxiv Preprint arxiv, p. 1907.1116892

[5] V. ISANHH, Distilbert, inguqulo ehlikiwe yeBert: incane, esheshayo, eshibhile futhi ekhanyayo (2019), I-Arxiv Preprint arxiv, p. 1910.01108

[6] B. zhang, ukunciphisa ukucwasana okungafuneki ngokufunda okuphikisana (2018), ku Ingqungquthela ye-Aaai / ACM ku-AI, Ethics, nasemphakathini, I-PP. 335-340

[7] A. Mehrabi, inhlolovo ekuhlolweni nasekulunga ekufundeni komshini (2021), phakathi Acm computing izinhlolovo, vol. 54, cha. 6, PP. 1-35

[8] R. binns, ubuhle ekufundeni komshini: Izifundo ezivela kwifilosofi yezepolitiki (2018), ngaphakathi Ingqungquthela ye-PMLR ekulunga, ukuziphendulela kanye nokusobala, I-PP. 149-159

[9] US. Jhaver, A. UBruckman, no-E. Gilbert, ukusebenzisana komshini womuntu wokuqukethwe kokuqukethwe: icala leRedDit Autodeoderator (2019), Ukuthengiselana kwe-ACM kokuxhumana kwabantu ngekhompyutha (iTochi), vol. 26, Cha. I-5, PP. 1-35, 2019

[10] N. Lee, P. Resnick, no-G. Barton, Ukutholwa Nokuncishiswa Nokunciphisa Ukutholwa: Imikhuba Engcono Kakhulu Nezinqubomgomo Zokunciphisa Ukulimala Kwabathengi (2019), ku I-BROOCIngs InstituteEWashington, DC