{"id":14395,"date":"2026-05-14T15:10:36","date_gmt":"2026-05-14T13:10:36","guid":{"rendered":"https:\/\/sbunker.org\/?p=14395"},"modified":"2026-05-14T15:15:52","modified_gmt":"2026-05-14T13:15:52","slug":"paragjykimet-gjinore-ne-thjerrezat-e-inteligjences-artificiale-llm","status":"publish","type":"post","link":"https:\/\/sbunker.org\/sr\/opinion\/paragjykimet-gjinore-ne-thjerrezat-e-inteligjences-artificiale-llm\/","title":{"rendered":"Paragjykimet gjinore n\u00eb thjerr\u00ebzat e Inteligjenc\u00ebs Artificiale (LLM)"},"content":{"rendered":"

M\u00ebnyra se si LLM-t\u00eb ofrojn\u00eb informacione \u00ebsht\u00eb \u00e7\u00ebshtje e r\u00ebnd\u00ebsishme p\u00ebr faktin se k\u00ebto modele p\u00ebrdoren gjithnj\u00eb e m\u00eb shum\u00eb n\u00eb situata konkrete.<\/em><\/p>\n

Prezantimi i gjinis\u00eb n\u00eb modelet e m\u00ebdha gjuh\u00ebsore (LLM) mund t\u00eb pasqyroj\u00eb dhe p\u00ebrforcoj\u00eb pabarazit\u00eb ekzistuese sociokulturore. Megjithat\u00eb, natyra e paragjykimeve t\u00eb tilla gjinore mund t\u00eb ndryshoj\u00eb ndjesh\u00ebm n\u00eb t\u00eb gjitha gjuh\u00ebt, e ndikuar nga tiparet gjuh\u00ebsore dhe t\u00eb dh\u00ebnat e trajnimit t\u00eb nj\u00eb modeli. N\u00eb nj\u00eb studim shkencor t\u00eb publikuar s\u00eb fundi<\/a> n\u00eb revist\u00ebn Computational Communication Research<\/em>, \u00ebsht\u00eb hulumtuar prezantimi gjinor n\u00eb gjasht\u00eb LLM t\u00eb ndrysh\u00ebm (LLaMA, Mistral, Qwen, Gemma, DeepSeek dhe GPT) dhe gjuha shqipe \u00ebsht\u00eb nj\u00eb nga gjasht\u00eb gjuh\u00ebt e p\u00ebrfshira n\u00eb studim.<\/p>\n

LLM-t\u00eb zmadhojn\u00eb stereotipat gjinor\u00eb n\u00eb gjuh\u00ebn shqipe?<\/strong><\/p>\n

Duke analizuar modele t\u00eb ndryshme t\u00eb Inteligjenc\u00ebs Artificiale n\u00eb disa gjuh\u00eb, p\u00ebrfshir\u00eb edhe shqipen, studiuesit gjet\u00ebn se lidhjet mes gjinis\u00eb dhe tipareve t\u00eb caktuara p\u00ebrs\u00ebriten n\u00eb m\u00ebnyr\u00eb t\u00eb q\u00ebndrueshme. Rezultatet jan\u00eb arritur duke p\u00ebrdorur emra t\u00eb ndrysh\u00ebm p\u00ebr burra, gra dhe persona jobinar\u00eb, n\u00eb nj\u00eb k\u00ebrkes\u00eb\/prompt standard p\u00ebr LLM-t\u00eb. K\u00ebrkesat u jan\u00eb dh\u00ebn\u00eb modeleve t\u00eb m\u00ebdha gjuh\u00ebsore (LLM) dhe m\u00eb pas jan\u00eb gjeneruar p\u00ebrgjigje duke lidhur k\u00ebta emra me tipare apo koncepte t\u00eb ndryshme.<\/p>\n

Gjuha shqipe rezultoi si nj\u00eb gjuh\u00eb ku modelet riprodhojn\u00eb stereotipat tradicional\u00eb gjinor\u00eb (burra = dominanc\u00eb\/agjenci, gra = komunalitet\/dob\u00ebsi), nd\u00ebrsa p\u00ebr dhunti akademike, n\u00eb shqip u pa nj\u00eb shp\u00ebrndarje m\u00eb e balancuar n\u00eb krahasim me gjuh\u00ebt e tjera. Gjithashtu, p\u00ebrfshirja e jo-binar\u00ebve ishte m\u00eb e kufizuar sesa n\u00eb gjuh\u00ebt me m\u00eb shum\u00eb burime si anglishtja apo gjermanishtja. Pra, kur LLM-t\u00eb gjenerojn\u00eb tekste, burrat shpesh p\u00ebrshkruhen si udh\u00ebheq\u00ebs, t\u00eb fort\u00eb, t\u00eb aft\u00eb p\u00ebr t\u00eb marr\u00eb vendime. Kjo pasqyron stereotipin tradicional se burrat jan\u00eb \u201cdrejtuesit\u201d dhe \u201cambicioz\u00ebt\u201d. Nga ana tjet\u00ebr, grat\u00eb shfaqen m\u00eb shpesh si t\u00eb kujdesshme, t\u00eb dashura, t\u00eb buta, ose t\u00eb brishta. Kjo pasqyron stereotipin e zakonsh\u00ebm se grat\u00eb jan\u00eb \u201ckujdestaret\u201d dhe \u201cemocionalet\u201d.<\/p>\n

Paragjykimet gjinore n\u00eb Inteligjenc\u00ebn Artificiale mund t\u00eb mos jen\u00eb t\u00eb dukshme n\u00eb shikim t\u00eb par\u00eb. Ato nuk shfaqen si gabime t\u00eb qarta apo p\u00ebrgjigje t\u00eb pasakta, por si modele t\u00eb p\u00ebrs\u00ebritura n\u00eb m\u00ebnyr\u00ebn se si p\u00ebrshkruhen njer\u00ebzit dhe rolet e tyre. Kur nj\u00eb model lidh burrat me pushtetin dhe grat\u00eb me emocionin, kjo nuk duket si problem teknik. Megjithat\u00eb, \u00ebsht\u00eb pik\u00ebrisht kjo q\u00eb e b\u00ebn t\u00eb qart\u00eb se Inteligjenca Artificiale nuk \u00ebsht\u00eb neutrale, por reflekton m\u00ebnyr\u00ebn se si shoq\u00ebria ka nd\u00ebrtuar kuptimet p\u00ebr gjinin\u00eb. N\u00eb gjuh\u00ebn shqipe, kjo situat\u00eb nd\u00ebrlikohet edhe m\u00eb shum\u00eb. Si gjuh\u00eb me m\u00eb pak burime digjitale, shqipja \u00ebsht\u00eb m\u00eb pak e p\u00ebrfaq\u00ebsuar n\u00eb t\u00eb dh\u00ebnat mbi t\u00eb cilat trajnohen k\u00ebto modele. Kjo do t\u00eb thot\u00eb se m\u00ebnyra se si funksionon IA-ja n\u00eb shqip shpesh \u00ebsht\u00eb rezultat i p\u00ebrkthimit ose i adaptimit nga gjuh\u00eb t\u00eb tjera, kryesisht anglishtja. N\u00eb k\u00ebt\u00eb proces, jo vet\u00ebm q\u00eb bartet struktura e stereotipave ekzistues, por ata mund t\u00eb b\u00ebhen edhe m\u00eb t\u00eb theksuar p\u00ebr shkak t\u00eb munges\u00ebs s\u00eb alternativave n\u00eb t\u00eb dh\u00ebna. M\u00ebnyra se si LLM-t\u00eb ofrojn\u00eb informacione \u00ebsht\u00eb \u00e7\u00ebshtje e r\u00ebnd\u00ebsishme p\u00ebr faktin se k\u00ebto modele p\u00ebrdoren gjithnj\u00eb e m\u00eb shum\u00eb n\u00eb situata konkrete. Tekstet q\u00eb LLM-t\u00eb gjenerojn\u00eb nuk jan\u00eb thjesht p\u00ebrgjigje t\u00eb izoluara, por pjes\u00eb e komunikimit t\u00eb p\u00ebrditsh\u00ebm. N\u00eb k\u00ebt\u00eb m\u00ebnyr\u00eb, Inteligjenca Artificiale kontribuon n\u00eb normalizimin e m\u00ebnyrave t\u00eb caktuara t\u00eb t\u00eb menduarit, duke i b\u00ebr\u00eb ato t\u00eb duken t\u00eb zakonshme dhe t\u00eb padiskutueshme. Problemi nuk q\u00ebndron vet\u00ebm tek ajo q\u00eb gjenerohet nga Inteligjenca Artificiale, por te m\u00ebnyra se si ajo p\u00ebrforcon at\u00eb q\u00eb tashm\u00eb ekziston<\/p>\n

A mund t\u00eb m\u00ebnjanohen k\u00ebto paragjykime?<\/strong><\/p>\n

Ndryshe nga sa mund t\u00eb mendohet, k\u00ebto paragjykime nuk zbuten leht\u00ebsisht me ndryshime t\u00eb vogla teknike. Edhe kur ndryshojn\u00eb parametrat e gjenerimit t\u00eb tekstit, modelet vazhdojn\u00eb t\u00eb prodhojn\u00eb modele t\u00eb ngjashme. Kjo tregon se paragjykimet nuk jan\u00eb sip\u00ebrfaq\u00ebsore, por t\u00eb nd\u00ebrtuara n\u00eb vet\u00eb struktur\u00ebn e tyre. N\u00eb k\u00ebt\u00eb kuptim, zgjidhja nuk mund t\u00eb jet\u00eb vet\u00ebm teknike. N\u00ebse duam q\u00eb Inteligjenca Artificiale t\u00eb jet\u00eb m\u00eb pak paragjykuese n\u00eb gjuh\u00ebn shqipe, duhet t\u00eb ndryshoj\u00eb m\u00ebnyra se si prodhohet dhe p\u00ebrdoret gjuha vet\u00eb. Tekefundit, Inteligjenca Artificiale nuk b\u00ebn gj\u00eb tjet\u00ebr ve\u00e7se na kthen nj\u00eb pasqyr\u00eb.<\/p>\n

Zbutja efektive e paragjykimeve k\u00ebrkon nd\u00ebrhyrje n\u00eb nivelin e t\u00eb dh\u00ebnave t\u00eb trajnimit, arkitektur\u00ebs s\u00eb modelit dhe procedurave t\u00eb harmonizimit. Pra, n\u00ebse tekstet mbi t\u00eb cilat trajnohet modeli p\u00ebrmbajn\u00eb stereotipa ose paragjykime, modeli do t\u2019i m\u00ebsoj\u00eb dhe do t\u2019i riprodhoj\u00eb. Nd\u00ebrhyrja k\u00ebtu do t\u00eb thot\u00eb pastrimi, balancimi ose shtimi i t\u00eb dh\u00ebnave q\u00eb p\u00ebrfaq\u00ebsojn\u00eb m\u00eb mir\u00eb diversitetin gjinor e kulturor. Struktura teknike e modelit (si \u00ebsht\u00eb nd\u00ebrtuar rrjeti i tij) ndikon n\u00eb m\u00ebnyr\u00ebn si p\u00ebrpunon gjuh\u00ebn. Ndryshimet n\u00eb arkitektur\u00eb mund t\u00eb ndihmojn\u00eb q\u00eb modeli t\u00eb mos e p\u00ebrforcoj\u00eb aq shum\u00eb paragjykimin, p.sh. duke p\u00ebrdorur mekanizma q\u00eb kontrollojn\u00eb shp\u00ebrndarjen e cil\u00ebsive mes gjinive. N\u00eb procedurat e harmonizimit (alignment) p\u00ebrfshihen rregullat dhe udh\u00ebzimet q\u00eb i vendosen modelit pas trajnimit, p\u00ebr ta b\u00ebr\u00eb sjelljen e tij m\u00eb t\u00eb p\u00ebrshtatshme p\u00ebr p\u00ebrdoruesit. K\u00ebtu futen teknika si \u201cfine-tuning\u201d ose \u201creinforcement learning with human feedback\u201d, ku modeli m\u00ebsohet t\u00eb shmang\u00eb p\u00ebrgjigjet paragjykuese dhe t\u00eb jap\u00eb rezultate m\u00eb neutrale. Pra, p\u00ebr t\u00eb luftuar paragjykimet n\u00eb modelet gjuh\u00ebsore, nuk mjafton vet\u00ebm nj\u00eb hap. Duhet t\u00eb punohet te t\u00eb dh\u00ebnat, arkitektura dhe harmonizimi. Vet\u00ebm kombinimi i k\u00ebtyre niveleve mund ta b\u00ebj\u00eb modelin m\u00eb t\u00eb drejt\u00eb dhe m\u00eb t\u00eb barabart\u00eb.<\/p>","protected":false},"excerpt":{"rendered":"

M\u00ebnyra se si LLM-t\u00eb ofrojn\u00eb informacione \u00ebsht\u00eb \u00e7\u00ebshtje e r\u00ebnd\u00ebsishme p\u00ebr faktin se k\u00ebto modele p\u00ebrdoren gjithnj\u00eb e m\u00eb shum\u00eb n\u00eb situata konkrete. Prezantimi i gjinis\u00eb n\u00eb modelet e m\u00ebdha gjuh\u00ebsore (LLM) mund t\u00eb pasqyroj\u00eb dhe p\u00ebrforcoj\u00eb pabarazit\u00eb ekzistuese sociokulturore. Megjithat\u00eb, natyra e paragjykimeve t\u00eb tilla gjinore mund t\u00eb ndryshoj\u00eb ndjesh\u00ebm n\u00eb t\u00eb gjitha gjuh\u00ebt, […]<\/p>","protected":false},"author":667,"featured_media":14397,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[363,9],"tags":[2582,2590,989,624,2768,2769],"ppma_author":[2743],"class_list":["post-14395","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-disinfo","category-opinion","tag-ai","tag-ia","tag-inteligjenca-artificiale","tag-kryesore","tag-llm","tag-stereotipat-gjinore"],"authors":[{"term_id":2743,"user_id":667,"is_guest":0,"slug":"dren-gerguri","display_name":"Dren G\u00ebrguri","avatar_url":{"url":"https:\/\/sbunker.org\/wp-content\/uploads\/2026\/03\/IMG_1743.jpeg","url2x":"https:\/\/sbunker.org\/wp-content\/uploads\/2026\/03\/IMG_1743.jpeg"},"user_url":"","last_name":"G\u00ebrguri","first_name":"Dren","description":"Dren G\u00ebrguri, Ph.D., \u00ebsht\u00eb ligj\u00ebrues n\u00eb Departamentin e Gazetaris\u00eb t\u00eb Universitetit t\u00eb Prishtin\u00ebs, Fulbrighter dhe ish-gazetar. Agjenda e tij k\u00ebrkimore \u00ebsht\u00eb n\u00eb kryq\u00ebzimin e dezinformimit, gazetaris\u00eb dhe komunikimit politik. Ai ka sh\u00ebrbyer si redaktor i asociuar n\u00eb revist\u00ebn shkencore, \u201cCentral European Journal of Communication\u201d n\u00eb periudh\u00ebn 2023-2025. G\u00ebrguri \u00ebsht\u00eb i angazhuar si konsulent n\u00eb projekte t\u00eb ndryshme t\u00eb organizatave vendore e nd\u00ebrkomb\u00ebtare. G\u00ebrguri ka botuar kat\u00ebr libra, tre kapituj librash, mbi dhjet\u00eb artikuj shkencor\u00eb dhe dhjet\u00ebra op-ed n\u00eb fush\u00ebn e medias. Disa nga op-ed jan\u00eb p\u00ebrkthyer n\u00eb rreth 10 gjuh\u00eb t\u00eb huaja, duke p\u00ebrfshir\u00eb anglishten, gjermanishten, fr\u00ebngjishten, rusishten, japonishten, etj."}],"_links":{"self":[{"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/posts\/14395","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/users\/667"}],"replies":[{"embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/comments?post=14395"}],"version-history":[{"count":2,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/posts\/14395\/revisions"}],"predecessor-version":[{"id":14398,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/posts\/14395\/revisions\/14398"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/media\/14397"}],"wp:attachment":[{"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/media?parent=14395"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/categories?post=14395"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/tags?post=14395"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/sbunker.org\/sr\/wp-json\/wp\/v2\/ppma_author?post=14395"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}