- Ukufana kwe-Cosine kulinganisa ukuhambisana kwe-vectors ezimbili ngokubala i-cosine ye-engeli ephakathi kwazo, kungawanaki ubukhulu bazo bonke.
- Le metric ibalulekile kwi-AI yesimanje, ivumela ukusesha okunencazelo, izinhlelo zokuncoma ezenzelwe wena, kanye nokucutshungulwa kokushumeka okunezinga eliphezulu.
- Nakuba ukufana kwe-cosine okujwayelekile kuphatha izici njengezizimele, izinguqulo ezithuthukisiwe ezifana ne-Soft Cosine zihlanganisa ubudlelwano bezici ukuze kuthuthukiswe ukunemba.
Wake wazibuza ukuthi iSpotify ibonakala kanjani iyazi kahle ukuthi iyiphi ingoma ezofika endaweni noma ukuthi i-Google iyaqonda kanjani ukuthi umbuzo wakho wokusesha usho okuthile okuqondile noma ungasebenzisi amagama aqondile? Umlingo omningi wenzeka ngemuva kwezigcawu kusetshenziswa ama-vector embeddings . Esikhundleni sokuphatha amagama noma izinto njengombhalo olula, i-AI iwaguqula abe amaphuzu esikhaleni esikhulu, esinobukhulu obuningi, futhi yilapho umqondo wokufana kwe-cosine uqala khona ukuze uqonde konke.
Ngokuyisisekelo, leli qhinga lezibalo livumela amakhompyutha ukuthi athole ukuthi izinto ezimbili “zisondelene” kangakanani ngokubheka i -engeli ephakathi kwama-vector awo . Akunandaba ukuthi i-vector eyodwa inde kakhulu kunenye; kukhathalela kuphela ukuthi ikhomba ohlangothini olufanayo. Kuyinguquko ephelele ye- Natural Language Processing (NLP) kanye nezinjini zokuncoma ngoba igxila encazelweni yesimantiki kunokufanisa izinhlamvu.
Ama-Nuts nama-Bolts okuBalwa
Ukuze uthole indlela yokusebenza kwalokhu, udinga ukuqonda ukuthi yonke idatha—kungaba yigama noma i-movie—imelelwa njengevektha lapho ubukhulu ngabunye buyimfanelo ethile. Ukuze sithole ukufana, silandela izinyathelo ezimbalwa ezithile. Okokuqala, sibala umkhiqizo wamachashazi , ohilela ukuphindaphinda amanani ahambisanayo kusuka kumavektha womabili bese siwafingqa ukuze sibone ukuthi ahambisana kangakanani. Okulandelayo, sithola ubukhulu (noma ubude) bevektha ngayinye ngokuthatha impande yesikwele yesamba sezingxenye zayo eziyisikwele.
Isinyathelo sokugcina ifomula yokufana kwe-cosine yangempela : uthatha lowo mkhiqizo wechashazi bese uwuhlukanisa ngomkhiqizo wobukhulu obubili. Ngokwezibalo, kubukeka sengathi i-Cosine Similarity = (A · B) / (||A|| × ||B||) . Umphumela uba amaphuzu avame ukushintsha phakathi kuka--1 no-1. Amaphuzu angu -1 asho ukuthi ama-vector aqondaniswe kahle , u-0 usho ukuthi angama-orthogonal (awahlobene nhlobo), kanti u--1 usho ukuthi aphikisana kakhulu.
Ukukubeka Embonweni: Amakhosi, Izindlovukazi, nama-apula
Ake senze lokhu kube ngokoqobo. Cabanga nge-LLM icubungula amagama athi “inkosi” kanye “nendlovukazi.” Ngenxa yokuthi la magama avame ukuvela eduze kwamagama anjengokuthi “isihlalo sobukhosi” noma “ubukhosi,” ukushumeka kwawo kwe-vector kuzokhomba cishe ohlangothini olufanayo, okuholela emaphuzu aphezulu okufana kwe-cosine . Manje, faka igama elithi “i-apula” engxubeni. Noma ngabe likumbhalo ofanayo, lihlala namagama anjengokuthi “izithelo” noma “insimu yezithelo,” ngakho-ke i-vector yayo izokhomba ohlangothini oluhlukile kakhulu, okuholela emaphuzu aphansi kakhulu okufana.
Ukuze izinto zihlale zishesha, izinkampani azigcini nje ngokubala lokhu ngokushesha kuyo yonke into. Zisebenzisa izizindalwazi ze-vector . Lawa mathuluzi akhethekile akhelwe ukukhomba ama-vector anezilinganiso eziphakeme, okuvumela ukutholwa okusheshayo kokufana okufanayo ngaphandle kokuskena yonke idatha ngesandla.
Ukudlula Eziyisisekelo: I-Soft Cosine kanye Nezinye Izindlela Zokulinganisa
Ukufana kwe-cosine okujwayelekile kunephutha: kucabanga ukuthi yonke imingcele izimele. Kodwa-ke, ezweni langempela, amagama anjengelithi “play” kanye nelithi “game” ayizilinganiso ezahlukene kodwa ahlobene ngokwesisho . Yilapho i-Soft Cosine Similarity ingena khona. Yethula i-matrix yokufana (ngokuvamile isebenzisa ibanga le-Levenshtein noma i-WordNet) ukuze ilandise ubudlelwano phakathi kwezici, okuvumela imodeli ukuthi ihlanganise imiqondo ngempumelelo enkulu, noma ngabe izici ezisemthethweni zihlukile.
Kufanelekile futhi ukuqhathanisa lokhu nezinye izindlela zokulinganisa ezivamile. Isibonelo, i-Euclidean Distance (L2) ilinganisa ibanga eliqondile phakathi kwamaphuzu amabili, okuyinto enhle kakhulu ekusondeleni kwendawo. I-Manhattan Distance (L1) ibala ibanga lilandela indlela efana negridi. Ngenkathi lezi zilinganisa ibanga eliphelele , ukufana kwe-cosine kugxile kuphela ekuqondisweni . Kukhona futhi i -Dot Product Similarity , ebheka kokubili i-engeli kanye nobukhulu. Uma ulungisa ama-vector akho abe ubude beyunithi, umkhiqizo we-dot kanye nokufana kwe-cosine kuba yinto efanayo ngempumelelo, kodwa umkhiqizo we-dot ushibhile ukubalwa.
Izicelo Eziwusizo Kudatha Yegrafu Nokusesha
Ezweni ledathabheyisi yamagrafu njenge- Amazon Neptune nezinye izinhlelo zamagrafu , ukufana kwe-cosine kusetshenziselwa ukuthola ubumbano phakathi kwamaqembu noma ukuhlonza abasebenzisi abafanayo. Isibonelo, uma unegrafu yabantu kanye nokudla abakuthandayo, ungamelela izintandokazi zabo njengezimpawu zamaphuzu. Ngokusebenzisa i -algorithm yokufana kwe-cosine , ungabeka ohlwini ukuthi yibaphi abasebenzisi abanokunambitheka okufanayo kakhulu, kungakhathaliseki ukuthi umuntu oyedwa ulinganisa izindawo zokudlela eziningi kunomunye.
Izinjini zokusesha zanamuhla nazo ziyasuka ekusesheni okunezincazelo kuphela (njenge-Ctrl+F) ziye ekusesheni kwe-vector . Ngenkathi ukusesha okunezincazelo kuhle kakhulu ekusetshenzisweni kwe-tokenization, kuyaphuthelwa iphuzu lombhalo. Ukusesha kwe-vector kubamba inhloso eyisisekelo . Ezinhlelweni ezifana ne-Elasticsearch, lokhu kwenziwa ngokuguqula imibuzo ibe yi-embeddings nokuthola omakhelwane abaseduze kusetshenziswa amamethrikhi esixoxe ngawo, ngokuvamile kuguqula ububanzi be--1 kuya ku-1 bube amaphuzu amahle ukuze kube ngcono.
Kungakhathaliseki ukuthi wakha umncomi we-movie noma i-ejenti ye-AI eyinkimbinkimbi, ukukhetha isilinganiso sokufana esifanele kuncike ekutheni ubukhulu be-vector bunencazelo yini. Uma ukhathalela kuphela "isihloko" noma "isiqondiso" sedatha, ukufana kwe-cosine kuyindlela engcono kakhulu. Kulabo abadinga ibanga elingakahleleki lesikhala, i-Euclidean iyindlela okufanele uhambe ngayo. Ngokuhlanganisa la mathuluzi ezibalo nama -algorithms asebenzayo okukhomba , singaguqula idatha engahlelekile ibe imephu ehlelekile, engaseshekayo yencazelo yomuntu.
