Nlele na-adịghị mma na ntule dị iche iche nke mkpọtụ
Negative sampling na Noise Contrastive Estimation (NCE) bụ aghụghọ na-eme ka ụdị dị iche iche mụta ọtụtụ okwu na-enweghị ịgbakọ softmax zuru oke.
Nchịkọta
Instead of scoring every possible output, they teach the model to tell real (positive) examples from a handful of fake (negative) ones.
Ime miri emi
Mgbe okwu nwere ọtụtụ narị puku okwu, softmax nkịtị ga-emezigharị okwu ọ bụla maka usoro ọzụzụ ọ bụla - dị oke nwayọ. Noise Contrastive Estimation reframes nsogbu ahụ dị ka ọnụọgụ abụọ nkewa: nyere ebumnuche na ihe nlele ole na ole 'mkpọtụ' ewepụtara site na nkesa ama ama, mụta ịmata ọdịiche dị n'ezie na ụda mkpọtụ, nke na-eweghachite ihe ndị a chọrọ n'ezoghị ọnụ na-enweghị nhazi nke ọma. Nlele na-adịghị mma, nke word2vec's skip-gram na-ewu ewu, bụ nwa nwanne nne nke dị mfe: maka eziokwu ọ bụla (okwu, okwu) jikọtara ya sample k negatives ma zụọ ihe nlereanya ahụ iji kenye akara dị elu na ezigbo ụzọ abụọ na akara dị ala na fakes, na-eji ebumnuche sigmoid. Ha abụọ na-atụgharị nsogbu ọtụtụ klaasị dị oke ọnụ n'ime ọtụtụ ọnụọgụ abụọ dị ọnụ ala, na-eme ka ọzụzụ ntinye buru ibu bara uru. Nhọrọ nke nkesa mkpọtụ (mgbe unigram na-ebuli elu na ike 3/4) na-emetụta àgwà.
Nghọta nka nka
NCE na-eme atụmatụ ihe nlereanya site n'ịkewa data na mkpọtụ, na ka ọnụ ọgụgụ nke nlele mkpọtụ na-etolite, ọ na-enwe ike ịdị nso kacha-yiri na softmax kwesịrị ekwesị. Sampling na-adịghị mma na-atụba usoro nhazi NCE kpamkpam, na-eme ka log σ(akara ziri ezi) + Σ log σ(-akara ọjọọ). Nke ahụ na-eme ya ngwa ngwa mana ọ bụghịzi onye nleba anya njupụta na-agbanwe agbanwe - a na-ege ya ntị maka ịmụta ezigbo ntinye karịa ka enwere ike ịhazi ya. Nlele ihe na-adịghị mma site na nkesa unigram emeziri (ugboro ^ 0.75) na-edozi okwu nkịtị na nke na-adịghị ahụkebe.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke nlele na-adịghị mma na ntule dị iche iche nke mkpọtụ
Isi echiche - mụta site n'iche ọdịiche dị mma megide ihe ndị na-adịghị mma - ugbu a na-akwado nkuzi onwe onye nke ọgbara ọhụrụ yana ihe nnọchiteanya dị iche n'ofe ọhụụ, asụsụ, na nkwanye. Ọrụ n'ọdịnihu na-elekwasị anya na Ngwuputa ihe na-adịghị mma (ịhọrọ ihe ndị na-adịghị mma kama ịbụ ndị na-enweghị ihe ọ bụla), na-eleda anya maka ihe ọjọọ na-adịghị mma, na ịmepụta ihe ọjọọ dị ọnụ ala site na nnukwu ebe nchekwa ma ọ bụ nlele n'ime batch. Ka ụdị na-eto eto, ebumnobi atụpụtara nke ọma ka dị mkpa ebe ọ bụla oghere mmepụta ma ọ bụ ihe ndoro-ndoro anya dị nnukwu, dị ka iweghachite na ndị na-akwado nnukwu.
Mmejuputa n'ezie n'ụwa
word2vec skip-gram nwere ntinye okwu mmụta na-adịghị mma sitere na ijeri token na-enweghị softmax zuru oke.
Ụdị asụsụ n'akụkọ ihe mere eme na-eji NCE zụọ usoro okwu nke narị puku kwuru puku okwu nke ọma.
Nkwanye na usoro iweghachite na-enyocha ihe 'adịghị mma' nke onye ọrụ emekọrịtaghị iji zụọ ụdị ntinye ụlọ elu abụọ.
Ihe eserese na ihe mgbakwunye ihe ọmụma (dịka ọmụmaatụ, imebi isi ma ọ bụ ọdụ atọ) na-eji ihe nlele na-adịghị mma iji mụta mmekọrịta.
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Negative Sampling and Noise Contrastive Estimation quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Ngwuputa Ngwuputa n'ịntanetị na nke siri ike
Ajụjụ a na-ajụkarị
What is Negative Sampling and Noise Contrastive Estimation?
Negative sampling na Noise Contrastive Estimation (NCE) bụ aghụghọ na-eme ka ụdị dị iche iche mụta ọtụtụ okwu na-enweghị ịgbakọ softmax zuru oke. Kama ịdepụta nsonaazụ ọ bụla nwere ike ime, ha na-akụziri ihe nlereanya ka ọ gwa ezigbo ihe atụ sitere na ole na ole adịgboroja (adịghị mma).
Kedu nsogbu nlele na-adịghị mma na NCE na-edozi isi?
Ụzọ abụọ a na-ezere ịhazigharị n'elu nnukwu okwu site na ịmegharị ọzụzụ dị ka ịkpa ókè ọnụọgụ abụọ dị ọnụ ala.
Kedu otu nleba anya nke Noise Contrastive si hazigharịa ọrụ mmụta?
NCE na-azụ ihe nlere anya ịmata ọdịiche dị n'ezie na nlele nke ewepụtara na nkesa mkpọtụ ama ama.
Kedu ihe nlere mere ka nlele adịghị mma maka ịmụta ntinye okwu?
Ewebata nlele na-adịghị mma site na ụdị skip-gram nke word2vec.
Kedu ka nlele adịghị mma si dị iche na NCE zuru oke?
Nlele na-adịghị mma na-eme ka NCE dị mfe site n'iwepụ normalization, ịzụ ahịa nke ọma maka ọsọ na ezigbo ntinye.
Kedu ihe kpatara a na-ejikarị atụpụta ihe na-adịghị mma site na nkesa unigram na ike 3/4?
Ihe ngosi 0.75 na-eme ka nkesa ugboro ugboro ka okwu nkịtị ghara achịkwa ma okwu ndị na-adịghị ahụkebe ka na-apụta.