Ƙwaƙwalwar Fassara (TM) ta kasance ginshiƙi a cikin fassarar inji, tana ba da muhimman fassarorin tunani. Haɗin kai na baya-bayan nan na TM tare da Fassarar Injin Jijiya (NMT) ya nuna gagarumin ci gaba a cikin saitunan da ke da yalwar albarkatu. Duk da haka, wani sabani ya bayyana: NMT da aka ƙarfafa da TM yana yi kyau tare da yalwar bayanai amma ya kasa NMT na asali a cikin yanayin ƙarancin albarkatu. Wannan takarda tana binciken wannan sabani ta hanyar hangen nesa na yiwuwa da ƙa'idar rarraba bambanci da karkacewa, tana gabatar da sabuwar hanyar haɗaɗɗun tsari don magance matsalar bambanci.
Jigon wannan binciken shine sake nazari na asali kan yadda samfuran NMT da aka ƙarfafa da TM suke koyo da kuma gama gari.
Marubutan sun tsara NMT da aka ƙarfafa da TM a matsayin kusantar samfurin maɓalli mai ɓoye, inda ma'anar ƙwaƙwalwar fassara $z$ ke aiki a matsayin maɓalli mai ɓoye. Ana ƙirƙirar yuwuwar fassara a matsayin $P(y|x) \approx \sum_{z \in Z} P(y|x, z)P(z|x)$, inda $Z$ shine saitin zaɓaɓɓun TM masu yuwuwa. Wannan tsari ya nuna cewa aikin samfurin ya dogara da inganci da kwanciyar hankali na ma'anar $z$ da aka samo.
Aiwatar da rarraba karkacewa-bambanci na gargajiya daga ka'idar koyo, kuskuren tsinkaya da ake tsammani $E[(y - \hat{f}(x))^2]$ za a iya raba shi zuwa Karkacewa$^2$, Bambanci, da Hayaniya maras ragewa. Binciken takardar ya bayyana wani muhimmin ciniki:
Wannan babban bambanci yana bayyana sakamakon sabani: a cikin saitunan ƙarancin albarkatu, ƙarar bambanci ya zarce fa'idar ƙananan karkacewa, yana haifar da mafi munin gama gari.
Don rage babban bambanci, marubutan sun ba da shawarar hanyar haɗaɗɗun tsari mai sauƙi. Maimakon dogaro da ma'anar TM guda ɗaya, hanyar tana tattara tsinkaya daga nau'ikan NMT da aka ƙarfafa da TM da yawa ko bambance-bambance. Wata hanyar ƙira ko nauna mai sauƙi tana koyon haɗa waɗannan tsinkaya, yana rage bambancin samfurin gabaɗaya da kuma daidaita sakamako yadda ya kamata. Wannan hanya ba ta da alaƙa da samfurin kuma ana iya amfani da ita a saman tsarin NMT da aka ƙarfafa da TM da ke akwai.
An gudanar da gwaje-gwaje akan ma'auni na yau da kullun kamar JRC-Acquis (Jamusanci→Turanci) a cikin yanayin bayanai daban-daban.
Aiki: JRC-Acquis De→En
Hanyar haɗaɗɗun da aka gabatar ta yi nasarar magance lamarin gazawa, ta sami ci gaba mai dorewa akan NMT na asali da kuma samfurin NMT da aka ƙarfafa da TM na asali. Wannan ya tabbatar da hasashen cewa sarrafa bambanci shine mabuɗi a cikin yanayin ƙarancin bayanai.
Hanyar haɗaɗɗun ta kuma nuna ci gaba a cikin saitunan yalwar albarkatu, yana nuna ƙarfin hali. A cikin yanayin shigar da kuma kunnawa (amfani da wani TM na waje da ba a gani yayin horar da NMT), tasirin rage bambanci na haɗaɗɗun ya zama mai ƙima musamman, yana haifar da aiki mai dogaro.
Mahimmin Fahimta: Babbar gudummawar takardar ba sabon samfurin SOTA ba ce, amma hangen nesa mai kaifi. Ta gano babban bambanci da tsarin ma'ana ya haifar a matsayin ƙafar Achilles na NMT da aka ƙarfafa da TM, musamman a cikin yanayin ƙarancin albarkatu ko hayaniya. Wannan yana motsa tattaunawa daga "shin yana aiki?" zuwa "me ya sa ya kasa wani lokaci?"
Kwararar Ma'ana: Hujja tana da kyau. 1) Tsara matsalar ta hanyar yiwuwa (samfurin maɓalli mai ɓoye). 2) Aiwatar da ƙa'idar ƙididdiga marar iyaka (cinikin karkacewa-bambanci) don bincike. 3) Gano tushen dalili (babban bambanci). 4) Ba da takamaiman magani (haɗaɗɗu don rage bambanci). Ma'ana tana da ƙarfi kuma tana ba da tsarin ƙira don bincika sauran samfuran da aka ƙarfafa da ma'ana.
Ƙarfi da Aibobi: Ƙarfinsa yana cikin bincikensa na asali da kuma sauƙi, ingantacciyar mafita. Hanyar haɗaɗɗun tana da ƙarancin farashi kuma ana iya amfani da ita ko'ina. Duk da haka, aibin takardar shine mayar da hankali kan dabarun. Duk da yake haɗaɗɗun faci ne mai kyau, ba ya sake ƙirƙirar tsarin ma'ana don zama mai ƙarfi sosai. Yana magance alamar (bambanci) maimakon cutar (ma'ana mai hankali ga hayaniya). Idan aka kwatanta da hanyoyi kamar kNN-MT (Khandelwal et al., 2021) waɗanda ke haɗa kai da ma'ajin bayanai a hankali, wannan hanyar ba ta da haɗin kai sosai.
Fahimta Mai Aiki: Ga masu aiki: Yi amfani da haɗaɗɗun idan kuna amfani da NMT da aka ƙarfafa da TM, musamman tare da ƙarancin bayanai. Ga masu bincike: Wannan aikin ya buɗe hanyoyi da yawa. 1) Ma'ana Mai Ƙa'idar Bambanci: Shin za mu iya ƙirƙirar manufofin ma'ana waɗanda ke rage bambancin tsinkaya na gaba a fili? 2) Koyon Zurfin Bayesian don TM: Shin cibiyoyin jijiya na Bayesian, waɗanda ke ƙirƙira rashin tabbas na halitta, za su iya magance matsalar bambanci da kyau? 3) Binciken Tsarin Tsarin: Aiwatar da wannan tsarin bambanci-karkacewa ga sauran dabarun haɓakawa (misali, taswirar ilimi, bayanan harshe ɗaya) don tsinkaya yanayin gazawarsu.
Wannan bincike yana haɗuwa da wani babban yanayi a cikin ML zuwa ƙarfin hali da dogaro. Kamar yadda bincike a cikin hangen nesa ya wuce inganci kawai don yin la'akari da ƙarfin hali na adawa (kamar yadda aka gani a cikin aikin CycleGAN da sauran GANs game da rugujewar yanayi da kwanciyar hankali), wannan takarda tana tura NMT don yin la'akari da kwanciyar hankali a cikin tsarin bayanai. Alama ce ta fagen da ya girma.
Mahimmin fahimtar lissafi ya samo asali ne daga rarraba karkacewa-bambanci. Ga samfurin $\hat{f}(x)$ da aka horar da shi akan samfurin bazuwar na rarraba bayanai, kuskuren murabba'in da ake tsammani akan ma'anar gwaji $x$ shine:
$$ \mathbb{E}[(y - \hat{f}(x))^2] = \text{Bias}(\hat{f}(x))^2 + \text{Var}(\hat{f}(x)) + \sigma^2 $$ Ina:
Takardar ta ƙiyasta cewa ga NMT da aka ƙarfafa da TM, $\text{Var}(\hat{f}_{TM}(x)) > \text{Var}(\hat{f}_{Vanilla}(x))$, yayin da $\text{Bias}(\hat{f}_{TM}(x)) < \text{Bias}(\hat{f}_{Vanilla}(x))$. Hanyar haɗaɗɗun tana rage tasirin bambanci ta hanyar matsakaicin tsinkaya da yawa.
Yanayi: Kamfani ya tura tsarin NMT da aka ƙarfafa da TM don sabon nau'in harshe tare da jimlar jimloli 50,000 kawai (ƙarancin albarkatu).
Matsala: Tura farko ya nuna samfurin NMT da aka ƙarfafa da TM ba shi da kwanciyar hankali—makin BLEU yana canzawa sosai tsakanin rukunin gwaji daban-daban idan aka kwatanta da samfurin asali mai sauƙi.
Aiwatar da Tsarin: