Koyon Harshe Na Biyu na Tsarin Harshe na Jijiyoyi: Nazarin Harshe na Canja Harshe Tsakanin Harsuna

Table of Contents

1. Gabatarwa & Bayyani

Wannan bincike yana bincika tsarin Koyon Harshe Na Biyu (L2) a cikin Tsarin Harshe na Jijiyoyi (LMs), yana mai da hankali daga binciken da aka saba yi na Koyon Harshe Na Farko (L1). Babbar tambaya ita ce yadda ilimin L1 na baya yake tasiri akan inganci da yanayin koyon ilimin nahawu a cikin sabon harshe (L2). Binciken ya tsara yanayin koyon L2 mai kama da na ɗan adam don tsarin harshe masu amfani da harsuna biyu, yana horar da su a kan L1 (Faransanci, Jamusanci, Rashanci, Japananci) kafin a fallasa su da Turanci (L2). Babban ma'aunin kimantawa shine fahimtar harshe a cikin L2, wanda aka tantance ta hanyar gwaje-gwajen yanke hukunci na nahawu, da nufin fayyace abubuwan da (ba) suke kama da na ɗan adam na canja harshe na LM.

2. Tsarin Gwaji & Hanyoyin Bincike

Hanyar bincike ta bi tsari mai matakai uku da aka tsara don yin kama da koyon L2 na ɗan adam:

1. Horar da L1 (Koyon Harshe Na Farko): Ana horar da tsarin harshe mai rufe fuska (misali, tsarin BERT) daga farko akan tarin bayanai na harshe guda (L1).
2. Horar da L2 (Koyon Harshe Na Biyu): Tsarin da aka horar da L1 ana ci gaba da horar da shi akan bayanan Turanci a ƙarƙashin ƙayyadaddun yanayi, masu ƙarancin bayanai, don kwaikwayon koyon L2 mai ƙarancin albarkatu.
3. Kimantawa & Nazari: Ilimin L2 da tsarin ya samu ana bincika shi ta amfani da ma'aunin BLiMP, jerin gwaje-gwaje don kimanta iyawar haɗin kalmomi ta hanyar yanke hukunci na karɓuwar nahawu.

Mahimman masu canji da aka sarrafa sun haɗa da zaɓin L1 (wanda ke bambanta da nisan nau'in harshe daga Turanci) da tsarin bayanan horar da L2 (rubutu guda ɗaya vs. rubutun da suka yi daidai).

3. Ra'ayoyin Karkata a Hanyoyin Horar da Harshe Na Biyu (L2)

Gwaje-gwajen farko sun kwatanta saitunan bayanan L2 daban-daban don fahimtar ra'ayoyin karkata na tsarin. Wani muhimmin bincike shine cewa horarwa akan biyun fassarar L1-L2 ya rage saurin koyon nahawun L2 idan aka kwatanta da horarwa akan rubutun L2 guda ɗaya da aka gabatar a tsaka-tsaki (misali, kowane zamani biyu). Wannan yana nuna cewa don manufar musamman ta samun tsarin nahawun L2, fallasa kai tsaye ga tsarin L2 ya fi inganci fiye da koyo ta hanyar daidaita fassara a cikin wannan tsari, yana nuna bambanci tsakanin hanyoyin koyo na tsarin da na ɗan adam inda bayanan da suka yi daidai na iya zama mafi fa'ida.

4. Tasirin Horar da Harshe Na Farko (L1) akan Koyon Nahawun Harshe Na Biyu (L2)

4.1 Ilimin L1 Yana Haɓaka Fahimtar L2

Binciken ya gano cewa tsare-tsare tare da horar da L1 sun nuna mafi kyawun fahimtar harshe a cikin L2 idan aka kwatanta da tsare-tsaren da aka horar da L2 daga farko tare da jimlar bayanai daidai. Wannan yana nuna cewa ilimin harshe na baya, ko da daga wani harshe daban, yana ba da ra'ayi mai fa'ida don samun ƙa'idodin tsarin sabon harshe.

4.2 Zaɓin L1 Yana Tasiri Akan Ingantacciyar Canja Harshe

Kusancin nau'in harshe na L1 zuwa Turanci (L2) ya yi tasiri sosai akan ingantacciyar canja harshe. Tsare-tsare tare da Faransanci ko Jamusanci a matsayin L1 (harsunan Jamusanci/Romance mafi kusa da Turanci) sun sami mafi kyawun fahimtar L2 fiye da waɗanda ke da Rashanci ko Japananci (harsunan Slavic da Japananci, mafi nisa). Wannan ya yi daidai da binciken koyon harshe na biyu na ɗan adam, kamar waɗanda Chiswick da Miller (2004) suka ambata, waɗanda ke rarraba wahalar canja harshe bisa ga nisan harshe.

4.3 Tasiri Daban-daban akan Nau'ikan Nahawu

Fa'idar daga horar da L1 ba ta kasance iri ɗaya ba a duk abubuwan nahawu. Ribar ta fi girma ga abubuwan ilimin siffofi da haɗin kalmomi (misali, yarjejeniya tsakanin mai magana da fi'ili, tsibiran haɗin kalmomi) idan aka kwatanta da abubuwan ma'ana da haɗin ma'ana da nahawu (misali, iyakar ƙididdiga, tilastawa). Wannan yana nuna ilimin L1 da farko yana taimakawa fannoni na tsari na harshe, maimakon abubuwan da suka shafi ma'ana ko haɗin kai.

5. Nazarin Tsarin Koyon Harshe Na Biyu (L2)

5.1 Ci Gaba da Rashin Ingantaccin Amfani da Bayanai

Nazarin lanƙwan koyo ya bayyana cewa samun ilimin L2 a cikin waɗannan tsare-tsaren yana da rashin ingantaccin amfani da bayanai. Muhimman haɓaka na fahimta sau da yawa yana buƙatar tsarin ya ga duk ƙarancin bayanan L2 sau da yawa (misali, zamani 50-100). Bugu da ƙari, tsarin ya nuna katangar tsangwama ko lalacewar ilimi a yankin L1 yayin horar da L2, yana nuna tashin hankali tsakanin samun sabon ilimin harshe da riƙe na tsoho—ƙalubalen da aka lura a cikin wallafe-wallafen ci gaba da koyo don hanyoyin sadarwa na jijiyoyi.

6. Cikakkiyar Fahimta & Ra'ayi na Mai Bincike

Cikakkiyar Fahimta: Wannan takarda tana ba da gaskiya mai mahimmanci, wacce ake yawan yin watsi da ita: tsarin harshe na zamani ba su zama soso na harsuna da yawa ba ba. "Ƙwarewar L2" nasu tana da babbar jinginar ta hanyar "tarbiyyar L1" da bashin gine-ginen horon su na farko. Gano cewa bayanan da suka yi daidai na iya hana samun haɗin kalmomi wani abu ne mai ban mamaki, yana ƙalubalantar manufar masana'antu na "ƙarin bayanai, kowane bayanai" na AI mai harsuna da yawa. Yana bayyana rashin daidaituwa na asali tsakanin manufar fassara (taswira) da manufar koyon harshe (shigar da tsari).

Tsarin Hankali: Hankalin binciken yana da tsafta kuma yana da wahayi daga ilimin halin ɗan adam: 1) Kafa tushen harshe (L1), 2) Gabatar da ƙwaƙƙwaran L2 mai sarrafawa, 3) Bincika tasirin canja harshe. Wannan yana kama da hanyoyin bincike daga binciken SLA na ɗan adam, yana ba da damar kwatanta daidai (ko da yake ba cikakke ba) tsakanin ɗan adam da koyon na'ura. Amfani da BLiMP yana ba da tabarma mai ƙima, mai cike da ka'ida, yana wucewa daga ma'auni gabaɗaya kamar rudani, wanda sau da yawa yana rufe hanyoyin gazawa masu zurfi.

Ƙarfi & Aibobi: Ƙarfinsa shine ƙaƙƙarfan ƙirar gwajinsa mai ƙuntatawa da kuma mai da hankali kan fahimtar harshe maimakon aikin aiki. Yana tambaya "me suke koya?" ba kawai "yaya suke yi?" ba. Babban aibi, duk da haka, shine sikelin. Gwada ƙananan tsare-tsare akan ƙarancin bayanai, yayin da yake da kyau don sarrafawa, yana barin babbar alamar tambaya akan ko waɗannan binciken sun kai ga tsare-tsaren zamani na sigogi 100B+ da aka horar da su akan tarin bayanai na tiriliyan. Shin "Fa'idar L1" ta tsaya tsayin daka ko ma ta juyawa? Manta mai ban tsoro na L1 kuma ba a bincika shi sosai ba—wannan ba damuwa na ilimi kawai ba ne amma aibi mai mahimmanci ga tsarin harsuna da yawa na zahiri waɗanda dole su kiyaye duk harsuna.

Fahimta Mai Aiki: Ga masu haɓaka AI, wannan umarni ne don horon farko na dabarun. Kar ku yi tunani kawai "harsuna da yawa"; yi tunani "harsuna da yawa masu tsari." Zaɓin harshe(n) na tushe shine babban siga mai tasiri mai zurfi. Don tsara bayanai, jinkirin bayanan da suka yi daidai yana nuna buƙatar tsarin horo na matakai—watakila nutsewar L2 guda ɗaya da farko don haɗin kalmomi, sannan bayanan da suka yi daidai don daidaita ma'ana. A ƙarshe, fannin dole ne ya haɓaka jerin kimantawa waɗanda, kamar BLiMP, za su iya gano yadda tsare-tsaren suke da harsuna da yawa, ba kawai ko suna da su ba ba. Neman ba na mai yaren harsuna da yawa ba ne, amma don hankali mai harsuna da yawa cikin na'ura.

7. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Tsarin tsakiya ya dogara ne akan tsarin Transformer da manufar Koyon Harshe Mai Rufe Fuska (MLM). Yayin horar da L1, tsarin yana koyo ta hanyar annabta alamomin da aka rufe bazuwar $w_t$ a cikin jerin $W = (w_1, ..., w_n)$, yana ƙara yuwuwar: $$P(w_t | W_{\backslash t}; \theta)$$ inda $\theta$ suke sigogin tsarin kuma $W_{\backslash t}$ shine jerin tare da alamar a matsayi $t$ da aka rufe.

Yayin samun L2, tsarin, yanzu tare da sigogi $\theta_{L1}$ daga horar da L1, ana daidaita shi akan bayanan L2 $D_{L2}$ ta hanyar rage asarar giciye: $$\mathcal{L}_{L2} = -\sum_{(W) \in D_{L2}} \sum_{t \in M} \log P(w_t | W_{\backslash t}; \theta)$$ inda $M$ shine saitin wuraren da aka rufe. Binciken tsakiya ya ƙunshi kwatanta aikin tsare-tsaren da aka fara da $\theta_{L1}$ da tsare-tsaren da aka fara da bazuwar ($\theta_{random}$) bayan horarwa akan $D_{L2}$, ana auna ribar canja harshe $\Delta G = G(\theta_{L1}) - G(\theta_{random})$, inda $G$ shine daidaito akan ma'aunin BLiMP.

8. Sakamakon Gwaji & Fassarar Jadawali

Yayin da abin da aka ba da na PDF bai ƙunshi takamaiman jadawali ba, sakamakon da aka bayyana ana iya fassara su ta hanyar gani:

• Jadawali 1: Daidaiton L2 vs. Zamani na Horar da L2 don L1s daban-daban. Wannan zai nuna lanƙwan koyo guda huɗu (Fr, Ge, Ru, Ja). Lanƙwan Faransanci da Jamusanci za su tashi da sauri kuma zuwa mafi girman matakin ƙarshe fiye da na Rashanci da Japananci, suna nuna tasirin "nisan harshe". Duk lanƙwan za su nuna dogon hawan, suna nuna rashin ingantaccin amfani da bayanai.
• Jadawali 2: Ribar Aiki daga Horar da L1 ta Nau'in Nahawu. Jadawali mai sanduna tare da rukuni: Ilimin Siffofi, Haɗin Kalmomi, Ma'ana, Haɗin Ma'ana da Nahawu. Sandunan Ilimin Siffofi da Haɗin Kalmomi za su fi girma sosai fiye da na Ma'ana da Haɗin Ma'ana da Nahawu, suna tabbatar da tasirin canja harshe daban-daban ta hanyar gani.
• Jadawali 3: Ƙwarewar L1 vs. Zamani na Horar da L2. Jadawalin layi mai yiwuwa yana nuna raguwar yanayin daidaiton kimantawa na L1 yayin da horar da L2 ke ci gaba, yana nuna al'amarin katangar tsangwama ko lalacewar ilimi a yankin L1.

Mahimman abin da za a ɗauka daga waɗannan sakamakon da ake zato shine cewa canja harshe yana da kyau amma zaɓaɓɓe kuma mara inganci, kuma yana zuwa da yuwuwar farashi ga ilimin da aka samu a baya.

9. Tsarin Nazari: Nazarin Lamari

Yanayi: Nazarin samun L2 na tsarin Turanci (L2) da aka horar da shi a kan Japananci (L1).

Aiwatar da Tsarin:

1. Hasashe: Saboda babban nisan nau'in harshe (Tsari na Batun-Abu-Fi'ili vs. Batun-Fi'ili-Abu, hadaddun ɓangarorin bayan magana vs. karin magana), tsarin zai nuna raunin canja harshe akan abubuwan haɗin kalmomi na Turanci, musamman waɗanda suka haɗa da tsarin kalmomi (misali, Yarjejeniyar Anaphor a cikin BLiMP), idan aka kwatanta da tsarin da aka horar da shi a kan Jamusanci.
2. Bincike: Bayan horar da L2, gudanar da gwaje-gwajen BLiMP masu dacewa (misali, "Yarjejeniyar Anaphor," "Tsarin Hujja," "Haɗawa") ga duka tsarin Ja->En da De->En.
3. Ma'auni: Lissafa Ingantacciyar Canja Harshe ta Dangantaka (RTE): $RTE = (Acc_{L1} - Acc_{No-L1}) / Acc_{No-L1}$, inda $Acc_{No-L1}$ shine daidaiton tsarin da aka horar da Turanci daga farko.
4. Hasashe: RTE don tsarin Ja->En akan gwaje-gwajen haɗin kalmomi masu kula da tsarin kalmomi zai yi ƙasa fiye da na tsarin De->En, kuma mai yiwuwa ya yi ƙasa fiye da nasa RTE akan gwaje-gwajen ilimin siffofi (misali, jujjuyawar lokaci na baya).
5. Fassara: Wannan lamari zai nuna cewa ra'ayin karkata daga L1 ba "iyawar koyon harshe" gabaɗaya ba ce amma an siffanta ta ta takamaiman kaddarorin tsarin L1, wanda zai iya sauƙaƙa ko hana samun takamaiman gine-ginen L2.

10. Aikace-aikace na Gaba & Hanyoyin Bincike

• Ingantaccin Horon Tsarin Harsuna da Yawa: Ƙirƙirar dabarun koyon tsarin karatu inda ake zaɓar L1s da oda su bisa siffofi na nau'in harshe don ingantaccin tsarin samun saitin harsuna da aka yi niyya.
• Kayan Aikin Koyon Harshe Na Musamman: Malamai na AI waɗanda ke gano L1 na ɗalibi na ɗan adam kuma suka annabta wuraren wahala a cikin L2 bisa ga tsarin canja harshe na tsari, suna ba da atisaye masu niyya.
• Rage Manta Mai Ban Tsoro: Haɓaka algorithms na ci gaba da koyo don LMs waɗanda ke ba da damar samun kwanciyar hankali na harsuna da yawa ba tare da lalata aikin a cikin waɗanda aka koya a baya ba, tare da wahayi daga ƙarfafa nauyin roba ko hanyoyin sadarwa masu ci gaba.
• Gano Nau'in Harshe: Yin amfani da "ingantacciyar canja harshe" tsakanin harsuna a matsayin ma'auni mai ƙima, mai dogaro da bayanai don haɓaka nau'in harshe na al'ada, mai yiwuwa yana bayyana sabbin iyalai ko alaƙar harshe.
• Ƙaddamar da Harshe Mai Ƙarancin Albarkatu: Zaɓar da dabara na "harshen tsari" mai yawan albarkatu (L1) wanda zai fi amfana ga samun takamaiman harshe mai ƙarancin albarkatu (L2), yana rage bayanan da ake buƙata don aiwatar da ingantaccin tsari sosai.

11. Nassoshi

1. Oba, M., Kuribayashi, T., Ouchi, H., & Watanabe, T. (2023). Second Language Acquisition of Neural Language Models. arXiv preprint arXiv:2306.02920.
2. Chiswick, B. R., & Miller, P. W. (2004). Linguistic Distance: A Quantitative Measure of the Distance Between English and Other Languages. Journal of Multilingual and Multicultural Development.
3. Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems.
4. Papadimitriou, I., & Jurafsky, D. (2020). Pretraining on Non-English Data Improves English Syntax. Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics.
5. Warstadt, A., et al. (2020). BLiMP: The Benchmark of Linguistic Minimal Pairs. Proceedings of the Society for Computation in Linguistics.
6. Kirkpatrick, J., et al. (2017). Overcoming Catastrophic Forgetting in Neural Networks. Proceedings of the National Academy of Sciences. (Tushen waje akan ci gaba da koyo).
7. Ruder, S. (2021). Challenges and Opportunities in NLP Benchmarking. The Gradient. (Ra'ayi na waje akan kimantawa).