Teburin Abubuwan Ciki
- 1. Gabatarwa
- 2. Tarihi da Ayyukan da suka Danganta
- 3. Hanyoyin Bincike
- 4. Sakamako da Bincike
- 5. Bayanan Fasaha da Tsarin Lissafi
- 6. Misali: Tambaya don Matakin A1
- 7. Bincike na Asali
- 8. Hanyoyi na Gaba da Aikace-aikace
- 9. Manazarta
1. Gabatarwa
ChatGPT, a matsayin babban samfurin harshe (LLM), yana ba da dama mara misaltuwa don koyon harshe na musamman. Wannan bincike yana nazarin yadda tambayoyin da aka tsara a hankali zasu iya daidaita fitowar ChatGPT da ka'idojin CEFR da EBCL na Sinanci a matsayin harshe na biyu. Yana mai da hankali kan matakan A1, A1+, da A2, kuma yana magance kalubalen rubutun Sinanci ta hanyar sarrafa kalmomi da haruffa.
2. Tarihi da Ayyukan da suka Danganta
2.1 Juyin Halittar Chatbots a Koyon Harshe
Daga ELIZA (1966) zuwa ALICE (1995) da kuma na'urorin AI na zamani, chatbots sun samo asali daga tsarin tushen dokoki zuwa wakilai masu iya daidaitawa. Binciken da Wang (2024) ya yi na tasiri 70 daga nazarce-nazarce 28 ya tabbatar da tasiri mai kyau na chatbots akan aikin koyon harshe. Duk da haka, canjin yanayin da LLMs kamar ChatGPT suka kawo bayan 2020 bai kama a cikin bita na baya ba (Adamopoulou, 2020).
2.2 Tsarin CEFR da EBCL
CEFR yana ba da ma'auni mai matakai shida (A1 zuwa C2) don ƙwarewar harshe. Aikin EBCL musamman yana auna Sinanci, yana bayyana jerin haruffa da kalmomi don kowane mataki. Ga A1, ana sa ran haruffa 150 da kalmomi 300; A1+ yana ƙara haruffa 100; A2 yana nufin haruffa 300 da kalmomi 600. Waɗannan jerin sun zama tushen ƙuntatawa na tambayoyi.
3. Hanyoyin Bincike
3.1 Tsara Tambayoyi don Matakan A1-A2
An tsara tambayoyi don haɗa umarni bayyanannu: "Yi amfani da haruffa kawai daga jerin EBCL A1" da "Ƙayyade kalmomi zuwa kalmomi 300 masu yawan amfani." Tambayoyin kuma sun ƙayyade yanayin tattaunawa (misali, yin odar abinci, gabatar da kai) don tabbatar da dacewa da mahallin.
3.2 Saitin Gwaji
Mun gudanar da gwaje-gwaje na tsari ta amfani da samfuran ChatGPT-3.5 da ChatGPT-4. An gwada kowace tambaya sau 50, kuma an bincika fitowar don bin jerin haruffa, bambancin kalmomi, da daidaiton nahawu. An ayyana maki bin ka'ida $C$ a matsayin rabon haruffa a cikin fitowar da ke cikin jerin EBCL da aka yi niyya.
4. Sakamako da Bincike
4.1 Bin Ka'idojin Kalmomi
Haɗa jerin haruffa bayyanannu a cikin tambayoyi ya ƙara bin ka'ida daga 62% (asali) zuwa 89% ga matakin A1. Ga A1+, bin ka'ida ya kai 84%. Ingantaccen ya kasance mai mahimmanci a kididdiga ($p < 0.01$).
4.2 Maimaita Haruffa
Sarrafa maimaita haruffa (maimaita haruffa a cikin tattaunawa) ya inganta riƙewa. Matsakaicin yawan maimaita haruffa ya ƙaru daga 1.2 zuwa 2.4 a kowane haruffa 100, wanda ya dace da ka'idojin ilmantarwa na maimaita tazara.
5. Bayanan Fasaha da Tsarin Lissafi
An ayyana maki bin ka'ida $C$ kamar haka:
$$C = \frac{N_{\text{target}}}{N_{\text{total}}} \times 100\%$$
inda $N_{\text{target}}$ shine adadin haruffa daga jerin EBCL da aka yi niyya, kuma $N_{\text{total}}$ shine jimillar adadin haruffa a cikin fitowar. Bambancin kalmomi $D$ ana auna shi ta amfani da rabon nau'in-alamu (TTR):
$$D = \frac{V}{N}$$
inda $V$ shine adadin kalmomi na musamman kuma $N$ shine jimillar adadin kalmomi. Tambayoyin da suka fi dacewa sun sami $C > 85\%$ da $D \approx 0.4$ ga matakin A1.
6. Misali: Tambaya don Matakin A1
Tambaya: "Kai malamin Sinanci ne na mai farawa (matakin A1). Yi amfani da haruffa kawai daga jerin EBCL A1: 我, 你, 好, 是, 不, 了, 在, 有, 人, 大, 小, 上, 下, 来, 去, 吃, 喝, 看, 说, 做. Ƙirƙiri ɗan gajeren tattaunawa game da yin odar abinci a gidan cin abinci. Ka sanya jimloli masu sauƙi kuma ka maimaita mahimman haruffa."
Misalin Fitowa: "你好！我吃米饭。你喝什么？我喝水。好，不吃了." (Sannu! Ina cin shinkafa. Me kake sha? Ina sha ruwa. To, na gama ci.)
Wannan fitowar tana amfani da haruffa 100% na manufa kuma tana nuna maimaitawa ta halitta.
7. Bincike na Asali
Mahimmanci: Wannan takarda ita ce gada mai amfani tsakanin tsauraran manhajoji (CEFR/EBCL) da ikon haifar da rikice-rikice na LLMs. Ba wai kawai tana tambaya "Shin ChatGPT zai iya koyar da Sinanci?" amma "Ta yaya za mu tilasta ChatGPT ya koyar da Sinanci daidai?" Wannan canji ne mai mahimmanci daga sabon abu zuwa amfani.
Tsarin Tunani: Marubuta sun ci gaba a hankali daga mahallin tarihi (ELIZA zuwa ChatGPT) zuwa takamaiman matsala (sarrafa fitowar haruffa), sannan zuwa mafita (tsara tambayoyi tare da jerin bayyanannu), kuma a ƙarshe zuwa tabbatarwa na zahiri. Tsarin yana da ƙarfi, kodayake iyakar gwaji ya kasance kunkuntar (A1-A2 kawai).
Ƙarfi da Rauni: Ƙarfin shine hanyar aiki mai amfani—kowane malami zai iya maimaita waɗannan tambayoyi. Raunin shine rashin bayanan sakamako na dogon lokaci na ɗalibai. Shin bin ka'ida mafi girma yana haifar da ingantaccen koyo? Takarda ta ɗauka haka, amma ba ta tabbatar ba. Har ila yau, binciken ya yi watsi da haɗarin ruɗin LLM (misali, ƙirƙirar haruffa). Kamar yadda Bender da sauransu (2021) suka lura a cikin sukar su na LLMs, "stochastic parrots" na iya samar da fitowar da ta dace amma ba daidai ba, wanda ke da haɗari ga masu farawa.
Shawarwari masu Amfani: Ga masu aiki, mahimmin abin da aka samu shine cewa tsara tambayoyi shine shiga tsakani mai rahusa kuma mai tasiri. Ga masu bincike, mataki na gaba shine gudanar da gwaji mai sarrafawa wanda ya kwatanta ChatGPT mai tambaya da mara tambaya don ingantaccen koyo. Fannin yana buƙatar motsawa daga ma'aunin bin ka'ida zuwa ma'aunin ƙwarewa.
8. Hanyoyi na Gaba da Aikace-aikace
Ayyuka na gaba yakamata su faɗaɗa wannan hanya zuwa matakan CEFR mafi girma (B1-C2) kuma su haɗa da abubuwan shigarwa da yawa (misali, gane magana don sautuna). Haɓaka "Laburaren Tambayoyi" don malaman Sinanci, kamar jerin tunani na EBCL, zai ba da damar samun dama ga kowa. Bugu da ƙari, daidaita ƙaramin LLM akan bayanan EBCL na musamman zai iya rage dogaro da tsara tambayoyi. Babban burin shine malami mai daidaitawa wanda zai canza rikitattun haruffa bisa ga aikin ɗalibi, ta amfani da koyon ƙarfafawa daga ra'ayin ɗan adam (RLHF).
9. Manazarta
- Adamopoulou, E., & Moussiades, L. (2020). Chatbots: History, technology, and applications. Machine Learning with Applications, 2, 100006.
- Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of FAccT 2021.
- Li, B., et al. (2024). ChatGPT in education: A systematic review. Computers and Education: Artificial Intelligence, 6, 100215.
- Wang, Y. (2024). Chatbots for language learning: A meta-analysis. Language Learning & Technology, 28(1), 1-25.
- Weizenbaum, J. (1966). ELIZA—a computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36-45.