Magance Tambayoyin Cikar Jumloli na ESL ta Hanyar Tsarin Harshe na Jijiya da aka Horar da Shi Tun Da Farko

Table of Contents

1. Gabatarwa

Tambayoyin Cikar Jumla (Sentence Completion - SC) wani muhimmin kayan aiki ne wajen tantance ƙwarewar Turanci a matsayin Harshe na Biyu (ESL). Suna gabatar da jumla mai ɗaya ko fiye da guraben aiki (blank) da kuma tarin zaɓuɓɓukan kalmomi ko jimloli. Yin amfani da na'ura (automation) don warware waɗannan tambayoyin yana ba da fa'idodi masu mahimmanci ga masu koyon harshe (amsa nan take), malamai (tantance ingancin tambaya), da kuma haɓaka tsarin koyarwa na hankali (intelligent tutoring systems).

Hanyoyin lissafi na baya, kamar tsarin harshe na n-gram ko na'urorin harshe na musamman don guraben aiki (blank LMs), suna fuskantar ƙalubale a cikin saitunan ilimi na ainihi: ɓangarorin damuwa masu ruɗani da ƙwararru suka ƙirƙira, buƙatar zurfin ilimin harshe (nahawu, tsarin jumla, ma'anar kalmomi), da kuma bambancin adadin guraben aiki da alamomin kalmomi (tokens) a kowane gurabe.

Wannan aikin yana ba da shawarar tsarin jijiya (neural framework) wanda ke amfani da manyan tsarin harshe da aka horar da su tun da farko don magance waɗannan ƙalubalen, yana nuna babban aiki mai girma akan bayanan ainihi na ESL K-12.

2. Hanyarmu

2.1 Tsarin Matsala

Ana ayyana tambayar SC a matsayin tuple $(q, O)$, inda $q$ shine jumlar da ke da guraben aiki $m$ da aka nuna ta alamomin `[MASK]`, kuma $O = \{o_1, o_2, ..., o_n\}$ shine tarin zaɓuɓɓukan $n$ (yawanci 3-5). Kowane zaɓi $o_i$ jerin alamomin kalmomi ne da aka yi niyya don cika duk guraben aiki tare. Manufar ita ce zaɓar zaɓin $o^* \in O$ wanda ya sa cikakkiyar jumlar ta zama mafi dacewa.

2.2 Tsarin Tsarin (Model)

Jigon hanyar shine tsarin jeri-zuwa-jeri (sequence-to-sequence model) wanda ya dogara ne akan tsarin Transformer, wanda aka horar da shi tun da farko ta amfani da manufar mai gyarar sauti ta atomatik (denoising autoencoder objective) (misali, BART ko T5). An daidaita tsarin (fine-tuned) don aikin SC. Ga wata tambaya $q$ da zaɓi $o_i$, an ba tsarin aikin gina ainihin jumlar, cikakkiyar.

Abin da aka shigar da shi cikin na'urar shigarwa (encoder) shine jerin da aka lalata (tambayar da ke da guraben aiki). Na'urar fitarwa (decoder) tana dogara ne akan wannan kuma dole ne ta samar da ainihin jumlar. Ana shigar da zaɓin $o_i$ cikin guraben aikin $q$ don ƙirƙirar jerin manufa (target sequence) don na'urar fitarwa. Ana ƙididdige aikin tsarin ta hanyar mummunan log-likelihood na samar da jerin manufa bayan an ba da abin shigarwa.

2.3 Horarwa da Hukunce-hukuncen Tsarin

Yayin horarwa, tsarin yana koyon sake gina jumloli daga sigoginsu da aka rufe (masked versions). Don hukunce-hukuncen tsarin (inference), bayan an ba da tambaya $q$ da zaɓuɓɓukanta $O$, tsarin yana ƙididdige maki $s_i$ ga kowane zaɓi $o_i$: $$s_i = -\sum_{t=1}^{T} \log P(w_t | w_{

3. Gwaje-gwaje & Sakamako

3.1 Bayanan Gwaji (Dataset)

An yi amfani da bayanan ainihi da aka tattara daga dandamalin ilimi na kan layi na K-12. Ya ƙunshi dubban tambayoyin SC waɗanda ƙwararrun malaman Turanci suka ƙirƙira don masu koyon ESL na Sinawa. Bayanan suna da tambayoyi tare da guraben aiki 1-3 da ɓangarorin damuwa masu inganci, masu kama da ma'ana.

3.2 Ma'auni na Asali (Baselines)

An kwatanta tsarin da aka ba da shawara da ma'auni na asali masu ƙarfi da yawa:

1. Tsarin Harshe na N-gram (LM): Tsarin ƙididdiga na al'ada da aka horar a kan babban tarin rubutu (corpus).
2. Tsarin Harshe na Gurabe (Blank LM) [Shen et al.]: Na'urar harshe ta musamman mai maimaitawa don cika guraben aiki.
3. Tsarin Harshe da aka Rufe (Masked LM) (misali, BERT): Yin amfani da tsarin harshe da aka horar da shi tun da farko da aka rufe (masked) don ƙididdige yuwuwar alamomin kalmomin zaɓi a wuraren guraben aiki.
4. Tsarin Harshe na Jeri-zuwa-Jeri (ba a horar da shi tun da farko ba): Tsarin Transformer na al'ada da aka horar daga farko akan aikin SC.

3.3 Sakamako Mafi Muhimmanci

Tsarin jeri-zuwa-jeri da aka ba da shawara wanda aka horar da shi tun da farko ya fi duka tsarin ma'auni na asali aiki sosai dangane da daidaiton hasashe akan saitin gwaji da aka keɓe (held-out test set). Babban fa'ida ya samo asali ne daga horonsa na farko akan manyan tarin rubutu, wanda ya sanya shi da zurfin ilimin harshe da ilimin duniya masu mahimmanci don warware ɓangarorin damuwa masu sauƙi. Tsarin jeri-zuwa-jeri shima yana magance guraben aiki da yawa da zaɓuɓɓuka masu alamomin kalmomi da yawa a zahiri.

3.4 Nazarin Daidaito da Tunawa (Precision-Recall)

Takardar ta gudanar da nazarin ciniki na daidaito da tunawa (precision-recall trade-off analysis) don tattauna aiwatarwa a aikace. Ta hanyar daidaita kofa (threshold) na maki don karɓar amsa, ana iya daidaita tsarin don babban daidaito (ba da ra'ayi kawai lokacin da aka yi kwarin gwiwa sosai, rage kurakurai) ko babban tunawa (ƙoƙarin amsa ƙarin tambayoyi, mai yuwuwa tare da ƙarin kurakurai). Wannan yana da mahimmanci ga aikace-aikacen ilimi na ainihin rayuwa inda farashin ra'ayin da ba daidai ba yake da girma.

4. Muhimman Fahimta & Nazari

Mahimman Fahimta: Babban nasarar takardar ba kawai amfani da tsarin da aka horar da shi tun da farko akan sabon aiki ba ne; shine gane cewa manufar gyara sauti ta jeri-zuwa-jeri (sequence-to-sequence denoising objective) kusan cikakkiyar wakilci ce ga tsarin fahimi da ke bayan warware tambayoyin SC. Tsarin ba kawai yana zaɓar kalma ba; yana "cika" jumlar a hankali kuma yana duba don haɗin kai—wani tsari da ke kwatanta ta hanyar sake gina cikakkiyar jumla daga sigar da aka rufe. Wannan hanya ce mafi kyau da ƙarfi fiye da kawai amfani da Tsarin Harshe da aka Rufe (Masked LM) don ƙididdige alamomin kalmomi ɗaya-ɗaya, wanda ya kasa ɗaukar dogaro tsakanin guraben aiki da yawa.

Kwararar Ma'ana: Hujjar tana da sauƙi mai gamsarwa: 1) Tambayoyin ESL na ainihi suna da wahala saboda ɓangarorin damuwa da ƙwararru suka ƙirƙira da ƙaƙƙarfan ƙa'idodin harshe. 2) Hanyoyin gargajiya har ma da na farko na jijiya ba su da fahimtar da za su magance wannan. 3) Manyan Tsarin Harshe da aka horar da su tun da farko, musamman waɗanda aka horar da su tare da manufar gyara sauti (kamar BART ko T5), suna da wannan fahimtar. 4) Don haka, tsara SC a matsayin aikin sake gina jeri ta amfani da waɗannan tsare-tsaren ya kamata ya haifar da sakamako na zamani. Gwaje-gwaje sun tabbatar da wannan kwararar da ƙarfi.

Ƙarfi & Kurakurai: Babban ƙarfi shine kyawun ra'ayi da nasarar aikin. Amfani da bayanan ainihi na K-12, ba tarin ilimi da aka tsabtace ba, yana ƙara babban amincin aiki. Nazarin daidaito da tunawa yana nuna tunani mai zurfi don aiwatarwa. Babban aibi, gama gari ga yawancin takardun AI-a-cikin-ilimi, shine yanayin akwatin baƙi (black box) na mafita. Ba ya ba da ra'ayi mai bayyanawa (explainable)—dalibi ya sami "D daidai ne" amma ba "saboda 'dole' yana nuna tabbataccen hukunci a cikin sashe na farko, kuma 'ba zai iya ba' shine madaidaicin ƙin yarda a cikin sashe na biyu bisa ga shaidar 'ya ƙi launin baƙi'." Kamar yadda aka lura a cikin bita na 2022 "Explainable AI for Education" (XAIED), wannan rashin fassarar yana iyakance amfanin koyarwa kai tsaye. Bugu da ƙari, aikin tsarin yana da alaƙa da bayanan horonsa na farko, waɗanda zasu iya ƙunsar son kai ko rashin ɗaukar wasu ƙirar kuskuren ESL.

Fahimta Mai Aiki: Ga kamfanonin EdTech, wannan binciken shiri ne da aka riga aka yi. Mataki na farko shine daidaita tsarin kamar T5 ko BART akan bankunan tambayoyi na musamman. Duk da haka, ainihin fa'idar gasa ba za ta zo daga daidaito kawai ba amma daga bayyanawa (explainability). Juzu'i na gaba ya kamata ya haɗa dabarun daga AI mai fassara—watakila ta amfani da ma'aunin kulawa don haskaka sassan jumlar da suka fi dacewa da zaɓaɓɓen amsa ko samar da hujjojin harshe na halitta. Na biyu, babban aikace-aikacen wannan fasahar ba a cikin gwaji mai tsanani ba ne amma a cikin aiki da tantancewa na tsari (formative assessment). Haɗa shi cikin dandamalin koyo masu daidaitawa don samar da tambayoyin aiki marasa iyaka, na keɓance (ta hanyar rufe kalmomi a cikin rubutun gaskiya) hanya ce mai ma'ana da ƙima, tafawa daga mai warwarewa zuwa mai samarwa, kamar yadda aka nuna a cikin gabatarwa.

5. Cikakkun Bayanai na Fasaha

Tsarin yana amfani da tsarin na'urar shigarwa da fitarwa (encoder-decoder framework) na tsarin Transformer. Manufar horarwa ta farko tana da mahimmanci. Ga tsarin kamar BART, ana horar da shi ta hanyar lalata rubutu tare da aikin ƙara hayaniya (noising function) na sabani (misali, rufe alamomin kalmomi, canza tsarin jumloli, jujjuya takarda) sannan kuma koyon sake gina ainihin rubutun. Wannan ya sa ya dace da aikin SC, wanda wani nau'i ne mai sarrafawa na lalata rubutu da sake ginawa.

Manufar daidaitawa (fine-tuning objective) ita ce rage asarar giciye (cross-entropy loss) tsakanin rarraba fitarwa na na'urar fitarwa (decoder) da jerin manufa (jumlar da aka cika tare da madaidaicin zaɓi). Ga tarin bayanai (batch), aikin asara shine: $$\mathcal{L} = -\frac{1}{N} \sum_{j=1}^{N} \sum_{t=1}^{T_j} \log P(w_t^{(j)} | w_{

6. Misalin Tsarin Nazari

Yanayi: Tantance tsarin da ake zaɓa don aikin SC.

Aiwatar da Tsarin:

1. Rarraba Aiki (Task Decomposition): Rarraba tambayar SC: Gano adadin guraben aiki, nau'in sashi ko rawar tsarin jumla da ake buƙata ga kowanne, da kuma alaƙar ma'ana tsakanin alamun jumla da madaidaicin amsa.
2. Ƙididdigar Tsarin (Model Scoring): Ga kowane zaɓi, yi amfani da tsarin don ƙididdige makin jerin $s_i$. Misali, ga tambayar "Ya _ zuwa kantin jiya," tare da zaɓuɓɓuka {tafi, ya tafi, yana tafiya}, tsarin zai ƙididdige jerin "Ya tafi zuwa kantin jiya" mafi girma saboda daidaiton lokaci na baya.
3. Nazarin Kuskure (Error Analysis): Idan tsarin ya gaza, bincika yanayin gazawar. Shin ya zaɓi "tafi"? Wannan yana nuna rauni a fahimtar lokacin nahawu. Shin ya zaɓi "yana tafiya"? Wannan yana nuna rauni a cikin yarjejeniyar abu da fi'ili. Wannan nazarin yana jagorantar ƙarin tattara bayanai ko daidaita tsarin.
4. Tantance Ƙarfin ɓangarorin damuwa (Distractor Strength Assessment): Yi amfani da rarraba makin tsarin a kan zaɓuɓɓuka. Babban maki ga madaidaicin amsa da ƙananan maki sosai ga ɓangarorin damuwa yana nuna tambaya mai sauƙi. Idan zaɓuɓɓuka biyu suna da makamantan, manyan maki, yana nuna ɓangarorin damuwa masu inganci, masu ruɗani, waɗanda ke da ƙima don tantancewa na bincike (diagnostic assessment).

7. Aikace-aikace na Gaba & Jagorori

1. Haɗa AI Mai Bayyanawa (Explainable AI - XAI): Jagora mafi mahimmanci shine haɓakawa daga "mai warwarewa na akwatin baƙi" zuwa "malami mai bayyanawa." Tsare-tsaren gaba yakamata su samar da dalilai, haskaka mahimman shaidun jumla, ko ma gano takamaiman ƙa'idar nahawu da ake gwadawa.
2. Samar da ɓangarorin damuwa na Keɓance (Personalized Distractor Generation): Ana iya amfani da tsarin don samar da ɓangarorin damuwa masu ma'ana amma ba daidai ba waɗanda suka dace da tsarin kuskure na gama gari na ɗalibi, ƙirƙirar aiki na keɓance mai ƙarfi (hyper-personalized practice).
3. Samar da Tambaya ta Atomatik (Automated Question Generation - AQG): Juya tsarin. Bayan an ba da rubutu, tsarin zai iya gano mahimman kalmomi don rufewa da samar da ɓangarorin damuwa masu ma'ana, ƙirƙirar sabbin tambayoyin SC ta atomatik don bankunan aiki, haɓaka ƙirƙirar abun ciki da yawa.
4. Ƙara Nau'i-nau'i (Multimodal Extension): Ga ƙananan masu koyo ko takamaiman yanayi, tambayoyin SC na iya haɗawa da hotuna. Aikin gaba zai iya haɗawa da tsare-tsare da aka horar da su tun da farko na nau'i-nau'i (multimodal pre-trained models) (kamar VL-T5) don warwarewa ko samar da tambayoyin da ke haɗa rubutu da alamun gani.
5. Canja Harsuna (Cross-lingual Transfer): Aiwatar da tsarin zuwa wasu harsuna ta hanyar amfani da tsare-tsaren da aka horar da su tun da farko na harsuna da yawa (multilingual pre-trained models) (kamar mT5), taimaka wa masu koyon ESL waɗanda harshensu na farko ba Sinanci ba ne.

8. Nassoshi

1. Liu, Q., Liu, T., Zhao, J., et al. (2021). Solving ESL Sentence Completion Questions via Pre-trained Neural Language Models. arXiv:2107.07122.
2. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of NAACL-HLT.
3. Lewis, M., Liu, Y., Goyal, N., et al. (2020). BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. Proceedings of ACL.
4. Shen, L., Allauzen, C., & Ji, H. (2015). Blank Language Models. Proceedings of EMNLP.
5. Zweig, G., & Burges, C. J. (2012). A Challenge Set for Advancing Language Modeling. Proceedings of the NAACL-HLT Workshop.
6. Holstein, K., McLaren, B. M., & Aleven, V. (2022). Explainable AI for Education (XAIED). In The Handbook of Artificial Intelligence in Education.
7. Raffel, C., Shazeer, N., Roberts, A., et al. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research.