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

Teburin Abubuwan Ciki

1. Gabatarwa

Tambayoyin Cikar Jumloli (Sentence Completion - SC) wata hanya ce ta asali wajen tantance ƙwarewar Turanci a matsayin Harshe na Biyu (ESL). Suna gabatar da jumla mai ɗaya ko fiye da gogewa da kuma jerin kalmomi/jumloli da za a zaɓa, suna gwada fahimtar ɗalibi game da nahawu, tsarin jumla, da ma'anar kalmomi. Sarrafa warware waɗannan tambayoyin ta atomatik yana da matuƙar muhimmanci ga tsarin koyarwa mai hankali, yana ba da ra'ayi nan take, kimanta ingancin tambaya, da kuma samar da kayan aiki na horo.

Hanyoyin gargajiya, kamar tsarin harshe na n-gram, suna fuskantar wahala tare da ƙalubalen tambayoyin ESL na ainihi: ɓangarorin da ƙwararru suka ƙera waɗanda ke daɗa ruɗani, buƙatun ilimin harshe mai zurfi, da kuma bambancin adadin gogewa/kalmomi. Wannan takarda tana gabatar da tsarin jijiya wanda ke amfani da manyan tsarin harshe da aka horar da su tun da farko don magance waɗannan ƙalubalen yadda ya kamata.

2. Hanyarmu

Jigon tsarin da aka gabatar shine daidaita tsarin harshe da aka horar da su tun da farko na jeri-zuwa-jeri (sequence-to-sequence), musamman tsarin Transformer, don aikin Cikar Jumloli (SC).

2.1 Tsarin Matsala

Ana ayyana tambayar SC a matsayin tuple $(q, O)$, inda $q$ shine jumlar da ke da gogewa $k$ wanda aka nuna ta alamar musamman `[MASK]`, kuma $O = \{o_1, o_2, ..., o_m\}$ shine jerin zaɓuɓɓuka $m$ (kowane zaɓi na iya cika gogewa ɗaya ko fiye). Manufar ita ce zaɓar zaɓin $o^* \in O$ wanda ya sa jumlar da aka cika ta fi dacewa.

2.2 Tsarin Tsarin (Model Architecture)

Tsarin ya dogara ne akan tsarin mai shigar da bayanai (encoder) da mai fitar da bayanai (decoder) da aka horar da su tun da farko (misali, BART ko T5). Abin da aka shigar shine jumlar da aka rufe gogewa $q$. Ga kowane zaɓi na yuwu $o_i$, tsarin yana samar da jumla da aka cika ta hanyar maye gurbin alamomin `[MASK]`. Tsarin yana ƙididdige maki kowane cikakken jumla bisa ga yuwuwar samarwarsa ko kuma shugaban mai rarraba (classifier head) da aka daidaita. Ana iya samun makin $S(o_i | q)$ daga mummunan log-yuwuwar (negative log-likelihood) na samar da jerin da aka cika:

$S(o_i | q) = -\sum_{t=1}^{T} \log P(w_t | w_{

inda $w_t$ su ne kalmomin jumlar da aka cika. Ana zaɓar zaɓin da ya fi girma maki (mafi ƙarancin ruɗani - lowest perplexity).

2.3 Dabarar Horarwa

Ana daidaita tsarin akan bayanan tambayoyin SC ta amfani da manufar mai sarrafa kansa mai kawar da hayaniya (denoising autoencoder) da farko, sannan kuma a bi da daidaitawa na musamman ga aikin. Aikin asarar (loss function) yawanci yana haɗa asarar tsarin harshe mai rufewa (masked language modeling loss) da asarar rarrabuwa ta jeri (sequence classification loss) don inganta duka sauƙin jumla da bambancewar zaɓi mai kyau.

3. Gwaje-gwaje & Sakamako

3.1 Bayanan Gwaji (Dataset)

An gudanar da gwaje-gwaje akan bayanan ainihi na tambayoyin ESL SC na K-12 da aka tattara daga dandalin ilimi na kan layi. Bayanan sun ƙunshi dubban tambayoyi tare da ɓangarori masu inganci, waɗanda ƙwararru suka ƙera, suna rufe batutuwa daban-daban na nahawu da ƙamus.

Ƙididdiga na Bayanan Gwaji

Tushe: Dandalin Ilimi na K-12 na Ainihi na Kan layi
Adadin Tambayoyi: Dubu da yawa
Gogewa a Kowane Tambaya: 1 ko fiye
Zaɓuɓɓuka a Kowane Gogewa: 3 zuwa 5
Mai Da Hankali: Nahawu, Tsarin Jumla, Ma'anar Kalmomi

3.2 Ma'auni na Asali (Baselines)

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

N-gram LM: Tsarin harshe na ƙididdiga na gargajiya.
Blank LM [10]: Tsarin harshe mai maimaitawa don cika gogewa.
BERT (Masked LM): Amfani da yuwuwar hasashen alamar da aka rufe ta BERT kai tsaye.
BERT da aka Daidaita (Mai Rarrabawa - Classifier): BERT tare da matakin rarrabawa akan alamar `[CLS]`.

3.3 Sakamako Mafi Muhimmanci

Tsarin jeri-zuwa-jeri da aka horar da shi tun da farko da aka gabatar ya fi duka hanyoyin ma'auni na asali a daidaiton hasashe akan saitin gwaji da aka keɓe. Babban fa'ida ta samo asali ne daga ikonsa na ƙirƙirar haɗin kai na dukan jumla bayan shigarwa, maimakon kawai mahallin gida, yana magance tambayoyi masu gogewa da yawa da zaɓuɓɓukan jumloli yadda ya kamata.

Mahimman Fahimta daga Sakamako

Tsarin da aka horar da su tun da farko (BERT, wanda aka gabatar) sun fi na gargajiya na n-gram LMs girma sosai.
Hanyar samarwa ta jeri-zuwa-jeri ta fi ta rufaffen LM da hanyoyin rarrabawa, musamman ga zaɓuɓɓuka masu kalmomi da yawa.
Tsarin yana nuna ƙarfi a kan ɓangarori masu ruɗani waɗanda ƙwararru suka ƙera.

3.4 Nazarin Daidaito da Tunawa (Precision-Recall)

Takarda ta gabatar da nazarin ciniki tsakanin daidaito da tunawa (precision-recall trade-off), wanda ke da mahimmanci ga aiwatarwa a duniyar ainihi. Ta hanyar daidaita bakin kofa na maki don karɓar amsa, ana iya daidaita tsarin don yanayin babban daidaito (mai ra'ayin mazan jiya, yana amsawa kawai lokacin da ya tabbata sosai) ko babban tunawa (yana ƙoƙarin amsa ƙarin tambayoyi). Wannan sassauci yana da mahimmanci ga tsarin koyo mai daidaitawa inda kimanta tabbaci ke da muhimmanci.

4. Nazarin Fasaha & Fahimta

Fahimta ta Asali: Wannan takarda ba game da sabon tsari ba ce; ta zama darasi mai zurfi a cikin injiniyan AI mai aiki. Marubutan sun gano daidai cewa ƙarfin tilasta na zamani na LMs da aka horar da su tun da farko, musamman tsarin jeri-zuwa-jeri kamar BART ko T5, shine kayan aiki mafi inganci ga matsala mai rikitarwa, iyakancewa, amma mai wadatar ma'ana ta cikar jumloli na ESL. Ainihin ƙirƙira yana cikin tsarawa da dabarar daidaitawa don yanki na ilimi na musamman.

Kwararar Hankali: Hankalin yana da sauƙi mai gamsarwa: 1) Tambayoyin ESL SC suna da wahala saboda ɓangarori masu matakin ƙwararru da ƙaƙƙarfan iyakoki. 2) LMs da aka horar da su tun da farko suna da ilimin duniya da na harshe mai yawa. 3) Don haka, daidaita LM mai ƙarfi, mai amfani da gabaɗaya (tsarin seq2seq) akan bayanan yanki na musamman don warware aikin. Sakamakon gwaje-gwaje ya tabbatar da wannan tsari yadda ya kamata, yana nuna fifikon hanyar seq2seq akan rufaffen LMs kawai (kamar BERT) waɗanda ke fuskantar wahala tare da haɗin kai na kalmomi da yawa.

Ƙarfi & Kurakurai: Babban ƙarfi shine aikace-aikacen kai tsaye na NLP na zamani zuwa matsala ta ilimi ta ainihi, mai tasiri tare da ingantaccen kimantawa. Amfani da bayanan ainihi na K-12 ya ƙara gaskiya sosai, kamar yadda aka lura a cikin wallafe-wallafen hako bayanan ilimi (misali, aikin daga Ƙungiyar Hako Bayanan Ilimi ta Duniya). Duk da haka, kuskuren takarda shine na gama gari a cikin AI da aka yi amfani da shi: rashin bayyana a cikin "yaya". Duk da yake ya ambaci daidaita mai sarrafa kansa mai kawar da hayaniya, cikakkun bayanai game da ainihin ayyukan asara, ma'auni masu girma (hyperparameters), da dabarun haɓaka bayanai don samar da samfuran horo na `[MASK]` suna da ƙaranci. Wannan yana sa kwafi ya yi wahala. Bugu da ƙari, bai yi nazari mai zurfi ba dalilin da ya sa tsarin ya kasa a kan wasu tambayoyi—wani mataki mai mahimmanci ga tsarin bincike na ilimi. Kwatanta wannan da ƙoƙarin fassara a cikin tsarin kamar CycleGAN, inda ake amfani da taswirar hankali ko nunin fasali don bayyana sakamako.

Fahimta Mai Aiki: Ga kamfanonin EdTech, abin da za a ɗauka a bayyane yake: daina gina tsarin tushen ƙa'ida ko ƙididdiga masu sauƙi don tantance harshe. Dawowar zuba jari (ROI) yana cikin amfani da kuma daidaita tsarin tushe a hankali. Nazarin daidaito da tunawa yana ba da tsari don haɗa samfur: gina tsarin yanayi biyu inda yanayin babban daidaito yana taimakawa tantancewa na yau da kullun, kuma yanayin babban tunawa yana motsa aikin horo na bincike. Mataki na gaba, kamar yadda aka gani a cikin binciken tsarin koyarwa mai ci gaba (misali, dandamalin Koyon Carnegie), shine tsawaita wannan daga "ƙididdigar amsa" zuwa "nazarin ɓangarori" da "samar da alamun bayani na sirri," ta amfani da makin tabbacin tsarin da wakilcin ciki don gano takamaiman kuskuren ɗalibi.

5. Misalin Tsarin Nazari

Yanayi: Nazarin dalilin da ya sa tsarin zai iya kasa akan wata tambaya ta SC.

Tambaya: "Ita _____ zuwa kantin jiya ta sayi madara."
Zaɓuɓɓuka: (A) tafi (B) tafiya (C) ta tafi (D) tana tafiya

Aikace-aikacen Tsarin:

Wakilcin Shigarwa: Tsarin yana karɓar: "Ita [MASK] zuwa kantin jiya ta sayi madara."
Ƙididdigar Zaɓi: Ga kowane zaɓi, tsarin yana samarwa/cika jumlar kuma yana ƙididdige maki.
- Maki("ta tafi") = -log P("Ita ta tafi zuwa kantin...") // Ya kamata ya zama mafi ƙanƙanta (mafi kyau).
- Maki("tafiya") = -log P("Ita tafiya zuwa kantin jiya...") // Ya fi girma saboda rashin daidaituwar lokaci.
Binciken Rashin Nasarar: Idan tsarin ya zaɓi "tafiya" ba daidai ba, muna bincika:
- Bambance-bambancen Bayanai (Data Bias): Shin "tafiya" ta yi yawa sosai a cikin bayanan horo a irin wannan mahallin?
- Taga Mahalli (Context Window): Shin tsarin ya kasa ba da fifiko isa ga alamar lokaci "jiya"?
- Ƙarfin ɓangaro (Distractor Strength): Shin "tafiya" ƙwararren ɓangaro ne saboda daidai ne a nahawu ga batun "Ita" a sarari?
Gyara: Ƙara bayanan horo da ƙarin misalai da ke jaddada daidaitawar lokaci-bayan lokaci da fi'ili, ko kuma daidaita manufar daidaitawa don hukunta rashin daidaituwar lokaci da ƙarfi.

Wannan tsarin nazari ya wuce ma'auni na daidaito kawai zuwa ingantaccen tsari.

6. Aikace-aikace na Gaba & Hanyoyi

Hanyoyin Koyo na Sirri: Amfani da tabbacin tsarin da tsarin kuskure don gano takamaiman raunin nahawu na ɗalibi da ba da shawarar ayyukan horo da aka yi niyya.
Samar da Tambaya ta Atomatik: Juya tsarin don samar da sabbin tambayoyin SC masu inganci tare da ɓangarori masu ma'ana ta hanyar rufe kalmomi a cikin jumloli na ainihi da amfani da tsarin don ba da madadin, kama da hanyoyin da aka bincika a arXiv:2005.05909.
Haɗin Nau'i-nau'i (Multimodal Integration): Haɗa tsarin tushen rubutu tare da gane magana don tantance cikar jumloli da aka faɗa, yana ba da kimanta ƙwarewar harshe gabaɗaya.
AI Mai Bayyanawa don Ilimi (XAI-Ed): Haɓaka dabarun sanya "tunani" na tsarin ya zama bayyane—misali, haskaka waɗanne kalmomi a cikin jumla suka kasance mahimmanci don ƙin ɓangaro—don gina amana da samar da ra'ayi mai zurfi.
Canja Harsuna (Cross-lingual Transfer): Aiwatar da tsarin ga tambayoyin SC na wasu harsuna, ta amfani da tsarin da aka horar da su tun da farko na harsuna da yawa kamar mT5 ko mBART.

7. Nassoshi

Zweig, G., et al. (2012). SAT Sentence Completion. Microsoft Research Tech Report.
Shen, L., et al. (2015). Blank Language Model. EMNLP.
Donahue, J., et al. (2020). Pre-training with Masked Text. NeurIPS.
Liu, Y., et al. (2019). RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv:1907.11692.
Lewis, M., et al. (2020). BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. ACL.
Raffel, C., et al. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. JMLR.
Koedinger, K.R., et al. (2012). The Knowledge-Learning-Instruction Framework: Bridging the Science-Practice Chasm to Enhance Robust Student Learning. Cognitive Science.
Zhu, J.Y., et al. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. ICCV. (An ambata a matsayin misalin ƙoƙarin fassara).
International Educational Data Mining Society (IEDMS). Resources on Real-world Educational Datasets. https://educationaldatamining.org/