Bayyana Kansa a cikin Na'urorin AI na Zamantakewa: Hanyar Haɗakar AI mai Ƙirƙira da Ilimi

1. Gabatarwa & Bayyani

Wannan takarda tana magance wata ƙalubale mai mahimmanci a cikin aiwarta na na'urorin AI na zamantakewa, musamman a cikin fagage masu mahimmanci kamar ilmantarwa ta kan layi. Marubutan sun mai da hankali kan SAMI (Hulɗar Tsakanin Mataimakan AI na Zamantakewa), wani mataimakin AI da aka ƙera don haɓaka alaƙar zamantakewa tsakanin ɗalibai a cikin manyan azuzuwan kan layi. Duk da cewa irin waɗannan na'urorin na iya rage sanannen matsalar ƙarancin kasancewar zamantakewa, suna haifar da sabuwar matsala: rashin bayyana. Daliban da ke hulɗa da SAMI a zahiri suna tambayar yaya da dalilin da ya sa yake ba da takamaiman shawarwari (misali, haɗa ɗalibai biyu). Babbar tambayar bincike ita ce: Ta yaya mataimakin AI na zamantakewa zai iya ba da bayyanai masu bayyana, masu fahimta game da tunaninsa na ciki don gina amincin mai amfani?

Magani da aka gabatar shine sabuwar fasahar bayyana kai. An tsara wannan a matsayin tsarin amsa tambayoyi ta hanyar yare na halitta inda na'urar ke bincika ciki akan tsarin samfurin kai na burinta, iliminta, da hanyoyinta. Babban ƙirƙira shine tsarin gine-gine na haɗakar da ke haɗa tsarin bayyanai masu bayyana na AI mai dogaro da ilimi tare da ƙarfin samar da yare na halitta na AI mai ƙirƙira (musamman, ChatGPT).

2. Tsarin Aiki & Tsarin Gine-gine

Tsarin bayyana kai wani tsari ne mai matakai da yawa da aka ƙera don fassara dabaru na ciki na na'urar zuwa labarai masu sauƙin fahimta ga mai amfani.

2.1. Samfurin Kai: Tsarin Aiki, Hanyar, Ilimi (TMK)

Tushen bayyana kai shine samfurin kai mai ƙididdigewa. Marubutan sun daidaita tsarin TMK, inda aikin na'urar ya rabu zuwa:

Ayyuka (T): Manyan manufofi (misali, "Ƙara haɗin kai na zamantakewa").
Hanyoyi (M): Hanyoyin ko algorithms don cim ma ayyuka (misali, "Nemo ɗalibai masu ra'ayoyi iri ɗaya").
Ilimi (K): Bayanai ko imani da hanyoyi ke amfani da su (misali, "Sha'awar ɗalibi A: Koyon Injina").

Wani daidaitawa mai mahimmanci shine wakilcin abubuwan TMK ba a matsayin shawarwari na hankali ba amma a matsayin gajerun bayyanai na yare na halitta. Wannan yana haɗa gibin tsakanin tsarin alamar na'urar da sararin yare na samfurin ƙirƙira.

2.2. Haɗakar Bayyanar: Haɗa AI mai Dogaro da Ilimi da AI mai ƙirƙira

Tsarin samar da bayyanai ya ƙunshi matakai biyar masu mahimmanci:

Shigarwa: Mai amfani ya gabatar da tambaya ta yare na halitta (misali, "Me ya sa ka haɗa ni da Alex?").
Dawowa: Ana yin binciken kama-karya tsakanin tambayar da bayanan Turanci a cikin samfurin kai na TMK don gano ɓangarorin ilimin kai masu dacewa.
Bincika Ciki: Ana amfani da tsarin Jerin Tunani (CoT) don "tafiya" ta ɓangarorin da suka dace na samfurin TMK, sake gina matakan hankali da na'urar ta ɗauka.
Ƙirƙira: Ana tsara sakamakon CoT da aka tsara da ɓangarorin ilimin da aka dawo da su zuwa wani umarni don babban samfurin yare (ChatGPT).
Fitowa: ChatGPT yana samar da bayyani mai ma'ana, yare na halitta wanda aka mayar da shi ga mai amfani.

Wannan hanyar haɗakar tana amfani da daidaito da tabbatarwa na samfurin kai mai dogaro da ilimi don kafa bayyanin, yayin amfani da AI mai ƙirƙira don sauƙi da daidaitawa na labarin ƙarshe.

3. Aiwarta ta Fasaha & Cikakkun Bayanai

3.1. Tsarin Lissafi na Binciken Kama-karya

Matakin dawowa yana da mahimmanci don inganci. Idan aka ba da tambayar mai amfani $q$ da saitin $N$ na vectors na bayanin TMK $\{d_1, d_2, ..., d_N\}$ (misali, daga samfurin saka kalmomi kamar Sentence-BERT), tsarin yana dawo da manyan-$k$ bayanan da suka fi dacewa. Ana ƙididdige makin dacewa yawanci ta amfani da kamancen cosine:

$\text{kamance}(q, d_i) = \frac{q \cdot d_i}{\|q\| \|d_i\|}$

inda $q$ da $d_i$ su ne wakilcin vectors a cikin sararin ma'ana guda ɗaya. Manyan-$k$ bayanan tare da mafi girman makin kamance ana wuce su zuwa mataki na gaba. Wannan yana tabbatar da cewa bayyanin ya mai da hankali kan tunanin na'urar da ya dace da tambayar, ba duka samfurinta ba.

3.2. Jerin Tunani don Bincika Ciki

Tsarin CoT yana canza ɓangarorin TMK da aka dawo da su zuwa alamar tunani da aka tsara. Don aikin da aka dawo da shi $T_1$, hanyar $M_1$, da abubuwan ilimi $K_1, K_2$, ana iya ƙera umarnin CoT kamar:

"Manufar na'urar (Aiki) ita ce: [Bayanin T_1].
Don cim ma wannan, ta yi amfani da hanyar: [Bayanin M_1].
Wannan hanyar tana buƙatar sanin: [Bayanin K_1] da [Bayanin K_2].
Saboda haka, shawarar na'urar ta dogara ne akan..."

Ana ciyar da wannan alamar da aka tsara zuwa ChatGPT tare da umarni kamar: "Dangane da matakan tunani masu zuwa, ka samar da bayyani mai haske, a takaice ga ɗalibi."

4. Kimantawa ta Gwaji & Sakamako

4.1. Ma'auni na Kimantawa: Cikakke & Daidaito

Marubutan sun kimanta bayyana kai tare da manyan fuskoki guda biyu:

Cikakke: Shin bayyanin ya ƙunshi duk matakan da suka dace a cikin tsarin yanke shawara na na'urar kamar yadda samfurin TMK ya ayyana? An kimanta wannan ta hanyar dawo da abubuwan cikin bayyanin zuwa abubuwan TMK.
Daidaito: Shin bayyanin yana nuna ainihin tsarin na'urar daidai, ba tare da gabatar da ruɗi ko sabani ba? Wannan yana buƙatar tabbatarwa ta ƙwararru da lambar/rajistan na'urar.

Mahimman Hasashe na Kimantawa

Hanyar haɗakar ta nuna manyan maki a cikin daidaito saboda samfurin ƙirƙira ya kasance cikin takurawa sosai ta bayanan TMK da aka dawo da su. Cikakke ya kasance mafi canzawa, ya danganta da ingancin binciken kama-karya da ƙera umarni don CoT.

4.2. Sakamako daga Aiwarta a cikin Ajin Kai tsaye

An aiwatar da tsarin a cikin aji na kan layi kai tsaye. Duk da yake ba a yi cikakken bayani game da takamaiman sakamako na ƙididdiga ba a cikin abin da aka ba da shaci, takardar ta ba da rahoto game da wannan aiwarta, tana nuna mai da hankali kan inganci ko tabbatarwa na farko a duniyar gaske. Aiwarta kanta wani sakamako ne mai mahimmanci, yana nuna yuwuwar aikin a cikin yanayin ilimi mai motsi. Aikin gaba zai amfana daga gwajin A/B wanda ke auna ma'auni na aminci (misali, binciken mai amfani game da bayyana da ake gani, amincin) tsakanin ƙungiyoyin da suka karɓi bayyanai da waɗanda ba su karɓa ba.

Bayanin Zane mai Hasashe: Zane mai shafe-shafe wanda ke kwatanta makin "Ingancin Bayyani" (Cikakke da Daidaito akan ma'auni 1-5) don hanyar Haɗakar TMK+ChatGPT da aka kwatanta da tushen amfani da ChatGPT kaɗai tare da tambayar mai amfani kawai. Hanyar haɗakar za ta nuna manyan sanduna don Daidaito, yana nuna tasirin kafa tushe na samfurin kai.

5. Tsarin Nazari & Misalin Lamari

Yanayi: SAMI ya haɗa ɗalibi Bob (mai sha'awar "Kiɗan Jazz" da "Shirye-shiryen Python") da ɗaliba Alice (mai sha'awar "Kiɗan Blues" da "Kimiyyar Bayanai").

Tambayar Mai Amfani: "Me ya sa aka haɗa ni da Alice?"

Ƙananan Bayanan Samfurin Kai na TMK na Ciki (An dawo da su):

T1: "Haɓaka alaƙar zamantakewa bisa kamancen bayanan martaba."
M1: "Ƙididdige juzu'in sha'awa ta amfani da kamancen Jaccard akan alamun sha'awa."
K1 (Bob): "Sha'awa: Kiɗan Jazz, Shirye-shiryen Python."
K2 (Alice): "Sha'awa: Kiɗan Blues, Kimiyyar Bayanai."
K3 (Ƙididdiga): "Alamun da aka raba: 'Kiɗa' (daga Jazz/Blues) da 'Shirye-shirye' (daga Python/Kimiyyar Bayanai). Makin kamance: 0.4."

Jerin Tunani Umarni don ChatGPT:
"Aikin na'urar shine haɓaka alaƙa bisa kamancen bayanan martaba. Hanyarta ita ce ƙididdige juzu'in sha'awa. A gare ka (Bob), ta san sha'awarka: Kiɗan Jazz, Shirye-shiryen Python. Ga Alice, ta san sha'awarta: Kiɗan Blues, Kimiyyar Bayanai. Ta ƙididdige makin kamance bisa rukunoni masu juzu'i (Kiɗa, Shirye-shirye). Makin ya kasance 0.4, wanda ya cika ƙa'idarta don ba da shawarar haɗin kai. Saboda haka..."

Bayyanin Ƙarshe da aka Ƙirƙira:
"Na haɗa ka da Alice saboda tsarinmu ya gano sha'awar da aka raba wanda zai iya zama tushe mai kyau don tattaunawa. Dukanku kuna da sha'awa a cikin kiɗa (Jazz/Blues) da batutuwan fasaha masu alaƙa da shirye-shirye (Python/Kimiyyar Bayanai). Bisa wannan juzu'in, na yi tunanin za ku iya samun ginshiƙi guda don fara tattaunawa."

6. Nazari Mai Zurfi & Hanyoyin Kwararru

Mahimman Hasashe: Wannan takarda ba kawai game da sanya AI ya bayyana ba; shiri ne na dabarun ƙera aminci a cikin na'urori da aka saka a cikin zamantakewa. Marubutan sun gano daidai cewa a cikin fagage kamar ilimi, tasirin na'urar ba ya zuwa daga aikin aiki kawai amma daga matsayinta na ɗan wasa na zamantakewa mai aminci. Hanyarsu ta haɗakar—ta amfani da samfurin kai na alama a matsayin "tushen gaskiya" don kubutar da AI mai ƙirƙira daga halin ƙirƙira—wata dabara ce mai amfani kuma wajibi a cikin zamanin LLM na yanzu. Tana magance kai tsaye abin da masu bincike kamar Cynthia Rudin suke jayayya: muna buƙatar samfuran da ke da fahimta a cikin su, ba bayyanai bayan haka ba. A nan, samfurin TMK yana ba da wannan tsari na ciki.

Kwararar Hankali & Gudunmawa: Hankali yana da ban sha'awa: 1) Na'urorin zamantakewa suna buƙatar aminci, 2) Aminci yana buƙatar bayyana, 3) Bayyana yana buƙatar bayyana kai, 4) Bayyana kai mai dogaro yana buƙatar samfurin kai mai tushe, 5) Bayyanai masu amfani suna buƙatar yare na halitta, 6) Saboda haka, haɗa samfurin mai tushe (TMK) tare da mai ƙirƙira yare (LLM). Babban gudunmawar shine takamaiman tsarin gine-gine wanda ke aiwatar da wannan kwararar, musamman amfani da binciken kama-karya akan bayanan TMK na halitta a matsayin hanyar dawowa. Wannan ya fi kyau fiye da faɗakar da aka ƙera.

Ƙarfi & Kurakurai: Babban ƙarfin shine ƙirar haɗakar ta aiki, tare da guje wa rashin bayyana na cikakken koyon zurfi da rashin ƙarfi na tsarin alama. Yana da kyakkyawan aikace-aikacen ƙa'idodin ƙirƙira da aka haɓaka (RAG), amma an yi amfani da su ga ilimin kai maimakon takardun waje—ra'ayi mai ƙafafu. Duk da haka, kurakurai suna da mahimmanci. Na farko, samfurin kai yana tsaye kuma an yi shi da hannu. Ba ya koyo ko sabuntawa daga hulɗa, yana haifar da nauyin kulawa da haɗarin karkata daga ainihin lambar na'urar. Na biyu, kimantawa yana da sirara. Ina manyan lambobi game da amincin mai amfani, fahimta, ko canjin hali? Ba tare da waɗannan ba, shaidar ƙira ce, ba kayan aikin gina aminci da aka tabbatar ba. Na uku, yana ɗauka cewa samfurin TMK cikakkiyar wakilci ce na "gaskiyar" tunanin na'urar, wanda bazai yi riƙe da na'urori masu rikitarwa, masu daidaitawa ba.

Hanyoyin Aiki masu Aiki: Ga masu aiki, abin da za a ɗauka a bayyane yake: Fara tsara tsarin AI ɗinku tare da samfurin kai mai tambaya tun daga ranar farko. Wannan takarda tana ba da samfuri mai yuwuwa. Mataki na gaba shine sarrafa ƙirƙira da sabuntawa na wannan samfurin kai, watakila ta amfani da fasahohi daga AI na neuro-symbolic ko fahimtar injiniya. Ga masu bincike, ƙalubalen shine matsawa sama da samfuran kai masu tsayi zuwa wakilcin kai mai motsi, mai koyo. Shin na'urar za ta iya koyon tsarin TMK dinta daga abubuwan da ta fuskanta da lambarta? Ƙari ga haka, fannin dole ne ya haɓaka ma'auni na daidaitawa don kimanta tasirin zamantakewa na bayyanai, ba kawai cikakkun bayanansu na fasaha ba. Shin bayyani kamar wanda aka ƙirƙira a zahiri yana ƙara son ɗalibi don yin hulɗa da takwaransa da AI ta ba da shawara? Wannan shine ma'auni na ƙarshe wanda ke da mahimmanci.

7. Ayyukan Gaba & Hanyoyin Bincike

Koyon Samfurin Kai ta Atomatik: Haɗa fasahohi daga haɗakar shirye-shirye ko nazarin lamba na tushen LLM don samar da sabunta samfurin kai na TMK ta atomatik daga lambar tushe da rajistan aiki na na'urar, rage aikin injiniyan hannu.
Tsarin Na'urori Da Yawa Masu Bayyanawa: Faɗaɗa tsarin don bayyana halayen tarin na'urori ko tururuwa, inda bayyanai na iya haɗawa da ka'idojin haɗin kai da halayen da suka taso.
Salon Bayyanai Na Musamman: Daidaita ɓangaren ƙirƙira don daidaita rikitarwar bayyanai, sautin murya, da mai da hankali bisa bayanan martabar mutum ɗaya (misali, sabo da ƙwararren, mai shakka da mai aminci).
Bayyanai Masu Gabatarwa & Kwatance: Matsawa sama da amsa QA zuwa samun na'urar ta ba da bayyanai don ayyukan da ba a zata ba ko bayar da bayyanai masu kwatance ("Na haɗa ka da Alice maimakon Charlie saboda...").
Aikace-aikace a cikin Fagage Masu Matsala: Aiwarta irin wannan tsarin gine-gine na bayyana kai a cikin AI na kiwon lafiya (bayyana shawarwarin magani), fintech (bayyana ƙin lamuni), ko tsarin mai cin gashin kansa (bayyana yanke shawara na kewayawa), inda bayyana ya zama dole a doka ko a ɗabi'a.
Binciken Daidaita Aminci: Nazari na dogon lokaci don auna yadda bayyanar irin waɗannan bayyanai a tsawon lokaci ke shafar amincin mai amfani, dogaro, da ingancin tsarin gaba ɗaya wajen cim ma manufofinsa na zamantakewa.

8. Nassoshi

Goel, A. K., & Joyner, D. A. (2017). Using AI to teach AI: Lessons from an online AI class. AI Magazine.
Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence.
Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems.
Muller, M., et al. (2019). Principles for Explainable AI. Communications of the ACM.
Confalonieri, R., et al. (2021). A historical perspective of explainable AI. WIREs Data Mining and Knowledge Discovery.
Goodfellow, I., et al. (2014). Generative Adversarial Nets. Advances in Neural Information Processing Systems. (A matsayin misali na fasahar AI ta asali, wacce duk da haka sau da yawa ba ta bayyana ba, wacce ke buƙatar hanyoyin bayyana bayan haka).
Georgia Institute of Technology, Interactive Computing - Design & Intelligence Lab. (https://dilab.gatech.edu/) – Don mahallin yanayin binciken da ke samar da wannan aikin.
OpenAI. (2023). ChatGPT. (https://openai.com/chatgpt) – Bangaren AI mai ƙirƙira da aka ambata a cikin takardar.