Hankali Na Gama Gari A Cikin AI Mai Tattaunawa: Nazari Akan Matsayin Fasaha Na Yanzu

1. Gabatarwa

Wannan takardar nazari ta magance babbar ƙalubale ta haɗa hankali na gama gari cikin tsarin AI mai tattaunawa na zamani. Duk da cewa ƙirar ƙira kamar BERT, GPT, da T5 sun sami nasara mai ban mamaki wajen fahimtar tsarin harshe da ma'anar mahallin, har yanzu suna fuskantar wahala a ayyukan da suke buƙatar ilimin gama gari—wato ilimin duniya da mutane suka saba ɗauka a matsayin abin da ba a buƙatar bayyanawa. Takardar ta nuna cewa wannan gibi yana hana ci gaban tsarin tattaunawa na gaske wanda ya dace da yanayin ɗan adam.

Muhimmancin hankali na gama gari ga hankalin injina an riga an gane shi, amma har yanzu ba a sami tsari gama gari na tsarawa da haɗa wannan ilimin ba. Wannan nazari ya mai da hankali kan haɗuwar hankali na gama gari da AI mai tattaunawa, yana bitar bayanan da suka dace, hanyoyi, da ma'auni na ƙima.

2. Hankali Na Gama Gari A Cikin Matsalolin AI Mai Tattaunawa

Hankali na gama gari yana da muhimmanci a ko'ina cikin AI mai tattaunawa. Takardar ta gano wasu muhimman fagage inda rashinsa ya fi bayyana.

3. Hanyoyin Haɗa Hankali Na Gama Gari

Nazarin ya rarraba manyan hanyoyin zuwa manyan dabarai guda uku don haɗa hankali na gama gari cikin ƙirar ƙirar AI mai tattaunawa.

4. Ma'auni da Ma'aunin Ƙima

Ƙimar hankali na gama gari a cikin tattaunawa yana da rikitarwa. Takardar tana tattauna ma'auni da yawa:

• Ma'auni Na Musamman: Bayanan da aka keɓe don ƙimar ƙwarewar tunani ta musamman (misali, tunanin zahiri a cikin PIQA, tunanin zamantakewa a cikin SocialIQA).
• Ma'auni Na Haɗaɗɗen Tattaunawa: Ƙima a cikin manyan ayyukan tattaunawa, kamar bayanan Tattaunawar Gama Gari wanda ke gwada ko amsoshin ƙirar ƙira sun dace da gaskiyar gama gari.
• Ƙimar Mutum: A ƙarshe, dacewar tattaunawa da daidaituwa, wanda mutane suka yi hukunci, ya kasance ma'auni mai mahimmanci, ko da yake na zahiri.

Ma'auni na atomatik na gama gari sun haɗa da daidaito akan tambayoyi masu zaɓi da yawa, BLEU/ROUGE don ingancin amsa, da sabbin ma'auni da aka ƙera don auna daidaiton gaskiya ko yuwuwar tunani.

5. Abubuwan Lura Na Farko Akan Ƙirar SOTA

Takardar ta gabatar da nazari na farko na manyan ƙirar ƙirar tattaunawa guda biyu: BlenderBot 3 da LaMDA. Duk da ƙwarewarsu na ci gaba, duka ƙirar ƙirar sun nuna gazawa mai mahimmanci a cikin hankali na gama gari. Misalai sun haɗa da:

• Samar da amsoshi waɗanda suka saba wa dokokin zahiri na asali (misali, ba da shawarar cewa abu zai iya kasancewa a wurare biyu lokaci guda).
• Rashin fahimtar alamun zamantakewa ko ka'idoji da ba a bayyana ba.
• Samar da maganganun da ba su dace da gaskiya ba a cikin juzu'in tattaunawa guda ɗaya.

Waɗannan abubuwan lura suna ƙarfafa buƙatar bincike mai mai da hankali a wannan fanni, saboda irin waɗannan gazawar suna lalata amincewar mai amfani da kuma yanayin dacewar hulɗa.

6. Cikakkun Bayanai Na Fasaha da Tsarin Lissafi

Haɗa taswirar ilimi sau da yawa yana ƙunsar tsarin samarwa wanda aka ƙarfafa da dawo da bayanai. Idan aka ba da mahallin tattaunawa $C$ da taswirar ilimi $\mathcal{K}$, manufar ƙirar ƙira za a iya tsara ta a matsayin samar da amsa $R$ wanda ya ƙaru:

$P(R | C, \mathcal{K}) = \sum_{k \in \mathcal{K}_C} P(k | C) \cdot P(R | C, k)$

Inda $\mathcal{K}_C$ wani yanki ne na maɗaukakin ilimi masu dacewa da aka dawo daga $\mathcal{K}$ dangane da mahallin $C$. Kalmar $P(k | C)$ tana wakiltar yuwuwar zaɓin maɗaukakin ilimi $k$ na ƙirar dawo da bayanai, kuma $P(R | C, k)$ shine yuwuwar amsa idan aka ba da mahallin da ilimin da aka zaɓa. Ƙirar ƙira kamar COMET suna aiwatar da wannan ta hanyar gyara transformer (misali, GPT-2) akan maɗaukakin ilimin taswira da aka tsara su azaman $(head, relation, tail)$, yana ba ta damar samar da cikakkun $tail$ masu yuwuwa don sabbin tambayoyin $(head, relation)$.

7. Tsarin Nazari: Nazarin Lamari

Yanayi: Ƙimar fahimtar chatbot game da labari mai sauƙi.

Shigarwar Mai Amfani: "Na zuba kaina gilashin ruwan lemu, amma sai wayar ta yi ringi. Lokacin da na dawo, gilashin ya kasance babu kowa."

Tsarin Nazari:

1. Dawo da Ilimi: Ya kamata tsarin ya dawo da gaskiyar gama gari masu dacewa: Ana iya sha ruwa. Dabbobin gida (kamar kuliyoyi) suna iya sha ruwa. Mutane suna amsa wayoyi.
2. Samar da Hasashe: Amfani da ƙirar ƙira kamar COMET, samar da yuwuwar hasashe don lamarin "gilashin ruwan lemu da aka bar ba tare da kulawa ba": "Idan X ya bar abin sha ba tare da kulawa ba, to dabbar gida na iya sha shi" (dangantakar ATOMIC: xEffect).
3. Ƙimar Hasashe: Ƙima wane bayanin da aka yi hasashe ("wani ya sha shi", "ya ƙafe", "dabbar gida ta sha shi") ya fi dacewa da mahallin da yuwuwar zahiri. Daidaitaccen hasashe ya dogara da ilimin duniya da ba a bayyana ba game da abubuwan da suka faru na gida na yau da kullun.
4. Tsara Amsa: Samar da tambaya ko bayani mai daidaituwa: "Haba, shin kyanwarka ta isa gare shi?" sabanin wanda ba zai yuwu ba: "Shin ya zama gas?"

Wannan tsarin yana nuna matakan tunani da ake buƙata, daga dawo da bayanai zuwa hasashe zuwa haɗa mahallin.

8. Ayyuka Na Gaba da Hanyoyin Bincike

Hanyar gaba don AI mai tattaunawa mai sane da hankali na gama gari ta ƙunshi manyan hanyoyi da yawa:

• Hankali Na Gama Gari Mai Nau'i Daban-daban (Multimodal): Haɗa ilimin gani, ji, da azanci tare da harshe, kamar yadda ƙirar ƙira kamar OpenAI's CLIP da DALL-E suka fara, waɗanda ke haɗa rubutu da ra'ayoyin gani. Wakilan tattaunawa na gaba na iya buƙatar yin tunani game da fage da aka kwatanta a cikin tattaunawa.
• Taswirar Ilimi Mai Ƙarfi (Dynamic): Ƙaura daga KGs masu tsayayye zuwa tsarin da zai iya koyo da sabunta ilimin gama gari ci gaba da hulɗa, kamar yadda mutane suke yi.
• Tunani Na Dalili (Causal): Zurfafa fahimtar ƙirar ƙira game da dalili da sakamako, babban ɓangare na hankali na gama gari. Bincike daga matakin dalili na Judea Pearl yana nuna motsawa daga haɗin gwiwa zuwa shiga tsakani da tunanin akasin haka yana da mahimmanci don AI mai ƙarfi.
• Hankali Na Gama Gari Na Keɓaɓɓu da Al'adu: Haɓaka ƙirar ƙira waɗanda suka fahimci ka'idojin gama gari waɗanda suka bambanta tsakanin mutane, al'ummomi, da al'adu.
• Haɗin Neuro-Symbolic: Haɗa ƙarfin gane tsarin cibiyoyin jijiyoyi (kamar transformers) tare da ƙwarewar tunani na zahiri, na ma'ana na tsarin AI na alama. Wannan hanya ta gauraye, kamar yadda ƙirar ƙirar MIT's Probabilistic Symbolic (PS) suka bincika, hanya ce mai ban sha'awa don hankali na gama gari mai yuwuwa da fassara.

9. Nassoshi

• Richardson, C., & Heck, L. (2023). Commonsense Reasoning for Conversational AI: A Survey of the State of the Art. Workshop on Knowledge Augmented Methods for NLP, AAAI 2023.
• Speer, R., Chin, J., & Havasi, C. (2017). ConceptNet 5.5: An Open Multilingual Graph of General Knowledge. Proceedings of AAAI.
• Sap, M., et al. (2019). ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning. Proceedings of AAAI.
• Bosselut, A., et al. (2019). COMET: Commonsense Transformers for Automatic Knowledge Graph Construction. Proceedings of ACL.
• Gao, J., et al. (2018). Neural Approaches to Conversational AI. Foundations and Trends® in Information Retrieval.
• Pearl, J., & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.
• Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. Proceedings of ICML (CLIP).