Wannan takardar nazari tana magance babbar ƙalubale ta haɗa hankali na gargajiya cikin tsarin AI na tattaunawa na zamani. Duk da cewa manyan ƙirar harshe da aka riga aka horar da su (misali, BERT, GPT, T5) sun sami nasara mai ban mamaki wajen fahimtar nahawu da mahallin, a zahiri ba su da ilimin duniya na ɓoye da mutane suka ɗauka a matsayin abin da ya saba. Takardar ta yi iƙirarin cewa wannan gibi shine babban toshewa da ke hana AI shiga cikin tattaunawa ta gaske, mai daidaituwa, da hankali. Marubutan, Christopher Richardson da Larry Heck daga Georgia Tech, sun sanya aikinsu a matsayin taswira mai mahimmanci na yanayin yanzu—hanyoyi, tarin bayanai, da ƙima—don jagorantar bincike na gaba a cikin wannan fagen da ba a saba da shi ba amma yana da mahimmanci.
Takardar ta bayyana takamaiman ayyukan tattaunawa inda gazawar hankali na gargajiya ta fi bayyana.
Kiyaye tattaunawa mai ma'ana da dacewa da jigo a cikin juzu'i da yawa. Ba tare da hankali na gargajiya ba, ƙirar ƙirar suna samar da amsoshi waɗanda suke daidai a nahawu amma ba su da ma'ana ko ba su da alaƙa a zahiri.
Amsa tambayoyi ko kammala umarni waɗanda ke buƙatar zato da ba a bayyana ba. Misali, fahimtar cewa "dafa tukunyar shayi" yana nuna cewa mataki na gaba shine "zuba ruwan," ko da ba a bayyana shi ba.
Fahimtar barkwanci, zagi, tausayi, da ka'idojin zamantakewa. Wannan yana buƙatar zurfin ƙirar ilimin halayyar ɗan adam da al'adun zamantakewa waɗanda ƙirar ƙirar na yanzu galibi suna ƙididdige su ta hanyar ƙididdiga maimakon fahimta.
Nazarin ya rarraba manyan hanyoyin fasaha da aka bincika a cikin wallafe-wallafen.
Ƙara horar da manyan ƙirar harshe (LLMs) akan tarin bayanai masu wadatar da ilimin gargajiya (misali, ATOMIC, SocialIQA). Wannan hanyar tana nufin sanya hankali na gargajiya cikin sigogin ƙirar a ɓoye.
Haɗa ƙirar ƙirar da tushen ilimi mai tsari kamar ConceptNet ko ATOMIC a sarari. Ƙirar ƙirar tana dawo da ko yin tunani akan waɗannan taswirori yayin ƙididdiga. Babban misali shine COMET (Bosselut et al., 2019), ƙirar transformer da aka horar don samar da sabbin rukunin ilimi daga waɗannan taswirori.
Horar da ƙirar ƙirar don samar da ba kawai amsa ba har ma da bayanin tunani ko bayani a cikin harshe na halitta. Wannan yana tilasta wa ƙirar ƙirar bayyana matakan ɓoye, yana iya inganta ƙarfi.
Bayan daidaitaccen daidaito, fagen yana amfani da ma'auni kamar:
Marubutan sun gabatar da bincike mai mahimmanci, na hannu akan manyan ƙirar tattaunawa na buɗe ido, BlenderBot 3 da LaMDA. Abubuwan da suka lura suna da laifi: duk da girman da haɓakar waɗannan ƙirar ƙirar, sau da yawa suna kasawa a cikin ayyukan gargajiya marasa mahimmanci. Misalai sun haɗa da samar da maganganu masu karo da juna a cikin tattaunawa ko kasa fahimtar ƙayyadaddun ƙayyadaddun zahiri. Wannan shaida ta zahiri tana ƙarfafa babban jigon takardar: aikin ma'auni bai daidaita da ƙarfin hankali na gargajiya, mai amfani, a cikin hulɗar buɗe ido ba.
Fahimta ta Asali: Fagen AI na tattaunawa yana fama da "bashi na hankali na gargajiya" mai tsanani. Mun gina manyan gine-gine (manyan LLMs) akan shaky, tushe na ɓoye. Nazarin ya gano daidai cewa babbar matsala ba rashin fasaha ba ce, amma rashin daidaituwa na asali tsakanin yanayin ƙididdiga, daidaitawar tsari na zamani na NLP da yanayin alama, dalili, da kwatankwacin hankalin ɗan adam. Kamar yadda aka lura a cikin babban aikin "A kan Ma'aunin Hankali" na Chollet (2019), haƙiƙanin hankali yana buƙatar samun ƙwarewa da haɓakawa a cikin sabbin yanayi—wani gagarumin aiki da ba zai yiwu ba tare da ƙirar duniya mai wadata ba.
Tsarin Ma'ana: Tsarin takardar yana da ma'ana kuma yana gamsarwa. Yana motsawa daga ayyana matsala da bayyanuwarta (Sashe na 1-2), zuwa lissafin mafita na injiniya da aka gwada (Sashe na 3), zuwa binciken yadda muke auna ci gaba (Sashe na 4), kuma a ƙarshe yana ba da shaida ta zahiri cewa mafita na yanzu ba su isa ba (Sashe na 5). Wannan kwarara yayi daidai da hanyar kimiyya: hasashe (hankali na gargajiya ya ɓace), gwaji (hanyoyin haɗawa daban-daban), ma'auni (ma'auni), da ƙarshe (ba a warware ba).
Ƙarfi & Kurakurai: Babban ƙarfin takardar shine ƙimar ta ta zahiri, mai mahimmanci akan ƙirar ƙirar SOTA. Ya wuce ra'ayoyin ilimi don nuna ainihin yanayin gazawa. Babban laifinsa, gama gari ga nazari, shine yanayin bayyana maimakon tsari. Yana taswira yankin amma yana ba da jagora mai iyaka akan waɗanne hanyoyi suka fi ban sha'awa. Yana rage iyakokin gine-ginen ƙirar ƙirar na transformer kawai don tunanin dalili, wani batu da aka ƙarfafa sosai a cikin bincike daga cibiyoyi kamar MIT's CSAIL akan haɗin neuro-symbolic.
Fahimta Mai Aiki: Ga masu aiki da masu bincike, abin da za a ɗauka a bayyane yake: daina ɗaukar hankali na gargajiya a matsayin wani tarin bayanai kawai don gyara. Fagen yana buƙatar canjin tsari. 1) Saka hannun jari a cikin Gine-ginen Neuro-Symbolic: Ƙirar ƙirar haɗin gwiwa waɗanda suka haɗa hanyoyin sadarwa na jijiya tare da bayyanannun wakilcin ilimi, waɗanda za a iya sarrafa su (kamar aikin akan Differentiable Inductive Logic Programming) hanya ce mai mahimmanci. 2) Haɓaka Mafi kyawun Muhalli na Kwaikwayo: Kamar Gym na OpenAI don ƙarfafa koyo, muna buƙatar masu kwaikwayo masu wadata, masu hulɗa (waɗanda aka yi wahayi zuwa gare su ta dandamali kamar AllenAI's THOR) inda wakilai za su iya koyon hankali na gargajiya ta hanyar gogewa da sakamako, ba kawai rubutu ba. 3) Sake Tunani Ƙima: Matsar daga ma'auni na QA na tsaye zuwa ƙima mai ƙarfi, mai hulɗa inda ƙirar ƙirar dole su nuna daidaitaccen fahimtar duniya akan lokaci, kama da ƙa'idodin da ke bayan ƙalubalen ARC (Abstraction and Reasoning Corpus).
Hanyar dogaro akan taswirar ilimi sau da yawa ta ƙunshi tsarin samarwa wanda aka ƙarfafa da dawo da bayanai. A bisa ƙa'ida, idan aka ba da mahallin tattaunawa $C$, ƙirar ƙirar tana dawo da tarin rukunin ilimin gargajiya masu dacewa $K = \{(h_i, r_i, t_i)\}$ daga taswirar ilimi $\mathcal{G}$, inda $h$ ke zama abu na kai, $r$ alaƙa, kuma $t$ abu na wutsiya. Ana samar da amsa ta ƙarshe $R$ ta hanyar yanayin duka $C$ da $K$:
$P(R | C) \approx \sum_{K} P_{\text{retrieve}}(K | C) \cdot P_{\text{generate}}(R | C, K)$
Ƙirar ƙirar kamar COMET suna aiwatar da wannan ta hanyar gyara transformer (misali, GPT-2) don hasashen abu na wutsiya $t$ idan aka ba da $(h, r)$, yana koyon ketare taswira a cikin sararin ɓoye: $t = \text{COMET}(h, r)$.
Duk da cewa samfotin PDF ba su ƙunshi cikakkun chati ba, abubuwan da aka bayyana na farko suna nuna babban gibin aiki. Za mu iya tunanin chati mai ma'ana na mashaya wanda ke kwatanta aikin ɗan adam da BlenderBot3 da LaMDA akan jerin ayyukan tattaunawa na gargajiya (misali, Daidaituwa, Tunani na Zahiri, Tunani na Zamantakewa). Y-axis zai wakilci maki (0-100). Chatin zai nuna:
Yanayi: Ƙimar fahimtar wakilin tattaunawa game da labari mai sauƙi.
Mahallin Tattaunawa: Mai amfani: "Na sami sabon ɗan kwikwiyo! Yana da ƙarfi sosai. Na bar shi a cikin falo tare da takalmina na fi so yayin da nake amsa ƙofar."
Amsar Wakili A (Rashin Hankali na Gargajiya): "Yana da kyau. Takalmin ku launin me ne?"
Amsar Wakili B (Tare da Hankali na Gargajiya): "Haba, watakila kana son duba waɗannan takalman! ƴan kwikwiyo suna son tauna."
Binciken Tsarin:
Warware tunanin hankali na gargajiya zai buɗe aikace-aikace masu canzawa: