Teburin Abubuwan Ciki
1. Gabatarwa
Mataimakan kama-da-kai (VAs) suna canza mu'amalar mutum da kwamfuta, duk da haka aikace-aikacensu a fagage na musamman kamar injiniyan software ya kasance mai iyaka. Babban matsalar shi ne karancin ingantattun tarin bayanan tattaunawa na musamman da ake buƙata don horar da samfuran AI na asali. Wannan takarda ta magance wannan gibi ta hanyar gabatar da nazarin Wizard of Oz (WoZ) wanda aka tsara don kwaikwayi da tattara tattaunawa tsakanin masu shirye-shirye da mataimakin kama-da-kai don amfani da API. Binciken ya ƙunshi masu shirye-shirye ƙwararru 30 waɗanda suka yi imanin cewa suna mu'amala da AI, yayin da a zahiri, ƙwararrun mutane (“wizards”) suka samar da amsoshi. Sakamakon tarin bayanan an yi masa bayanan karin bayani ta fuskoki da yawa don fahimtar tsari da manufar tattaunawar neman taimako a cikin mahallin shirye-shirye.
2. Hanyar Bincike & Tsarin Gwaji
Jigon wannan binciken shine gwajin WoZ da aka tsara sosai, hanyar da aka tabbatar a cikin HCI don kwaikwayi tsarin hankali kafin a gina su gaba ɗaya.
2.1. Ka'idar Wizard of Oz
An yi amfani da tsarin WoZ don ƙirƙirar kwaikwaiyon simulatin mataimakin API mai aiki. Masu shirye-shirye sun yi mu'amala ta hanyar mu'amalar hira, ba tare da sanin cewa ƙwararrun mutane a bayan fage suka ƙirƙira amsoshi a cikin ainihin lokacin ba. Wannan hanyar tana ba da damar tattara bayanan tattaunawa na yanayi wanda ke nuna ainihin buƙatun mai amfani da dabarun, wanda ke da mahimmanci don horar da tsarin AI na gaba, kamar yadda aka jaddada a cikin wallafe-wallafen tsarin tattaunawa na asali kamar na Rieser da Lemon.
2.2. Daukar Mahalarta & Ayyuka
Binciken ya ɗauki masu shirye-shirye ƙwararru 30. An ba kowane ɗan takara ayyukan shirye-shirye da ke buƙatar amfani da API daban-daban guda biyu. An tsara ayyukan don su zama masu mahimmanci, suna ƙarfafa buƙatar taimako don haka samar da tarin tattaunawa mai wadata.
2.3. Tsarin Tattara Bayanai & Bayanan Karin Bayani
Tattaunawar da aka tattara an yi musu bayanan karin bayani tare da mahimman fuskoki huɗu:
- Manufar Illocutionary: Manufar mai magana (misali, buƙata, sanarwa, tabbatarwa).
- Nau'in Bayanin API: Rukunin bayanin da ake nema (misali, tsari, siga, misali).
- Aiki Mai Fuskantar Baya: Yadda furuci ke da alaƙa da tattaunawar da ta gabata (misali, amsa, bayani).
- Bin Sawu zuwa Abubuwan API: Taswirar abubuwan tattaunawa zuwa takamaiman aji/hanyoyin API.
Kididdigar Gwaji
- Mahalarta: Masu Shirye-shirye Ƙwararru 30
- API da aka Yi Amfani da su: API daban-daban 2
- Girman Bayanan Karin Bayani: Muhimman Girma 4
- Tarin Bayanai: Ana Samun su a Bude akan GitHub
3. Sakamako & Muhimman Bincike
3.1. Nazarin Ayyukan Tattaunawa
Bayanan karin bayani ya bayyana sassa daban-daban na ayyukan tattaunawa. Masu shirye-shirye sau da yawa suna ba da buƙatun rikitattun, sassa da yawa waɗanda suka haɗa tambayoyi game da tsari, ma'ana, da misalan amfani. Amsoshin “wizard” sau da yawa suna buƙatar rarraba waɗannan buƙatun kuma su ba da bayanai masu tsari, mataki-mataki, suna nuna buƙatar ci-gaba da sarrafa tattaunawa a cikin VAs na gaba.
3.2. Bayanin Kididdiga
Yayin da takardar ba ta ba da cikakkun ƙididdiga ba, tana nuna cewa tarin bayanan yana da girma kuma ya bambanta don tallafawa koyon inji. Rarraba ayyuka a cikin girman bayanan karin bayani huɗu yana ba da tushe na ƙididdiga don ƙirar yanayin tattaunawa da manufa a cikin mataimakin kama-da-kai.
3.3. Muhimman Fahimta daga Mu'amala
Muhimman Fahimta 1: Halin neman taimako na masu shirye-shirye yana da mahallin sosai kuma mai maimaitawa, ba sauƙaƙan tambaya da amsa ba.
Muhimman Fahimta 2: Nasara taimako yana buƙatar haɗa tambayoyi na zahiri zuwa takamaiman abubuwan API masu bin sawu.
Muhimman Fahimta 3: Dabarun tattaunawar da aka lura sune tushe don ƙirar dabaru na tattaunawa na mataimakin da ke da ƙarfin AI.
4. Tsarin Fasaha & Samfurin Lissafi
Binciken a ɓoye yayi daidai da samfurin Tsarin Ƙaddarawar Markov Wanda Ba a Gani Gabaɗaya (POMDP) wanda aka saba da shi a cikin tsarin tattaunawa. Manufar mataimakin ita ce zaɓi aiki $a$ (misali, ba da misali, nemi bayani) bisa ga yanayin imaninsa $b(s)$ akan ainihin yanayin mai amfani $s$ (misali, gibin ilimin mai amfani, matakin aiki na yanzu) don haɓaka lada $R$ (misali, kammala aiki).
Ana iya ƙirar sabunta imani kamar haka: $b'(s') = \eta \cdot O(o | s', a) \sum_{s \in S} T(s' | s, a) b(s)$ inda $T$ shine aikin canzawa, $O$ shine aikin lura (fassarar furucin mai amfani $o$), kuma $\eta$ shine mai daidaitawa. Tarin bayanan da aka yi bayanin karin bayani yana ba da bayanan don koyon waɗannan ayyuka $T$ da $O$ don yankin API.
5. Tsarin Nazari: Misalin Nazarin Shari'a
Yanayin: Mai shirye-shirye yana ƙoƙarin amfani da hanyar API DataFrame.merge() amma ya ci karo da kuskure.
Guntun Tattaunawa (An Yi Bayanin Karin Bayani):
- Mai Amfani: "Haɗina yana kasawa tare da kuskuren maɓalli. Ta yaya zan iya tantance maɓallan haɗin?"
- Manufa: Buƙata
- Nau'in Bayani: Tsari/Siga
- Bin Sawu:
DataFrame.merge(), sigogi `on`/`left_on`/`right_on`
- Wizard/Mataimaki: "Hanyar `merge()` na iya amfani da sigogi `on`, `left_on`, da `right_on`. Idan DataFrame ɗin ku suna da sunan ginshiƙi gama gari, yi amfani da `on='sunan_ginshiƙi'`. Idan sun bambanta, yi amfani da `left_on` da `right_on`. Za ka iya nuna mini sunayen ginshiƙi na DataFrame ɗinka biyu?"
- Manufa: Sanarwa + Fito
- Nau'in Bayani: Bayani + Misalin Ƙarfafawa
- Aiki na Baya: Amsa + Bayani
6. Hangen Nesa na Aikace-aikace & Hanyoyin Gaba
Gajeren Lokaci: Tarin bayanan shine kayan aikin horo kai tsaye don gina ƙirar mataimakan API ta amfani da samfuran jeri-zuwa-jeri ko na canzawa (misali, daidaita samfura kamar Codex ko CodeT5).
Matsakaicin Lokaci: Haɗawa cikin Muhallin Ci Gaban Haɗaɗɗu (IDEs) azaman kwamitin taimako mai himma, rage canza mahallin zuwa takaddun bayanai.
Dogon Lokaci & Bincike na Gaba:
- Keɓancewa: Ƙirar matakin ƙwarewar mai shirye-shirye don daidaita bayani.
- Taimakon Nau'i-nau'i: Haɗa tattaunawa tare da samar da lamba, kamar GitHub Copilot, amma tare da iyawar bayani.
- Gabaɗaya na API: Haɓaka samfuran da za su iya koyon dabarun taimako masu canzawa a cikin ɗakunan ajiya da tsare-tsare daban-daban, suna motsawa fiye da horon API guda ɗaya.
- AI Mai Bayyanawa don Lamba: Yin amfani da tsarin tattaunawa don sa shawarwarin samfuran samar da lamba su zama masu fassara.
7. Nassoshi
- McTear, M., Callejas, Z., & Griol, D. (2016). The Conversational Interface: Talking to Smart Devices. Springer.
- Rieser, V., & Lemon, O. (2011). Reinforcement Learning for Adaptive Dialogue Systems: A Data-driven Methodology for Dialogue Management and Natural Language Generation. Springer.
- Serban, I. V., et al. (2015). A survey of available corpora for building data-driven dialogue systems. arXiv preprint arXiv:1512.05742.
- OpenAI. (2021). Codex. [https://openai.com/blog/openai-codex]
- Google AI. (2021). Conversational AI. [https://ai.google/research/teams/language/conversational-ai]
- Chen, M., et al. (2021). Evaluating Large Language Models Trained on Code. arXiv preprint arXiv:2107.03374.
8. Asalin Nazari & Sharhin Kwararru
Muhimman Fahimta: Wannan takarda ba kawai game da tattara bayanai ba ce; yana da dabarun tono tsarin aikin fahimi na mai shirye-shirye da ya makale akan API. Ainihin darajar yana cikin bayyana gibin tsakanin abin da masu shirye-shirye suke tambaya (“Me yasa wannan kuskuren ke faruwa?”) da abin da suke buƙata a zahiri (hanyar bin sawu daga ƙirar tunaninsu mara kyau zuwa ma'anar API daidai). Hanyar WoZ ta yi fice ta ƙetare iyakokin NLP na yanzu don ɗaukar wannan ƙaramin bayani, wani abu da keɓaɓɓen loda na binciken Stack Overflow zai rasa gaba ɗaya. Hanyar HCI ce ta tsohuwar makaranta da aka yi amfani da ita don magance matsalar bayanan AI na zamani.
Tsarin Ma'ana & Gudunmawa: Marubutan sun gano daidai hamada bayanai a cikin haɓaka VA na musamman, wani batu da aka maimaita a cikin binciken fadi kamar na Serban et al. Maganinsu yana da inganci ta hanyar bincike: 1) Kwaikwayi manufar ƙarshe (mataimaki mai aiki) ta hanyar WoZ don samun mu'amala na gaske, 2) Rushe tattaunawar tare da tsarin bayanan karin bayani mai fuskoki da yawa wanda ya wuce sauƙaƙan rabe-raben manufa, da 3) Ƙirƙiri kadari na jama'a (tarin bayanan) don farawa da al'umma. Wannan aikin tushe ne na gargajiya—gina bututun kafin samfurin. Girman bayanan karin bayani huɗu, musamman 'bin sawu,' su ne miyar sirrin takardar, suna haɗa tattaunawa kai tsaye zuwa abubuwan lamba, wanda ya zama dole ga kowane mataimaki wanda ke nufin zama fiye da chatbot.
Ƙarfi & Kurakurai: Ƙarfin yana cikin ingantacciyar hanyar bincike, mai maimaitawa da ƙirƙirar tarin bayanai mai ƙima, da wuya. Yana da amfani nan take ga duk wanda ke horar da samfurin tattaunawa na musamman. Duk da haka, kurakurai—wanda aka yarda amma yana da mahimmanci—shine sikelin da farashi. Mahalarta talatin da wizards na ɗan adam aikin bincike ne, ba injin samar da bayanai mai sikelin ba. Ilimin “wizard” kuma shine toshewa; ƙwarewarsu ta ayyana iyakar “cikakken” mataimaki. Shin dabarun za su bambanta idan wizards sun kasance manyan vs. ƙananan masu haɓakawa? Bugu da ƙari, yayin da samfurin POMDP yana a ɓoye, takardar ta tsaya kafin ta ba da manufar da aka horar da ita ko takamaiman ma'auni na ML akan sabon tarin bayanan, ta bar “to me” na bayanan karin bayani a matsayin mai ban sha'awa maimakon tabbatacce.
Fahimta Mai Aiki & Tasirin Kasuwa: Ga masu binciken AI, wannan filin horo da gwaji ne da aka riga aka yi. Mataki na gaba shine amfani da wannan tarin bayanan don yin ma'auni ga samfura kamar Codex ko CodeT5 akan iyawar tattaunawar su, ba kawai samar da lamba ba. Ga masu gina kayan aiki (misali, JetBrains, Microsoft VS Code), fahimtar ita ce cewa taimakon cikin-IDE dole ne ya zama mai mu'amala da bincike, ba kawai juji na takaddun bayanai ba. Gaba ba chatbot ne wanda ke amsa tambayoyi ba; yana da wakili na haɗin gwiwa wanda ke shiga cikin tattaunawar maimaitawa, mai bin sawu wanda wannan binciken ya zana taswira. Gaskiyar gasa ba kawai game da wanda ke da mafi kyawun samfurin kammala lamba ba, amma wanda zai iya haɗa Layer na bayani wanda wannan binciken ya yi taswira yadda ya kamata. Wannan aikin yana canza mayar da hankali daga “samar da amsa” zuwa “sarrafa tattaunawar bayani,” wanda shine inda za a sami ainihin ribar yawan aiki don ayyuka masu sarƙaƙi kamar injiniyan software.