Bayanan DICES: Bambancin a cikin Kimantawar Tsaro na AI na Tattaunawa

1. Gabatarwa

Yaduwar tsarin AI na tattaunawa da aka gina akan Manyan Samfuran Harshe (LLMs) ya sa kimantawar tsaro ya zama babban abin damuwa. Hanyoyin gargajiya sau da yawa sun dogara da bayanai masu rabe-raben "tsaro" da "rashin tsaro" a sarari, wanda ke sauƙaƙe yanayin tsaro na asali mai ra'ayi da kuma al'adu. Wannan takarda ta gabatar da bayanan DICES (Bambancin a cikin Kimantawar AI na Tattaunawa don Tsaro), wanda aka tsara don ɗauka da bincika bambancin fahimtar tsaro a cikin al'ummomin mutane daban-daban.

Babbar matsalar da aka magance ita ce rashin kula da bambancin al'umma da ra'ayi a cikin bayanan tsaro da ake da su, wanda zai iya haifar da samfuran da ba su dace da ka'idojin ƙungiyoyin masu amfani na musamman ba kuma suna da "sakamako maras so ko ma mai ban tsoro a cikin yanayin duniya na gaske."

1.1. Gudunmawa

Babban gudunmawar bayanan DICES da wannan aikin sune:

• Bambancin Masu Kimantawa: Ya karkata hankali daga "rage son zuciya" zuwa karɓa da auna "bambancin" ra'ayoyin masu kimantawa.
• Bayanan Al'umma cikin Zurfi: Ya haɗa da cikakkun bayanan al'umma (ƙabila/kabila, shekaru, jinsi) ga kowane mai kimantawa.
• Maimaitawa Mai Girma a kowane Abu: Kowane abu na tattaunawa yana samun adadi mai yawa na kimantawa don tabbatar da ƙarfin ƙididdiga don binciken ƙananan ƙungiyoyi.
• Wakilci dangane da Rarraba: Yana sanya kuri'un tsaro a matsayin rarraba a cikin ƙungiyoyin al'umma, yana ba da damar bincika dabarun haɗawa daban-daban fiye da kuri'un mafi rinjaye.
• Tsarin Bincike: Yana ba da tushe don kafa sabbin ma'auni waɗanda suka haɗa kimantawar masu kimantawa da rukunin al'umma.

2. Tsarin Bayanan DICES

An gina DICES a matsayin albarkatu da ma'auni na raba don mutunta ra'ayoyi daban-daban yayin kimantawar tsaro. Ya wuce alamar gaskiya guda ɗaya.

2.1. Ka'idoji na Zane na Asali

• Bambancin da aka Yi niyya: An tsara tafkin masu kimantawa don samun daidaiton adadi daga manyan ƙananan ƙungiyoyin al'umma.
• Ƙarfin Ƙididdiga: Maimaitawar kimantawa mai yawa a kowane abu na tattaunawa yana ba da damar ingantaccen bincike na yarjejeniya, rashin yarda, da bambanci a ciki da tsakanin ƙungiyoyi.
• Tsaro na Mahallin: Kimantawa ya dogara ne akan tattaunawar mutum-da-robot, yana ɗaukar tsaro a cikin mahallin mai ƙarfi, mai mu'amala maimakon akan faɗakarwa masu zaman kansu.

2.2. Tsarin Bayanai & Ƙididdiga

3. Hanyar Fasaha & Tsarin Bincike

Ƙirƙirar fasaha ta ta'allaka ne a cikin kula da tsaro ba a matsayin ma'auni ba amma a matsayin rarraba mai yawa.

3.1. Wakiltar Tsaro a matsayin Rarraba

Ga wani abu na tattaunawa $i$, tsaro ba a wakilta shi da alamar guda ɗaya $y_i$ ba amma ta hanyar rarraba kimantawa a cikin ƙungiyoyin al'umma $K$. Bari $R_{i,g}$ ya zama saitin kimantawa don abu $i$ daga masu kimantawa a cikin ƙungiyar $g$. Bayanan tsaro don abu $i$ shine vector: $\mathbf{S}_i = (\bar{R}_{i,1}, \bar{R}_{i,2}, ..., \bar{R}_{i,K})$, inda $\bar{R}_{i,g}$ shine tsarin tsakiya (misali, ma'ana, matsakaici) na kimantawa a cikin ƙungiyar $g$.

Ana iya ƙididdige ma'auni na bambanci kamar $\sigma^2_{i,g}$ (bambanci a cikin ƙungiyar) da $\Delta_{i, g1, g2} = |\bar{R}_{i,g1} - \bar{R}_{i,g2}|$ (rashin yarda tsakanin ƙungiyoyi) don ƙididdige shubuha da bambancin ra'ayi.

3.2. Dabarun Haɗawa & Ma'auni

DICES yana ba da damar kwatanta hanyoyin haɗa lakabi daban-daban:

• Kuri'un Mafi Rinjaye (Tushe): $y_i^{maj} = \text{mode}(\bigcup_{g=1}^{K} R_{i,g})$
• Haɗawa Mai Girman Al'umma: $y_i^{weighted} = \sum_{g=1}^{K} w_g \cdot \bar{R}_{i,g}$, inda $w_g$ zai iya zama daidai da girman yawan jama'a ko wasu ma'auni masu mai da hankali kan adalci.
• Tsaro Mafi ƙarancin (Mai Tsattsauran Ra'ayi): $y_i^{min} = \min(\bar{R}_{i,1}, ..., \bar{R}_{i,K})$ yana ba da fifiko ga ra'ayin ƙungiyar mafi hankali.

Ana iya samun sabbin ma'auni kamar Fihirisar Rashin Yardar Al'umma (DDI) ko Makin Daidaitawar Ƙananan Ƙungiyoyi don auna yadda aikin samfurin ya bambanta a cikin ƙungiyoyi.

4. Sakamakon Gwaji & Babban Abubuwan da aka Gano

Yayin da gajeren fayil ɗin da aka bayar yana ƙarƙashin nazari kuma bai ƙunshi cikakkun sakamako ba, tsarin da aka gabatar yana haifar da abubuwan da ake tsammani da yawa:

• Bambanci Mai Girma: Matsayin rashin yarda mai yawa a cikin ƙungiyar da tsakanin ƙungiyoyi akan alamun tsaro don wani yanki mai mahimmanci na abubuwan tattaunawa, yana ƙalubalantar ra'ayin ma'aunin tsaro na duniya.
• Haɗin Al'umma: Ana lura da bambance-bambance na tsari a cikin kimantawar tsaro a cikin shekaru, kabilanci/kabilanci, da layin jinsi don takamaiman batutuwa ko sautunan tattaunawa (misali, barkwanci, kai tsaye, nassoshin al'adu).
• Tasirin Haɗawa: Zaɓin dabarun haɗawa (mafi rinjaye vs. ma'auni vs. mafi ƙanƙanta) yana haifar da alamun tsaro na ƙarshe daban-daban ga kashi 15-30% na abubuwan, yana tasiri sosai kan waɗanne tattaunawar za a horar da samfurin don gujewa ko yarda.
• 
  Tazarar Kimantawar Samfuri: Samfurin da aka ɗauka "mai tsaro" ta hanyar saitin gwaji da aka haɗa mafi rinjaye na iya nuna ƙimar kuskure mai yawa (misali, +20% kuskuren marasa kyau/masu kyau) lokacin da aka kimanta shi da abubuwan da ake so na takamaiman ƙananan ƙungiyoyin al'umma.

Bayanin Chati (Ra'ayi): Chati mai fuskoki da yawa zai kasance a tsakiya don gabatar da sakamako. Panel A yana nuna taswirar zafi na matsakaicin makin tsaro (ma'auni 1-5) don abubuwan tattaunawa 100 (layuka) a cikin ƙungiyoyin al'umma 4 (ginshiƙai), yana bayyana alamu na daidaitawa da rashin yarda. Panel B ginshiƙi ne da ke kwatanta kiran "mai tsaro/mara tsaro" na ƙarshe don abubuwa 20 masu shubuha a ƙarƙashin dabarun haɗawa guda uku, yana nuna sakamakon zaɓin haɗawa a zahiri. Panel C yana tsara daidaiton samfurin don ƙungiyar mafi rinjaye da daidaitonsa ga takamaiman ƙananan ƙungiyar, tare da maki da yawa sun faɗi ƙasa da layin daidaito, yana kwatanta bambance-bambancen aiki.

5. Tsarin Bincike: Nazarin Lamari na Aiki

Yanayi: Ƙungiyar ci gaba tana daidaita mataimakin AI na tattaunawa don aikace-aikacen sabis na abokin ciniki na duniya. Suna amfani da daidaitaccen bayanan tsaro don tace bayanan horo. Yanzu suna son amfani da DICES don duba daidaitawar tsaron samfurinsu don tushen masu amfani daban-daban.

Matakan Bincike:

1. Binciken Aikin Ƙananan Ƙungiyoyi: Kunna samfurin akan faɗakarwar tattaunawar DICES. Tattara amsoshinsa da aka samar. Sami sabon tafkin masu kimantawa mai bambancin al'umma (ko kuma amfani da ainihin kimantawar DICES idan faɗakarwar iri ɗaya ce) su kimanta tsaron waɗannan tattaunawar da samfurin ya samar. Ƙididdige daidaito/maimaitawa/F1 don gano tsaro daban ga masu kimantawa a cikin Ƙungiyar A (misali, shekaru 18-30, Arewacin Amurka) da Ƙungiyar B (misali, shekaru 50+, Kudu maso Gabashin Asiya).
2. Gano Wuraren Rashin Yardar da suka fi yawa: Ware batutuwan tattaunawa ko salo inda tazarar aiki tsakanin Ƙungiyar A da Ƙungiyar B ya fi girma (misali, bambanci >30% a cikin ƙimar tsaro da aka gane). Wannan yana nuna takamaiman wuraren inda daidaitawar tsaron samfurin ba ta da ƙarfi.
3. Bincika Dabarun Haɗawa: Yi kwaikwayon daidaita samfurin ta amfani da alamun tsaro da aka samo daga DICES ta amfani da: a) Kuri'un mafi rinjaye, b) Tsarin ma'auni wanda ya wuce gona da iri na al'ummar yanki da aka yi niyya (Ƙungiyar B). Kwatanta halayen samfuran da aka samu. Tsarin DICES yana ba da bayanan da za a yi wannan zaɓin cikin ilimi maimakon komawa ga mulkin mafi rinjaye.
4. Sakamako: Ƙungiyar ta gano samfurinsu na yanzu yana da kashi 25% mafi yuwuwar samar da amsoshi da masu kimantawa daga Kudu maso Gabashin Asiya masu tsufa suka ɗauka a matsayin "mai tursasawa" ko "mara tsaro" a cikin mahallin shawarwari. Sun yanke shawarar yin amfani da aikin asarar ma'auni na al'umma yayin zagayon daidaitawa na gaba don inganta daidaitawa don wannan yanki mai mahimmanci na mai amfani.

6. Ayyukan Gaba & Hanyoyin Bincike

• Daidaitawar Tsaro mai Ƙarfi: Samfuran da zasu iya ƙididdige mahallin mai amfani/al'umma (tare da kariyar sirri da ta dace) kuma su daidaita tsaronsu/tsaronsu na tattaunawa a cikin ainihin lokaci, ta amfani da tsare-tsare kamar DICES a matsayin bincike don bambancin da aka yarda da shi.
• Daidaitawar AI Na Sirri: Tsawaita tsarin daga tsaro zuwa wasu halaye masu ra'ayi (taimako, barkwanci, ladabi) yana ba da damar masu amfani su daidaita halayen AI a cikin kewayon abubuwan da ake so waɗanda al'umma ta tabbatar.
• Manufofi & Ƙirƙirar Ma'auni: Sanar da masana'antu da ma'auni na doka don kimantawar tsaron AI. DICES yana ba da hanyar ayyana bakin kofa na "rashin yarda mai ma'ana" da kuma tilasta tantance tasirin ƙananan ƙungiyoyi, kama da binciken adalci a cikin algorithms na ɗaukar ma'aikata.
• Horo na Samfurin Tsakanin Al'adu: Yin amfani da bayanai kamar DICES a zahiri don horar da samfuran waɗanda suka san bambancin ra'ayi a sarari, mai yuwuwa ta hanyar koyo mai ayyuka da yawa ko tsarin ƙirar abin da ake so wanda aka yi wahayi ta hanyar ƙarfafawa daga amsawar ɗan adam (RLHF) amma tare da samfuran lada na ƙungiyoyi da yawa na musamman.
• Nazarin Tsawon Lokaci: Yin lura da yadda fahimtar tsaro a ciki da tsakanin al'umma ke tasowa akan lokaci don mayar da martani ga canje-canjen fasaha da zamantakewa, yana buƙatar sabunta sigogin bayanan DICES.

7. Nassoshi

1. Aroyo, L., et al. (2023). Bayanan DICES: Bambancin a cikin Kimantawar AI na Tattaunawa don Tsaro. arXiv preprint arXiv:2306.11247.
2. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). Kan Hadarin Makuɗaɗɗen Aku: Shin Manyan Samfuran Harshe Zasu iya Zama Manya? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
3. Gehman, S., Gururangan, S., Sap, M., Choi, Y., & Smith, N. A. (2020). Faɗakarwar Guba na Gaske: Kimanta Lalacewar Guba ta Jijiya a cikin Samfuran Harshe. Findings of the Association for Computational Linguistics: EMNLP 2020.
4. Ouyang, L., et al. (2022). Horar da samfuran harshe don bin umarni tare da amsawar ɗan adam. Advances in Neural Information Processing Systems, 35.
5. Prabhakaran, V., Denton, E., Webster, K., & Conover, A. (2022). Ƙirƙira, Taka tsantsan, da Haɗin Kai: Fahimta da Tallafawa Haɗin gwiwar Haɗin gwiwar Mutum-AI. Proceedings of the ACM on Human-Computer Interaction.
6. Xu, J., et al. (2020). RECAST: Ba da damar mai amfani da kuma fassarar samfuran gano guba tare da nunin nunin mu'amala. Proceedings of the ACM on Human-Computer Interaction.

8. Nazarin Kwararru: Fahimtar Asali, Tsarin Ma'ana, Ƙarfafawa & Kurakurai, Fahimtar Aiki