1. Gabatarwa
Wannan takarda ta gabatar da wani bincike na farko da tsarin MODOMA ya yi, wani yanayin dakin gwaje-gwaje na kwaikwayo na multi-agent don gwaje-gwajen samun harshe ba tare da kulawa ba. Tsarin yana kwaikwayon mu'amalar iyaye da yaro inda dukkanin agents suke samfuran harshe tare da wakilcin ilimin nahawu a bayyane. Ba kamar manyan samfuran harshe (LLMs) waɗanda suka dogara da hanyoyin sadarwa na jijiyoyi masu duhu ba, MODOMA yana ba da tsarin nahawu na gaskiya, mai iya dawo da shi.
2. Babban Ra'ayi: Tsarin MODOMA
Tsarin MODOMA (moeder-dochter-machine) yanayi ne na kwaikwayo wanda aka daidaita shi gabaɗaya. Agent uwa yana samar da kalmomi ta amfani da ƙa'idodin harshe na gaskiya, yayin da agent yaro ke amfani da hanyoyin ƙididdiga don ƙaddamar da samfurin tushen ƙa'ida na harshen da ake nufi. Wannan hanya ta haɗa tsarin tushen ƙa'ida da tsarin ƙididdiga.
2.1 Tsarin Multi-Agent
Tsarin yana aiwatar da zagayowar mu'amalar iyaye da yaro. Agent uwa yana samar da misalai, kuma agent yaro yana sabunta wakilcin nahawunsa bisa ga abin da ya karɓa. Dukkan hanyoyin ana rubuta su, wanda ke ba da damar gano cikakken tsarin samun harshe.
2.2 Wakilcin Ilimi a Bayyane
Dukkanin agents suna riƙe da wakilcin nau'ikan nahawu (misali, suna, fi'ili, mai tantancewa) da ƙa'idodi a bayyane. Wannan ya bambanta MODOMA daga samfuran jijiyoyi waɗanda ke ɓoye ilimi a cikin ma'aunin nauyi.
3. Tsarin Tunani: Tsarin Gwaji
Binciken yana bincika ko agent 'yar za ta iya samun nau'ikan aiki da abun ciki daga bayanan horo da agent babba ya samar. Gwaje-gwaje sun bambanta yawan misalan da aka bayar.
3.1 Bayanan Horo da Gwaji
Agent babba yana samar da kalmomi masu rikitarwa daban-daban. Agent yaro yana karɓar waɗannan kalmomi kuma yana ƙoƙarin ƙaddamar da nau'ikan nahawu. Bayanan gwaji suna kimanta daidaiton nahawun da aka samu.
3.2 Ma'aunin Kimantawa
An auna nasarar samun harshe ta hanyar ikon agent yaro na rarraba kalmomi daidai da samarwa/fassara sabbin kalmomi. Sakamako ya nuna alamu masu kama da samun harshe na ɗan adam, tare da inganta aiki yayin da yawan misalai ya ƙaru.
4. Ƙarfi da Rauni: Bincike Mai Zurfi
Ƙarfi: Wakilcin ilimin nahawu a bayyane babbar fa'ida ce akan LLMs masu duhu. Tsarin da aka daidaita yana ba da damar gwaje-gwaje masu sarrafawa. Mu'amalar multi-agent tana kwaikwayon koyo na halitta.
Rauni: Gwaje-gwajen na yanzu sun iyakance ga tsarin nahawu masu sauƙi. Ƙarfin haɓaka zuwa harshe mai rikitarwa na duniya bai tabbata ba. Dogaro da ƙa'idodin da aka yi da hannu don agent uwa na iya haifar da son zuciya.
5. Ra'ayoyi Masu Aiki: Abubuwan da ke Tattare da NLP
MODOMA yana ba da madadin gaskiya ga samfuran harshe na jijiyoyi don nazarin samun harshe. Masu bincike za su iya amfani da shi don gwada ka'idodin harshe ta hanyar kwamfuta. Ana iya faɗaɗa tsarin don kwaikwayon harshe biyu ko cututtukan harshe.
6. Cikakkun Bayanai na Fasaha da Tsarin Lissafi
Ana iya tsara algorithm na samun harshe a matsayin matsalar ƙaddamar da nahawu mai yiwuwa. Bari $G$ ya zama nahawu tare da nau'ikan $C$ da ƙa'idodin $R$. Agent yaro yana sabunta imaninsa akan $G$ bisa ga kalmomin da aka lura $U$:
$$P(G|U) \propto P(U|G) P(G)$$
inda $P(U|G)$ shine yiwuwar samar da $U$ a ƙarƙashin $G$, kuma $P(G)$ shine fifiko akan nahawu. Agent yaro yana amfani da tsarin bincike na Bayesian don ƙididdige bayanai.
7. Sakamakon Gwaji da Bayanin Zane
Hoto na 1 (tunani): Ginshiƙi mai nuna daidaiton samun harshe (y-axis) vs. yawan misalan horo (x-axis). Daidaito yana ƙaruwa daga ~40% tare da misalai 50 zuwa ~85% tare da misalai 500, tare da tsayawa bayan misalai 300. Sandunan kuskure suna nuna bambancin aiki.
Tebur na 1: Daidaiton samun nau'ikan kalmomi daban-daban: sunaye (92%), fi'ili (88%), masu tantancewa (95%), prepositions (78%). Agent yaro yana yin mafi kyau akan nau'ikan aiki masu yawan amfani.
8. Misalin Tsarin Bincike: Nazari
Yi la'akari da harshe mai sauƙi kamar Ingilishi tare da nau'ikan: D (mai tantancewa), N (suna), V (fi'ili). Agent uwa yana samar da kalmomi kamar "the cat runs" (D N V). Agent yaro yana karɓar wannan kuma yana hasashen nau'ikan. Bayan misalai da yawa, yana koyon cewa "the" mai tantancewa ne, "cat" da "dog" sunaye ne, kuma "runs" da "sleeps" fi'ili ne. Nahawun da aka samu zai iya fassara sabon shigarwa kamar "a dog sleeps".
9. Aikace-aikace da Hanyoyi na Gaba
Ana iya faɗaɗa MODOMA don kwaikwayon samun harshe na biyu, canza harshe, da rawar mu'amalar zamantakewa a cikin koyo. Haɗin kai tare da sassan jijiyoyi na iya haɗa mafi kyawun tsarin biyu. Tsarin kuma yana da yuwuwar a cikin fasahar ilimi don koyar da harshe na musamman.
10. Bincike na Asali
Tsarin MODOMA yana wakiltar wani babban canji daga manyan samfuran harshe na jijiyoyi ta hanyar ba da fifiko ga gaskiya da wakilcin nahawu a bayyane. Yayin da LLMs kamar GPT-3 (Brown et al., 2020) ke samun nasara mai ban sha'awa, ayyukansu na ciki sun kasance a ɓoye. Hanyar MODOMA ta yi daidai da kira mai girma don AI mai fassara a ilimin harshe (Baroni, 2022). Nasarar samun nau'ikan keɓantacce yana nuna sakamako a cikin ci gaban harshen yara (Tomasello, 2003), yana tabbatar da ingancin yanayin kwaikwayon. Duk da haka, dogaron tsarin akan ƙa'idodin da aka yi da hannu don agent uwa yana iyakance ƙarfin haɓakarsa. Aikin gaba ya kamata ya bincika ƙaddamar da ƙa'ida ta atomatik daga tarin bayanan halitta. Wakilcin ilimin nahawu a bayyane kuma yana buɗe hanyoyi don kwatanta harsuna daban-daban, saboda harsuna daban-daban na iya buƙatar tsarin nau'ikan daban-daban. Wannan aikin yana cika bincike kan ƙaddamar da nahawu ta amfani da samfuran Bayesian (Perfors et al., 2011) kuma yana ba da filin gwaji don ka'idodin harshe. Tsarin MODOMA zai iya zama mai mahimmanci musamman don nazarin hasashen lokaci mai mahimmanci da rawar yawan shigarwa a cikin samun harshe.
11. Manazarta
- Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. NeurIPS.
- Baroni, M. (2022). On the proper role of linguistically-oriented deep net analysis in linguistic theorizing. In Algebraic Structures in Natural Language.
- Tomasello, M. (2003). Constructing a Language: A Usage-Based Theory of Language Acquisition. Harvard University Press.
- Perfors, A., Tenenbaum, J. B., & Regier, T. (2011). The learnability of abstract syntactic principles. Cognition, 118(3), 306-338.
- Devlin, J., et al. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL.