Binciken Ra'ayin Duniya na Samfuran Harshe don Ƙirƙirar Labarun Ƙagaggun

1. Gabatarwa

Samfuran Harshe Manya (LLMs) sun zama kayan aiki na yau da kullun a cikin kirkire-kirkire na kwamfuta, tare da ƙarin aikace-aikace a cikin ƙirƙirar labarun ƙagaggun. Duk da haka, labarun ƙagaggun suna buƙatar fiye da ƙwarewar harshe—suna buƙatar ƙirƙira da kiyaye duniyar labari mai daidaituwa wacce ta bambanta da gaskiya yayin riƙe daidaiton ciki. Wannan takarda tana bincika ko LLMs na yanzu suna da "ra'ayin duniya" ko yanayin ciki da ake buƙata don ƙirƙirar labarun ƙagaggun masu jan hankali, suna motsawa bayan kammala rubutu mai sauƙi zuwa ga ginin labari na gaskiya.

Kalubalen asali yana cikin bambanci tsakanin dawo da ilimin gaskiya da gina duniyar ƙagaggun. Yayin da LLMs suka yi fice wajen daidaita tsari da haɗa bayanai, suna fuskantar wahalar kiyaye duniyoyin madadin da suka dace—abu na asali da ake buƙata don rubuce-rubucen labarun ƙagaggun. Wannan binciken yana tantance LLMs tara bisa ma'auni na daidaito da ayyukan ƙirƙirar labari, yana bayyana manyan iyakoki a cikin gine-ginen na yanzu.

2. Tambayoyin Bincike & Hanyoyin Bincike

Binciken yana amfani da tsarin ƙima mai tsari don tantance dacewar LLMs don ƙirƙirar labarun ƙagaggun, yana mai da hankali kan iyakoki biyu masu mahimmanci.

3. Sakamakon Gwaji & Bincike

Sakamakon gwaji ya bayyana iyakoki na asali a cikin iyawar LLMs na yanzu na aiki azaman masu ƙirƙirar labarun ƙagaggun.

4. Tsarin Fasaha & Tsarin Lissafi

Ana iya tsara babban kalubalen a matsayin matsalar kula da yanayi. Bari $W_t$ ya wakilci yanayin duniya a lokacin $t$, yana ɗauke da duk gaskiyar ƙagaggun da aka kafa, halayen halaye, da ƙayyadaddun labari. Ga LLM da ke ƙirƙirar labarun ƙagaggun, muna tsammanin:

$P(response_{t+1} | prompt, W_t) \neq P(response_{t+1} | prompt)$

Wato, amsar samfurin ya kamata ta dogara da umarnin nan take da yanayin duniya da aka tara $W_t$. Duk da haka, gine-ginen na yanzu na tushen mai canzawa suna fifita da farko don:

$\max \sum_{i=1}^{n} \log P(w_i | w_{

inda $\theta$ ke wakiltar sigogin samfurin kuma $w_i$ sune alamun rubutu. Wannan manufar hasashen token na gaba ba ta ƙarfafa kiyaye $W_t$ a bayan taga mahallin nan take ba.

Binciken ya nuna cewa ƙirƙirar labarun ƙagaggun mai nasara yana buƙatar hanyoyi masu kama da waɗanda ke cikin tsarin jijiyoyi-alama ko gine-ginen ƙwaƙwalwar ajiya na waje, inda yanayin duniya $W_t$ ke nan a fili ana kiyaye shi da sabunta shi, kamar yadda aka tattauna a cikin ayyuka kamar Kwamfutar Jijiyoyi Mai Bambanta (Graves et al., 2016).

5. Nazarin Hali: Rashin Bin Yanayin Duniya

Labari: Ana umurci samfurin da ya ƙirƙiri labari game da "duniya inda nauyi ke aiki a gefe." Bayan kafa wannan jigo, umarni na gaba suna tambaya game da rayuwar yau da kullum, gine-gine, da sufuri a cikin wannan duniyar.

Lura: Yawancin samfuran suna komawa cikin sauri zuwa zato na nauyi na yau da kullum a cikin juyi 2-3 na amsa, suna saba wa jigon da aka kafa. Misali, bayan bayyana "gidaje da aka gina a cikin fuskokin dutse," samfurin na iya ambaton "faɗuwa daga gini" ba tare da gane sabani a cikin duniyar nauyi a gefe ba.

Tsarin Bincike: Ana iya samfurin wannan a matsayin gazawar bin diddigin yanayi inda wakilcin ciki na samfurin $W_t$ bai sabunta ko dage ƙayyadaddun ƙagaggun $C_{gravity} = \text{gefe}$ yadda ya kamata ba. Rarraba yuwuwar akan amsoshi a hankali yana komawa ga rarraba horo $P_{train}(\text{ra'ayoyin nauyi})$ maimakon ya kasance yana dogara da $C_{gravity}$.

Ma'ana: Ba tare da hanyoyi na musamman don kula da ƙayyadaddun ƙagaggun ba, LLMs ba za su iya zama masu ƙirƙirar labarun ƙagaggun amintattu ba, ba tare da la'akari da ƙwarewarsu ta harshe ba.

6. Ayyukan Gaba & Jagororin Bincike

Binciken ya nuna zuwa ga jagororin bincike masu ban sha'awa da yawa don inganta iyawar LLMs na ƙirƙirar labarun ƙagaggun:

• Ƙungiyoyin Yanayin Duniya na Musamman: Gine-gine waɗanda ke raba bin diddigin yanayin labari daga ƙirƙirar harshe, mai yuwuwar ta amfani da ƙwaƙwalwar ajiya na waje ko wakilcin alama.
• Horo Mai Da Hankali kan Daidaito: Manufofin daidaitawa waɗanda ke ba da lada a fili don kiyaye ƙayyadaddun ƙagaggun a cikin mahallin da aka faɗaɗa.
• Tsarin Mutum-a-cikin-Madauki: Musanya haɗin gwiwa inda mutane ke sarrafa yanayin duniya yayin da LLMs ke kula da gane harshe, kama da tsarin haɗin gwiwar da aka bincika a Yuan et al. (2022).
• Samfuran Labarun Ƙagaggun na Musamman: Horon musamman na yanki akan tarin labarun ƙagaggun da aka tsara tare da bayyana bayanin abubuwan gina duniya da baka na labari.
• Ma'aunin Ƙima: Haɓaka ma'auni na daidaitattun ma'auni don daidaiton ƙagaggun, wucewa bayan ma'aunin samfurin harshe na al'ada don tantance daidaiton labari da kula da yanayin duniya.

Waɗannan hanyoyin za su iya cike gibin tsakanin iyawar LLM na yanzu da buƙatun ƙirƙirar labarun ƙagaggun na gaske, mai yuwuwar ba da damar sabbin nau'ikan kirkire-kirkire na kwamfuta da ba da labari mai ma'amala.

7. Nassoshi

1. Graves, A., et al. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626), 471-476.
2. Patel, A., et al. (2024). Large Language Models for Interactive Storytelling: Opportunities and Challenges. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment.
3. Riedl, M. O., & Young, R. M. (2003). Character-focused narrative generation for storytelling in games. Proceedings of the AAAI Spring Symposium on Artificial Intelligence and Interactive Entertainment.
4. Tang, J., Loakman, T., & Lin, C. (2023). Towards coherent story generation with large language models. arXiv preprint arXiv:2302.07434.
5. Yuan, A., et al. (2022). Wordcraft: A Human-AI Collaborative Editor for Story Writing. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems.
6. Yang, L., et al. (2023). Improving coherence in long-form story generation with large language models. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics.

8. Ra'ayin Manazarta: Gibin Ƙirƙirar Labarun Ƙagaggun