AudioBoost: Inganta Gano Littattafan Audio a cikin Binciken Spotify ta hanyar Tambayoyin Ruhani da LLM ke Samarwa

1. Gabatarwa & Bayanin Matsala

Faɗaɗa Spotify zuwa littattafan audio ya gabatar da matsala ta gargajiya ta farawa a cikin yanayin bincikenta. Tsarin samun bayanai na dandamali ya kasance mai son waƙa da faifan sauti saboda shekaru da yawa na tarin bayanan hulɗar masu amfani. Sabbin abubuwan littattafan audio sun sha wahala da ƙarancin samuwa—yiwuwar a dawo da su don tambayoyi masu dacewa—saboda ba su da alamun shiga na tarihi. Masu amfani, waɗanda suka saba bincika takamaiman waƙoƙi ko faifan sauti, ba sa tsara tambayoyi masu faɗi, na bincike (misali, "labaran ban tsoro na tunani da aka saita a cikin shekarun 80") da ake buƙata don fitar da abubuwan littattafan audio daban-daban. Wannan ya haifar da madauki mara kyau: ƙarancin ganuwa ya haifar da ƙananan hulɗa, wanda ya ƙara ƙarfafa matsayinsu na ƙasa a cikin samfuran samun bayanai.

2. Tsarin AudioBoost

AudioBoost wani tsari ne da aka tsara don karya wannan zagayowar farawa ta hanyar amfani da Manyan Samfuran Harshe (LLMs) don tayar da sararin tambaya don littattafan audio.

3. Aiwatar da Fasaha & Kimantawa

4. Fahimtar Asali

AudioBoost na Spotify ba kawai fasaha ce ta injiniya ba; juyawa ce ta dabarun yadda dandamali ya kamata suyi tunani game da gano abun ciki. Fahimtar asali ita ce a cikin tsarin sifili ko ƙananan bayanai, ba za ku iya dogara ga masu amfani su koya wa tsarin ku abin da ya dace ba. Dole ne ku yi amfani da AI mai samarwa don cika sararin niyya. Maimakon jiran tambayoyin halitta su yi kwarara—tsarin da ke son abubuwan da aka sani—AudioBoost ta fayyace abin da "tambaya mai dacewa" ga littafin audio zai iya zama. Wannan yana jujjuya tsarin bincike na gargajiya: maimakon kawai daidaita tambayoyi zuwa takardu, kuna amfani da LLMs don samar da rarraba tambaya mai yuwuwa ga kowane sabon takarda, ta haka yana tabbatar da matakin tushe na samuwa daga ranar farko. Wani nau'i ne na inganta injin bincike (SEO) da dandamali kansa ke yi, a lokacin shigarwa.

5. Tsarin Ma'ana

Tsarin ma'ana yana da sauƙi mai kyau, wanda shine dalilin da yasa yake aiki:

1. Gano Matsala: Sabon nau'in abun ciki (littattafan audio) yana da kusan sifili na samuwa saboda son hulɗa zuwa tsofaffin nau'ikan (kiɗa/faifan sauti).
2. Hasashe: Gibin yana nan a cikin sararin tambaya, ba kawai samfurin matsayi ba. Masu amfani ba su san abin da za su bincika ba, kuma tsarin ba shi da sigina don taswira tambayoyi masu faɗi zuwa sababbin abubuwa.
3. Shiga Tsakani: Yi amfani da LLM azaman "injin tunanin tambaya" dangane da metadata na abu.
4. Aiwatar da Aiki Biyu: Ciyar da tambayoyin ruhani ga duka Kammala Tambaya ta atomatik (don jagorantar masu amfani) da fihirisar samun bayanai (don tabbatar da wasanni).
5. Ƙirƙirar Zagayowar Nagarta: Ƙaruwar abubuwan da ake gani/danna suna samar da bayanan hulɗa na gaske, waɗanda sannu a hankali suka maye gurbin su kuma suka inganta sigina na ruhani, suna dumama farawar sanyi.

6. Ƙarfafawa & Kurakurai Masu Muhimmanci

Ƙarfafawa:

• Sauƙi Mai Kyau: Yana magance matsala mai sarƙaƙƙiya ta kasuwa tare da madaidaicin aikace-aikacen zamani na LLMs.
• Tunani Cikakke: Magance duka halayen mai amfani (ta hanyar QAC) da kayan aikin tsarin (ta hanyar ƙididdigewa) hanya ce ta gaba ɗaya da sau da yawa ake rasa a cikin samfuran bincike.
• Sakamako Mai Ƙarfi, Mai Aunawa: Haɓakar kusan 2% a cikin tambayoyin bincike a cikin gwajin A/B mai rai babban nasara ne don ma'aunin ɗabi'a.
• Dandamali Mai zaman kansa: Hanyar tana canzawa kai tsaye zuwa kowane dandamali na abun ciki da ke fuskantar matsalolin farawa (misali, sabbin nau'ikan samfura akan shafukan kasuwanci na e-commerce, sabbin nau'ikan bidiyo akan sabis na yawo).

• Mafarki na LLM & Rashin Daidaituwa: Babban haɗari shine LLM yana samar da tambayoyi marasa ma'ana, marasa dacewa, ko ma tambayoyi masu cutarwa. Takardar ta ambaci "inganci mai girma" amma ta ba da ƙaramin cikakken bayani kan bututun tabbatarwa. Shawarar tambaya ɗaya ta banƙyama ko ban mamaki na iya haifar da raguwar amincewar mai amfani.
• Kafaffen Wucin Gadi: Tsarin gada ne, ba makoma ba. Dogaro da bayanan ruhani na iya haifar da "kumfa na ruhani," yana jinkirta ikon tsarin don koyo daga ɗabi'ar ɗan adam ta gaske, mai sauƙi. Takardar daga Binciken Google akan "Tafarkin Bayanan Ruhani don Tsarin Shawarwari" (2023) ta yi gargadin irin waɗannan batutuwan canjin rarraba.
• Dogaro da Metadata: Ingancin tambayoyin ruhani gaba ɗaya ya dogara da wadatar da daidaiton metadata na shigarwa. Don littattafan audio tare da metadata mara yawa ko mara kyau, dabarar na iya kasawa.
• Girma & Farashi: Samar da tambayoyi masu inganci da yawa kowace abu don kasida na miliyoyin yana buƙatar farashi mai yawa na LLM. An nuna nazarin farashi-amfani amma ba a yi cikakken bayani ba.

7. Fahimta Mai Aiki

Ga shugabannin samfura da injiniyoyi, AudioBoost yana ba da littafin wasa bayyananne:

1. Bincika Samfuran Farawar ku: Nan da nan gano inda sabbin abubuwa/ƙungiyoyi a cikin tsarin ku ke kasawa saboda ƙarancin tambaya, ba kawai rashin matsayi ba.
2. Samfuri tare da LLMs na Kasuwa: Ba kwa buƙatar samfuri na al'ada don gwada wannan. Yi amfani da GPT-4 ko APIs na Claude akan samfurin kasidar ku don samar da tambayoyin ruhani kuma ku auna yuwuwar haɓakar samuwa a layi.
3. Ƙirƙirar Layer na Tabbatarwa mai Ƙarfi: Kafin fara aiki, saka hannun jari a cikin mataki mai matakai da yawa: ƙa'idodin dabara (jerin toshe), binciken kamanceceniya na tushen saka, da ƙaramin madauki na bitar ɗan adam don kama mafarkai.
4. Tsara Faɗuwar Rana: Tsara tsarin daga ranar farko don kawar da sigina na ruhani. Ai wata ma'aunin amincewa wanda ke haɗa maki na tambaya-abu na ruhani da na halitta, a hankali yana rage nauyin ɓangaren ruhani yayin da hulɗar gaske ke girma.
5. Faɗaɗa Bayan Rubutu: Gaba gaba shine samar da tambaya mai yanayi da yawa. Don littattafan audio, shin samfurin LLM-vision zai iya bincika zane-zane don samar da tambayoyi? Shin za a iya amfani da guntun sauti don samar da tambayoyi na tushen yanayi? Yi tunani fiye da metadata na rubutu.

8. Zurfin Fasaha: Kalubalen Samuwa

Takardar ta tsara matsala ta hanyar hangen nesa na samuwa, ra'ayi daga Maido da Bayanai wanda ke auna damar abu na a dawo da shi don kowace tambaya mai yuwuwa. A cikin tsarin son kai, samuwa $R(d)$ don sabuwar takarda $d_{new}$ (littafin audio) ya yi ƙasa sosai fiye da na kafaffen takarda $d_{old}$ (waƙa mai shahara). A hukumance, idan sararin tambaya $Q$ ya mamaye tambayoyi $q_i$ waɗanda ke da alaƙa da tsofaffin abubuwa, to: $$R(d_{new}) = \sum_{q_i \in Q} P(\text{dawo da } d_{new} | q_i) \cdot P(q_i) \approx 0$$ Shiga tsakani na AudioBoost yana faɗaɗa sararin tambaya mai tasiri $Q'$ don haɗa tambayoyin ruhani $q_{syn}$ waɗanda aka fayyace taswira zuwa $d_{new}$, ta haka yana haɓaka $R(d_{new})$: $$R'(d_{new}) = R(d_{new}) + \sum_{q_{syn} \in Q_{syn}} P(\text{dawo da } d_{new} | q_{syn}) \cdot P_{syn}(q_{syn})$$ inda $P_{syn}(q_{syn})$ shine yiwuwar kiyasin tambayar ruhani da aka bayar ko aka ba da shawara. Tsarin fihirisa biyu yana tabbatar da $P(\text{dawo da } d_{new} | q_{syn})$ yana da girma ta hanyar gini.

9. Sakamakon Gwaji & Jaridu

Abin da aka cire na PDF ya nuna sakamakon gwajin A/B mai rai. Za mu iya ƙididdige sakamakon mahimmanci an gabatar da su a cikin ginshiƙi ko tebur wanda ke nuna haɓakar dangi don ƙungiyar jiyya da ƙungiyar kulawa a cikin ma'auni uku na asali:

• Jadawali 1: Haɓakar Ma'auni Mai Muhimmanci: Wataƙila ginshiƙi ya nuna sanduna uku: "Abubuwan Littattafan Audio" (+0.7%), "Danna Littattafan Audio" (+1.22%), da "Kammala Tambayoyin Bincike" (+1.82%), duk tare da haɓaka mai kyau. Sandar "Kammala Tambayoyin Bincike" za ta kasance mafi tsayi, ta jaddada tasirin ɗabi'a na farko a zahiri.
• Jadawali 2: Rarraba Samuwa: Wataƙila jadawalin kimantawa na layi ya nuna rarraba tarawa na makin samuwa don littattafan audio kafin da bayan ƙara tambayoyin ruhani. "Bayan" lanƙwasa zai karkata zuwa dama, yana nuna ƙarin littattafan audio tare da mafi girman makin samuwa na tushe.
• Jadawali 3: Haɗin Nau'in Tambaya: Wataƙila ginshiƙi ko ginshiƙi mai tarawa zai iya nuna rabon nau'ikan tambaya (misali, na tushen take, na tushen marubuci, na jigo, na tushen nau'i) don littattafan audio a cikin ƙungiyoyin kulawa da jiyya, yana nuna haɓakar tambayoyin jigo/nau'i.

10. Tsarin Nazari: Madauki na Rage Farawa

AudioBoost yana aiwatar da tsarin da za a iya amfani da shi gabaɗaya don matsalolin farawa: Mataki 1 - Nazarin Gibi: Gano rukunin bayanan da aka ɓace da ke haifar da farawa (misali, nau'i-nau'i na tambaya-abu, hulɗar mai amfani-abu, fasalin abu). Mataki 2 - Shigar da Samarwa: Yi amfani da samfurin samarwa (LLM, GAN, VAE) don ƙirƙirar bayanan ruhani masu yuwuwa don rukunin da aka ɓace, dangane da bayanan gefe da ake da su (metadata). Mataki 3 - Allurar Tsarin Biyu: Allurar bayanan ruhani cikin duka fuskar mai amfani (don jagorantar ɗabi'a) da tsarin samun bayanai/matsayi na baya (don tabbatar da iyawa). Mataki 4 - Mataki na Tushen Ma'auni: Ayyana ma'aunin nasara (misali, ƙimar hulɗar halitta) da aikin lalacewa don tasirin bayanan ruhani. Yayin da ma'aunin ya inganta, a hankali rage nauyin siginar ruhani. Mataki 5 - Gyara Mai Maimaitawa: Yi amfani da sabbin bayanan halitta da aka tattara don daidaita samfurin samarwa, ƙirƙirar madauki mai inganta kansa. Wannan tsarin za a iya amfani dashi fiye da bincike: yi tunanin samar da bita na mai amfani na ruhani don sabbin samfura, ko tallan wasan kwaikwayo na ruhani don sabbin wasannin bidiyo, don tayar da gano.

11. Aikace-aikace na Gaba & Hanyoyin Bincike

Tsarin AudioBoost yana buɗe hanyoyi da yawa:

• Samar da Tambaya mai Yanayi da yawa: Yin amfani da LLMs masu yanayi da yawa don samar da tambayoyi daga guntuwar sauti (sautin mai ba da labari, yanayi), hoton zane-zane, ko ma tallan bidiyo don wasu kafofin watsa labarai.
• Tambayoyin Ruhani Na Musamman: Ƙaddamar da samar da tambaya ba kawai akan metadata na abu ba, amma akan abubuwan da mai amfani ya fi so na tarihi, samar da faɗakarwar gano na musamman (misali, "Idan kuna son Marubuci X, gwada wannan...").
• Ciyarwar Gano Mai Ƙarfafawa: Matsawa sama da bincike don fitar da nau'i-nau'i na tambaya-sakamako na ruhani a cikin ciyarwar shawarwari ("Gano littattafan audio game da...") azaman cibiyoyin bincike masu danna.
• Rage Son Kai a cikin Haɗin kai: Hanyar bincike mai mahimmanci ita ce tabbatar da cewa LLM baya ƙara ƙarfafa son kai na al'umma da ke cikin bayanan horonsa ko metadata. Dole ne a haɗa fasaha daga ML mai adalci da kuma kawar da son kai na samfuran harshe.
• Ƙwarewar Samfuri Mai Tattalin Arziki: Haɓaka ƙananan samfura, daidaitattun samfura musamman don samar da tambaya don rage farashin aiki idan aka kwatanta da amfani da manyan LLMs na gabaɗaya ga kowane abu.
• Haɗin kai tare da Binciken Tattaunawa: Yayin da binciken murya ke girma, ana iya inganta tambayoyin ruhani don tsarin yaren magana da tsayin daka, ƙarin "tambayoyi" na tattaunawa.

12. Nassoshi

1. Azad, H. K., & Deepak, A. (2019). Query-based vs. session-based evaluation of retrievability bias in search engines. Journal of Information Science.
2. White, R. W., & Drucker, S. M. (2007). Investigating behavioral variability in web search. Proceedings of WWW.
3. Boldi, P., et al. (2009). Query suggestions using query-flow graphs. Proceedings of WSDM.
4. Goodfellow, I., et al. (2014). Generative Adversarial Nets. Advances in Neural Information Processing Systems.
5. Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. Proceedings of ICML.
6. Google Research. (2023). The Pitfalls of Synthetic Data for Recommender Systems. arXiv preprint arXiv:2307.xxxxx.
7. Palumbo, E., et al. (2025). AudioBoost: Increasing Audiobook Retrievability in Spotify Search with Synthetic Query Generation. Proceedings of the EARL Workshop@RecSys.
8. OpenAI. (2023). GPT-4 Technical Report. arXiv preprint arXiv:2303.08774.