Teburin Abubuwan Ciki
1. Gabatarwa & Bayyani
Wannan takarda tana bincikin hanyoyin fahimi da ke haifar da wahalar masu magana na baƙi wajen sarrafa kalmomin magana. A al’ada, ana danganta waɗannan ƙalubalen ga rashin daidaitaccen tsarin sauti (phonological encoding) a cikin ƙwaƙwalwar ajiyar ƙamus. Marubutan sun ba da shawara kuma sun gwada wata hasashe: cewa yawancin abubuwan da aka lura ana iya bayyana su ta hanyar fahimtar sauti (phonetic perception) kawai, wanda ke tasowa daga daidaitawar mai magana da tsarin sautin yarensa na asali, ba tare da buƙatar wakilcin sauti na zahiri ba.
Binciken ya yi amfani da tsarin lissafi na koyon sauti, wanda aka ƙera shi da farko don fasahar magana (Kamper, 2019), don kwaikwayon sarrafa kalmomin baƙi. An horar da tsarin akan magana ta halitta, wacce ba a raba ta ba, daga yare ɗaya ko biyu kuma an kimanta shi akan ayyukan bambance-bambance na sauti da sarrafa kalmomi.
2. Babban Bincike & Hanyoyin Bincike
2.1. Tsarin Koyon Sauti
Tsarin wata hanyar sadarwar jijiyoyi ce ta kai-da-kai wacce ke koyo daga shigarwar sauti ɗaya ba tare da alamun matakin sauti ko rarrabuwa ba. Yana gina sararin wakilci na ɓoye daga bayanan magana. Muhimmanci, ba shi da wata hanya ta ciki don koyon ilimin sauti (phonology); wakilcinsa sun samo asali ne kawai daga kamancen sauti da ƙididdiga na rarrabawa.
2.2. Horar da Tsari & Bayanai
An horar da tsarin a cikin yanayi biyu: Yare Guda (Monolingual) (wanda ke kwaikwayon mai magana na asali) da Yare Biyu (Bilingual) (wanda ke kwaikwayon mai magana na baƙi mai tushen Yare na Farko). Horarwa ta yi amfani da tarin magana ta halitta. Bayanan horar da tsarin yare biyu sun haɗa yaruka biyu, wanda ya tilasta masa koyon haɗin sararin sauti.
2.3. Ayyukan Gwaji
An gwada halayen tsarin a fuskoki uku:
- Bambance-bambance a Matsakin Sauti (Phone): Shin zai iya bambanta tsakanin sautuka masu kama da juna (misali, Ingilishi /r/ da /l/)?
- Sarrafa Kalmomin Magana: Shin yana nuna tsarin "ruɗani" mai kama da na masu magana na baƙi a cikin ayyukan gane kalmomi?
- Nazarin Sararin Ƙamus: Ta yaya aka tsara kalmomi daga yaruka daban-daban a cikin sararin wakilcinsa na ciki?
3. Sakamako & Bincike
3.1. Bambance-bambance a Matsakin Sauti (Phone)
Tsarin ya yi nasara wajen maimaita sanannun wahalolin fahimtar ɗan adam. Misali, tsarin da aka horar da shi akan yare wanda ba shi da bambanci tsakanin /r/ da /l/ ya nuna rashin iya bambanta tsakanin waɗannan sautuka, yana kwatanta ƙalubalen da ɗaliban Japanawa na Ingilishi ke fuskanta.
3.2. Sarrafa Kalmomi a Matsakin Kalma
Babban binciken: Tsarin, wanda ba shi da ilimin sauti (phonology), ya nuna tasirin ruɗanin kalmomi da aka lura a cikin masu magana na baƙi. Misali, ya kunna duka "rock" da "lock" lokacin da ya ji "rock," kuma ya nuna ruɗani tsakanin kalmomin Rasha kamar "moloko" (madara) da "molotok" (guduma), ko da lokacin da bambancin sauti (/k/ da /t/) ba shi da wahala a asalinsa. Wannan yana nuna cewa kamancen sauti a cikin sararin sauti ya isa ya haifar da waɗannan tasirin.
3.3. Nazarin Sararin Wakilcin Ƙamus
Nazarin wakilcin ciki na tsarin ya bayyana cewa kalmomi daga yarukan horarwa biyu ba a raba su gaba ɗaya zuwa gungu daban-daban. A maimakon haka, sun mamaye sarari mai haɗuwa, wanda aka tsara shi ta hanyar kamancen sauti-sauti fiye da ta alamar yare. Wannan yayi daidai da binciken da aka samu a cikin ƙwaƙwalwar ajiyar ƙamus na masu yare biyu.
Mahimman Fahimta
- Fahimtar sauti, wacce aka koya daga bayyanawa, na iya bayyana wasu wahalolin sarrafa kalmomin baƙi ba tare da kiran ilimin sauti na zahiri ba.
- Halin tsarin yayi daidai da bayanan ɗan adam, yana goyan bayan ra'ayi mafi ci gaba, mai tushen misali na wakilcin ƙamus.
- Haɗin sararin ƙamus na tsarin yare biyu yana ƙalubalantar ra'ayoyin ƙa'idodi na rarrabuwar yare a cikin zuciya.
4. Cikakkun Bayanai na Fasaha & Tsarin Aiki
4.1. Tsarin Lissafi
Jigon tsarin ya ƙunshi koyon aikin saka ciki $f_\theta(x)$ wanda ke zana sashi na sauti $x$ zuwa wakilcin vector mai yawa $z \in \mathbb{R}^d$. Manufar horarwa sau da yawa ta ƙunshi asarar bambanci, kamar InfoNCE (Oord et al., 2018), wanda ke jawo tare wakilcin sassan daga kalma ɗaya (ma'aurata masu kyau) kuma ya raba sassan daga kalmomi daban-daban (ma'aurata marasa kyau):
$\mathcal{L} = -\mathbb{E} \left[ \log \frac{\exp(z_i \cdot z_j / \tau)}{\sum_{k} \exp(z_i \cdot z_k / \tau)} \right]$
inda $z_i$ da $z_j$ sune saka cikin ma'aurata masu kyau, $z_k$ sune samfuran marasa kyau, kuma $\tau$ sigar zafin jiki ce.
4.2. Misalin Tsarin Nazari
Nazarin Shari'a: Kwaikwayon Tasirin /r/-/l/ na Japan-Ingilishi
- Shigarwa: Siffofin igiyoyin sauti na kalmomin Ingilishi waɗanda ke ɗauke da /r/ da /l/.
- Matsayin Tsari: Tsarin da aka riga aka horar da shi akan Japananci kawai (wanda ba shi da wannan bambanci).
- Tsari: Tsarin yana sarrafa kalmar "rock." Aikinsa na saka ciki $f_\theta(x)$ yana zana siginar sauti zuwa wani batu $z_{rock}$ a cikin sararinsa na ɓoye.
- Nazari: Lissafa kamancen cosine tsakanin $z_{rock}$ da saka cikin wasu kalmomi ($z_{lock}$, $z_{sock}$, da sauransu).
- Sakamako: An gano kamancen tsakanin $z_{rock}$ da $z_{lock}$ ya fi na kalmomin da ba su da alaƙa sosai, yana nuna ruɗanin da sauti ya haifar. Ana iya amfani da wannan tsarin ga kowane ma'auratan kalmomi don hasashen tsarin ruɗanin baƙi.
5. Nazari Mai Zurfi & Fassarar Ƙwararru
Babban Fahimta: Wannan takarda tana ba da ƙalubale mai ƙarfi ga mulkin ilimin sauti (phonology) a cikin ilimin harshe na fahimi. Tana nuna cewa tsarin lissafi mai sauƙi, wanda ba shi da ilimin sauti, zai iya sake maimaita hadaddun halayen baƙi. Ainihin fahimta ba cewa ilimin sauti ba shi da muhimmanci ba ne, amma cewa an yi karin bayani game da bukatunsa na bayyana wasu abubuwan. Burdin hujja yanzu yana kan masu goyon bayan bayanan sauti na ƙa'ida don nuna inda tsarin sauti ya gaza a zahiri.
Tsarin Ma'ana: Hujjar tana da kyau kuma tana da sauƙi. 1) Gano rarrabuwa a cikin bayanan ɗan adam (aikin matakin sauti da na matakin kalma). 2) Yi hasashen dalili gama gari, ƙasa-ƙasa (fahimtar sauti). 3) Gina tsarin da ke tabbatar da wannan dalili kawai. 4) Nuna tsarin ya sake haifar da rarrabuwar. Wannan wata hanya ce ta "tabbatar da ra'ayi" na ƙirar ƙira, mai kama da yadda sauƙaƙan hanyoyin sadarwar jijiyoyi suka ƙalubalanci AI na alama ta hanyar nuna cewa za a iya fitar da hadaddun hali daga ƙa'idodi na asali.
Ƙarfi & Aibobi: Babban ƙarfinsa shine bayyananniyar ra'ayi da ƙaƙƙarfan ƙira. Yin amfani da tsarin da ke da iyakancefin iyawa (babu ilimin sauti) wani bincike ne mai ƙarfi na cirewa. Duk da haka, aibin yana cikin iyakar da'awar. Tsarin ya yi fice wajen bayyana ruɗani bisa kamancen sauti, amma ya yi shiru game da manyan ayyuka, halayen sauti na ƙa'ida (misali, fahimtar cewa "dogs" jam'in "dog" ne duk da bambance-bambancen sauti na zahiri). Kamar yadda masana irin su Linzen da Baroni (2021) suka yi jayayya, nasarar tsarin akan aiki ɗaya ba ta tabbatar da cewa ya kama cikakken iyawar ɗan adam ba. Takardar tana da haɗarin yin ƙarin gabaɗaya daga nasararta ta musamman.
Fahimta Mai Aiki: Ga masu bincike, wannan aikin ya tilasta sake kimanta ayyukan bincike. Idan tsarin sauti ya wuce gwaje-gwajen "sauti" na al'ada, muna buƙatar sabbin gwaje-gwaje masu tsauri waɗanda ke buƙatar taƙaice a zahiri. Ga masu haɓaka aikace-aikace a fasahar magana da koyon harshe (misali, Duolingo, Babbel), fahimtar tana da zurfi: mayar da hankali kan horon bambance-bambancen sauti mai laushi. Kayan aiki yakamata su jaddada horon fahimta akan bambance-bambancen da ke da wahala a cikin kalmomi na gaske, ba kawai gano sauti na zahiri ba. Tsarin ginin kansa, mai kama da tsarin kai-da-kai kamar Wav2Vec 2.0 (Baevski et al., 2020), ana iya daidaita shi don ƙirƙirar ƙarin kimantawa na koyon harshe na musamman da na sirri waɗanda ke nuna takamaiman matsalolin sauti ga ɗalibai ɗaya ɗaya.
6. Aikace-aikace & Hanyoyin Gaba
- Ingantattun Kayan Aikin Koyon Harshe: Haɓaka tsarin daidaitawa waɗanda ke gano takamaiman tsarin ruɗanin sauti na ɗalibi (ta amfani da tsarin kamar wannan) kuma su haifar da ayyukan sauraro da aka yi niyya.
- Fasahar Magana don Canza Lamba (Code-Switching): Inganta gane magana ta atomatik (ASR) ga masu magana da yare biyu ta hanyar ƙirar haɗin sararin sauti, maimakon tilasta raba tsarin harshe daban-daban.
- Bincike na Neurolinguistic: Yi amfani da hasashen tsarin (misali, maki kamancen tsakanin kalmomi) a matsayin masu lissafi a cikin binciken fMRI ko EEG don gwada ko aikin kwakwalwa yana da alaƙa da kamancen sauti, maimakon kamancen sauti na zahiri.
- Haɓaka Tsarin Gaba: Haɗa wannan tsarin sauti na ƙasa-zuwa-sama tare da ƙuntatawa na sauti na sama-zuwa-ƙasa a cikin gine-ginen haɗin gwiwa. Bincika ko kuma ta yaya taƙaicewar sauti ke fitowa daga irin wannan hulɗar, yana iya haɗa rata tsakanin ka'idojin misali da na zahiri.
- Aikace-aikace na Asibiti: Daidaita tsarin don ƙirar fahimtar magana a cikin jama'a masu cututtukan sauti, yana iya bambanta tsakanin nakasar sauti da na sauti na zahiri.
7. Nassoshi
- Cutler, A., & Otake, T. (2004). Pseudo-homophony in non-native listening. Proceedings of the 26th Annual Conference of the Cognitive Science Society.
- Cook, S. V., et al. (2016). The role of phonological input in second language lexical processing. Studies in Second Language Acquisition, 38(2), 225-250.
- Kamper, H. (2019). Unsupervised neural and Bayesian models for zero-resource speech processing. PhD Thesis, Stellenbosch University.
- Matusevych, Y., et al. (2020b). Modeling infant phonetic learning from natural data. Proceedings of the 42nd Annual Conference of the Cognitive Science Society.
- Oord, A. v. d., et al. (2018). Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.
- Baevski, A., et al. (2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33.
- Linzen, T., & Baroni, M. (2021). Syntactic structure from deep learning. Annual Review of Linguistics, 7, 195-212.
- Pierrehumbert, J. B. (2002). Word-specific phonetics. Laboratory Phonology VII, 101-139.