Nazarin Koyon Zurfin (Deep Learning) don Rarrabe Halayen Tausayi a cikin Gajerun Rubutun Turanci: Nazari & Tsarin Aiki

Tsarin Abubuwan Ciki

• 1. Gabatarwa & Bayyani
• 2. Hanyoyin Bincike & Tsarin Fasaha
  • 2.1 Tsarin Koyon Zurfi
  • 2.2 Bayanan SmallEnglishEmotions
  • 2.3 Horar da Model & Koyon Canja Wuri
• 3. Sakamakon Gwaji & Nazari
  • 3.1 Ma'aunin Aiki
  • 3.2 Nazarin Kwatance
• 4. Muhimman Fahimta & Tattaunawa
• 5. Cikakkun Bayanan Fasaha & Tsarin Lissafi
• 6. Tsarin Nazari: Misalin Nazarin Shari'a
• 7. Aikace-aikace na Gaba & Hanyoyin Bincike
• 8. Nassoshi

1. Gabatarwa & Bayyani

Wannan binciken ya magance babban kalubalen gano halayen tausayi a cikin gajerun rubutun Turanci, wani fanni da ke da wahala saboda ƙarancin bayanan mahallin da kuma ƙayyadaddun harshe. Yaduwar kafofin sada zumunta da sadarwar dijital sun haifar da dimbin bayanan rubutu na gajeren lokaci inda fahimtar halayen tausayi ke da muhimmanci ga aikace-aikace tun daga sa ido kan lafiyar hankali har zuwa nazarin ra'ayoyin abokan ciniki da kuma hako ra'ayin jama'a. Nazarin ra'ayi na al'ada sau da yawa ya kasa kama ƙayyadaddun halayen tausayi kamar farin ciki, baƙin ciki, fushi, tsoro, da mamaki a cikin taƙaitaccen rubutu.

Binciken ya gabatar da kuma tantance ƙwararrun dabarun koyon zurfi, tare da mai da hankali musamman kan samfuran da suka dogara da transformer kamar BERT (Bidirectional Encoder Representations from Transformers) da dabarun koyon canja wuri. Babbar gudummawar shi ne gabatar da bayanan SmallEnglishEmotions, wanda ya ƙunshi gajerun rubutu 6,372 da aka yiwa lakabi a cikin manyan nau'ukan halayen tausayi guda biyar, wanda ke aiki a matsayin ma'auni ga wannan takamaiman aiki.

2. Hanyoyin Bincike & Tsarin Fasaha

2.1 Tsarin Koyon Zurfi

Binciken ya yi amfani da mafi kyawun tsarin koyon zurfi. Babban samfurin ya dogara ne akan BERT, wanda ke amfani da tsarin transformer don samar da haɗakar mahalli don kowace alama a cikin rubutun shigarwa. Ba kamar haɗakar kalmomi na tsaye ba (misali, Word2Vec, GloVe), BERT yana la'akari da cikakken mahallin kalma ta hanyar duba kalmomin da suka gabata da kuma bayanta. Wannan yana da ƙarfi musamman ga gajerun rubutu inda dangantakar kowace kalma ke da muhimmanci. An daidaita samfurin akan aikin rarrabe halayen tausayi, yana daidaita iliminsa na harshe da aka riga aka horar da shi don gane alamun tausayi.

2.2 Bayanan SmallEnglishEmotions

Don rage rashin ƙayyadaddun albarkatu don nazarin halayen tausayi na gajerun rubutu, marubutan sun tsara bayanan SmallEnglishEmotions. Ya ƙunshi samfura 6,372, kowanne jimla ko jumla gajere ce ta Turanci, wanda aka yiwa lakabi da hannu ɗaya daga cikin alamun halayen tausayi guda biyar. An tsara bayanan don nuna iri-iri da taƙaitaccen lokaci da ake samu a cikin tushen duniya na gaske kamar tweets, bita na samfura, da saƙon hira. Wannan bayanan ya magance gibi da aka lura a cikin aikin da ya gabata, wanda sau da yawa yakan yi amfani da bayanan da ba a inganta su don ƙalubalen musamman na gajeren tsawon rubutu ba.

2.3 Horar da Model & Koyon Canja Wuri

Koyon canja wuri shine ginshiƙin hanyar. Maimakon horar da samfuri daga farko, wanda ke buƙatar dimbin bayanan da aka yiwa lakabi, tsarin yana farawa da samfurin BERT da aka riga aka horar da shi akan babban tarin rubutu (misali, Wikipedia, BookCorpus). Wannan samfurin ya riga ya fahimci tsarin harshe na gaba ɗaya. Daga nan sai a daidaita shi akan bayanan SmallEnglishEmotions. Yayin daidaitawa, ana ɗan gyara sigogin samfurin don ƙware wajen bambanta tsakanin halayen tausayi guda biyar da aka yi niyya, yana yin amfani da ingantaccen bayanan da aka yiwa lakabi da aka samu.

3. Sakamakon Gwaji & Nazari

3.1 Ma'aunin Aiki

An kimanta samfuran ta amfani da ma'auni na rarrabuwa na yau da kullun: daidaito, daidaito, tunawa, da maki-F1. Samfurin da ya dogara akan BERT ya sami mafi kyawun aiki a duk ma'auni idan aka kwatanta da samfuran tushe kamar masu rarrabe injin koyon gargajiya (misali, SVM tare da fasalin TF-IDF) da hanyoyin sadarwar jijiyoyi masu sauƙi (misali, GRU). Maki-F1, wanda ke daidaita daidaito da tunawa, ya fi girma musamman ga BERT, yana nuna ƙarfinsa wajen sarrafa rashin daidaiton aji da kuma ƙayyadaddun bayyanar halayen tausayi.

3.2 Nazarin Kwatance

Gwaje-gwajen sun nuna tsarin aiki bayyananne:

1. BERT tare da Daidaitawa: Mafi girman daidaito da maki-F1.
2. Sauran Samfuran Transformer (misali, XLM-R): Masu gasa amma aikin ɗan ƙasa kaɗan, mai yiwuwa saboda ƙarancin ingantaccen horo na farko don wannan takamaiman fanni.
3. Hanyoyin Sadarwar Jijiyoyi na Maimaitawa (GRU/LSTM): Matsakaicin aiki, suna fama da dogon lokaci na dogaro a wasu gine-gine.
4. Samfuran ML na Gargajiya (SVM, Naive Bayes): Mafi ƙarancin aiki, yana nuna iyakacin jakar kalmomi da fasalin n-gram don kama ma'anar tausayi a cikin gajerun rubutu.

Bayanin Ginshiƙi (An yi tunani daga Mahallin Rubutu): Taswirar ginshiƙi za ta iya nuna "Daidaiton Model" akan Y-axis da sunayen samfura daban-daban (BERT, XLM-R, GRU, SVM) akan X-axis. Ginshiƙin BERT zai fi sauran girma sosai. Wani jadawalin layi na biyu zai iya nuna maki-F1 kowane aji na tausayi, yana nuna cewa BERT yana ci gaba da samun maki masu girma a duk halayen tausayi guda biyar, yayin da sauran samfuran za su iya raguwa sosai ga azuzuwan kamar "Tsoro" ko "Mamaki" waɗanda ba su da yawa ko kuma sun fi sauƙi.

4. Muhimman Fahimta & Tattaunawa

Babban Fahimta: Gaskiyar da ba a faɗi ba amma bayyananne a cikin takardar ita ce, zamanin ƙwararrun ƙira na ƙira don ayyukan NLP masu ƙayyadaddun ƙira kamar gano tausayi ya ƙare tabbatacce. Dogaro da TF-IDF ko ma haɗakar kalmomi na tsaye don gajeren rubutu kamar yin amfani da taswirar layin ƙasa don jagorar GPS na ainihin lokaci—yana ba da ma'auni amma ya rasa duk mahallin. Mafi kyawun aikin BERT ba kawai ci gaba ne kawai ba; canjin tsari ne, yana tabbatar da cewa fahimtar zurfin ma'ana, mai sanin mahalli, ba za a iya sasantawa ba don fassara halayen tausayi na ɗan adam a cikin rubutu, musamman lokacin da kalmomi suka yi ƙaranci.

Kwararar Hankali & Ƙarfuka: Hankalin binciken yana da inganci: gano gibi (bayanan halayen tausayi na gajerun rubutu), ƙirƙirar albarkatu (SmallEnglishEmotions), da kuma amfani da kayan aiki mafi ƙarfi na yanzu (BERT/daidaitawa). Ƙarfinsa yana cikin wannan ingantacciyar hanyar, ƙarshen-zuwa-ƙarshe. Bayanan, ko da yake matsakaici ne, gudummawa ce mai mahimmanci. Zaɓin BERT yana da hujja mai kyau, yana daidaitawa da babban yanayin a cikin NLP inda samfuran transformer suka zama ma'auni na zahiri, kamar yadda suka yi mulki a cikin ma'auni kamar GLUE da SuperGLUE.

Kurakurai & Ra'ayi Mai Muhimmanci: Duk da haka, takardar tana sanye da makafi. Tana ɗaukar BERT a matsayin harsashi na azurfa ba tare da isasshen fuskantar babban farashin lissafinsa da jinkiri ba, wanda ke da muhimmanci ga aikace-aikace na ainihin lokaci kamar chatbots ko daidaita abun ciki. Bugu da ƙari, samfurin halayen tausayi guda biyar yana da sauƙi. Halayen tausayi na duniya na gaske sau da yawa suna haɗuwa (misali, farin ciki mai ɗaci), wani sarkakiya da samfuran kamar EmoNet ko samfuran ma'auni (valence-arousal) ke ƙoƙarin kama. Takardar kuma ta kau da kai daga batun mahimmanci na son zuciya—samfuran BERT da aka horar da su akan faɗin bayanan intanet na iya gada da haɓaka son zuciya na al'umma, matsala da aka rubuta sosai a cikin binciken ɗa'a na AI daga cibiyoyi kamar Cibiyar AI Now.

Fahimta Mai Aiki: Ga masu aiki, saƙon yana bayyananne: fara da tushen transformer (BERT ko zuriyarsa mafi inganci kamar DistilBERT ko ALBERT) kuma a daidaita shi akan bayanan takamaiman fanninku. Duk da haka, kada ku tsaya a nan. Mataki na gaba shine gina bututun kimantawa waɗanda ke gwada musamman don son zuciya a cikin ƙungiyoyin alƙaluma da kuma bincika ƙarin ƙayyadaddun harajin halayen tausayi. Gaba ba kawai game da mafi girman daidaito akan matsalar aji 5 ba ne; yana game da gina samfuran da za a iya fassara su, masu inganci, da adalci waɗanda suka fahimci cikakken yanayin halayen tausayi na ɗan adam.

5. Cikakkun Bayanan Fasaha & Tsarin Lissafi

Jigon rarrabuwar BERT ya ƙunshi ɗaukar yanayin ɓoye na ƙarshe na alamar [CLS] (wanda ke tattara bayanan jeri) da wucewa ta hanyar layin hanyar sadarwar jijiyoyi na ciyarwa don rarrabuwa.

Don wani tsarin rubutun shigarwa da aka bayar, BERT yana samar da haɗakar mahalli don alamar [CLS], wanda aka nuna shi da $\mathbf{C} \in \mathbb{R}^H$, inda $H$ shine girman ɓoye (misali, 768 don BERT-base).

Yuwuwar cewa rubutun ya kasance cikin aji na halayen tausayi $k$ (daga cikin azuzuwan $K=5$) ana ƙididdige shi ta amfani da aikin softmax: $$P(y=k | \mathbf{C}) = \frac{\exp(\mathbf{W}_k \cdot \mathbf{C} + b_k)}{\sum_{j=1}^{K} \exp(\mathbf{W}_j \cdot \mathbf{C} + b_j)}$$ inda $\mathbf{W} \in \mathbb{R}^{K \times H}$ da $\mathbf{b} \in \mathbb{R}^{K}$ su ne nauyin da son zuciya na layin rarrabuwa na ƙarshe, waɗanda aka koya yayin daidaitawa.

An horar da samfurin ta hanyar rage asarar giciye-entropy: $$\mathcal{L} = -\frac{1}{N} \sum_{i=1}^{N} \sum_{k=1}^{K} y_{i,k} \log(P(y_i=k | \mathbf{C}_i))$$ inda $N$ shine girman rukuni, kuma $y_{i,k}$ shine 1 idan samfurin $i$ yana da ainihin lakabin $k$, kuma 0 in ba haka ba.

6. Tsarin Nazari: Misalin Nazarin Shari'a

Yanayi: App ɗin lafiyar hankali yana son tantance shigarwar mujallar mai amfani don alamar rikice-rikice masu yuwuwa ta hanyar gano halayen tausayi mara kyau.

Aikace-aikacen Tsarin:

1. Shirya Bayanai: Tattara kuma yiwa lakabin saitin gajerun shigarwar mujalla tare da lakabi kamar "babban damuwa," "matsakaicin baƙin ciki," "tsaka tsaki," "tabbatacce." Wannan yayi daidai da ƙirƙirar bayanan SmallEnglishEmotions.
2. Zaɓin Model: Zaɓi samfurin da aka riga aka horar da shi kamar bert-base-uncased. Idan aka yi la'akari da hankalin fannin, samfurin kamar MentalBERT (wanda aka riga aka horar da shi akan rubutun lafiyar hankali) zai iya zama mafi tasiri, bin hankalin koyon canja wuri na takardar.
3. Daidaitawa: Daidaita samfurin da aka zaɓa akan sabon bayanan shigarwar mujalla. Madauki na horo yana rage asarar giciye-entropy kamar yadda aka bayyana a Sashe na 5.
4. Kimantawa & Tura: Kimanta ba kawai akan daidaito ba, amma a mahimmanci akan tunawa ga aji na "babban damuwa" (rashin siginar rikici yana da tsada fiye da ƙararrawar ƙarya). Tura samfurin a matsayin API wanda ke ƙididdige sabbin shigarwa a ainihin lokaci.
5. Saka idanu: Ci gaba da saka idanu kan tsinkayar samfurin da tattara ra'ayoyi don sake horarwa da rage karkata, tabbatar da cewa samfurin ya kasance daidai da harshen mai amfani akan lokaci.

Wannan nazarin shari'ar yana nuna yadda hanyar takardar ke ba da takamaiman tsari mai aiki don gina aikace-aikacen duniya na gaske.

7. Aikace-aikace na Gaba & Hanyoyin Bincike

Aikace-aikace:

• Tallafin Lafiyar Hankali na Ainihin Lokaci: Haɗa shi cikin dandamalin kiwon lafiya ta hanya da app ɗin walwala don ba da nazarin yanayin tausayi nan take da kuma haifar da albarkatun tallafi.
• Ingantaccen Ƙwarewar Abokin Ciniki: Nazarin rajistan hira na tallafi, bita na samfura, da ambaton kafofin sada zumunta don auna halayen tausayi na abokin ciniki a ma'auni, yana ba da damar sabis na gaggawa.
• Daidaita Abun Ciki & Tsaro: Gano maganganun ƙiyayya, cin zarafi ta kan layi, ko niyyar cutar da kai a cikin al'ummomin kan layi ta hanyar fahimtar tashin hankali ko yanke ƙauna a cikin saƙonni.
• Nishadantarwa Mai Mu'amala & Wasanni: Ƙirƙirar NPCs (Halayen da Ba 'yan wasa ba) ko labarai masu mu'amala waɗanda ke amsawa da sauri ga sautin tausayi na ɗan wasa da aka bayyana a cikin shigarwar rubutu.

Hanyoyin Bincike:

• Gane Halayen Tausayi ta Hanyoyi Daban-daban: Haɗa rubutu tare da sautin sauti (a cikin saƙon murya) da bayyanar fuska (a cikin sharhin bidiyo) don cikakkiyar ra'ayi, kama da ƙalubale da hanyoyin da ake gani a cikin binciken koyo ta hanyoyi daban-daban.
• AI Mai Bayyanawa (XAI) don Samfuran Halayen Tausayi: Haɓaka dabarun haskaka waɗanne kalmomi ko jimloli ne suka fi ba da gudummawa ga tsinkayar halayen tausayi, gina aminci da ba da fahimta ga likitoci ko masu daidaitawa.
• Samfuran Mai Sauƙi & Mai Inganci: Bincike cikin tace manyan samfuran transformer zuwa ƙananan sigogi, masu sauri waɗanda suka dace da na'urorin hannu da na gefe ba tare da asarar aiki mai mahimmanci ba.
• Canja Harshe & Daidaitawar Ƙarancin Albarkatu: Ƙara nasarar koyon canja wuri zuwa harsuna masu ƙarancin albarkatu na gaske tare da ƙaramin bayanan da aka yiwa lakabi, mai yiwuwa ta amfani da dabarun koyo kaɗan ko sifili.

8. Nassoshi

1. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805.
2. Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., & Bowman, S. R. (2018). GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. Proceedings of the 2018 EMNLP Workshop BlackboxNLP.
3. AI Now Institute. (2019). Disability, Bias, and AI. Retrieved from https://ainowinstitute.org/
4. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. Proceedings of the IEEE International Conference on Computer Vision (ICCV). (An ambata a matsayin misali na ingantaccen tsarin koyon zurfi a wani fanni daban).
5. Poria, S., Cambria, E., Bajpai, R., & Hussain, A. (2017). A review of affective computing: From unimodal analysis to multimodal fusion. Information Fusion, 37, 98-125.
6. Bhat, S. (2024). Emotion Classification in Short English Texts using Deep Learning Techniques. arXiv preprint arXiv:2402.16034.