Tsarin Robotic don Koyon Turanci tare da Samar da Rubutu na DNN

1. Gabatarwa

Yayin da Hankalin Wucin Gadi (AI) ya ƙara samun kayan aiki don fahimtar hanyoyin sadarwar ɗan adam, ƙarin cibiyoyi suna amfani da wannan fasahar a fagagen da Sarrafa Harshe na Halitta (NLP) zai iya kawo canji mai mahimmanci. Wannan takarda ta gabatar da ƙirar aiki na tsarin mutum-mutumi wanda aka ƙera don taimaka wa masu koyon Turanci ta hanyar samar da rubutu ta amfani da Hanyoyin Sadarwa na Ƙwaƙwalwar Ƙwaƙwalwa na Dogon Lokaci (LSTM).

Tsarin ya haɗa da Fuskar Mai Amfani da Hoto (GUI) wanda ke samar da rubutu bisa matakin iya Turancin mai amfani. Sakamakon gwaji da aka auna ta amfani da tsarin Gwajin Turanci na Duniya (IELTS) ya nuna kyakkyawan ci gaba a cikin kewayon nahawu tsakanin masu koyo waɗanda suka yi hulɗa da tsarin.

2. Bayanan Baya

3. Hanyar Bincike

4. Aikin Gwaji

5. Sakamako

6. Ƙarshe da Aikin Gaba

Ƙirar ta nuna yuwuwar tsare-tsaren mutum-mutumi waɗanda suka haɗa da samar da rubutu na tushen DNN don koyon harshen Turanci. Duk da yake sakamakon farko yana da ban sha'awa, ana buƙatar ƙarin gwaji don gano binciken da kuma inganta tsarin don faɗaɗa aikace-aikacen ilimi.

Aikin gaba zai mai da hankali kan faɗaɗa ƙarfin tsarin don haɗa da ƙarin fannoni na harshe, inganta daidaitawar samar da rubutu, da gudanar da manyan bincike a cikin al'ummomin masu koyo daban-daban.

7. Bincike na Asali

Wannan bincike yana wakiltar wani muhimmin haɗuwa na mutum-mutumi, sarrafa harshe na halitta, da fasahar ilimi wanda ke magance wasu muhimman kalubale a cikin tsare-tsaren koyon harshe mai cin gashin kansa. Haɗa mutum-mutumi na zahiri tare da samar da rubutu na tushen LSTM ya haifar da yanayin koyo mai yawa wanda ke amfani da alamun gani da na harshe, yana iya haɓaka riƙon ilimi ta hanyar ƙa'idodin fahimtar jiki. Kama da yadda CycleGAN (Zhu et al., 2017) ya nuna ikon koyo mara kulawa a cikin fassarar hoto, wannan tsarin yana amfani da zurfin koyo ga yankin samar da abun ciki na ilimi, kodayake tare da horo mai kulawa akan tarin harshe.

Hanyar fasaha ta amfani da hanyoyin sadarwa na LSTM tana da tushe mai kyau, kamar yadda waɗannan gine-ginen suka nuna kyakkyawan aiki a cikin ayyukan samar da jeri a cikin yankuna da yawa. Bisa ga bincike daga Ƙungiyar Sarrafa Harshe ta Lissafi, hanyoyin sadarwa na LSTM sun kasance masu inganci musamman a aikace-aikacen ilimi saboda ikon su na yin samfuri na dogon lokaci a cikin harshe. Duk da haka, fagen yana haɓaka cikin sauri zuwa gine-ginen tushen canji kamar GPT da BERT, waɗanda suka nuna mafi girman aiki a yawancin ayyukan NLP. Zaɓin LSTM a cikin wannan ƙirar na iya wakiltar sasantawa mai amfani tsakanin buƙatun lissafi da aiki, musamman idan aka yi la'akari da ƙarancin albarkatun tsare-tsaren mutum-mutumi.

Sakamakon gwaji da ke nuna haɓaka a cikin kewayon nahawu ya yi daidai da binciken da aka samu daga wasu tsare-tsaren koyon harshe da aka inganta da fasaha. Kamar yadda aka lura a cikin binciken meta daga Kimantawar Harshen Turanci ta Cambridge, tsare-tsare masu ma'amala waɗanda ke ba da amsa nan take, na mahallin sukan samar da sakamako mafi kyau a cikin samun nahawu fiye da hanyoyin gargajiya. Haɓakar band 0.5-1.0 da aka lura a cikin wannan binciken yana da mahimmanci musamman idan aka yi la'akari da ɗan gajeren lokacin shiga tsakani, yana nuna cewa ƙirar mutum-mutumi na iya haɓaka haɗin kai da ƙwaƙƙwara.

Ta fuskar aiwatarwa, tsarin yana fuskantar kalubale iri ɗaya da sauran kayan aikin ilimi masu ƙarfin AI, gami da buƙatar faɗaɗa, ingantaccen bayanin horo da daidaita matakan wahala a hankali. Ci gaba na gaba zai iya amfana daga haɗa hanyoyin koyo na canja wuri, yuwuwar daidaita ingantattun samfuran harshe akan tarin ilimi, kama da yadda kamfanonin fasahar ilimi kamar Duolingo suka sanya tsarin AI nasu. Binciken yana ba da gudummawa ga girma na shaida masu goyan bayan tsare-tsaren koyo na keɓance, masu daidaitawa, kodayake za a yi bincike na dogon lokaci don tabbatar da riƙewa na dogon lokaci da canja wurin koyo.

8. Aiwar Fasaha

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Embedding

def create_text_generation_model(vocab_size, embedding_dim, lstm_units):
    model = Sequential([
        Embedding(vocab_size, embedding_dim, input_length=50),
        LSTM(lstm_units, return_sequences=True),
        LSTM(lstm_units),
        Dense(lstm_units, activation='relu'),
        Dense(vocab_size, activation='softmax')
    ])
    
    model.compile(
        optimizer='adam',
        loss='categorical_crossentropy',
        metrics=['accuracy']
    )
    return model

# Saitin samfurin bisa matakin iyawa
MODEL_CONFIGS = {
    'mai farawa': {'embedding_dim': 128, 'lstm_units': 256},
    'matsakaici': {'embedding_dim': 256, 'lstm_units': 512},
    'cigaba': {'embedding_dim': 512, 'lstm_units': 1024}
}

def generate_text(model, tokenizer, seed_text, num_words, temperature=1.0):
    """
    Samar da rubutu ta amfani da samfurin LSTM da aka horar da zafin jiki
    """
    generated_text = seed_text
    
    for _ in range(num_words):
        # Alama da cika rubutun iri
        token_list = tokenizer.texts_to_sequences([seed_text])[0]
        token_list = tf.keras.preprocessing.sequence.pad_sequences(
            [token_list], maxlen=50, padding='pre'
        )
        
        # Hasashen kalma mai zuwa tare da zafin jiki
        predictions = model.predict(token_list, verbose=0)[0]
        predictions = np.log(predictions) / temperature
        exp_preds = np.exp(predictions)
        predictions = exp_preds / np.sum(exp_preds)
        
        # Samfurin daga rarraba yuwuwar
        probas = np.random.multinomial(1, predictions, 1)
        predicted_id = np.argmax(probas)
        
        # Canza ID zuwa kalma da haɗawa
        output_word = ""
        for word, index in tokenizer.word_index.items():
            if index == predicted_id:
                output_word = word
                break
                
        seed_text += " " + output_word
        generated_text += " " + output_word
    
    return generated_text

9. Aikace-aikacen Gaba

Fasahar da aka nuna a cikin wannan bincike tana da aikace-aikace masu ban sha'awa na gaba:

• Tsare-tsaren Koyo na Harsuna da Yawa: Miƙa hanyar zuwa harsuna da yawa ta amfani da koyo na canja wuri da haɗakar harsuna da yawa
• Ilimi na Musamman: Daidaita tsarin don masu koyo masu buƙatu na musamman, haɗa ƙarin hanyoyi kamar harshen kurame
• Horo na Kamfani: Aikace-aikace a cikin mahallin ƙwararru don harshen kasuwanci da horar da ƙwarewar sadarwa
• Koyo na Nesa: Haɗa kai tare da dandamali na zahiri da haɓakawa don gogewar koyo mai zurfi na harshe
• Kimantawa mai Daidaitawa: Amfani da bayanan hulɗa don haɓaka hanyoyin kimantawa masu zurfi da ci gaba

Hanyoyin bincike na gaba sun haɗa da haɗa gine-ginen canji, inganta hankalin tsarin ta hanyar lissafin tasiri, da haɓaka ingantattun algorithms na keɓancewa bisa binciken mai koyo.

10. Bayanan

1. Morales-Torres, C., Campos-Soberanis, M., & Campos-Sobrino, D. (2023). Ƙirar tsarin mutum-mutumi don taimaka wa tsarin koyon harshen Turanci tare da samar da rubutu ta hanyar DNN. arXiv:2309.11142v1
2. Zhu, J. Y., Park, T., Isola, P., & Efros, A. A. (2017). Fassarar hoto zuwa hoto mara haɗin gwiwa ta amfani da hanyoyin sadarwa na juyi masu dacewa. Gabatarwar taron kwamfuta na IEEE na duniya.
3. Hochreiter, S., & Schmidhuber, J. (1997). Ƙwaƙwalwar ƙwaƙwalwa na dogon lokaci. Lissafin jijiyoyi, 9(8), 1735-1780.
4. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017) Hankali shine duk abin da kuke buƙata. Ci gaba a cikin tsarin sarrafa bayanai na jijiyoyi.
5. Kimantawar Harshen Turanci ta Cambridge. (2021). Fasaha da koyon harshe: Binciken meta. Cambridge University Press.
6. Ƙungiyar Sarrafa Harshe ta Lissafi. (2022). Matsayin fasaha a cikin NLP na ilimi. Tarin tarihin ACL.
7. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., ... & Amodei, D. (2020). Samfuran harshe masu koyo kaɗan. Ci gaba a cikin tsarin sarrafa bayanai na jijiyoyi.