Browsing by Author "Bezayt Yewondwossen Awlachew"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Machine Learning Models for Amharic Clinical Chatbot
    (Addis Ababa University, 2024-12-01) Bezayt Yewondwossen Awlachew; Demeke Asres
    This master’s thesis focuses on experimenting with and evaluating deep learning techniques alongside various classical machine learning models for the development of an Amharic chatbot. The models examined include decision trees, support vector machines, random forests, logistic regression, Naive Bayes, multi-layer perceptron , and Bi-LSTM. Additionally, the research aims to identify the most effective feature extraction method. The methodology encompasses data collection, preprocessing, feature extraction, model training, and evaluation, with accuracy serving as the primary performance metric. The results demonstrated that using TF-IDF with hyperparameters 'n_estimators': 100, 'min_samples_split': 2, the Random Forest, SVM, and Decision Tree models achieved an accuracy of 0.9286, while Naive Bayes and Logistic Regression had accuracies of 0.6964 and 0.7679, respectively. Using CountVectorizer with the same hyperparameters, the SVM, Naive Bayes, and Logistic Regression models achieved the highest accuracy of 0.9286. The Decision Tree model followed with an accuracy of 0.8929, while the Random Forest model had an accuracy of 0.8214. Precision, recall, and F1 scores were also evaluated, with SVM, Naive Bayes, and Logistic Regression models showing consistently high performance across these metrics. These findings suggest that the choice of feature extraction technique and hyperparameter tuning significantly impact the performance of certain models. Interestingly, the MLP Classifier model outperformed the other models when using the TF-IDF feature extraction technique, achieving an accuracy of 0.9643, In General, from the experiment, we observed that the Bi-LSTM model performance is lower than the other models.