http://hdl.handle.net/123456789/3074 2026-04-15T10:26:56Z http://hdl.handle.net/123456789/8784 Title: Detection and Correction of Amharic Grammar Errors Using Deep Learning Authors: Fantaye, Tewodros Abstract: Effective communication in natural or human language relies heavily on grammatical accuracy. As a result, natural language processing (NLP) has emerged as a critical area of research, aiming to enhance computer’s ability to comprehend and interact using human language. A sentence is considered grammatically correct when its word structure adheres to rules governing number, person, gender, tense, and other grammatical agreements. Numerous studies have explored various languages and grammatical frameworks to develop methods for verifying the grammatical accuracy of sentences. The main aim of this study is to create and execute a system based on deep learning for identifying and rectifying grammatical mistakes in the Amharic language. The suggested method employs a Bidirectional Long Short-Term Memory (BiLSTM) Recurrent Neural Network (RNN), developed using Python 3.7, with Keras and TensorFlow serving as the backend. The evaluation of the BiLSTM model revealed an accuracy of 88.89%, along with a recall of 88.89%, precision of 89%, and an F1 score of 89%. A significant challenge faced during this research was dealing with the complexity of Amharic words, which can have multiple meanings or reflect different levels of respect, thereby introducing ambiguity into the error detection and correction process. To enhance the effectiveness of detecting and correcting grammatical errors, it is crucial to include a comprehensive dataset of morphologically annotated sentences. Furthermore, future investigations should aim to refine the model and examine alternative methodologies to improve the system's overall performance and accuracy. 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/8783 Title: ATM TRANSACTION FAILURE DETECTION USING MACHINE LEARNING Authors: Yeshambele, Zemenu Abstract: Automated Teller Machine (ATM) services represent a critical component of modern banking strategies aimed at enhancing service quality. However, the performance of these services often falls short of expectations, leading to customer dissatisfaction and revenue losses for financial institutions. This research focuses on creating a predictive model to identify failures in ATM transactions through the application of machine learning techniques. Transaction failures can arise from various factors, including technical issues with the ATM itself and customer-related problems, resulting in frustration for users and the necessity for frequent reconciliation by banks. Prior studies have not sufficiently tackled the issue of detecting transaction failures using machine learning methods. Therefore, this research utilizes machine learning algorithms to recognize and forecast transaction failures. Data for this research was collected from Zemen Bank's ATM reconciliation records and CR2 database, encompassing a total of 20,516 transaction datasets for a selected month in 2024. The research employs supervised machine learning techniques, applying a range of classification algorithms such as Logistic Regression, Decision Trees, Random Forest, K-Nearest Neighbors, Gaussian Naïve Bayes, and Support Vector Classifier (SVC). The experiments were carried out using Python 3.7 within the Jupyter Notebook environment (Anaconda distribution). The Random Forest algorithm was employed to create the predictive model for identifying ATM transaction failures, which achieved an accuracy of 96.78%, while the Naïve Bayes classifier demonstrated the least performance with an accuracy of 86.94%. The findings indicate that machine learning techniques can effectively predict ATM transaction statuses, thereby enhancing the management of ATM operations. Key factors influencing transaction failures include the available balance, the amount of cash requested, the ATM's location, the time of the transaction, the type of transaction, and whether the transaction occurred on a working day. This research highlights the promise of machine learning techniques in enhancing the reliability and efficiency of ATM services, providing important insights for financial institutions aiming to streamline their transaction operations. 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/8782 Title: ASPECT-BASED SENTIMENT ANALYSIS FOR AMHARIC AND ENGLISH COMMENTS FROM ETHIO TELECOM FACEBOOK AND TWITTER PAGES USING DEEP LEARNING Authors: Abawa, Wudie Abstract: A vital tool for comprehending public opinion in a variety of fields, such as customer service and business decision-making, is sentiment analysis. In this study, user comments from Ethio Telecom Facebook and Twitter pages in both Amharic and English are analyzed for sentiment. The primary aim is to classify these comments into distinct sentiment categories such as positive, negative, or neutral, providing actionable insights to improve customer satisfaction and service delivery. This work was to develop a bilingual sentiment analysis model using written comments from Ethiopian telecom platforms on Facebook and Twitter in both Amharic and English To address the unique linguistic and morphological challenges of Amharic, the study incorporates specialized preprocessing steps, tokenization methods, and embedding’s. A balanced dataset of annotated comments in both languages is compiled for training and evaluation. The results demonstrate the effectiveness of deep learning models in capturing sentiment across both languages, achieving high accuracy and robustness. A total of 13,389 comments were collected, preprocessed, and manually labeled. In terms of language distribution, 52.91% (7,084 comments) were in pure Amharic, 28.75% (3,850 comments) in pure English, and 18.34% (2,455 comments) were mixed-language comments. Data sampling techniques, feature extraction using word representation techniques like Word2Vec, GloVe, and FastText, and deep learning architectures like Long Short-Term Memory networks (LSTMs) and Gated Recurrent Units (GRUs) were all used in the study. Metrics like accuracy, precision, recall, and F1-score were used to evaluate the models, and by achieving an accuracy of 74.38% and an F1-score of 74.12% in the train test split, LSTM was the best performer. While GRU models showed lower performance with accuracies of 73.67% and an F1-score of 70.62% in the 80% training and 20% of the dataset test set. The LSTM model demonstrated the most consistent and robust performance train-test splitting methods, making it the best choice for this bilingual sentiment analysis task. Based on these experimental results, the LSTM model with train test split is recommended for analyzing the sentiment of bilingual social media comments, ensuring consistent and generalizable results. 2025-01-01T00:00:00Z http://hdl.handle.net/123456789/8781 Title: Sentiment analysis based- Hate speech detection of Amharic social media post and comments Authors: Shita, Tsion Abstract: Social media allows the user to post, comment and communicate freely. And this led to an increasing amount of online hate speech. Online hate speech has different offline repercussions, according to studies. In recent years, hate speech have led to internal violence, relocation, and human rights violations against specific social groups around the world. And Ethiopian societies are among the victims. To lessen the spread of hate speech, this study develops Amharic hate speech detection. The study's main goal is to create a model for detecting hate speech by taking into account sentiment analysis of the relevant datasets and proving a link between hate speech and sentiment analysis. Peacemakers can take action when hate speech comments are being circulated online by using an Amhariclanguage hate speech detection system. Additionally, it will assist owners of social media platforms by automatically reporting hate speech remarks before they are seen by a wider audience Comments were gathered from Facebook, TikTok, and YouTube channels in order to create a labeled large Amharic dataset. Following data cleaning, 79991 hate and hate-free annotated datasets along with their sentiment were obtained. To label the dataset as hate and hate-free, new annotation guidelines were created. Despite previous related work, recent and large dataset were collected and their sentiment were also considered. To construct the model, CNN and GRU deep learnings were used in conjunction with Word embedding features. Negative sentiment was revealed to be the source of hate speech content. And most of the hate free dataset were found to be having positive sentiment. Using datasets that have been annotated by humans as a hate and hate free, the GRU and CNN models demonstrated respective accuracies of 0.90 and 0.72. And when both hate and non-hate annotated datasets along with their sentiment were used in the hate speech detection model, the models' respective accuracies become 0.75 and 0.74. As a result, in both model GRU outperform CNN model, and the CNN approach shows good performance for the hate speech detection model that was developed by integrating sentiment analysis. 2025-01-01T00:00:00Z