Aspect-Level Sentiment Analysis of Amharic Social Media Comments on the Ethiopian Civil War using Deep Learning

This paper presents the sentiment analysis of Amharic social media comments pertaining to the Ethiopian Civil War using two sophisticated deep learning models: the convolutional neural network coupled with bidirectional gated recurrent units (CNN-Bi-GRU) and the convolutional neural network coupled with bidirectional long short-term memory (CNN-Bi-LSTM). Specifically, the study focuses on the subtleties of public sentiment in a critical Ethiopian Civil War, with the goal of addressing the difficulties associated with sentiment categorization in languages with limited resources. The experiments were conducted using a dataset of 3,507 Amharic comments, which were split into training and testing sets. The CNN-Bi-LSTM model achieved a training accuracy of 95.79% and a test accuracy of 73.21%, respectively. In comparison, the CNN-Bi-GRU model demonstrated similar performance metrics (achieved a training accuracy of 95.72% and a test accuracy of 74.5%, respectively), indicating its effectiveness in capturing sentiment-related features while also highlighting the importance of model architecture in handling sequential data. According to the findings, both models successfully extract features from convolutional layers; nevertheless, over-fitting and generalization to new data present problems. This work emphasizes how expanding labeled datasets and improving data quality are essential for improving model performance. These results add to the increasing amount of literature on sentiment analysis in low-resource languages and lay the groundwork for further study into creating more reliable sentiment analysis systems.