POLITICAL STANCE DETECTION FOR AFAAN OROMO TEXT USING DEEP LEARNING APPROACH

Advisor: Dr. Obsa Gilo (Ph.D.) Social media and other technology advancements are becoming more and more important tools for connecting with people around the globe, including governments, social media activists, and political figures. Social media has recently been used by people to voice their thoughts on a variety of topics or issues. In the area of natural language processing, stance detection or classification has several applications. For example, automatically determining if a community supports or opposes a specific position on political and religious issues, either in support of or against the stated objectives. We created our own dataset for this study, which included 10,136 Afaan Oromo comments, targeting the OFC, OLF, and Oromia Prosperity Party. The data were preprocessed and subjected to morphological analysis following their collection and annotation in accordance with annotation rules. This research utilized a dataset comprising 10,136 Afaan Oromo text comment sourced from three political parties’ verified Facebook pages OLF, OFC, and Oromia Prosperity parties. To address the challenges of Afaan Oromo text stance categorization, we proposed an attention-based RNN and an attention-based LSTM in our work. Our experimental results show that the Attention-based Long Short-Term Memory (LSTM) model achieves 98.6% training accuracy and 97.5% validation accuracy. In comparison, the second model in this study, the Attention-based Recurrent Neural Network (RNN) model, achieved 95.82% training accuracy and 93.47% validation accuracy. The greatest training accuracy of 98.6% and the highest validation accuracy of 97.5% show that the Attention-based LSTM model outperforms the others. We used previously unseen data to assess our model's effectiveness by applying accuracy, recall, and F-Score measures to categorized text using validation data.<br> A precision, recall, and f-score of 0.83, 0.83, and 0.83 were attained by the attention-based RNN, respectively. The precision, recall, and f-scores of the attention-based LSTM f-measure were higher than those of the attention-based RNN, with respective values of 0.94, 0.94, and 0.93. Consequently, this implies that the stance classification performance of the attention-based LSTM model is superior.<br> Keywords: Natural language processing, Stance classification, Deep learning, Attention mechanism.