Explainable Amharic Emotional Text Classification Using Transfer Learning

The widespread use of digital platforms has led to a surge in Amharic emotional comments, highlighting the critical need for robust emotion detection systems in this under-resourced language. To address this, our paper introduces a novel Amharic emotion detection model, fine-tuned from a large language model pre-trained on 17 African languages, including Amharic. Crucially, we integrate Explainable Artificial Intelligence (XAI) into our pipeline to enhance the transparency of the model's decisions. The model was trained, validated, and tested on 14,191, 1,577, and 1,752 expert-labeled Amharic social media comments, categorized into seven emotions: neutral, fear, sadness, joy, anger, surprise, and disgust. Utilizing Local Interpretable Model-agnostic Explanations (LIME) as our XAI framework, the experimental results demonstrate excellent performance, achieving an accuracy of 68.15%, with comparable precision (68.24%), recall (68.15%), and F1-scores (68.17%). The XAI algorithms clearly show how the model arrives at its classifications by calculating probabilities for each category and highlighting relevant textual terms. When compared against several state-of-the-art text classification techniques, including Bidirectional Encoder Representations from Transformers (base variant), Support Vector Machine, Bidirectional Long Short-Term Memory (Bi-LSTM), LSTM, and Convolutional Neural Network (trained from scratch), our fine-tuned model consistently outperforms these approaches. While promising, these results indicate avenues for further improvement. Future research will focus on enhancing performance through training on larger emotional datasets and incorporating code-switching.