Dataset and Source Code for "Explainable Amharic Emotional Text Classification Using Transfer Learning"

This repository contains the Amharic Emotion Text Dataset (AETD) and the source code developed for the study “Explainable Amharic Emotional Text Classification Using Transfer Learning.” The dataset consists of 17,520 manually annotated Amharic social media comments categorized into seven emotion classes: neutral, fear, sadness, joy, anger, surprise, and disgust. Data were collected from multiple online platforms, including YouTube, Facebook, TikTok, blogs, and X (formerly Twitter), to capture diverse emotional expressions in contemporary Amharic digital communication. The repository provides the complete experimental pipeline used in the study, including data preprocessing, text normalization, dataset preparation, model fine-tuning, evaluation, and explainability analysis. The proposed emotion classification model was developed by fine-tuning the Davlan/afro-xlmr-base transformer model, while Local Interpretable Model-agnostic Explanations (LIME) were employed to enhance model transparency by identifying the textual features contributing to individual predictions. The resources in this repository are intended to support reproducibility, transparency, and future research in Amharic Natural Language Processing (NLP), emotion classification, low-resource language technologies, and explainable artificial intelligence (XAI). Researchers may use the dataset and source code to reproduce the reported results, benchmark new approaches, or extend the work to related tasks such as sentiment analysis, emotion intensity prediction, multilingual emotion classification, and explainable NLP.