Jaefer-Kemal/sentimental-analysis-in-amharic

# Sentiment Analysis in Amharic using BERT > **Adama Science and Technology University** > College of Electrical Engineering and Computing > Department of Computer Science and Engineering > **Course:** Introduction to Natural Language Processing (NLP) > **Submitted To:** Dr. Sintayehu Hirpassa > **Submission Date:** 20 April 2026 --- ## 👥 Team Members | Name | Student ID | |------|-----------| | Betsegaw Tesfaye | UGR/25502/14 | | Muhammed Samson | UGR/26320/14 | | Jaefar Mohammed | UGR/25447/14 | | Nikodimos Mekonen | UGR/25434/14 | | Samuel Fikadesilassie | UGR/25290/14 | | Sisay Derese | UGR/25512/14 | --- ## 📌 Project Overview This project develops a **production-level Amharic Sentiment Analysis pipeline** using fine-tuned BERT-based transformers. It addresses the unique challenges of Amharic — a morphologically rich, low-resource language written in the Ge'ez script — including agglutinative morphology, code-mixing (English–Amharic), cultural idioms, and severe class imbalance in social media data. ### Research Collaboration This project builds on two collaborative Colab notebooks developed by the team: | Notebook | Dataset | Description | |----------|---------|-------------| | AfriSenti Colab | **AfriSenti** (3-class: Negative / Neutral / Positive) | **Primary codebase** — real-world Twitter data, full pipeline with ensembling, threshold optimization, and error analysis | | Rasyosef Colab | Rasyosef/Amharic-Sentiment (binary: Negative / Positive) | Clean binary baseline — used to establish upper-bound performance and select the best backbone model | > **This repository is based on the AfriSenti Colab.** The `rasyosef/bert-medium-amharic-finetuned-sentiment` model (which achieved the highest F1 on the binary dataset) was carried forward as the pre-trained backbone for the AfriSenti pipeline, avoiding the need to re-train from scratch. --- ## 📊 Dataset: AfriSenti (SemEval 2023 Task 12) | Split | Total | Negative | Neutral | Positive | |-------|---- …