getnetbogale27/Aklil-amharic-sentiment-analysis

# Developing Deep Learning-Based Sentiment Analysis of Amharic Social Media for Public Policy Enhancement in Ethiopia > **Academic Research Project** · Addis Ababa University > Author: **Getnet Bogale** --- ## Overview This repository contains the full implementation for a research project that applies **deep learning** to classify the sentiment of Amharic-language social media posts, with the goal of informing public policy decisions in Ethiopia. Amharic is the official working language of the Ethiopian federal government and is spoken by over 30 million people. Despite its importance, Amharic remains severely under-resourced in the NLP community. This project addresses that gap by: 1. Leveraging the **AfriSenti** benchmark dataset for Amharic Twitter sentiment 2. Augmenting it with **additionally collected policy-related Amharic tweets** 3. Training and comparing deep learning models (Bi-LSTM, XLM-RoBERTa) 4. Deploying a **policy sentiment dashboard** for decision-makers --- ## Research Objectives | # | Objective | |---|-----------| | 1 | Collect and annotate a domain-specific Amharic corpus of public-policy tweets | | 2 | Develop an Amharic-specific text preprocessing pipeline (normalisation, tokenisation, stopword removal) | | 3 | Train and evaluate CNN, Bi-LSTM, and transformer-based sentiment classifiers | | 4 | Identify which deep learning architecture achieves the best macro-F1 on Amharic sentiment | | 5 | Build an interactive policy dashboard that visualises public sentiment trends for government stakeholders | --- ## Datasets ### AfriSenti — Amharic (`data/raw/amh/`) | Split | File | Size | |-------|------|------| | Train | `train.tsv` | ~8,000 tweets | | Dev | `dev.tsv` | ~1,000 tweets | | Test | `test.tsv` | ~1,000 tweets | Tab-separated format: `tweet label` where label ∈ {`positive`, `negative`, `neutral`}. AfriSenti also includes 13 other African languages used in transfer-learning experiments (`data/raw/arq`, `ary`, `hau`, `ibo`, …