meselesolomon/Tigrigna-Sentiment-Analysis

Tigrigna Sentiment Analysis Research: Achieving 82% Accuracy and 78% Macro F1-score using Bi-LSTM + FastText architectures. # Tigrigna Sentiment Analysis ### A Design Science Approach to Real-Time Sentiment Analysis for YouTube This repository contains the dataset and source code for the **Tigrigna Sentiment Tool**, a browser extension developed to analyze public opinion on YouTube in real-time. This research addresses the gap in NLP tools for low-resource languages, specifically focusing on the Tigrigna language. --- ## 📊 Dataset Overview The dataset consists of **30,353 manually annotated comments** collected via the YouTube API. ### Data Structure The dataset is provided in `.csv` format with the following columns: | Column | Description | | :--- | :--- | | **text** | The raw, original text as it appeared on YouTube. | | **label** | The sentiment class (Positive, Negative, or Neutral). | > **Data Quality Note:** This dataset represents a significant manual annotation effort. While every effort was made to ensure accuracy through a human-in-the-loop strategy, users should be aware of "label noise" inherent in social media text due to linguistic nuances, sarcasm, and subjectivity. --- ## 🛠 Model & Methodology Using a **Bi-LSTM** architecture combined with **FastText** embeddings, the model was optimized for the unique morphological structure of Tigrigna. * **Architecture:** Bi-LSTM (Bidirectional Long Short-Term Memory). * **Embeddings:** FastText (100-Dimensional) trained on a custom Tigrigna corpus. * **Hyperparameters:** * **44,000** Vocab Size (90% Coverage). * **32** Max Sequence Length. * **Evaluation:** 70/15/15 stratified split. * **Framework:** Design Science Research Methodology (DSRM). ### Performance Results | Metric | Score | | :--- | :--- | | **Accuracy** | **82%** | | **Macro F1-score** | **78%** | --- ## 🚀 The Browser Extension The **Tigrigna Sentiment Tool** allows for real-time inference. 1. **Fetch:** It grabs comments directly from the YouTube DOM. 2. **Process:** Data is sent to a Flask-based API for preprocessing. 3. **Display:** The sentiment distribut …