Bleztam/Multilingual-QA-Chatbot-for-Low-Resource-Languages

# Amharic AI Student Assistant **Live Demo:** Try the App ## Problem Statement Ethiopian university students often face challenges finding fast, accurate answers to questions regarding admissions, scholarships, fees, and requirements in both English and Amharic. This multilingual AI Student Assistant bridges that information gap by offering an accessible Streamlit interface that quickly retrieves relevant educational information. ## Dataset The dataset contains 1010 Q&A pairs focused on topics such as: - Admissions - Exams - Scholarships - Study tips Both English and Amharic entries are provided. Data is processed locally to yield fast similarity searches. ## Model Comparison The application features two different retrieval systems for the sake of comparison and academic study: 1. **TF-IDF (Baseline)**: Computes word frequencies and inverse document frequencies using `sklearn.feature_extraction.text.TfidfVectorizer`. It relies on exact keyword matching. 2. **Sentence Embeddings (Modern)**: Uses `sentence-transformers` (`paraphrase-multilingual-MiniLM-L12-v2`) to capture semantic meaning. It calculates cosine similarity over the dense vectors, providing more context-aware results, especially when phrasing differs heavily from the dataset. ## Key Insights - TF-IDF significantly outperformed embeddings in Amharic due to strong lexical overlap in the dataset. - Embedding models performed better in English when queries were paraphrased. - Leave-One-Out testing resulted in 0% accuracy, revealing a limitation in dataset design (no redundancy or clustering). - This project highlights how dataset structure and language constraints can impact model performance more than model complexity. ## UI Features - **Interactive Chat Interface**: Familiar messaging bubble format. - **Model Selection**: Evaluate TF-IDF, Embeddings, or compare both side by side. - **Confidence Scoring**: High, Medium, or Low thresholds based on cosine similarity scores. Provides graceful fallback text …