mzvirikuzhe/Miracle-Zvirikikuzhe-COS802-Project

African Languages for Semantic Relatedness Using SemRel Datasets. # Comparing Baseline vs. Fine-Tuned Models for Semantic Relatedness in African Languages # 1. Project Overview. The objective of this current project is to explore semantic relatedness in three African languages Hausa (hau), Kinyarwanda (kin) and Amharic (amh). The goal is to assess how multilingual baseline models perform against a fine-tuned transformer model (XLM-RoBERTa) on the Semantic relatedness task. The set of solutions comprises of a zero-shot semantic similarity model; a fine-tuned transformer model applied transfer learning, complete execution, evaluation, and results-saving experiments. # 2. Research Focus & Methodology. 2.1 Research Questions. 1. How do baseline multilingual models (like LaBSE and MPNet) compare to a fine-tuned model (XLM-RoBERTa) for this task? 2. Which transfer learning approach is best for Hausa, Kinyarwanda, and Amharic: zero-shot or fine-tuning? # 2.2 Methodological Overview. The study employs two modelling paradigms. Baseline (Zero-Shot) Models. We use the below sentence-embedding models without task-specific fine-tuning. • LaBSE: huggingface.co. • Multilingual MPNet Base V2 Rewriter: huggingface.co. Embeddings are taken and cosine similarity is used to compute similarity. Transfer Learning Model. • XLM-RoBERTa Base: huggingface.co. A regression head was added to the model to fine-tune it on SemRel2024. Dataset. SemRel / SemRel2024 Dataset. SemRel/SemRel2024 Characteristics. • Semantic relatedness dataset annotated in multiple languages with the help of human. • Includes a number of varieties, such as Hausa, Amharic, Kinyarwanda, Zulu etc. • It has sentence pairs with cosine similarity scores between 0 and 1. It has train/dev/test splits. Evaluation Metrics. • The primary metric for semantic similarity ranking is Spearman’s Rank Correlation (ρ). • Pearson Correlati …