The project goal is to match text chunks against a document set comprising 13 cleaned ebook texts. The process combined embedding similarity with relevance scoring functions built using a curated term cloud, with the goal of surfacing emphasizes distinctive passages. The rating step used BM25/centroid scores as weights together with embeddings as scores, deriving the rating of each chunk strongest-match, top-k, and 95th percentile evidence. It produces clearer results than plain cosine similarity when many documents share similar language.
Built with
- Machine Learning
- Natural Language Processing
- Latent Semantic Analysis
- Python
- Word Embedding