Explainable Semantic Book Recommendation System Using Sentence-BERT Embeddings and Large Language Model-Based Explanations

Abstract— With the rapid development of digital libraries and online book platforms, there is a growing need for recommender systems to be able to (i) identify suitable books and (ii) provide explanations regarding the recommendations. Traditional techniques based on keywords work well, but often do not capture the semantic links if two descriptions use different vocabularies. In this research, we propose an Explainable SBERT embeddings, cosine similarity and Gemini 2.5 Flash-based natural language explanations. The 15,147 books from the Google Books dataset are cleaned and converted to semantic vectors using SBERT.SBERT is used to convert 15,147 books from the Google Books dataset to semantic vectors after they have been cleaned. The most relevant titles, as judged by cosine similarity, are returned to the user for a given book, with the top five returned. Then, Gemini 2.5 Flash writes a brief overview of themes, genres, and narrative qualities. We experimentally analyse sample recommendations, evaluate similarity scores and compare with a TF-IDF baseline. The highest similarity score of SBERT in the case of “Sweet Revenge” is 0.6556 for “Public Secrets”. The average score of the top 5 recommendations is 0.62. The results indicate that semantic embeddings provide more context-aware suggestions and that LLM-generated explanations enhance transparency and interpretability.