BookForYou: Leveraging BART-Generated Narrative Tropes for Content-Based Book Recommendations
Abstract
This paper presents BookForYou, a content-based book recommendation system where a fine-tuned BART model (139M parameters) generates structured narrative trope tags from book descriptions while Sentence-BERT retrieves candidates from a 42,350-book catalog. Candidates are re-ranked via late fusion of semantic similarity and Asymmetric Trope Overlap (ATO). On a held-out test set of 489 queries whose descriptions were excluded from BART’s fine-tuning data (fusion weight selected on a disjoint 488-query validation split), the system achieves NDCG@10 of 0.789, outperforming SBERT-only (0.546, p < 0.001), an SBERT+BM25 fusion control (0.553), a cross-encoder re-ranker (0.518), BM25 (0.454), and popularity ranking (0.384). ATO is the dominant signal: trope re-ranking alone reaches 0.783, accounting for 97.8% of the gain over SBERT-only, while a generic lexical fusion over the same candidate pool adds a non-significant 0.007 (p = 0.33); fusing SBERT with ATO adds a small but significant 0.006 (0.7%, p = 0.017). The system also achieves a 91.0% trope match rate and 0.766 serendipity, providing matched tropes as natural-language explanations.