Multidimensional semantic retrieval of classical Chinese poetry for literary heritage discovery

Abstract Classical Chinese poetry expresses meanings that keyword searches often miss. We construct a retrieval dataset covering emotion, authorial intent, theme and thought, and develop an embedding optimization method for literary heritage discovery. Large language model analyses, semantic sampling, multi-model candidate pooling and four-level relevance annotation produce 29,080 queries. Human annotation provides a gold test set of 8,400 query–poem pairs. Gain-aware Listwise Ordinal Loss (GLO) combines graded-gain listwise supervision with constraints on similarity-score gaps between relevance levels. Using Qwen3-Embedding-0.6B, default GLO achieves the highest macro-averaged normalized discounted cumulative gain at 10 (nDCG@10) at relevance thresholds 1 and 2 across three evaluation resources. When only highly relevant, grade-3 poems qualify, the human gold score increases from 0.174 to 0.391; LambdaLoss@10 scores slightly higher. The dataset and method support the discovery and comparison of poems according to different aspects and degrees of semantic relevance.