Abstract: To address the challenges of semantic dilution and fine-grained information loss in long-text retrieval,this study proposes a text embedding optimization framework based on explicit knowledge extraction.Traditional methods that rely on single-vector representations of entire texts struggle to capture the multi-topic and hierarchical local semantic associations present in lengthy documents.It designs two knowledge-aware embedding strategies:1) Knowledge-aware separate embedding(KASE),which employs a large language model to extract key knowledge points from text segments and vectorizes them independently to preserve fine-grained semantics;2) Knowledge-aware concatenated embedding(KACE),which concatenates the extracted know-ledge points into a single passage and encodes it holistically to explore aggregation effects.Experimental results on Chinese datasets(CMRC,DRCD) and English datasets(SQuAD,NewsQA) demonstrate that the proposed knowledge-enhanced methods significantly outperform both traditional baselines and recent approaches such as QA-pair-augmented retrieval(QAEA-DR) and Hypothetical document embeddings(HyDE).KASE achieves the most substantial improvements,especially in long-text scenarios.Ablation studies reveal that independent knowledge-point representations are crucial for mitigating information loss,while hybrid strategies that combine original text embeddings with knowledge-aware vectors further optimize retrieval effectiveness.

Key words: Text retrieval, Knowledge extraction, Semantic embedding