Dense retrievers rely heavily on high-quality training triplets, yet existing data construction strategies remain inadequate for reasoning-intensive retrieval tasks involving multi-hop reasoning, entity relation tracing, and implicit evidence composition. Positive samples are often based on shallow semantic relevance and fail to capture explicit reasoning chains, while negative samples are typically sampled from lexical overlap or random candidates and therefore provide limited supervision for learning clear decision boundaries. To address these issues, we propose S-Gens, a structure-aware synthetic data generation framework for enhancing reasoning-intensive dense retrieval. S-Gens uses relation paths in an external knowledge graph to synthesize queries and structurally consistent positive samples, and further constructs semantically similar but structurally inconsistent hard negatives. To improve data reliability, we introduce a Siamese graph neural network-based consistency filtering mechanism. Because S-Gens operates entirely during offline supervision construction, it remains model-agnostic, preserves the original inference architecture, and is complementary to graph-guided retrieval or RAG pipelines that inject structure online. Experiments on five benchmark datasets show that S-Gens consistently improves multiple trainable retrievers, with the largest gains on multi-hop reasoning tasks such as WebQSP and HotpotQA. These results indicate that structure-aware synthetic supervision can effectively improve dense retrieval in reasoning-intensive settings.
Lei et al. (Sun,) studied this question.