EVENT DOSSIER
The research team proposed the NS-ST-GraphRAG framework, aimed at handling information retrieval and generation in long-form narrative texts. This framework integrates ontology-guided extraction, certainty constraint checking, and dynamic subgraph retrieval techniques. By matching query spatial and temporal ranges to select graph states, the generated answers are anchored on traceable evidence. Additionally, the team released Red-Chamber-QA, the first multi-hop question-answering benchmark for classical literature, covering time, space, and general question categories. In a test set containing 120 questions, NS-ST-GraphRAG performed exceptionally well: the reproduction rate of mechanical answers was 0.733, surpassing the freeze window baseline (0.675) and closed-book models (0.083); the semantic judgment accuracy reached 0.866, higher than the baseline of 0.850.
NS-ST-GraphRAG 框架提出用于长篇小说叙事检索增强生成，整合本体引导提取、确定性约束检查及动态子图检索等技术。该框架通过选择符合查询时空范围的图状态来 grounding 答案于可追溯证据。研究团队同时发布首个古典中文文学多跳…
NS-ST-GraphRAG 框架提出用于长篇小说叙事的信息处理，整合本体引导提取、确定性约束检查及动态子图检索。该框架选择与查询时空范围匹配的图状态，将生成答案锚定在可追溯证据上。研究团队还发布了首个古典文学多跳问答基准 Red-Cha…
NS-ST-GraphRAG 框架提出用于长篇小说叙事检索增强生成，整合本体引导提取、确定性约束检查及动态子图检索等技术。该框架通过选择符合查询时空范围的图状态来 grounding 答案于可追溯证据。研究团队同时发布首个古典中文文学多跳问答基准 Red-Chamber-QA，包含时间、空间及通用问题类别。在 120 题测试集上，NS-ST-GraphRAG 机械答案复现率为 0.733，高于冻结窗口基线的 0.675 和闭卷模型的 0.083；语义判断准确率为 0.866，高于基线的 0.850。
NS-ST-GraphRAG 框架提出用于长篇小说叙事的信息处理，整合本体引导提取、确定性约束检查及动态子图检索。该框架选择与查询时空范围匹配的图状态，将生成答案锚定在可追溯证据上。研究团队还发布了首个古典文学多跳问答基准 Red-Chamber-QA，涵盖时间、空间及通用问题类别。在 120 题测试集中，NS-ST-GraphRAG 机械答案复现率为 0.733，优于冻结窗口基线（0.675）和闭卷模型（0.083），语义判断准确率达 0.866。