1三条数据源汇进一条流水线Three sources, one pipeline
每日采集邮件订阅、PMC、期刊页，非结构化的文献文本经清洗、解析、富化后入库。Daily intake from mail subscriptions, PMC and journal pages; unstructured paper text is cleaned, parsed, enriched, then ingested.
2缺失元信息逐级补齐Missing metadata filled in, tier by tier
来源自带 → 落地页 → PubMed/PMC → Crossref 四级降级，任一级失败不中断后续。Source payload → landing page → PubMed/PMC → Crossref: four tiers of fallback, and a failure at any tier does not stop the rest.
3用 LLM 判断引用类型An LLM decides the citation type
一篇论文与一个数据集的关系有三种：提交、复用、只是引用。由 LLM 做语义识别并交叉验证，准确率由规则版 64.1% 提升至 94.2%。A paper relates to a dataset in one of three ways: deposited, reused, or merely referenced. The LLM reads for semantics and cross-checks itself — accuracy up from 64.1% (rule-based) to 94.2%.
4审核系统The review console
所有记录先进审核队列：校验通过的自动入库，其余转人工复核。Every record enters the review queue first: those that pass validation are ingested automatically, the rest go to a human.