PTR: Prompt Tuning with Rules for Text Classification
清华;liuzhiyuan;通过规则制定subpromptRelation Extraction as Open-book Examination: Retrieval-enhanced Prompt Tuning
Relation Extraction as Open-book Examination: Retrieval-enhanced Prompt Tuning
ACM SIGIR 22;浙大;chenhuajun & zhangningyu 团队;将关系抽取视为开卷,训练过程中通过[MASK]的向量,构建知识embedding库,以及KNN检索。Inference时,考虑KNN的topn结果,可以提升模型的鲁棒性和在长尾数据上的效果。
KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization for Relation Extraction;
WWW22;浙大;chenhuajun & zhangningyu 团队;我们在实体周围分配虚拟类型词,这些词使用潜在实体类型集合的聚合嵌入进行初始化;将MLM个具有额外可学习关系嵌入的头部层扩展为虚拟答案词集V‘,以完全表示对应的关系标签Y。