A Unified Multi-Task Learning Framework for Joint Extraction of Entities and Relations
2021-10-16 2021-10-16
来源:AAAI 2021
机构:北航
任务:Joint RE
动机:
先 entity 再 relation、先 relation 再 entity 都不好,先 subject 再 relation 再 object 最好,因为 1) The query question explicitly provides prior signals about the type information. 2) It en- hances the interactions between the query and the text based on the QA structure. 3) It presents a natural way to deal with overlapping entities and relations
三部分 ① the type-attentional subject extraction ② the subject-aware relation prediction (SRP) ③ the QA-based object extraction
整体模型架构是 bert + 两层 highway + GRU
其中 13 都是 input task specific info + 句子,抽取式 mrc;1 是用 entity type;3 是用 QG 生成的问题
2 是用头实体以及 cls 的信息 + 分类器,关系分类
QG 直接 seq2seq 生成,input 是 pseudo question “Find [object type] that [subject text] is [relation type]”,生成自然语言的问题
对于没有预先定义的 relation,psudo question 是 “What is the [relation type] of the [subject text]?”(所以预先定义 relation 是指给头尾实体的实体类别做定义吗?)
实验
ACE05 CoNLL04
先抽 relation 再抽头尾实体:在 ACE05 CoNLL04 上要比现在的模式 r 高,f1 低;DUIE 上全面比现在好。作者说 relation-middle method (UMT) is more qualified for concise datasets with fewer entity and relation types