Fine-grained relation extraction with focal multi-task learning
Published:
Relation extraction aims to identify relation facts for pairs of entities in raw texts to construct triplets such as [Arthur Lee, place born, Memphis]. To automatically extract relation facts, the distant supervision strategy has been proposed, which assumes that, if there exists a relation between two entities in a known knowledge base, all the sentences that mention these two entities will likely express the same relation. Recently, neural networks have been widely applied to distant supervised relation extraction and have achieved good performances by precisely extracting semantic features. However, most current studies have not paid sufficient attention to distinguish fine-grained relations. Relations, such as company/shareholder, company/advisors, and company/founders, contain similar relation features. All three relations can be easily recognized from others such as location/contain; however, it is difficult for relation extractors to distinguish them from each other. Meanwhile, previous studies have not focused on relations that are difficult to distinguish. Even though remarkable relations can be easily extracted, similar relations are difficult to be recognized precisely to improve the fine-grained relation extraction.
Recommended citation: Zhang X, Liu T, Jia W, et al. Fine-grained relation extraction with focal multi-task learning[J]. Science China Information Sciences, 2020, 63(6): 1-3.