Adaptive entity extraction method based on distant supervision

The traditional domain knowledge entity extraction algorithm mainly depends on the professional knowledge of experts, which requires a large amount of annotation workload and is difficult to apply in new fields. To solve this problem, this paper proposes an entity extraction algorithm based on remote supervision and applies it to the field of grain and oil storage. Under the framework of positive unlabeled learning, the algorithm performs entity extraction through two stages of entity determination and entity classification. First, a bidirectional Long Short-Term Memory neural network(BiLSTM) was used for two-class entity identification. Second, the fully connected network was used for entity type identification. Finally, the algorithm was used to extract entities to construct a knowledge graph in the field of grain and oil storage, which verified the feasibility of the algorithm. This algorithm is suitable for entity extraction tasks with few training entity samples and reduces the corpus size required for the BiLSTM-based algorithm entity extraction. Moreover, it achieves comparable results to those of the classical BiLSTM-based algorithm. Copyright ©2022 Journal of Harbin Engineering University.