BCSLinker: automatic method for constructing a knowledge graph of venous thromboembolism based on joint learning

Abstract

The outstanding performance of BCSLinker can be attributed to two main advantages: First, we extract entities and relations simultaneously using a multi-module one-step extraction method, effectively alleviating error propagation. Second, we adopt the multi-label cross-entropy loss to mitigate the impact of negative samples, which widely exist in multi-module one-step extraction methods. Baseline models, such as CasRel, BiRTE, and PRGC, extract entities and relations separately through multiple steps, essentially following pipelined extraction patterns and facing error propagation. As shown in