Abstract: Existing relation extraction (RE) methods suffer from insufficient task interaction and redundant computation between entity extraction and relation classification． To address these issues, we propose a Span-aware Task-Adaptive Relation-filtering Network with Cross-task Interaction (short for STAR) for joint entity and relation extraction． Specifically, a convolutional cross-attention mechanism is introduced to enhance semantic collaboration between subtasks and construct a decoupled entity-relation representation space． Furthermore, a relation filtering mechanism, boundary detection mechanism, and bidirectional gated recurrent unit (short for Bi-GRU) are integrated to perform local contextual modeling, enabling dynamic pruning of candidate relation types and candidate spans． Experimental results demonstrate that STAR achieves competitive overall performance on the CoNLL04 and ACE05 benchmarks, obtaining the best F1-scores of 73.59% and 63.14% for relation extraction, respectively． Moreover, STAR exhibits superior computational efficiency in terms of floating-point operations． Ablation studies and sensitivity analyses further validate the rationality of the proposed architecture and the effectiveness of its key components．