Abstract: 【Purpose/significance】In order to automatically extract argumentative relations in judgment documents and assist in under⁃ standing the legal reasoning and judgment process of cases, this is done to further promote research and development in the field of le⁃ gal text analysis.【Method/process】This paper proposes a Bidirectional Attention-Driven Semantic Perception Siamese Neural Net⁃ work Model (BAD-SPS) based on the constructed corpus of judgment documents for argumentative relationships. It models the argu⁃ mentative relationships extraction task as a text matching task. The model uses BERT to encode the input argumentation elements separately, then employs whitening operations to remove the correlation between features and simplify feature representations. Next, a bidirectional attention mechanism is introduced to further enhance the semantic interaction between argumentation elements, thereby improving the model's understanding of the structure of argumentation elements. Finally, the model simultaneously handles the two sub-tasks of argumentative relationships extraction in judicial documents: argumentation relation detection and argumentation relation classification.【Result/conclusion】The BAD-SPS model shows a significant improvement in F1 score compared to the baseline model on argumentative relation detection dataset 2, achieving a score of 11.02. Similarly, on the argumentative relation classification dataset D, it also achieves a significant improvement with an F1 score of 10.59. Experimental results demonstrate the model's effectiveness in accurately understanding and identifying argumentative relations.【Innovation/limitation】This study explores a semantic-aware Sia⁃ mese network model for argumentative relation extraction in judgment documents. It can automatically extract argumentative relations in judicial documents, and further experiments are needed to generalize it to more civil cases.