Abstract: To address the boundary ambiguity issues in academic theses caused by complex domain-specific terminology and nested logical argumentation structures, a method named MacBERT Dual-Level Dynamic-Convolution Global Pointer (MBDDGP) was proposed, which integrated a dual-level semantic attention mechanism with a Global Pointer (GP) network incorporating dynamic convolution. First, a pre-trained language model MacBERT was employed to obtain contextual semantic representations, and a dual-level attention module was designed to capture both character-level and sentence-level semantic information. Subsequently, Rotary Position Embedding (RoPE) from the GP network was introduced to effectively model long-range dependencies. Finally, a dynamic convolution mechanism was incorporated to enhance the model's sensitivity to local boundary features, thereby improving the recognition accuracy of argumentative component boundaries and laying the foundation for high-precision Argument Component Classification (ACC). Experimental results on the CLUENER2020 dataset and a self-constructed master's thesis dataset show that the proposed method achieves F1 scores of 81.5% and 78.6% on argument component classification tasks, outperforming baseline models by 1.2 and 1.6 percentage points, respectively, thus validating its effectiveness.

Key words: argumentation mining, Argument Component Classification (ACC), Global Pointer (GP) network, dynamic convolution, attention mechanism

摘要： 针对学位论文中因领域术语复杂、逻辑论证结构嵌套导致的边界模糊问题，提出一种双层次注意力机制与融合动态卷积的全局指针(GP)网络相结合的方法(MBDDGP)。首先，利用预训练语言模型(MacBERT)获取上下文语义表征，并设计双层次融合机制以捕捉字符级和句子级语义信息；其次，引入GP网络中的旋转位置编码(RoPE)以有效建模长距离依赖关系；最后，通过动态卷积机制增强模型对局部边界特征的敏感度，从而提升论辩部件边界的识别精度，为高精度的论辩部件分类(ACC)奠定基础。在CLUENER2020数据集和构建的硕士论文数据集的实验结果表明，所提方法在ACC任务上的F1值分别达到81.5%和78.6%，相较于基线模型提升了1.2和1.6个百分点，验证了方法的有效性。

关键词: 论辩挖掘, 论辩部件分类, 全局指针网络, 动态卷积, 注意力机制