Abstract:Argument pair extraction aims to identify corresponding argument unit pairs across texts and is a core task in argumentation mining To address challenges in argument pair extraction, including hierarchical confusion in cross-document semantic interaction modeling, difficulty in capturing fine-grained correlation information, and the isolated modeling of argument identification and pair reasoning, this study proposes a joint extraction model of argument pairs based on hierarchicalcross and multi-granularity hybrid attention mechanism——HMA-APE. Cross-text interactions are modeled at the paragraph-to-sentence levels through a hierarchical cross attention encoder, by which non-argument noise is filtered. A triple attention collaboration mechanism is designed to construct multi-granularity hybrid attention, enabling the capture of content relevance, sequence dependency, and stance contrast, respectively, so that multi-dimensional correlation information is comprehensively explored. A dual-path interaction decoder is adoptedto enhance information interaction efficiency between tasks through bidirectional transfer, in which pair reasoning is constrained by boundary features and boundary identification is corrected by pair features, thereby achieving deep collaboration between argument identification and pair reasoning. Experimental results show that the proposed model outperforms all the compared baseline models on two public datasets, which verifies its effectiveness.