采用双仿射注意力的英文AMR 解析模型

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号： 1000-5013（2026）02-0213-09

Abstract：To address the limitation abstract meaning representation （AMR） parsing in modeling complex syntax structures long-distance dependencies，especiall hierarchical relationships between main subordinate clauses，an end-to-end model that combines biafine attention with a multi-channel graph convolutional networks（GCN）is proposed.Firs，biaffine attention is employed to perform relation-specific scoring over word pairs，constructing an adjacency tensor with channels divided by relation types.Then，multi-channel graph convolution propagate message wihin each channel aggregate information across channels to generate structure-aware node representations. Experimental results show that the proposed model achieves Smatch scores 85.6 84.3 on the AMR 2.0 AMR 3.O datasets，respectively，outperforming the baseline model SPRING by 1.8% 2.6% . The results demonstrate that the proposed model more accurately capture hierarchical dependencies head-tail relationships，exhibiting strong robustness promising applicability in complex syntactic scenarios.

Keywords ： semantic parsing; abstract meaning representation; sequence-to-sequence model; natural language processing

随着自然语言处理技术向深层语义理解演进，抽象语义表示（AMR）作为一种将自然语言句子转换为有向图结构的语义表示方法，成为连接表层文本与深层含义的关键桥梁。（剩余13176字）