Coherence Assessment of Shared Subgraph Patterns of Articles Based on Graph Attention Networks

Text coherence is crucial for ensuring both the quality and comprehensibility of written material and is vital for various natural language processing tasks and applications. While existing research in feature engineering and neural networks has achieved commendable results, these studies have predominantly focused on content within individual documents, neglecting the structural similarities that may exist between different texts. This study explores a coherence model based on graph attention networks, aiming to capture structural similarities across texts and assess their coherence based on shared subgraph patterns. The approach involves first constructing a graph structure for each text, grounded in rhetorical structure theory and semantic relations, and then identifying different subgraph patterns. Subsequently, a document-pattern graph is built, representing shared subgraph patterns. Finally, the graph attention network is employed to model this graph. Experimental results indicate that this method outperforms baseline models in discourse coherence evaluation tasks, demonstrating notable advancements.