Converting Text Documents into Graph-based Documents with Local Discourse Structure and Semantic Hierarchy

Keiichiro Nii, Ryuta Arisaka, Shiyao Ding, Takayuki Ito

Abstract: Graph-based documents are anticipated to offer improved ease of creation and interpretation for both human and machine entities compared to textual counterparts, leading to an expected increase in societal adoption. However, graph-based documents tend to become more complex than textual documents, posing challenges in the transitioning from text to graph formats. The objective of this paper is to elucidate a methodology for efficiently creating large graph-based documents while managing their complexity. We propose a method to convert textual documents into graph-based documents through iterative localized operations, using directed discourse relations and a semantic hierarchy provided by a transformation algorithm. Empirical findings validate that this method effectively regulates complexity and enables graph creation for large texts. Additionally, we conducted extractive summarization from the generated graphs, verifying that the semantic structure is effectively integrated into the graph.