Multi-Level Paragraph Representation through Semantic States, Continuity, and Transitions

Abstract Paragraph similarity is commonly assessed by representing each paragraph as a single vector, which can capture overall semantic content but may obscure sentence-level organization and semantic changes across a paragraph. This paper proposes a multi-level paragraph representation that combines three complementary components: semantic states (L1), local semantic continuity (L2), and semantic transitions in embedding space (L3). The components are compared using optimal transport and combined through weighted fusion to evaluate both semantic similarity and structural sensitivity. Experiments across multiple datasets and evaluation settings show that the components capture different aspects of paragraph similarity. In a semantic similarity evaluation using SBERT embeddings, the fused representation achieves a Spearman correlation of 0.8881, comparable to 0.8876 for the SBERT mean-pooling baseline, although its Pearson correlation is lower. In a controlled sentence-reordering experiment on ROCStories, L2 and L3 achieve ROC-AUC scores of 1.000 for distinguishing original from reordered stories under the tested conditions. These findings suggest that multi-level representation can provide explicit sensitivity to sentence-level organization while maintaining competitive rank-based semantic similarity performance, rather than uniformly outperforming single-vector representations.