1) The document proposes a semantic similarity graph to model text coherence through cohesion between sentences. It represents text as a graph where vertices are sentences and edges represent semantic similarity between sentences. 2) It evaluates coherence by calculating the average weight of outgoing edges from each vertex. Previous methods only modeled local adjacency or entity repetition, while this approach captures relatedness between non-identical sentences. 3) Evaluation on discrimination and insertion tasks shows the proposed approach performs comparably or better than prior models, suggesting local and long-distance semantic relations both contribute to coherence.