To represent the text document more expressively, a kind of graph-based semantic model is proposed, in which more semantic information among keyphrases as well as the structural information of the text are incorporated. The method produces structured representations of texts by utilizing common, popular knowledge bases (e.g. DBpedia, Wikipedia) to acquire fine-grained information about concepts, entities, and their semantic relations, thus resulting in a knowledge-rich interpretation. We demonstrate the benefits of these representations in the task of document similarity measurement. Relevance evaluation between two documents is done by calculating the semantic similarity between two keyphrase graphs that represent them. Experimental results show that our approach outperforms standard baselines based on traditional document representations, and able to come close in performance to the specialized methods particularly tuned to this task on the specific dataset.