Motion recognition is an essential technology for social robots in various environments such as homes, offices and shopping center, where the robots are expected to understand human behavior and interact with them. In this paper, we present a system composed of three models: motion language model, natural language model and integration inference model, and achieved to generate sentences from motions using large high-order N-grams. We confirmed not only that using higher-order N-grams improves precision in generating long sentences but also that the computational complexity of the proposed system is almost the same as our previous one. In addition, we improved the precision by aligning the graph structure representing generated sentences into confusion network form. This means that simplifying and compacting word sequences affect the precision of sentence generation.