Summary: Computational efficiency is an important concern for machine learning algorithms, especially for applications on large test sets or in real-time scenarios. In this paper, a novel data structure and the corresponding algorithms for the execution system of the maximum entropy model are described. This data structure, called sparse feature tree, is used to represent the feature set to speed up the process of feature search (or feature matching), so that speed up the process of probability calculation and execution system. Experiments on chunking recognition and Part-of-Speech tagging are conducted to show that the new data structure greatly speeds up the feature matching process while keeping the same space complexity.