Word sense disambiguation (WSD) can be thought as a classification problem. Feature selection is of great importance in such a task. In general, features are selected manually, which requires a deep understanding of the task itself and the employed classification model. In this paper, the effect of feature template on Chinese WSD is studied, and an automatic feature selection algorithm based on maximum entropy model (MEM) is proposed, including uniform feature template selection for all ambiguous words and customized feature template selection for each word. Experimental result shows that automatic feature selection can reduce feature size and improve Chinese WSD performance. Compared with the best evaluation results of SemEval 2007: task #5, this method gets MicroAve (micro-average accuracy）) increase 3.10% and MacroAve (macro-average accuracy）) 2.96% respectively.