Understanding cross-cultural differences is an important application of natural language understanding. This problem is difficult due to the relativism between cultures. We present a computational approach to learn cultural models that encode the general opinions and values of cultures from multi-lingual Wikipedia. Specifically, we assume a language is a symbol of a culture and different languages represent different cultures. Our model can automatically identify statements that potentially reflect cultural differences. Experiments on English and Chinese languages show that on a held out set of diverse topics, including marriage, gun control, democracy, etc., our model achieves high correlation with human judgements regarding within-culture values and cultural differences.