In this paper, we propose a learning classifier based on maximum entropy (ME) for resolving ZA in Chinese. Besides regular grammatical, lexical, positional and semantic features, we develop two innovative Web-based features for extracting additional semantic information of ZA from the Web. Our study shows the Web as a knowledge source can be incorporated effectively in the learning framework and significantly improves its performance. In the application of ZA resolution in MT, it is viewed as a pre-processing module that is detachable and MT-independent. The experiment results demonstrate a signifcant improvement on BLEU/NIST scores after the ZA resolution is employed.