Meta-search that provides the capability for users to access and search all of the information sources in one query submission is one of the most important mechanisms to search deep web. One of the fundamental problems in building meta-search systems is to extract the semantic model of each query interface so that the system can automatically form and submit queries to each online source. In this paper, we develop a rule-based approach for parsing query interfaces. We classify query conditions into 5 categories of semantic structures, and develop parsing rules for each category. Our parsing rules use both structural and visual information. To alleviate the ambiguity, three parsing passes are adopted in our approach: null-path cluster parsing, inner-cluster parsing and inter-cluster parsing. Experiment shows that our approach works very well, achieving precision 86% and recall 92% for IW Random dataset, and precision 88.9% and recall 88.8% for ICQ dataset.