Query expansion aims to add related new terms to a query so as to improve the latter. The key to its success lies in the relatedness of the expansion terms to the query. When co-occurrence analysis is used to determine the expansion terms, we encounter a serious problem of ambiguity: a term may be related to another in a context different from that of the query. To solve this problem, we propose to add another context term in the relation to help specify the appropriate context of the relation. The extracted relations are used for query expansion in language models. Our experiments on several TREC collections clearly show that the new type of relation performs much better than the traditional co-occurrence relation.