Methods, systems, and apparatus, including computer program products, for generating visual media search results are provided. A first visual media search query in a first language is received. One or more bilingual refinements are identified for the first visual media search query. Each bilingual refinement associates an initial query in the first language and a subsequent query in a different second language. The initial query is equivalent to the first visual media search query. The subsequent query is a translation of the initial query in the second language. The subsequent query followed the initial query in a series of queries received from a user in a session. A first bilingual refinement is selected from the one or more bilingual refinements according to a first criterion. A set of visual media search results is generated based on the first search query and the selected first bilingual refinement's associated subsequent search query.