In one embodiment, a method includes generating a reconstructed embedding of a query based on one or more term embeddings associated with the one or more query terms, respectively, on receiving a query with the one or more query terms, formulating an evaluation model based at least on the reconstructed embedding of the query, where the evaluation model calculates a relevance score for posts with respect to the search query based at least on the classifier vectors of the posts, and calculating, for each of the retrieved posts, a relevance score for the post by applying the associated classifier vector to the formulated evaluation model.