The success of several knowledge retrieval experiments in some selected domains proved empirically that knowledge retrieval, if implemented in the restricted-domain, can yield performance gains over classical “bag-of-words” based retrieval systems. The key to their success is the availability of codified high quality domain knowledge accumulated in years. However, this knowledge base is rarely available in other domains. In this paper, we propose innovative and economical ways to generate knowledge-based structural metadata and demonstrate a prototype knowledge retrieval system, WikiBackyard, to serve user's knowledge oriented information need. The system benefits users by providing structure-level access to scientific papers and automatically inferring structural knowledge from their natural language queries by leveraging knowledge recommendation and knowledge feedback.