Office of Technology Transfer – University of Michigan

High Recall Retrieval with Query Pooling and Interactive Classification

Technology #6378

Searching for highly-specific data in a particular field, such as law or medicine is a challenging task. Most enterprise search algorithms focus on returning only a few results that are highly relevant, to avoid cluttering with irrelevant returns. But this leaves a huge number of relevant results unreturned in the interest of being conservative. Moreover, the most common complaint for this type of software is that the results are still not always relevant. This technology interacts with the user using a specialized double-loop algorithm to greatly enhance both the number of documents returned and the relevance of the documents to the user’s interest.

Empowering Users through Interaction

Because of the diversity of human language, it can be very difficult to choose the right search terms when looking for a specific kind of document in law, medicine, science, business, and government. This technology removes the requirement for extreme language precision. When a user searches, the technology returns a small number of documents for the user to “judge” on relevance. This teaches the technology what kind of documents the user is looking for. The technology can then, internally, enhance the search by changing or appending terms to the search, then repeating the process. By continuing this process for some time, users can get extremely large numbers of highly-relevant documents. The technology has been shown to outcompete several leading search algorithms under a variety of search conditions when users interact for even moderate amounts of time.


  • E-discovery in legal documents
  • Systematic literature review
  • Patent retrieval and review
  • Patient review in electronic medical records


  • Achieves a higher recall and precision than most current methods
  • Reduces computation time compared to many current methods
  • Can be integrated on top of other search methods, enhancing already good search engines