What we tried to do with it, is bypass any sort of keyword processing in order to find similar patents. The reason we’ve done this is to avoid the problems encountered by other systems that rely on natural language processing or semantic analysis simply because patents are built to avoid detection by similar keywords…we use network topology (specifically citation network topology) to mine the US patent database in order to predict similar documents.
When dealing with a large text dataset, most folks jump right into NLP and semantic analysis, it’s interesting to learn when that’s not such a good idea.
Check out the full video:
… and if you wanted to see just how exciting troll hunting can be: