Automated Hierarchical Patent Classification (Bachelor’s Thesis Proposal)

The number of patent applications is growing steadily over the years. In the year 2014, the U.S. Patent and Trademark Office registered 615,243 filed patent applications. In the year 2000, the number was only as high as 315,015 [1]. In order to efficiently process and search those applications, patent documents are annotated with the different areas of technology to which they belong. This process of annotating the documents can be automated with software systems. Given a patent document and a patent classification scheme, such systems automatically produce suggestions for annotations [2]. Typically, this is achieved by using a machine learning-based classification algorithm trained on a set of annotated patents. The predominantly used annotation scheme is the International Patent Classification (IPC) [3]. The IPC’s structure is hierarchical: It is comprised of a tree of categories with height 13 [4]. For example, a patent could belong to the categories (1) physical analysis of biological material, (2) physical analysis of liquid biological material, and (3) physical analysis of blood [3]. Here, (2) is a subcategory of (1) and (3) is a subcategory of (2). Other patent classification have a similar hierarchical layout [5]. This hierarchical structure poses an interesting challenge for automated patent classification. As opposed to many other classification problems, a classifier would have to take a hierarchy of target classes into account.