To What Extent is AI Learning to "Understand" Patents? - A Report from the 2020 European Patent Office Patent Information Conference

Since the beginning of IT, unstructured information in general – and text in particular – have posed an exciting challenge for machines. Starting from simple absence/presence of terms (Boolean search) followed by term frequencies and bag-of-words approaches, today's deep learning models are starting to understand patent language on a deeper, more meaningful level. In this session, we will discuss how AI is “machine” learning to handle patent syntax to guide and assist the patent searcher. The EPO will introduce some of its own work, explaining how examiners' knowledge and experience (a unique asset to IP offices) is contributing. The session will address three areas of patent searching and AI: unstructured text, patent figures and the problem of multiple languages, and aim to derive conclusions with respect to professional patent search and patent machine understanding. We will also think about the risks of biases in the system that can have unforeseen consequences.