Novelty detection in patent documents based on semantic annotation

Abstract

Novelty detection for a set of documents like patents or news articles plays a key role in many applications. While past research has mostly focused on detecting new stories from news a stream of news, this thesis deals with novelty detection in patent documents, which would help patent examiners determine novelty of a newly filed patent. Unlike past research in patent invalidation search, which relies on term-based comparisons, we propose a new method for novelty detection for patent invalidation using semantic information. This is based on our communication with a group of patent examiners who stated that in examining patents, some semantic categories such as components or methods would be used to distinguish among them. To annotate key phrases in patent claims with semantic information, we identified five categories: Main Component, Sub Component, Attribute, Function, and Method. For each category, a binary classifier was built based on features and patterns. Novelty is detected by comparing the semantically annotated phrases in the incoming patent document against those in the existing patent documents. We show in an experiment that the proposed approach outperforms a term-base approach in patent invalidation search.