Extracting Aspects and Polarity from Patents

We describe an approach to terminology extraction from patent corpora that follows from a view of patents as “positive reviews” of inventions. As in aspect-based sentiment analysis, we focus on identifying not only the components of products but also the attributes and tasks which, in the case of patents, serve to justify an invention’s utility. These semantic roles (component, task, attribute) can serve as a high level ontology for categorizing domain terminology, within which the positive/negative polarity of attributes serves to identify technical goals and obstacles. We show that bootstrapping using a very small set of domain-independent lexico-syntactic features may be sufficient for constructing domainspecific classifiers capable of assigning semantic roles and polarity to terms in domains as diverse as computer science and health.