Like many Patent Offices around the world, the UK Intellectual Property Office (UKIPO) is interested in AI-enabled prior art searches for patent examination. The World Intellectual Property Organisation’s Index of AI initiatives in IP offices lists investigations by the European Patent Office (EPO) and numerous national offices (Austria, Canada, Finland, Germany, Japan, Morocco, Philippines, Republic of Korea, Russian Federation, United Kingdom and United States of America).
The UKIPO has now published a feasibility study commissioned from Cardiff University: AI assisted patent prior art searching feasibility study.
It seems that the jobs of UK patent examiners are safe for now: the research concludes that it is not feasible with current AI tools to provide a fully automated solution. It found clear evidence that none of the available AI algorithms on their own can support every aspect of the prior art search process (e.g. classification, forming a search query, retrieval, ranking, identifying similarities and topic visualisation). The AI algorithms looked at included:
- Natural language processing: text segmentation, normalisation, lemmatisation, stemming, co-occurrences, multi-word terms;
- Supervised machine learning: support vector machine, naïve Bayesian learning, decision tree induction, random forest;
- Unsupervised machine learning: word embeddings, distributional semantics, neural networks, deep learning; and
- Semantic technologies: use of lexico-semantic knowledge, latent Dirchlet allocation.
Some aspects of the prior art search process were found to be more suited to the implementation of AI algorithms than others. For example, an automated classification task produced very high classification accuracy. Indeed, such classification could be imbedded in online pre-filing processes to allow applicants to improve their own due diligence checks. (For a similar approach by the UKIPO for trademarks, see The UKIPO launches AI-powered assessments of trademark applications).
The study also found that different state-of-the-art AI algorithms could be used to retrieve the closest documents, rank relevant documents, suggest synonyms, suggest classifications, cluster and visualise the retrieved documents/concepts. On the other hand, it found that drafting the search statement was less suitable for AI and that this should therefore remain a human task.