By Kathy Van Der Herten, Director, Product Management, CAS, a division of The American Chemical Society, Antwerp, Belgium
Research and development (R&D) investment is at an all-time high. It’s estimated that worldwide investment in R&D reached USD 2.476 trillion in 2022, driving the continued strong growth in the number of patent applications that is placing the global patent ecosystem, especially patent offices, under strain. The rise in volume and complexity of patent applications can lead to significant delays in granting patents, which results in legal uncertainty, which can stifle innovation, discourage R&D investment, and erode the competitiveness of national economies.
A study by the Japan Patent Office in 2018, estimated that its examiners spent 30 percent of their time searching for prior art (evidence the invention is already known) and another 10 percent of their time understanding it.
AI to assist in identifying potential prior art
One method being adopted by patent offices to improve application timeliness is to use artificial intelligence (AI) solutions to assist in identifying potential prior art during examinations. AI is able to mimic, at a fast rate, the human ability to compare data between patent applications and existing patents and non-patent publications to discover similarities that examiners can review in searching for prior art. While this does not replace the need for human examiners to review search results, it can significantly accelerate review in over 70 percent of applications.
AI is able to mimic, at a fast rate, the human ability to compare data between patent applications and existing patents and non-patent publications to discover similarities that examiners can review…
Increasing volume and complexity of patent applications
Based on WIPO data, CAS, a division of the American Chemical Society, estimates that the top five patent offices experienced a 4.4 percent compound annual growth rate in patent applications from 2012 to 2021. Added to the increase in application numbers is the growing complexity of patents, as seen in the number of claims per patent, patents cited per claim, prior art citations per patent, and other measures.
Searching for prior art is a complicated, iterative, and time-consuming process. For each application, searchers and examiners must devise a search strategy, select which databases to use, perform the search, evaluate the results, and, if necessary, fine-tune and repeat the search using different parameters.
The scale of these searches is staggering. According to a study by the European Patent Office , a comprehensive patent application search can draw on a prior art search of around 1.3 billion technical records in 179 databases, leading to about 600 million documents appearing in search results monthly.
A prior art search of around 1.3 billion technical records in 179 databases, leading to about 600 million documents appearing in search results monthly.
The growth in new technologies and complexity of patent applications require examiners to continually expand their levels of expertise in their field of art. If fueled by highly curated and structured data, AI can accelerate the process by sifting through millions of data sets and providing references in potential conflict with the target application.
Accelerating patent examination with AI-powered solutions
A number of patent offices are turning to AI-powered solutions to help them tackle the rising volume and complexity of patent applications. According to WIPO, 70+ AI related initiatives are underway in 27 national patent offices, including 13 that focus on prior art searches. While these are not complete solutions across the entire examination process, they do aim to accelerate examination times, which drives timeliness and ultimately customer satisfaction.
For instance, the Canadian Intellectual Property Office is utilizing commercially available AI search engines to discover linkages between citations, applications, and the current state of the art. The Japan Patent Office (JPO) is using AI for file indexing, suggesting relevant patent classifications and keywords and ranking prior art patent documents according to relevance. Meanwhile, the United States Patent and Trademark Office (USPTO) is using AI to help determine patentability, analyze patent prosecution history, and improve public access to USPTO data.
Machine learning is effective for searching text and indexed terms, but less effective for patents involving composition of matter, which often have important data contained inside structures.
Recently, the Instituto Nacional da Propriedade Industrial (INPI) of Brazil worked with CAS to complete a project using AI-enabled workflow optimization to accelerate chemistry prior art searches. Chemistry applications, which accounted for approximately 15 percent of INPI Brazil’s backlog, are extremely complex and require both text-based and structure-based searches of patent and non-patent publications. The AI component of the solution integrated four algorithm streams, performing different types of similarity and other analyses to ensure highly relevant results.
Each algorithm has its strengths. Machine learning is effective for searching text and indexed terms, but less effective for patents involving composition of matter (where two or more compounds are mixed) which often have important data contained inside structures. Likewise, a graph database can find similarities and connections that machine learning cannot. An ensemble algorithm then analyzed the results from the four streams and arrived at a single list of the publications that are most likely to have conflicts with the target application.
The benefits to productivity were substantial:
- Reduced examination times by up to 50 percent;
- Reduced search times for over 75 percent of applications processed;
- Contribution to an overall reduction of 80 percent in INPI’s patent backlog.
Quality data, quality outcomes
Quality data is critical for training AI algorithms. The more data machine learning algorithms can access, the more relevant, reliable, and trustworthy their results can be. Much of the publicly available non-curated data may include transcription errors, mislabeled units, and overly complex patent language, all of which hinder searches. This is especially challenging in chemistry and life sciences, where substances are described inconsistently across publications or have keywords embedded in tables or images. Using scientist-curated data that has been normalized, prepared, and connected in a structured format makes the information more easily searchable and improves the training of AI algorithms and the performance of prior art searches.
While training of sets can differ between technologies, industries, and applications, the fundamental approach to applying AI remains the same.
For the INPI Brazil project, we relied in large part on the CAS Content Collection™, the world’s largest collection of chemistry and life sciences data, which is extracted, indexed, and linked to simplify access and retrieval of relevant information. Additionally, we held out a random sample of patents from algorithm training that we used as a control set for measuring the accuracy of results and hit rates. These patents were evaluated by examiners in patent offices in China, Japan, the United States, and Europe, and validated for relevance by our own team of IP search professionals.
While training of sets can differ between technologies, industries, and applications, the fundamental approach to applying AI remains the same. For any technology area, the conflict citations used are identified during the examination process for every training set. Highly complex arts such as chemistry can do better with a topic-specific training set, but other arts might not show significant improvement using targeted training sets. In many areas, as long as the technology is represented in a general training set, the models perform well.
Nevertheless, the quality of the data is of significant importance.
Leveraging human expertise and workflow integration and technology
AI project teams require a wide range of subject matter expertise. The INPI Brazil project paired technology with experts in data analytics, workflow integration, high-performance computing, scientific searches, and many other disciplines.
Team members need cross-functional expertise in the challenges and outcomes being addressed. For instance, someone with experience in data science may not be able to develop fully effective algorithms if they do not understand the nuances of chemical structures. Computational scientists who create machine learning models must also understand chemistry data and structures.
Workflow integration is another important discipline in creating a comprehensive solution for patent offices. Examiners who have to navigate multiple systems and folders to find documents during a review can benefit from workflow improvements and technology enhancements that arrive at a single dashboard where all applications and supporting documents can be accessed and analyzed, and where they can see why certain reference documents are returned and how results are generated, providing the traceability they need to document decisions for prosecution and internal quality reviews.
Dramatic improvements in patent office productivity, efficiency, and customer service are possible when examiners can wield tools built around the latest technologies, such as AI. As innovation accelerates, so will the rise in volume and complexity of patent applications. This means that patent offices will continue to require new methods to optimize the patent examination process and thereby meet stakeholder expectations for higher levels of service satisfaction.
While AI solutions can help to tackle these ever-evolving challenges, they will still require expertise to implement tailored approaches. A one-size-fits-all approach will not work, as no two patent offices have the same needs. Patent offices have the same general activities but they differ in the level of staffing and technology needed to support each area. While algorithms may solve a common need, the way examiners interact with the output can be very different based on their existing technology environment.
Dramatic improvements in patent office productivity, efficiency, and customer service are possible when examiners can wield tools built around the latest technologies, such as AI.
Patent offices seeking to achieve their strategic outcomes will require custom innovations that meet stakeholder expectations despite resource constraints. Combining the right mix of data, technology, and human expertise can provide the flexibility needed to support sustainable improvements into the future.
Read more about how the productivity of the global patent system can be enhanced by AI in a white paper by CAS entitled “Sustainability of the Global Patent System: the role of AI in enhancing productivity.
Acknowledgements: Matthew Bryan and Andras Jokuti, Patents and Technology Sector, WIPO
Bruno Poulequen, Ulrike Till and Young-Woo Yun, Infrastructure and Platforms Sector, WIPO
Edited by: Catherine Jewell