Patenting is much about using natural language to define things according to established practices. Substance-related matters in the patent process are still strongly dependent on human expertise. However, recent developments in machine learning, a unique training dataset available and vast potential for increasing productivity make the domain an intriguing application area for artificial intelligence.
The patent process is a mixture of creativity, logic thinking, text analysis, verbal acrobatics and formalities. While there are tools to automate procedural formalities and routine quality checks, the most time-consuming parts of the process are still mainly manual work. The question is: can we automate these parts?
This blog post discusses some interesting possibilities with a focus on how the human intelligence stored in the patent data during the past years can serve the patent applicants of tomorrow.
Tens of thousands of patent engineers, searchers, attorneys and examiners around the world form a thinking machinery that turns hundreds of thousands of inventions to patent applications and further to granted patents each year.
Even small improvements in the process would bring enormous savings. With current filing rates, a cost reduction of 1 k€ for each patent application filed in the U.S. and Europe only would save almost 1 B€ annually. That would be only a small fraction of the costs the pipeline of actors mentioned above form. Only the attorney and examination costs for a single patent are typically in the order of 10 to 20 k€.
The estimate above does not even take into account Asia which has passed the Europe and U.S. in the number of filings long ago.
The potential for saving time and money by patent automation is huge.
Currently, prior art searches are typically made as separate steps and require special expertise. In addition, crawling through the results is still hard manual work. If some of this can be carried out automatically while filling an invention disclosure form or drafting a patent application, the overall process can be greatly improved.
Let’s take a look at the patent grant process. It aims at finding a verbal definition (claim) that sufficiently distinguishes an invention from everything that is known before filing of the patent application (the prior art). Much of the prosecution work is related to coping with documents that you weren’t familiar with when drafting the patent application. But you could have been.
So, what happens before filing of the patent application not only fundamentally determines the maximum scope of the patent but also largely defines how swift the grant process is. Regarding both quality and productivity, emphasis should be on timely finding and efficient analysis of the relevant prior art.
Besides prosecution, there are other information-intensive tasks, like competitor monitoring and freedom-to-operate analyses, which also need more powerful tools as the patent mass keeps growing.
Machine learning models love quality data. Let’s take a look at what we’ve got.
Patent data from the major patent systems is publicly available. The data contains not only full texts and drawings of granted patents, but also published patent applications. There is also a lot of patent examination information, most notably references to prior art. Typically these are other patent documents. The examination data is the most interesting — and most unused — part.
The dataset actually contains decades of experience and expertise of patent examiners and attorneys put into practice. The question is: how can you not learn something from that?
Despite the data is mostly inherently complex natural langugage, it has also many benefits from the machine learning perspective:
There’s lots of it. The data from the USPTO dates back to the 1920’s. The EPO offers data from four decades. That’s millions of documents.
It is relatively commensurate over time. The fundamentals of evaluating patentability have not changed much.
It is straight to the point. As real-world data, it reflects how the system works in practice. Each single entry is a result of hours of human intelligence work.
Patent documents have a certain structure. Although we are dealing with natural language, established practices and patent jargon make the problem simpler.
It’s continuously expanding. There are hundreds of thousands of new entries annually from the major patent jurisdictions. The level of details is also on the rise.
For these reasons, the dataset forms a unique training corpus for machine learning algorithms.
The dataset provides answers to questions like: If you file a claim with content X, what prior art citations C is the examiner likely to present? What limitations to claim X are sufficient to get rid of citations C, i.e. to get the patent granted? These questions can be formulated as neural networks. By training them with the dataset, they will be predicting most likely outcomes for new data inputs. For example new inventions.
Of course, there is still an interesting challenge: how to model patent text so that a computer is able to do something meaningful with it?
The work done around natural language processing (NLP) over the years certainly helps. For example part-of-speech tagging, dependency parsing, vectorization of words and recurrent neural networks (RNNs) can be used to increase a machine’s understanding of the actual content of the text.
Mapping of patent texts to graphs is also an interesting option. There has been some research work relating to the use of generic conceptual graphs for summarization of patents and expressing the content of patent claims. The results are promising.
What also helps is that patents are not just any pieces of text. Their sole purpose is to define technical concepts. Each part of the specification has its own function. There are reasons why certain things are expressed in certain ways. There are established practices for interpretation of text. All of this is domain-specific.
Therefore, incorporating domain knowledge into the data graphs and machine learning models so that they better reflect how things are seen in the patent world may be the key to success. This way, the models are able to learn the domain-specific principles faster. And what’s most important: to give sensible results that allow quick decision-making and smooth workflows.
A virtual patent specialist is perhaps not that far away after all.
There are several use cases for the technology.
Think about a smart invention disclosure form or assisted patent application drafting. You’d get instant search results and even estimate of patentability and be able to make quick decisions. Or how about an invention sieve or automated freedom-to-operate monitor? Just put in technical data and get it automatically ranked according to IPR possibilities or threats. It’s not so much about technology or data anymore, but putting it all together.
IPR and AI are bound to have a stronger relationship. The sooner the better.