AI is a core, transformative way by which we’re rethinking how we’re doing everything. – Google CEO Sundar Pichai
Patents were first digitized on a mass scale in the 90s. Since then most applications built on top of patent data have stayed more or less within the comfortable bounds of providing data access through keyword filters or fielded searches. With the advent of new and powerful AI methods in the last few years, however, many more things are being attempted. The possibilities are endless but there are many challenges to overcome too.
Patents Are Written Differently
A number of AI programs are available today that understand language to some basic extent. They are called language models. However, most language models are trained on data obtained from online resources such as Wikipedia and Reddit. What an AI learns depends on its training data, so these language models don’t perform as well on patent data.
Recently, a team at Google created language models trained specifically on patent text. These models are going to play an important role in the development of patent AI apps.
Patent Text is Hard
Patent attorneys are not famous for drafting patents in a way that facilitates legibility. They don’t call a spade a spade, they might call it a hand-operated earth moving instrument’. Using such inclusive terms is one thing that makes patents hard to read. Also, most patents also use some custom, made-up terminology. This might look like an unnecessary complication but it isn’t. Most patents describe new things and new things need to be named! This is a challenge for an AI model because how are you supposed to make sense of a term that you have never seen before? Even patent experts have to read (and re-read) patents with great deliberation to make sense of them. AI has to go a long way to reach that level.
Not Everything is in the Text
For a lot of patents, especially those related to mechanical engineering, it is hard to make sense of the text without looking at the images. The same is the case of a lot of chemical patents where a drawing of the molecular structure makes much more sense than its IUPAC name, which the AI cannot visualize. Even for other patents, images are an integral part. The specification is written in conjunction with the images. AI techniques that can understand text in view of its associated images are yet to be developed. This is a big and exciting challenge.
Low Signal to Noise Ratio
Most patents are like puffer-fish – they have a simple idea at the center that could be described in a few lines. But patents go on and on for pages describing it in excruciating detail. This is done for legitimate reasons, of course, but it is a huge challenge for AI algorithms because all the details throw the AI off-track. It is important for AI algorithms to separate the wheat from the chaff and get the point’.
Logical Structure
Patents describe inventions and most inventions are defined by a complex interaction of their components. Today’s AI algorithms don’t model such interactions and interrelationships very well. They might parse a piece of text and extract some semblance of the components being talked about, but they are not really good at understanding how those components are related to each other.
What’s Awaited?
We have already begun seeing quite a few groups trying to overcome these challenges of AI patent search separately. I feel it would be amazing if these groups could come together and collaboratively leverage AI to make prior art searches easily accessible for everyone. One not-for-profit initiative to overcome these challenges in AI patent search is PQAI – Patent Quality Through Artificial Intelligence. Do you know of any such initiatives?
