- Long predicted rise in artificial intelligence to improve efficiency of sorting big data
- Researchers find AI drastically aids similar searches during trademark examinations
- Exclusive guest post looks at how AI should be considered at national IP offices
There is little doubt that artificial intelligence (AI) will transform many industries across the globe. Now, researchers from the Ben-Gurion University of the Negev in Israel have looked at how advanced AI tools will transform how trademark searches are conducted in the future – and, in an exclusive guest post, expand on what it means for trademark practitioners.
WTR has written before about how artificial intelligence could change the status quo for those working in the trademark industry. Indeed, late last year, we cut through concerns about how the development of AI technology could impact legal, and specifically trademark, practice.
One of the primary areas that AI could significantly improve the trademark ecosystem is through improved search. In research released in the past few weeks, it was found that IP office examinations could be a game-changer, especially when it comes to more efficiently identifying trademark similarity.
Now, in the exclusive guest post below, the authors of the study, Idan Mosseri, Matan Rusanovsky and Gal Oren from Ben-Gurion University of the Negev, explain some of the key findings from their research and why IP offices should consider how AI could improve their internal processes. All three academics are co-founders of AI tool TradeMarker, while Rusanovsky is also a researcher at the Israel Atomic Energy Commission, and Mosseri and Oren are researchers at the Nuclear Research Center - Negev.
The main function of patent offices worldwide is to provide legal protection of industrial intellectual properties through the registration of patents, designs and trademarks. Granting a right to intellectual properties depends on the examination of the specific application. Therefore, an examination is essential to ensure the exclusivity for said property.
The current examination process at national IP offices is done manually and slowly, using human trademark examiners who are required to conduct a massive search in a large unordered database. On top of that, they are deciding whether there is any similarity between the trademark submitted via application and the already approved marks using old-fashioned techniques such as the Vienna classification. According to our study, automation of the examination process using artificial intelligence – with the supervision of trademark examiners – can provide a solution for this problem with greater ease and higher accuracy.
However, before delving into trademarks, we would like to discuss a larger question – indeed, it is a question that we have been thinking about for the past year or so: what are we really talking about when we talk about this seemingly unnatural connection between AI and government?
Well, surprisingly, the best example we could find in this context is the United States Postal Service’s AI-driven address recognition. Although it might sound crazy, the USPS is considered to be a pioneer in the field of machine-learning – it was one of the first organisations to start making substantial investments in the technology of AI in the age of so-called ‘Big Data’. Let’s just think about that for a minute: there are roughly 154 million addresses in the United States, and a lot of online shopping to deliver. In all, there are roughly 160 billion deliveries a year. And somehow the USPS cut a third of their employees since the 80s, and still manages to deliver virtually all mail successfully. This surprising outcome derives from the fact that the USPS founded and funded the Center of Excellence for Document Analysis and Recognition (CEDAR) at the State University of New York in order to find the precise machine-learning technique to digitise hand and typed addresses. Amazingly, they did this in 1998. So now, many people don’t need to do this tedious job anymore.
Fast forward 20 years, and this novel technology is now here for everyone. There is now an awareness of the huge amount of quality data being generated every day, together with an understanding that the government can lead the introduction of AI for public good. This means that even the most non-technological driven offices can grasp that AI is not as futuristic as we might think – it’s already here and it could already be used for solving today’s challenges. One of them is the IP arena, and especially around trademark similarity.
Currently, there are already few systems which provide a good solution for the image similarity problem (eg, Google’s reverse image search). However, trademark similarity search is a much more complicated problem as it requires searching for dissimilar images as opposed to the more common approach of searching similar (or identical) images. In the latter, as long as the amount of similar images is sufficient, one could try to train a neural network-based model to catch similarities between images. For example, in order to teach the machine to differentiate between cats and dogs we should supply it with many images of cats and dogs. Unfortunately, in a trademarks database, this is obviously not the case. Moreover, catching differences between trademarks is far more complex since it is much harder to find pairs of similar trademarks, and on top of that, there is no formal definition of similar trademarks, as trademarks are considered to be similar only if they are deceptively similar.
To add another dimension of complexity to this problem, let’s think about what happens when we have two trademarks with very similar figures but with different text. If we look only on the text similarity of these trademarks, we will get bad results. What about trademarks with similar text and different figures? We cannot ignore text, as it might be inscripted in the trademark image itself, or aside some other figure which resemble other trademark. Furthermore, what about trademarks that are not really visually similar, but do contain the same content? These questions led us to the notion that we have to split the similarity question to different aspects, each with its own focus, and its own solution. We find that the main metrics that the human eye uses to quantify similarities between trademarks are the following:
- Visual similarity: do the two trademarks look visually similar?
- Semantic/Content similarity: do the two trademarks contain the same semantic content?
- Text similarity: do the two trademarks contain similar text in this way or the other?
We concluded that an improvement to the automatic examination process needs to be done by examining the trademarks ordered by a range of similarity aspects. To do this, we created ‘TradeMarker’, a machine-learning and deep-learning based application that can recognise and distinguish similarities between a new trademark and those already in the official trademarks database. Furthermore, it identifies it across all three of those metrics. We use this separation in order to focus on the best results from each metric rather than searching through an unorderly mixture of them – since accuracy is our primary goal, we found that separate lists to be the optimal solution to the trademark similarity problem.
While working with the Israeli Ministry of Justice (MoJ) and Israel Patent Office, we learned that in order to make our research practical to IP examiners, lawyers and the public, such an AI-driven system must withstand four crucial rules:
- Don’t miss a single pair in comparison to the old systems;
- Find as many pairs in the first 200 results;
- Detect pairs which don’t have a Vienna match;
- Bring the extra value of finding similarities that the old systems could not even detect and do so as fast as possible.
Under this strictness, it is our belief that TradeMarker is a leading AI innovation for trademark search. Our research finds that it can elevate immediate success rate to almost 80%, and also saves time and resources by IP offices by a factor of five. Indeed, we’ve found that it can identify almost all similarities that old systems could not detect (as our video demo demonstrates). Most importantly of all, perhaps is that – thanks to our research and our work with the Israel Patent Office – its effectively has been proven academically and practically.
When we look back, old fashioned techniques (in computer science terms) such as visual information retrieval by image processing or even AI-driven automatic tags recognition is not enough – those methods are unable to fully grasp the richness of trademarks as the human eye does. However, with the rise of computing power and the introduction of deep-learning that can mimic – and even outperform – human capabilities by automatic feature extraction, we understand that this is the technology that can finally compete. The force of computational neural networks, which emulate the way our brain learns to differentiate and catalogue the world around us, is still far from full human capacity. However, in the case of trademarks, it seems to fit like a glove.
To date, then, TradeMarker has been selected by Google as one of seven leading AI innovations for social good, and has been presented at Google events in California, Bangkok, and Tel-Aviv. Furthermore, research into the software will be presented later this month at the International Conference of Human-Computer Interface in Florida. On top of that, the Israeli MoJ is working on the deployment of TradeMarker for examiners and also to make it accessible to the Israeli public for pre-ruling decisions.
Our research, and future plans with TradeMarker, reveal that there’s a bright future when it comes to AI-powered trademark similarity searches. Our hope, then, is that other national IP offices and government agencies pick up the baton and consider how AI could improve their trademark search processes for both examiners and users.