Is the patent industry ready for deep-learning analytics?

Screen Shot 2017-05-30 at 1.50.25 PM.png

“Patent search is more of an art than a science” is a saying commonly uttered in patent industry circles. The reason being, there are millions of patents worldwide, and accurately searching among terabytes of unstructured natural language and drawing expert level analysis requires a degree of comprehension that existing search technologies are not able to provide. Thus, the art of the search.

Today, the most sophisticated patent search memos are keyword or semantic based. Inevitably,  typical search results represent patents covering field of use similarities, not process similarities. This is because keyword driven search delivers results representing concepts instead of the process or execution behind the concepts. For example, an input containing a 500-word description on self-driving cars or aerial drones will return hundreds of search results relating directly to autonomous driving, and flying drones, respectively. However, rarely will you find in these results a list of patents limited to the enabling methodologies in the underlying disclosure, such as, for example, a self-driving car data array featuring similarities to a data array in a smart phone application that measures an individuals workout routine. This is where a deep-learning approach has the potential to make the biggest difference: it focuses less on field of use similarities, and more on the structure of the underlying disclosure and how it relates to the underlying disclosure in other documents.

How does deep-learning and machine learning in general achieve this? At a high level, deep-learning performs best on text if the corpus (the entire text database you wish to search) is clustered in a way that treats each word as a numerical value, versus clustering based on keywords.


Figure from Yoon Kim’s Convolutional Neural Networks for Sentence Classification

The process of training the machine to interpret the meaning of the words is based on how each numerical value relates to other words with different, yet potentially related numerical values. To put it simply, the computer program is learning measurements between numerically represented concepts, instead of simply drawing lines between words, and bundling similar words together. It is the difference between knowing a tomato is a fruit and knowing that a tomato doesn’t belong in a fruit salad.


Our Decision to Release the Neural Network into IPDealRoom

On May 19th, ClearAccessIP quietly launched it’s neural network patent search as a feature of IPDealRoom. We’ve received various questions relating to how the neural network operates, why we designed the search as we did, and how best to leverage it. Furthermore, below, we’re providing some insight on improvements we’re working on.

“What are the benefits of embedded patent search?”

ClearAccessIP’s mission is to seamlessly inform on both the status of an IP portfolio and infer on its value.

IPDealRoom is the first structured diligence room for organizing and structuring an IP portfolio around the market for goods and services. It is powered by the automated IP management feature, which continuously monitors and updates whole patent records. The decision to integrate the neural network into each individual IPDealRoom was made to ensure that IPDealRoom acted as a well-maintained, monitored, and trusted source for information on the commercial value of IP.

Given the speed of today’s technology sector, it is increasingly helpful for patent holders to understand the ecosystem of the technology they’ve invested in. Our first product launch of the neural network is “IP Map” which users can turn on by sliding the “enable machine learning” button into the “on” position. Once on, the machine learning “analyst” reads the contents of an IPDealRoom. Notably, the system reads any attached patent records as well as invention disclosures. This then enables weekly updates of the IP Map, which viewers with access to the IPDealRoom can link through to at anytime.

An example of this process is outlined below:

  1. Create an IPDealRoom.


2. Slide “on” the Enable Machine Learning Button.

Screen Shot 2017-05-31 at 4.09.21 PM.png

3. Visit the IPDealRoom and select the “IP Map” button to review related               technologies.


4. Explore a list of the most closely related disclosures to the IPDealRoom portfolio.


What else can we expect to see from the neural network?

Within the next few weeks we will be rolling out various analytics features, including a dynamic organization chart, a “field of use” filter that is paired with the IP Map results, as well as an inventor leader list. Stay tuned!