Machine Learning

Machine Learning

Why does using a Convolutional Neural Net give users better search results?

“It is the difference between knowing a tomato is a fruit and knowing that a tomato doesn’t belong in a fruit salad.”

“Patent search is more of an art than a science” is a saying in patent industry circles. 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.

There are three common methods of software-driven patent search marketed today, semantic analysis, supervised classifiers, and convolutional neural net training. ClearAccessIP believes that the convolutional neural net is the most powerful method. It can not only enable a hands-free search (no more Booleans!) but type of search that is flexible enough to work consistently and generate insights across industries.

*The most common patent search methodologies in order of their creation and sophistication.

Semantic Analysis

For decades, the most sophisticated patent search methodologies have been keyword or semantic based, yet they do little more than the most simple filter for relevance. Armed with a “bag of words”, this method compares keyword counts and ranks patents based on how many keywords they share. This simple match will inherently lack the context in which those key concepts were used and often creates over-broad, noisy patent lists with unclear review priorities.

Supervised Classifier

Supervised classifier-based systems have entered the market with the promise of more granular patent to technology mappings.

Developing a classifier-based system and taxonomy requires a multi-step process. A model owner starts with a sample of patents and some hypotheses about what they cover. Then, through a tagging process conducted by hand, they begin to organically define a set of classifiers and hierarchy in which those relate. Once a taxonomy has taken shape, each classifier becomes the basis for identifying similar patents across the patent corpus.

There are caveats to this approach. Because humans define every aspect of a taxonomy’s creation, there is inherent bias, such as over-emphasis on certain technical areas or trends. And, because it is a heavy and expensive process to create and maintain a taxonomy, the model owner must be diligent in covering all areas of technology carefully and execute reclassifications and updates in a timely fashion. This requires users to place significant faith in a black-box model.

Convolutional Neural Networks

ClearAccessIP’s Patent Intelligence AI™ provides users with insights to power strategic IP decision-making. ClearAccessIP’s Patent Intelligence AI™ relies on a convolutional neural network (“CNN”), that reads each paragraph and claim of each patent, breaks up the key inventive concept into tokens, and mathematically maps the relationships of overlapping concepts across more than 70 million worldwide filings.

*A portrayal of ClearAccessIP’s neural network.

This mapping finds results based on the process of how an invention works, rather than by searching across the plethora of words used to describe it. Finally, that neural network directly trains on the patent corpus and integrates new filings on a weekly basis, ensuring that users are provided with context rich and always up to date results.

Learn more about how the most advanced patent search on the market can help corporations, small-medium enterprises, universities, law firms, and investors unlock the value of their patent portfolio. Contact us with questions or to schedule a demo today at