Using Deep Neural Networks to Mine Your Patent Specification for Claims that Your Competitors Want
In my article, “Finding Forward Citations and Getting Claims Your Competitors Want Using Continuations,” I discuss the concept of obtaining claims that your competitors are seeking to obtain and that your patent specification pre-dates and supports.
In that article, the prerequisite to doing this was finding a forward citation to your patent where it was used to block the competitor:
If cited as a 102 reference (or a single reference 103) it means that the examiner has specifically determined that your patent is a "blocking patent" (or implicitly a "blocking patent" but for an obvious variation) to the claims sought by the later applicant. This is powerful and actionable information.
A patent owner that has its disclosure used to block later-filed claims by a competitor should consider substantially copying the competitor's later-filed rejected claims.
Sometimes, you may luck out and have an Examiner find and apply your patent to a competitor. Other times, you may not—even though your specification may be the best reference for the Examiner to apply. As I previously stated:
Not every patent portfolio will have these types of forward cations. However, where they are available, a patent owner can obtain the claims sought by their rejected competitor.
Wouldn’t it be nice if you could automate the discovery of competitor claims that read on to your prior patent specifications, even if an Examiner hasn’t appreciated this fact? This can be done using deep neural network sentence encoders.
How can we do this?
First, your patent specification is loaded into a vector containing the patent specification paragraphs or sentences.
Second, a search is performed against later filed published applications with priority dates after your specification. This search can be as narrow (e.g., limited to a specific competitor) or broad (e.g., encompassing as much art based on broad keyword searches or broad categories) as time and budget permits. A list of potential forward citation reference results is returned.
The pending claims of those applications are then each loaded into a vector of claim language elements.
The patent specification vector and claim language elements vector are both fed into the deep neural network sentence encoder, and then compared to form a similarity matrix heat map (as discussed in my prior posts available at Using Deep Neural Networks to Streamline Patent Prosecution and Litigation Tasks.)
A similarity / relevance score can be determined for each potential forward citation reference; optionally results below a threshold score level can be discarded.
The heat map can then be processed into a claim chart for each potential forward citation reference.
The results can be reviewed and, based on the results of the review, a continuation can be filed that substantially copies the claims sought by your competitor.
If you would like to explore this process further, contact JDB IP.
