Greatly improved machine learning models for patent proofreading

At Patent Bots, we've always used machine learning models (specifically, neural networks) to proofread your patent documents. Because we are a cloud-based service, we are able to use state-of-the-art machine learning tools to provide the most accurate results.

Our previous machine learning models were trained on normal English, but as you know, patent claims are quite different from normal English. Our previous models worked quite well, but now they work even better.

Our new machine learning models are trained specifically on patent claim language. Our new models thus understand patent claim language much better. We've actually decreased our error rate by more than 50%.  This means you'll have fewer false alarms, and more accurate proofreading than before.

Our new models provide other advantages to further improve our products:

• We are able to continuously improve proofreading performance.  As we come across errors, we can fix our training data or add more training data to prevent these errors from happening again.
• We are able to provide new services that depend on machine learning. For example, we are working to predict the art unit that your patent application will be assigned to.
• We are able to run our services on other platforms (e.g., AWS or Azure) and even allow on-premise solutions.

It is an exciting time to be in legal tech, and we are looking forward to improving your patent practice.

Patent Claim Grammar

As someone who has reviewed more than 10,000 claims and writes software for proofreading patent claims, I wish more attention was paid to claim grammar.

Even well-written claims are very hard to understand. Adding poor grammar to the mix makes claims painful to read and possibly more vulnerable to invalidity arguments in litigation.

Let's start with a couple of questions...

Is a patent claim a sentence?

Nope. It looks like a sentence. After all, it starts with a capital letter and ends with a period. But there is one important thing missing. A patent claim does not ever have a verb.

So what is a patent claim?

Each patent claim is a REALLY long noun. The noun is something like a method, system, or a non-transitory, computer-readable medium. All of the words that come after the noun just provide details of that noun.

Let's take an example:

5 Types of Patent Examiners – Master, Ideologue, Decider, Negotiator, and Granter

Any patent attorney knows that the behavior of patent examiners can vary greatly.  In this article, I describe 5 different types of patent examiners and suggest prosecution strategies for each to help you get better outcomes for your clients.  For each type, I include details of an actual patent examiner from Patent Bots examiner statistics.

The Master

The Master is a patent examiner with a low grant rate but who is also exceptionally good at his or her job.  Masters are able to find good prior art for just about every patent application that comes across their desk, and they thus have high affirmance rates on appeal.

Here is an example grant rate timeline for a Master:

This grant rate timeline shows the percentage of cases that are allowed (green), pending (yellow), and abandoned (red) at each month after the first office action.  This Master has allowed only 2.5% of applications at 3 years after the first office action and is in the 97th percentile for examiner difficulty across the USPTO.

Easy PAIR Access for Each Examiner's Applications

One useful strategy in responding to office actions is to see what arguments other attorneys have used to convince the patent examiner to allow claims. You can then try to use similar arguments to get claims allowed for your client. To facilitate this strategy, Patent Bots examiner statistics includes a list of recent dispositions for each examiner. Here is an example for one examiner:

Previously, you would use the publication number to find the IFW on the USPTO's PAIR website and then download office actions and responses. We are happy to report that we now do this for you! After each application, there is button to view the PAIR IFW right on the Patent Bots website (and a button to view it on Google Patents). Here is an example of a Patent Bots page with the IFW for a patent application:

The first column has the basic details of the patent application, and the second column has the IFW. By default, we hide some of the things that you probably don't care about, but you can see the entire IFW by using the button there.

In this case, there was a final rejection, the attorney filed after-final response, and the examiner then allowed the claims. If you had a case pending before this examiner, you might want to look at the arguments that attorney used to persuade the examiner to allow your claims.

We plan on expanding upon this in the future. For example, for each office action, we are planning to list the rejections (e.g., 101, 102, 103, or 112) in the table. That way, you can more quickly find successful attorney arguments on the issue you are facing.

Massive Speed Improvements in Processing Patent Applications

N. Tesla Speed Indicator Patent

Patent Bots was already the fastest gun in the west for automated patent proofreading. In a previous comparison with a competitor, we were 33 times faster over 100 random applications. Not being content with that, we've more than doubled our speed.

There are two ways we can make our processing faster: (1) hardware and (2) algorithms.

For hardware, we've upgraded are servers to double the processing speed. It costs us more money, but our customers are worth it. Our goal is to save you time and make you as efficient as possible.

On the algorithm side, we've greatly increased the efficiency of processing Microsoft Word documents. You'll especially notice the improvements when processing longer patent applications.

With these improvements, we were able to proofread the claims of a 543 page patent application in 23 seconds! By contrast, opening this patent application in Microsoft Word on my Mac took more than two and half minutes to completely open the document.

Our processing of reference labels is slower than processing the claims (since we need to do more processing of entire specification), but we are working on improving our speed of processing reference labels as well.

Should you interview your patent examiner?

We've added new statistics to Patent Bots to help you make better decisions regarding examiner interviews.  For most examiners, their grant rate increases for cases that have interviews over those that don't.  For some examiners, however, their grant rate actually goes down.

Here is an example:

You can see this examiner's grant rate for cases without any interviews, the grant rate for cases with at least one interview, and the benefit from doing an interview (along with the same stats for the art unit and the USPTO).

For this examiner, you are 8% less likely to get a patent if you conduct an interview!  By contrast, the USPTO average is that your chances increase by 15%.

This examiner also seems to avoid doing interviews.  Of the 157 cases before him, only 12 of them have had an interview.  If you have this examiner, you are better off not requesting an interview.

Here is a summary of some other improvements over the previous month:

• For reference label processing, we process figures in PDF format that include text ( we don't OCR yet)
• Improvements in our antecedent basis analysis.
• Better extraction of ref labels so that we don't count things like IP addresses as reference labels.
• Better handling of multiple dependent claims.

Visualizing Outcome Inconsistency at the USPTO

In an ideal world, your chance of getting a patent allowed is based on the merits of your patent application and independent of the largely random assignment of the patent examiner.  As any patent attorney knows, however, this is not the case.  Some examiners allow patents too easily and others seem predisposed against allowing any patents at all.

This ideal can be described as outcome consistency.  The outcome of a patent application should be largely the same regardless of the assigned patent examiner.  Outcome consistency is needed to ensure fairness.  It is unfair for an applicant to be denied a patent for a worthy invention because it was assigned a hard examiner, and it is unfair to the public for a patent to be granted for an unworthy invention because it was assigned to an easy examiner.  The lack of outcome consistency among patent examiners is a known issue that the USPTO is working on improving, and this article presents visualizations to help diagnose areas for improvement.

The patent application grant rate across the USPTO is 66% (computed as described here).  One would expect that a distribution of examiner grant rates would follow a bell-like curve with (i) the average examiner having a grant rate of 66% and (ii) a reasonably small standard deviation such that most examiners are close to the average.

Here is the actual distribution of examiner grant rates across the USPTO (this is a weighted histogram according to the number of cases handled by an examiner with SPEs excluded):

USPTO: