Using Deep Neural Networks to Find the “Best” Known Cited Art and Create Claim Charts

When initially reviewing a patent portfolio to understand its scope, it is helpful to understand the universe of art that is already known in the portfolio.  When there are a large number of known art references, it is extremely helpful to be able to focus on the "best" art that is of record.

One way to go about this is to read the prosecution history; theoretically the best art was applied by the Examiner during prosecution.  However, the best art (and portions of the art) may not have been applied for any number of reasons.  For example:

• The examiner applied the best art, but didn’t apply the best portions of the best art;

• There have been a large number of patent citations provided during prosecution and the Examination missed a particularly relevant piece of art for a claim element;

• A patent portfolio has grown larger over time, the pool of known art has grown, and older patents are being asserted where newly cited art had not been considered.

Using deep neural network sentence encoders, we can automate the discovery of the "best" parts of the known art in a portfolio.

In my article “Using Deep Neural Networks to Automate the Creation of Specification Support Charts,” we looked at the Google "PageRank" Patent No. 6,285,999 with respect to analyzing specification support under 35 USC 112(a).  Here, we will again use this patent as an example with respect to analyzing the "known" art cited in this and its related patents.

The known art for a particular patent may be obtained from a number of sources.  Here, we will look at the Google Patents "Patent citations" section, as shown below.  The list above "Family To Family Citations" were patents that were cited during the prosecution of the application that led to the '999 Patent.  The list below the "Family To Family Citations" are patents that were subsequently cited in continuing applications related to the '999 Patent.

Using the deep neural network sentence encoder, we can loop through the list of known references above, obtain their disclosure, turn the disclosure into a vector containing the specification paragraphs or sentences, encode the specification vector, and compare the similarity to the encoded specification vector of the claim elements for the '999 Patent, as illustrated below.

Based on this comparison, we can determine a similarity score between the claims of the claims of a subject patent (e.g., the '999 PageRank Patent) and each known art reference (e.g., those listed above).  Then, we can rank and order the art based on the similarity of the specification with the claim elements.  The ranking for the similarity of the '999 PageRank Patent Claim 1 against the known art as shown below.

Here, the three highest ranked known references are owned by Google, Stanford, or have Lawrence E. Page listed as an inventor. It makes sense that these would be the most similar. The specifications are likely nearly verbatim or cover the same system.

The highest ranked third-party patent is Patent No. 5,848,407, Ishakawa et al., assigned to Matsushita Electric Industrial Co.  A heat map (a red color scheme is used here to signify the heat map is for a reference that pre-dates the claims) and claim chart can be automatically generated to help analyze the strength of the patent relative to this art reference:

The Ishakawa et al. '407 Patent appears to provide support for a 102 rejection to the core Google patent claim elements (1-2) throughout its specification. The support for the first and last Google patent claim elements in the Ishakawa '407 Patent specification does not appear as strong.

We can then take this information, and generate an invalidity chart to further evaluate the reference:

Here, the art appears close for each of the claim elements.  The last element arguably has the weakest similarity between the claims and the prior patent depending on whether "processing the linked documents according to their scores" is anticipated by "displaying" hypertext documents with other information such as summaries.

This is interesting, but what happened during prosecution of the Google '999 PageRank Patent? The same reference was identified and used as a primary reference by the Examiner!

The Examiner rejected the Claim 18 (which later issued as Claim 1, amended) under 35 USC 102 in view of the same U.S. Patent 5,848,407 to Ishikawa et al.!

Interestingly, this is the same “known” reference found by the deep neural network. This shows that the Examiner and the deep neural network technique “agree” that of these references Ishikawa et al. is the “best” reference.

Non-Final Office Action Dated 2000-12-05, Application Serial No. 09/004,827, page 2.

It is interesting to compare the Examiner’s citations to those provided by the deep neural network method. The Examiner cited the Abstract and the first paragraph of the Summary of the Invention.  In comparison, the deep neural network method focused primarily on the description of FIG. 2, the fourth paragraph of the Summary of the Invention, column. 15 lines 60-68, and column 4 lines 14-17 for each '999 Patent Claim 1 element, respectively.

In response, the Applicant pointed out that the 5,848,407 Patent did not qualify as prior art. See Applicant Remarks Dated 2001-02-06, Application Serial No. 09/004,827, page 7. The Application was subsequently allowed in view of the argument that Ishikawa was not prior art.

Was the claim supported by the provisional application under 35 USC 112, and should the office have accepted the assertion that the claim was entitled to the provisional priority date?  See my prior article "Using Deep Neural Networks to Analyze Priority Claims to Provisional Applications."

Additional known patent references can be analyzed in this manner to further understand the relationship between the known art cited during prosecution and the claimed inventions.

This technique can be used to quickly and effectively assist in analyzing the strength of the claims of a patent portfolio. It can help ensure that you are aware of the scope of each patent claim in a portfolio; or, it can help you identify weakness in your competitors portfolio.

If you would like to explore this process further, contact JDB IP.