Comparison of Ambercite to other engines for patent invalidation

Ambercite is often benchmarked against other AI patent search engines for such tasks as patent invalidation.

This is a reasonable thing to do so, as both Ambercite and other AI search engines both claim to be able to find relevant patents in a different way to conventional keyword and class code patent searches.

However there is a major differences between Ambercite and semantic search engines:

  • Ambercite finds similar patents using a unique citation-based algorithm that does not directly take into account the keywords or class codes of these patents, and is therefore free of the risky assumptions keywords and class codes lead to.

  • In contrast, semantic based AI search engines look for similar patents based on the similarity of languages. In some cases this can quick and simple - but there is a risk of both of ‘false positives’ (listed results that are not relevant) and ‘false negatives’ (results that are not listed that are relevant).

Not surprisingly, people have systematically compared Ambercite to other search engines - including work done by the Austrian patent office in 2019, which found Ambercite to perform best or best equal in a series of comparisons based on a patentability search.

Can we test how well Ambercite does patent invalidation?

There are a number of services out there that run ‘contests’ to find new prior art for granted patents, including Patroll. Of note, Patroll are currently running a contest to help invalidate US7885981, filed by a Mr Kaufman for a System and method for generating automatic user interface for arbitrarily complex or large databases (in particularly for relational databases). This contest invited people to find new prior art, and even provided lists of potential prior art generated by other people or systems, including:

So how does Ambercite compare? The decision of which list is best is a big question, and perhaps best done by others. However a second question is:

Does Ambercite provide potentially relevant patents that are not provided by these other approaches?

If so, this would confirm the key value proposition of Ambercite, namely that it complements keyword and semantic searching, as per the image below.

 
Circle.jpg
 

How does Ambercite compare to the above lists?

To answer this question, firstly we will run an invalidation search in Ambercite. In the search below, we are looking for the 100 top-ranked patents with prior art dates earlier than the 31 October 2000 priority date of US7885981:

Invalidation query.jpg

Ambercite can filter patents on dates to the nearest month, which is fast and easy to use and works perfectly well in 99% of cases. In this particular case, the first two patent family returned by this query are closely related to the query patents (filed by the same applicant, but as separate patent applications), and have the same priority date - but these are easy to hide via the hide button, circled below:

 
Hide button.jpg
 

For the rest of the patents, we have reviewed them, and identified a list of 59 ‘Liked’ patents - in this case patents that meet both of these criteria:

1) Not previously cited against this patent (Ambercite refers to these uncited patents as ‘unknown’ citations. There are 83 ‘Unknown’ patents in this list, and the highest ranked of these was in the second position on this list.

2) Making a reference to both ‘display’ and similar concepts, and also to relational databases (or relational data).

In other words, potentially relevant and new documents.

The first two of the resulting set of 61 liked patents is shown below - and you can click on this image, to see the full list of results in a fully interactive form.

Many of the patents in the full list appear to be quite relevant, which is a good thing. But how many of these patents were not found in the other lists of prior art patents supplied by these other analysts or AI systems?

This is a fairly complex question, but what we have done is:

1) Obtained a top 100 list from the four different lists discussed above

2) For each patent in any of the lists, and also the Ambercite list, identified its Patseer simple family number - this is a unique identifier for each simple patent family. This allows rigorous and systematic data matching between the lists, regardless of patent number format or family member listed.

3) Cross-matched the these top 100 results to the top 100 Ambercite results. If the patent is found in any of these other lists, we have reported its ranking below (in columns L to O). If the patent is not found in these other lists, no ranking is reported.

The results are shown in part below, and if you click on this image, you can see the full list of results in spreadsheet form.

Some key columns include:

  • Column D shows where the patent is ‘unknown’ and relevant (either blue or orange), and not found in other any of the above lists of patents (orange). There are 44 of these previously not listed anyone, and yet still potentially relevant patents.

  • Column K shows the patents that are found in any of the other lists of results, and if so, how many times.

This data is summarised in the chart below:

 
 

We can also see that:

  • From the results show, there was little correlation between the results and rankings in the other four lists of results

  • And the same applies to the patents found the top 100 list of Ambercite results where they are found in the other lists - there is no correlation between these rankings and the Ambercite ranks.

Summary - How Ambercite can strengthen patent invalidation searching

This blog has shown that:

  • How easy it is use to Ambercite to find new and relevant prior art for a patent invalidation search. All we needed was the patent number and to set up a date filter - and we ended up of 83 previously uncited (‘Unknown’) citations. After a review, we identified 61 of these as being new documents in the right technical area.

  • How Ambercite will find relevant patents not listed by other manual or automatic search services. We end up with a list of 44 of these entirely new and yet potentially relevant patents.

This has helped confirm the Ambercite value proposition - “No search is complete with it”

 
Comparison of Patent Search Methods, Prior Art SearchingMike Lloyd