The power of patent search beyond boolean - Part II: Finding wrongly classified and accidental prior art

Patent publications are each assigned at least one classification term indicating the subject to which the invention relates. The taxonomy called International Patent Classification (IPC) is in use since 30 years and it has been updated and refined continuously leading to more than 60 000 classes that a patent can belong to. The number of classes in the newer CPC system is manyfold. Thus, wrong or incomplete classification occurs even for the most experienced examiners. 

The workflow within traditional Boolean patent search engines often includes using class filters to limit the results. However, either the prior art may be incorrectly or incompletely classified or the searcher may miss some relevant classes. Both result in relevant documents being missed. Even if the classifications would be more or less in order, there is a high probability that so-called “accidental prior art”, i.e. publications disclosing all the limitations of the claim, but relating to a distant field of technology, are missed. 

Can AI be of help in this case?

Let’s see it in practice: we start an invalidation search with publication number WO2018184655A1 (Claim 1), Dispensing box comprising a stack of sheet products. Below, you can see how the machine perceives the technology recited in the first claim (unedited, automatically generated knowledge graph of IPRally).

And this is how the unfolded box looks like in the drawings:

The original target publication has been assigned to IPC subclass B65D 83/08: “Containers or packages with special means for dispensing contents, for dispensing thin flat articles in succession”.

Almost all of the results we get in IPRally fall to this particular class, or close neighbours thereof. Also, based on the titles they unsurprisingly relate to “dispensers” and “sheets”. IPRally does not use the classifications in training or when doing the search, but makes the search only based on the technical content of the graphs. 

A closer inspection of the results reveals a strange hit titled “Educational toy” (US5720617A) among the top 50 hits. However, a look into the original document reveals that it is also about a foldable dispenser box:

‍

The document discloses many many features that correspond to those of the target claim, and these, as well as their relationships, are identified automatically by IPRally. Here's an automatically generated based on Claim 1 as a sample:

The reason why the Educational toy - despite its structural similarities - would not show up in a classical, restrictive Boolean search with filters, lies in its classification. 

The Educational toy is classified as follows:

• A63F 11/00 - Game accessories of general use
• A63F 9/00 - Games not otherwise provided for
• G09B 1/00 - Manually- or mechanically-operated educational appliances using elements forming or bearing symbols, signs, pictures, or the like which are arranged or adapted to be arranged in one or more particular ways
• G09B 19/00 - Teaching not covered by other main groups of this subclass
  

We see that these are from completely different sections A and G, not B as the target patent. Here, the Deep learning Graph-AI is of help because it pays attention only to the technical content and is not bound by the limitations of the current classification systems - neither IPC nor CPC classes had been used for training the algorithm.

Some of our top users’ comments are: “IPRally searches wider than Orbit and finds the relevant classes faster and in a simpler way” and “IPRally takes me immediately to the point where I’m normally with 7 rounds of Boolean iteration”. One user reported a particular strength of Graph-AI is in finding prior art when working in technical fields which he is was familiar with, involving new classifications, keywords and players.

Do you still lose precious time or miss crucial hits by limiting your search through classification codes? Test through our free invalidation search to see the difference!