This decision is closely related to T0872/19 (reported here) in that it uses the same kind of technology (a numeric embedding of resources) as part of a machine learning algorithm. As in the previous decision, advantages associated with doing this were not considered technical as they were not applied to solve a technical problem / in a field of technology.
In the previous decision, the machine learning algorithm was used to provide a relevance score to help pick content (advertisements) for a web page. This was found not to be technical, so whether the algorithm improved the processing or not was not relevant. It was also not relevant whether the resources were considered to be images. Processing images for a non-technical purpose was also found to be non-technical. The details of the ML processing were thus not counted for the inventive step analysis. Here, instead of relevance scores, the algorithm produced classification scores in one of the requests classifying whether a search result was spam or not. Classifying search results or other resources as spam or not was not considered technical by the Board. So, following the reasoning in T0872/19, the invention was found to lack inventive step as a mere computer implementation of something non-technical, and the appeal was dismissed. The fact that a more accurate representation of the data was generated in a new way did not save the application as this did not, in itself, produce a technical effect.
The appellant tried a creative argument that the claim had to be seen as a simulation of a hardware circuit that classifies inputs and outputs and thus had a technical purpose, in line with T1227/05. However, the Board dismissed this line of argument quickly by referring to G1/19, reason 133: “The often-quoted criterion of T 1227/05 that the simulation constitutes an adequately defined technical purpose for a numerical simulation method if it is functionally limited to that purpose should not be taken as a generally applicable criteria of the COMVIK approach for computer-implemented simulations, since the findings of T 1227/05 were based on specific circumstances which do not apply in general.” And so, without an application in what the EPO considers to be a field of technology, ML and mathematical methods per se remain hard to patent, and only in the limited circumstances of a specific technical implementation.