Patent claim and description feature mapping and inventive step argument generation

Abstract

Using a large language model 706 to generate inventive step arguments in response to an obviousness rejection 701 against a patent claim in a patent office action 702. The large language model is given an input prompt comprising text from the patent office action and instructions to generate identifiers 704 associated with passages of a cited prior art document 703, such as paragraph numbers, verbatim quotes, or descriptions of paragraphs. The text of the prior art document is split into chunks 707, such as paragraphs, and chunks 711 are selected based on the generated identifiers. The large language model is then given an input prompt comprising the selected chunks of the cited prior art, the patent claim, the obviousness rejection, and instructions to generate inventive step arguments against the obviousness rejection based on these three pieces of text. Additional chunks of the prior art document may be selected to fill the large language model’s context window length up to a threshold. Additional chunks may be selected based on their position in the prior art document, or based on a distance between vector embeddings of the first selected chunk and a search term generated by a large language model.

Images