US 12,462,099 B2
Language model-based multi-objective optimization
Robert Plotkin, Englishtown, NJ (US)
Assigned to Quabbin Patent Holdings, Inc., Marlborough, MA (US)
Filed by Quabbin Patent Holdings, Inc., Marlborough, MA (US)
Filed on Dec. 20, 2023, as Appl. No. 18/391,114.
Claims priority of provisional application 63/458,682, filed on Apr. 12, 2023.
Claims priority of provisional application 63/451,435, filed on Mar. 10, 2023.
Claims priority of provisional application 63/440,200, filed on Jan. 20, 2023.
Claims priority of provisional application 63/435,147, filed on Dec. 23, 2022.
Prior Publication US 2024/0211685 A1, Jun. 27, 2024
Int. Cl. G06F 40/20 (2020.01)
CPC G06F 40/20 (2020.01) 20 Claims
OG exemplary drawing
 
1. A method performed by at least one computer processor executing computer program instructions stored on at least one non-transitory computer-readable medium, the method comprising:
(A) receiving product data identifying a product;
(B) receiving patent data identifying a patent claim;
(C) receiving prior art data identifying a set of prior art;
(D) executing an optimization process to minimize an objective function, the objective function including: a first sub-objective function for evaluating infringement of the patent claim by the product, and a second sub-objective function for evaluating validity of the patent claim over the set of prior art, the executing the optimization process comprising:
(D)(1) generating a plurality of interpretations of the patent claim;
(D)(2) for each particular interpretation in the plurality of interpretations of the patent claim:
(D)(2)(a) using a first autoregressive language model to evaluate the first sub-objective function by:
(D)(2)(a)(i) generating a first query text comprising the product data and the particular interpretation;
(D)(2)(a)(ii) providing the first query text as input to the first language model to generate first autoregressive language model output text; and
(D)(2)(a)(iii) processing the first language model output text to produce an output of the first sub-objective function;
(D)(2) (b) using a second autoregressive language model to evaluate the second sub-objective function by:
(D)(2) (b) (i) generating a second query text comprising the prior art data and the particular interpretation;
(D)(2)(b)(ii) providing the second query text as input to the second autoregressive language model to generate second language model output text; and
(D)(2)(b)(iii) processing the second language model output text to produce an output of the second sub-objective function;
thereby producing a plurality of outputs of the first sub-objective function and a plurality of outputs of the second sub-objective function corresponding to the plurality of interpretations of the patent claim; and
(E) generating, based on the objective function and the plurality of outputs of the first sub-objective function and the plurality of outputs of the second sub-objective function, optimization output representing at least one interpretation of the patent claim selected by the optimization process.