DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
Applicant’s remarks filed 12 June 2026 have been fully considered.
The rejection under section 112(d) is withdrawn.
Applicant argues that the claims contain a particular set of rules for training a single unified embedding model. Examiner respectfully disagrees. If anything, the claimed rules are now less specific than they were before amendment. Before amendment, the rules were specifically that the parameters of the single unified embedding model were updated by backpropagating a loss function. This was, of course, an abstract calculation. Indeed, it is the only disclosed means of training the parameters of the transformer embedding model using the loss function, which is why amending “updating the parameters of the transformer embedding model by backpropagating one or more terms obtained by the loss function” to “training the parameters of the transformer embedding model using the loss function” represents new matter.
Applicant has attempted to amend the claims taking the example of Example 39 which states that “training the neural network in a first stage using the first training set” does not recite an abstract idea. However, the claims do not recite “using a training set” but rather using the loss function which was computed proportional to dot products of an estimated query entity embedding and an estimated target entity embedding. This therefore recites a process of setting numerical parameters of the unified embedding training model based on a numerical loss function. This recites a mathematical relationship between the parameters and the loss function, and a step of mathematical calculation of the parameters based on the loss function. As the training, as recited, is abstract, it cannot provide an improvement to technology. MPEP § 2106.05(a).
Applicant argues that “item embeddings are generated using a different model than the TEM.” Examiner respectfully disagrees. The TEM generates embeddings for queries and items based on words in the query and words in reviews associated with the item. Crucially, the same embedding model, i.e., TEM, is recited as computing the embeddings for items and for queries. As Bi states, Bi pg. 2 (Section 3.2, “As shown in previous studies [2]”; Section 3.3, “As in [2]””), this is as in Ai et al., Learning a Hierarchical Embedding Model for Personalized Product Search, pg. 3 (“Embedding-based User/Item Language Model” . . . “Formally, given e ∈ Rα as the latent representation of an entity (which could be either a user or an item) and w ∈ Rα as the embedding of a word w”).
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 1-6, 8-13, and 15-20 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
The limitation, “training the parameters of the transformer embedding model using the loss function,” is a functional limitation. MPEP § 2161.01. However, Applicant has only disclosed a single species of training: by backpropagating one or more terms obtained by the loss function. Specification ¶ [0037]. This is not a sufficient showing that Applicant had in its possession the full scope of the claim. There is no indication that any means other than by backpropagation of using the loss function to update the parameters of the transformer embedding model would result in the transformer embedding model being capable of performing various types of queries on external data in order to perform tasks such as question-answering, text summarization, text generation, and the like based on information contained in the external dataset. Applicant has therefore not disclosed species sufficient to support a claim to the functionally-defined genus. MPEP § 2161.01(I).
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-6, 8-13, and 15-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
As per claims 1, 8, and 15:
The claim(s) recites an abstract idea.
The limitation, “applying parameters of the transformer embedding model to descriptions of query entities for a first set of pairs for a current iteration to generate estimated query entity embeddings for the query entities,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., applying an attention algorithm to the query entities. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “separately applying parameters of the same transformer embedding model to descriptions of target entities for the respective set of pairs for the current iteration to generate estimated target entity embeddings for the target entities,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., applying an attention algorithm to the target entities. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “computing dot products between the estimated query entity embeddings and the estimated target entity embeddings corresponding to each pair in the first set of pairs,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., calculating a dot product. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “computing a loss function that is proportional to the dot products for the first set of pairs,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., calculating a loss function. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “training the parameters of the transformer embedding model using the loss function,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., determining parameters, which are variables or numbers, using a loss function, a mathematical method. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “generating query entity embeddings and target entity embeddings using the transformer embedding model,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., applying an attention algorithm to calculate embeddings. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “computing dot products between the query entity embedding and the target entity embeddings to generate a plurality of scores,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., computing a dot product. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “selecting a subset of target entities based on the plurality of scores,” as drafted, is a process that, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. For example, in the context of this limitation, “selecting” encompasses a person forming a judgment, e.g., as to which target entities correspond to target entity embeddings whose dot product between itself and the query entity embedding are the largest. This limitation therefore falls within the “Mental Processes” grouping of abstract ideas. MPEP § 2106.04(a)(2)(III).
Accordingly, the claim(s) recites abstract ideas. MPEP § 2106.04(a).
For the purposes of evaluating whether the claim(s) is directed to an abstract idea or is significantly more than an abstract idea, these recited abstract ideas can be considered together as a single abstract idea, namely the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function. MPEP § 2106.04(II)(B). The claims themselves do not recite how to achieve transformer embedding model parameters “trained” using a loss function. Rather, this important and distinguishing element of the claims is described in a result-oriented way. That is, the claims recite transformer embedding model parameters trained using a loss function such that it achieves the objective of obtaining a target entity related to the query entity better when compared to prior art transformer embedding model parameters, Specification ¶ [0038]. Therefore, the claims cover the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function. See Constellation Designs, LLC v. LG Elecs. Inc., No. 2024-1822, slip op. at 17-18 (Fed. Cir. 28 April 2026).
The abstract idea of the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function is not integrated into a practical application.
The additional element, “obtaining training data including a plurality of pairs, wherein a pair includes a respective query entity and a respective target entity,” is mere instruction to apply the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function because it recites the outcome of obtaining training data for calculation of the loss function without detail of how the training data is obtained, and is insignificant extra-solution activity as mere data gathering. MPEP §§ 2106.05(f), 2106.05(g).
The additional element, “accessing a transformer embedding model,” is mere instruction to apply the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function is insignificant extra-solution activity because it recites accessing the transformer embedding model being optimized without detail of how the transformer embedding model being optimized is accessed, and is insignificant extra-solution activity as mere data gathering. MPEP §§ 2106.05(f), 2106.05(g).
The additional element, “dividing the training data into one or more batches for one or more iterations of training the transformer embedding model,” is mere instruction to apply the recited abstract idea on a computer because it recites only the idea of batched training without details of how it is to be accomplished. MPEP § 2106.05(f).
The additional element, “storing the query entity embeddings and the target entity embeddings in a database,” is mere instruction to apply the abstract idea of archiving the embeddings for reuse because the outcome of storing the embeddings is recited without detail of how the embeddings are stored, and is insignificant extra-solution activity as mere data gathering. MPEP §§ 2106.05(f), 2106.05(g).
The additional element, “retrieving, from the database, a query entity embedding for a particular query entity and the target entity embeddings,” is mere instruction to apply the abstract idea of archiving the embeddings for reuse because the outcome of retrieving the embeddings is recited without detail of how the embeddings are retrieved, and is insignificant extra-solution activity as mere data gathering. MPEP §§ 2106.05(f), 2106.05(g).
The additional element, “transmitting instructions to a client device to cause display of the subset of one or more target entities corresponding to the target entity embeddings,” is insignificant post-solution activity as mere output of the target entity obtained by the transformer embedding model. MPEP § 2106.05(g).
As an ordered combination, the claims are mere instruction to apply the mathematical concept of optimizing a transformer embedding model that obtains a target entity related to a query entity using a loss function because the outcome of how to achieve the important and distinguishing element of training transformer embedding model parameters using a loss function is described in a result-oriented way and without detail of particular steps or procedure to achieve the result. MPEP § 2106.05(f).
Accordingly, the additional elements, individually or in combination, do not integrate the abstract idea into a practical application, even viewing the claim(s) as a whole, and therefore the claim is directed to an abstract idea. MPEP § 2106.04(d).
As discussed above with respect to integration of the abstract idea into a practical application, the conclusions for the additional elements being generic computer components and mere instructions to apply on a computer, insignificant extra-solution activity, and/or mere field of use limitations are carried over and these additional elements do not provide significantly more than the abstract idea. MPEP § 2106.05(II).
In re-evaluating the limitations that are insignificant extra-solution activity, the following limitations represent elements that have been recognized as well-understood, routine, conventional activity within the field of computer functions:
The additional element, “obtaining training data including a plurality of pairs, wherein a pair includes a respective query entity and a respective target entity,” is well-understood, routine, and conventional activity because it is storing or retrieving information that is recited at a high level of generality similar to the activity of storing and retrieving information in memory. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363 (Fed. Cir. 2015).
The additional element, “accessing a transformer embedding model,” is well-understood, routine, and conventional because transformer embedding models are described, Specification ¶¶ [0033]-[0035], in a manner that indicates that they are sufficiently well-known that the specification does not need to describe their particulars to satisfy 35 U.S.C. 112(a). MPEP § 2106.07(a)(III)(A); see MPEP § 2161.01.
The additional element, “dividing the training data into one or more batches for one or more iterations of training the transformer embedding model,” is well-understood, routine, and conventi0onal activity because batch training is described as the means for training in Vaswani et al., Attention Is All You Need, pg. 7, and at the time of filing this publication was understood to be well-understood, routine, and conventional, Wikipedia, Attention Is All You Need, pg. 1 (“It is considered a foundational paper in modern artificial intelligence, as the transformer approach has become the main architecture of large language models like those based on GPT.”), regarding transformer embedding models, Specification [0002].
The additional element, “storing the query entity embeddings and the target entity embeddings in a database,” is well-understood, routine, and conventional activity because it is storing and retrieving information in a manner that is recited at a high level of generality similar to the activity of storing and retrieving information in memory. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363 (Fed. Cir. 2015).
The additional element, “retrieving, from the database, a query entity embedding for a particular query entity and the target entity embeddings,” is well-understood, routine, and conventional activity because it is storing and retrieving information in a manner that is recited at a high level of generality similar to the activity of storing and retrieving information in memory. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363 (Fed. Cir. 2015).
The additional element, “transmitting instructions to a client device to cause display of the subset of one or more target entities corresponding to the target entity embeddings,” is well-understood, routine, and conventional activity because it is presenting information in a manner that is recited at a high level of generality similar to the activity of using a computer interface to present information. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363-64 (Fed. Cir. 2015).
As an ordered combination, the claim simply appends well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the abstract idea of a mathematical algorithm to calculate vector embeddings because the claim as a whole amounts to nothing more than generic computer functions merely used to implement the abstract idea. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see BASCOM Global Internet Servs. v. AT&T Mobility LLC, 827 F.3d 1341, 1349 (Fed. Cir. 2016).
Accordingly, the claim(s) does not recite additional elements, either individually or in combination, that amount to significantly more than the abstract idea. MPEP § 2106.05. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
As per claims 2, 9, and 16:
The claim(s) recites an abstract idea.
The limitation, “computing, for each pair in the second set of pairs, a dot product between the estimated query entity embedding and the corresponding estimated target entities embedding,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., calculating a dot product. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
Accordingly, the claim(s) recites an abstract idea. MPEP § 2106.04(a).
The abstract idea of computing, for each pair in the second set of pairs, a dot product between the estimated query entity embedding and the corresponding estimated target entities embedding is not integrated into a practical application.
The additional element, “obtaining a second set of pairs for the current iteration, wherein a pair in the second set of pairs includes the respective query entity and a negative target entity,” is mere instruction to apply the mathematical concept of computing, for each pair in the second set of pairs, a dot product between the estimated query entity embedding and the corresponding estimated target entities embedding because it recites the outcome of obtaining the claimed second set of pairs for calculation of the loss function without detail of how the second set of pairs is obtained, and is insignificant extra-solution activity as mere data gathering. MPEP §§ 2106.05(f), 2106.05(g).
As an ordered combination, the claims is mere instruction to apply computing, for each pair in the second set of pairs, a dot product between the estimated query entity embedding and the corresponding estimated target entities embedding because the computer is invoked merely as a tool to perform the algorithm. MPEP § 2106.05(f).
Accordingly, the additional elements, individually or in combination, do not integrate the abstract idea into a practical application, even viewing the claim(s) as a whole, and therefore the claim is directed to an abstract idea. MPEP § 2106.04(d).
As discussed above with respect to integration of the abstract idea into a practical application, the conclusions for the additional elements being generic computer components and mere instructions to apply on a computer, insignificant extra-solution activity, and/or mere field of use limitations are carried over and these additional elements do not provide significantly more than the abstract idea. MPEP § 2106.05(II).
In re-evaluating the limitations that are insignificant extra-solution activity, the following limitations represent elements that have been recognized as well-understood, routine, conventional activity within the field of computer functions:
The additional element, “obtaining a second set of pairs for the current iteration, wherein a pair in the second set of pairs includes the respective query entity and a negative target entity,” is well-understood, routine, and conventional because it is recited at a high level of generality similar to the activity of storing and retrieving information in memory. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363 (Fed. Cir. 2015).
As an ordered combination, the claim simply appends well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the abstract idea of a mathematical algorithm to calculate vector embeddings because the claim as a whole amounts to nothing more than generic computer functions merely used to implement the abstract idea. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see BASCOM Global Internet Servs. v. AT&T Mobility LLC, 827 F.3d 1341, 1349 (Fed. Cir. 2016).
Accordingly, the claim(s) does not recite additional elements, either individually or in combination, that amount to significantly more than the abstract idea. MPEP § 2106.05. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
As per claims 3, 10, and 17:
The claim(s) recites an abstract idea.
The limitation, “wherein the loss function is inversely proportional to the dot products for the second set of pairs,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., calculating a loss function. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
Accordingly, the claim(s) recites an abstract idea. MPEP § 2106.04(a).
As the claim(s) recites no additional elements, the abstract idea is not integrated into a practical application, the claim is directed to the abstract idea, and the claim(s) does not amount to significantly more than the abstract idea. MPEP § 2106.07. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
As per claims 4, 11, and 18:
The claim(s) recites an abstract idea.
The limitation, “wherein an estimated query entity embedding is generated by applying the parameters of the transformer embedding model to a description of a user of an online system and an estimated target entity embedding is generated by applying the parameters of the same transformer embedding model to a description of an item the user interacted with,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., “applying” is an act of calculating using mathematical methods, e.g., linear regression, Specification ¶ [0062], to determine an embedding, which is a variable or number. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
Accordingly, the claim(s) recites an abstract idea. MPEP § 2106.04(a).
As the claim(s) recites no additional elements, the abstract idea is not integrated into a practical application, the claim is directed to the abstract idea, and the claim(s) does not amount to significantly more than the abstract idea. MPEP § 2106.07. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
As per claims 5, 12, and 19:
The claim(s) recites an abstract idea.
The limitation, “wherein an estimated query entity embedding is generated by applying the parameters of the transformer embedding model to a description of an item and an estimated target entity embedding is generated by applying the parameters of another item that is known to be a replacement for the item,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., “applying” is an act of calculating using mathematical methods, e.g., linear regression, Specification ¶ [0062], to determine an embedding, which is a variable or number. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
Accordingly, the claim(s) recites an abstract idea. MPEP § 2106.04(a).
As the claim(s) recites no additional elements, the abstract idea is not integrated into a practical application, the claim is directed to the abstract idea, and the claim(s) does not amount to significantly more than the abstract idea. MPEP § 2106.07. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
As per claims 6, 13, and 20:
The claim(s) recites an abstract idea.
The limitation, “applying the parameters of the transformer embedding model to descriptions of a plurality of query entities to generate the query entity embeddings,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., applying an attention algorithm to the query entity descriptions. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The limitation, “applying the parameters of the transformer embedding model to descriptions of a plurality of target entities to generate the target entity embeddings,” as drafted, is a process that, under its broadest reasonable interpretation, covers a calculation, i.e., applying an attention algorithm to the target entity descriptions. This limitation therefore falls within the “Mathematical Concepts” grouping of abstract ideas. MPEP § 2106.04(a)(2)(I).
The abstract idea of a mathematical algorithm to calculate vector embeddings is not integrated into a practical application.
The additional element, “storing the query entity embeddings and the target entity embeddings in the datastore,” is insignificant extra-solution activity as mere data gathering. MPEP § 2106.05(g).
As an ordered combination, the claims is mere instruction to apply a mathematical algorithm for determining a target entity from a query entity on a computer because the computer is invoked merely as a tool to perform the algorithm. MPEP § 2106.05(f).
Accordingly, the additional elements, individually or in combination, do not integrate the abstract idea into a practical application, even viewing the claim(s) as a whole, and therefore the claim is directed to an abstract idea. MPEP § 2106.04(d).
As discussed above with respect to integration of the abstract idea into a practical application, the conclusions for the additional elements being generic computer components and mere instructions to apply on a computer, insignificant extra-solution activity, and/or mere field of use limitations are carried over and these additional elements do not provide significantly more than the abstract idea. MPEP § 2106.05(II).
In re-evaluating the limitations that are insignificant extra-solution activity, the following limitations represent elements that have been recognized as well-understood, routine, conventional activity within the field of computer functions:
The additional element, “storing the query entity embeddings and the target entity embeddings in a datastore,” is well-understood, routine, and conventional because it is recited at a high level of generality similar to the activity of storing and retrieving information in memory. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see OIP Techs., Inc., v. Amazon.com, Inc., 788 F.3d 1359, 1363 (Fed. Cir. 2015).
As an ordered combination, the claim simply appends well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the abstract idea of a mathematical algorithm to calculate vector embeddings because the claim as a whole amounts to nothing more than generic computer functions merely used to implement the abstract idea. MPEP §§ 2106.07(a)(III)(B), 2106.05(d)(II); see BASCOM Global Internet Servs. v. AT&T Mobility LLC, 827 F.3d 1341, 1349 (Fed. Cir. 2016).
Accordingly, the claim(s) does not recite additional elements, either individually or in combination, that amount to significantly more than the abstract idea. MPEP § 2106.05. Therefore, as the claim(s) is directed to an abstract idea and does not recite additional elements that amount to significantly more than the abstract idea, the claim(s) is not patentable. MPEP § 2106.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-4, 6, 8-11, 13, 15-18, and 20 is/are rejected under 35 U.S.C. 103 as being obvious over Bi et al., A Transformer-based Embedding Model for Personalized Product Search, in view of Rohrer, Transformers from Scratch.
As per claims 1, 8, and 15, Bi teaches:
obtaining training data including a plurality of pairs, wherein a pair includes a respective query entity and a respective target entity, Bi pg. 1523 (“70% of all the available queries into the training set”);
accessing a transformer embedding model, Bi pg. 1521 (“a transformer-based embedding model”);
dividing the training data into one or more batches for one or more iterations of training the transformer embedding model, Bi pg. 1523 (“We train our model and all the baselines for 20 epochs with 384 samples in each batch”); and
for each iteration of one or more iterations:
applying parameters of the transformer embedding model to descriptions of query entities for a first set of pairs for a current iteration to generate estimated query entity embeddings for the query entities, Bi pg. 1522, section 3.2;
separately applying parameters of the same transformer embedding model to descriptions of target entities for the respective set of pairs for the current iteration to generate estimated target entity embeddings for the target entities, Bi pg. 1522, section 3.3;
computing dot products between the estimated query entity embeddings and the estimated target entity embeddings corresponding to each pair in the first set of pairs, Bi pg. 1522, section 3.1, where a dot product between an item vector embedding and a query vector embedding is calculated;
generating query entity embeddings and target entity embeddings using the transformer embedding model, Bi pg. 1522, equation 1; Bi pg. 1524 (“TEM not only retrieves more ideal items in the top 20 results”), where embeddings are generated such that the softmax can be calculated, thereby yielding the top 20 probabilities;
storing the query entity embeddings and the target entity embeddings in a database, Bi pp. 1522-1523, where embeddings are implicitly stored such that they can be retrieved in order to perform a search for top 20 matches;
retrieving a query entity embedding for the particular query entity and the target entity embeddings for the plurality of target entities, Bi pp. 1522-1523, where embeddings are retrieved in order to perform a search for top 20 matches;
computing dot products between the query entity embedding and the plurality of target entity embeddings to generate a plurality of scores, Bi pg. 1522, equation 1;
selecting a subset of target entities based on the plurality of scores, Bi pg. 1523, where the top 20 most probable items are selected; and
transmitting instructions to a client device to cause display of the subset one or more target entities corresponding to the target entity embeddings, Bi pp. 1521, 1524 (“top 20 results”), where the results are for being presented to a user in an e-commerce search,
Bi, however, does not explicitly teach:
computing a loss function that is proportional to the dot products for the first set of pairs;
training the parameters of the transformer embedding model using the loss function.
The analogous and compatible art of Rohrer, however, teaches using a gradient, a loss function as claimed, and backpropagation to learn the query, value, and key matrices, transformer parameters under a broadest reasonable interpretation. Rohrer pp. 15, 29-33.
It would therefore have been obvious to one of ordinary skill in the art at the time of filing to use the gradient backpropagation of Rohrer to learn the query, key, and value matrices of Bi, Bi pg. 1523, equation 7, in order to train the transformer.
As per claims 2, 9, and 16, the rejection of claims 1, 8, and 15 is incorporated, but Bi does not explicitly teach:
obtaining a second set of pairs for the current iteration, wherein a pair in the second set of pairs includes the respective query entity and a negative target entity; and
computing, for each pair in the second set of pairs, a dot product between the estimated query entity embedding and the corresponding estimated target entities embedding.
One of ordinary skill in the art would understand that the dot product of the query vector and item vector represents a positive relationship between query and item, such that the inverse of an item vector would represent that the user does not want to purchase the item.
It would therefore have been obvious to one of ordinary skill in the art at the time of filing to modify Bi to incorporate negative purchase signals that a user is explicitly uninterested in an item to result in a better-targeted item recommendation algorithm, thereby making obvious obtaining pairs of users and negative items for training.
As per claims 3, 10, and 17, the rejection of claims 2, 9, and 16 is incorporated, but Bi does not explicitly teach:
wherein the loss function is inversely proportional to the dot products for the second set of pairs.
One of ordinary skill in the art would understand that the dot product of the query vector and item vector represents a positive relationship between query and item, such that the inverse of an item vector would represent that the user does not want to purchase the item.
It would therefore have been obvious to one of ordinary skill in the art at the time of filing to modify Bi to incorporate negative purchase signals that a user is explicitly uninterested in an item to result in a better-targeted item recommendation algorithm, and that applying the gradient descent backpropagation of Rohrer in this context would mean that gradient should operate inversely with respect to these negative items.
As per claims 4, 11, and 18, the rejection of claims 1, 8, and 15 is incorporated, and Bi further teaches:
wherein an estimated query entity embedding is generated by applying the parameters of the transformer embedding model to a description of a user of an online system and an estimated target entity embedding is generated by applying the parameters of the same transformer embedding model to a description of an item the user interacted with, Bi pg. 1522 (“We feed the sequence (q,Iu) as the input to a l-layer transformer encoder”).
As per claims 6, 13, and 20, the rejection of claims 1, 8, and 15 is incorporated, and Bi further teaches:
wherein responsive to performing the one or more iterations, further comprises:
applying the parameters of the transformer embedding model to descriptions of a plurality of query entities to generate query entity embeddings, Bi pp. 1523-24, where query embeddings are generated for testing;
applying the parameters of the transformer embedding model to descriptions of a plurality of target entities to generate target entity embeddings, Bi pp. 1523-24, where item embeddings are generated for testing; and
storing the query entity embeddings and the target entity embeddings in a datastore, Bi pp. 1523-24, where the test set is stored.
As per claims 7 and 14, the rejection of claims 1 and 8 is incorporated, and Bi further teaches:
identifying an opportunity to present a plurality target entities for a particular query entity, Bi page 1523, where the results indicate that a search was performed, where the search was performed in response to some triggering event, the claimed opportunity under a broadest reasonable interpretation;
retrieving a query entity embedding for the particular query entity and the target entity embeddings for the plurality of target entities, Bi pp. 1522-1523, where embeddings are retrieved in order to perform a search for top 20 matches;
computing dot products between the query entity embedding and the plurality of target entity embeddings to generate a plurality of scores, Bi pg. 1522, equation 1;
selecting a subset of target entities based on the plurality of scores, Bi pg. 1523, where the top 20 most probable items are selected; and
transmitting instructions to a client device of a user to display the selected subset of target entities on the client device, Bi pg. 1521, where the TEM is used in a e-commerce product search.
Conclusion
Bi teaches training an embedding model to minimize a loss function on a combination of user and purchase embeddings but not on a combination of item and other item embeddings. However, this is still abstract and therefore not patentable.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WILLIAM SPIELER whose telephone number is (571)270-3883. The examiner can normally be reached Monday-Friday, 11-3.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Ann Lo can be reached at 571-272-9767. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
WILLIAM SPIELER
Primary Examiner
Art Unit 2159
/WILLIAM SPIELER/ Primary Examiner, Art Unit 2159
Prosecution Insights