DETAILED ACTION
Response to Amendment
This action is in response to the response to the amendment filed on 05/20/2026. Claims 1 and 3 have been amended, claims 10-20 have been canceled, and claims 21-32 have been newly added. Claims 1, 3-9, and 21-32 are pending and currently under consideration for patentability.
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 .
Inventorship
This application currently names joint inventors. In considering patentability of the claims under pre-AIA 35 U.S.C. 103(a), the examiner presumes that the subject matter of the various claims was commonly owned at the time any inventions covered therein were made absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and invention dates of each claim that was not commonly owned at the time a later invention was made in order for the examiner to consider the applicability of pre-AIA 35 U.S.C. 103(c) and potential pre-AIA 35 U.S.C. 102(e), (f) or (g) prior art under pre-AIA 35 U.S.C. 103(a).
Examiner would like to note that claims 1, 3-9 are not rejected under 35 U.S.C. 101 for claiming signals per se because of ¶¶ [0080] [0081] of Applicant’s specification disclosing “The terms “computer storage media” and “computer storage medium” do not comprise signals per se.”
Claim Rejections - 35 USC § 102(a)(1)
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim(s) 1, 3-5, 7, 21-24, 26, and 29-32 are rejected under 35 U.S.C. 102(a)(1) as being unpatentable by U.S. Publication 2023/0252549 to Xie.
Claims 1-9, 21-28, and 29-32 are computer-readable media, method, and system claims, respectively, with substantially indistinguishable features between each group. For purposes of compact prosecution, the Office has grouped the common method, system and non-transitory computer readable storage medium claims in applying applicable prior art.
With respect to Claim 1:
Xie teaches:
One or more computer storage media storing computer-useable instructions that, when used by one or more computing devices, cause the one or more computing devices to perform operations, the operations comprising (Xie: ¶ [0079]):
generating a training dataset by: providing a plurality of pairs of keyphrases and item listing data for item listings on a listing platform to a search relevance model of a search system for the listing platform (i.e. generating a training dataset by providing search query terms and item embedding data to a machine learning model, wherein the machine learning model determines relevance between search query and item data) (Xie: ¶ [0055] “In various embodiments, the machine learning training module 230 obtains a training dataset for a model from searches for items previously received by the content presentation module 210. For example, the machine learning training module 230 retrieves previously received queries and items displayed by the interface module 210 in response to the queries. From the retrieved queries and items displayed in response to the queries, the machine learning training module 230 obtains a training dataset for a model that determines a measure of relevance between a query and an item. The training dataset includes training examples, with each training example including a query and an item. Additionally, each training example includes a label indicating whether a user preformed a specific interaction with the item after the online concierge system 140 received the query. For example, the specific interaction is including the item in an order, while other specific interactions may be identified in other embodiments.”), and
based on a relevance output from the search relevance model for each pair, assigning to each pair a relevant label or a non-relevant label to provide a plurality of positive training samples and a plurality of negative training samples, wherein each positive training sample comprises a pair from the plurality of pairs with the relevant label and each negative training sample comprises a pair from the plurality of pairs with the non-relevant label (i.e. based on the output of the machine learning model, assigning positive/negative labeling to provide training samples that include positive training data which represents a user performing an interaction with an item indicating the search query term was relevant and negative training data which represents user not performing interaction with item indicating the search query term was not relevant) (Xie: ¶¶ [0055] [0056] “From the retrieved queries and items displayed in response to the queries, the machine learning training module 230 obtains a training dataset for a model that determines a measure of relevance between a query and an item. The training dataset includes training examples, with each training example including a query and an item. Additionally, each training example includes a label indicating whether a user preformed a specific interaction with the item after the online concierge system 140 received the query. For example, the specific interaction is including the item in an order, while other specific interactions may be identified in other embodiments…As the data describing prior queries and items generally includes positive training examples where a label applied to a combination of an item and a query indicates the user performed the specific interaction with the item after the online concierge system 140 received the query, the machine learning training module 230 generates negative training examples to improve the accuracy of the model in various embodiments. A negative training example is a training example where a label applied to a combination of a query and an item indicates the user did not perform the specific interaction with the item after the online concierge system 140 receives the query. As further described below in conjunction with FIGS. 3 and 4, in various embodiments, the machine learning module 230 generates negative training examples from positive training examples in the training dataset. For example, the machine learning training module 230 leverages positive training examples for other queries to generate negative training examples for a specific query. For example, items included in positive training examples for other queries are combined with the specific query to generate negative training examples for the specific query, as further described below in conjunction with FIGS. 3 and 4. Further, in some embodiments, the machine learning training module 230 selects a set of the negative training examples for a query to reduce a number of negative samples used when training the model, as further described below in conjunction with FIGS. 3 and 4.” Furthermore, as cited in ¶ [0068] “The model is configured to receive a query and an item and to generate a measure of relevance of the item to the query. The weights comprise a set of parameters used by the desirability model to transform input data-an item and a query-received by the model into output data-a measure of relevance between the item and the query.”); and
training a keyphrase relevance model using the training data to provide a trained keyphrase relevance model, the training including, for a first training sample from the plurality of training samples (i.e. training machine learning model using training data of query terms related to item interaction) (Xie: ¶ [0055] “In various embodiments, the machine learning training module 230 obtains a training dataset for a model from searches for items previously received by the content presentation module 210. For example, the machine learning training module 230 retrieves previously received queries and items displayed by the interface module 210 in response to the queries. From the retrieved queries and items displayed in response to the queries, the machine learning training module 230 obtains a training dataset for a model that determines a measure of relevance between a query and an item. The training dataset includes training examples, with each training example including a query and an item. Additionally, each training example includes a label indicating whether a user preformed a specific interaction with the item after the online concierge system 140 received the query. For example, the specific interaction is including the item in an order, while other specific interactions may be identified in other embodiments.” Furthermore, as cited in ¶ [0068] “The online system trains a model determining a measure of relevance between an item and a query using the training dataset, which includes positive training examples as well as the negative training examples generated 310 as further described above. In various embodiments, the model includes a query encoder, an item encoder, and a fusion layer. The model comprises a set of weights stored on a non-transitory computer readable storage medium in various embodiments. For training, the online system initializes a network of a plurality of layers comprising the model, with each layer including one or more weights. The model is configured to receive a query and an item and to generate a measure of relevance of the item to the query. The weights comprise a set of parameters used by the desirability model to transform input data-an item and a query-received by the model into output data-a measure of relevance between the item and the query.”):
causing one or more encoders of the keyphrase relevance model to generate one or more embeddings using a first item listing data and a first keyphrase from the first training sample (i.e. encoders of the machine learning model generate embeddings of the item data search query terms) (Xie: ¶¶ [0068] [0069] “In various embodiments, the model includes a query encoder and an item encoder configured to generate a query embedding for a query and an item embedding for an item, respectively. A fusion layer included in the model receives the item embedding and the query embedding and determines the measure of relevance based on the item embedding and the query embedding…FIG. 5 shows an example of the model 500 according to one or more embodiments. In the example of FIG. 5, the model 500 includes a query encoder 505 configured to receive a query and to generate a query embedding 515 that represents the query in a multidimensional space. In various embodiments, the query encoder 505 receives sentences or terms comprising a query as input and generates the query embedding 515 corresponding to the received sentences or terms Similarly, the model 500 includes an item encoder 510 configured to receive an item (or attributes of an item) and to generate an item embedding 520 that represents the item in a multidimensional space. Example attributes received by the item encoder 510 include a name of the item, a brand of the item, a size of the item, a category of the item, a description of the item, or other information describing the item. From the attributes for an item, the item encoder 510 generates the item embedding 520 representing the item in the multidimensional space. While FIG. 5 shows an example where the model 500 includes a query encoder 505 and a separate item encoder 510, in other embodiments, the model 500 learns a set of comprehensive embeddings from a combined set of data including both query data and item data rather than having a query encoder 505 and a discrete item encoder 510.”),
causing a classifier of the keyphrase relevance model to generate a first relevance output using the one or more embeddings (i.e. causing weighting function or layers of model to generate relevance output using the training data embeddings) (Xie: ¶ [0068] “The online system trains a model determining a measure of relevance between an item and a query using the training dataset, which includes positive training examples as well as the negative training examples generated 310 as further described above. In various embodiments, the model includes a query encoder, an item encoder, and a fusion layer. The model comprises a set of weights stored on a non-transitory computer readable storage medium in various embodiments. For training, the online system initializes a network of a plurality of layers comprising the model, with each layer including one or more weights. The model is configured to receive a query and an item and to generate a measure of relevance of the item to the query. The weights comprise a set of parameters used by the desirability model to transform input data-an item and a query-received by the model into output data-a measure of relevance between the item and the query. In various embodiments, the model includes a query encoder and an item encoder configured to generate a query embedding for a query and an item embedding for an item, respectively. A fusion layer included in the model receives the item embedding and the query embedding and determines the measure of relevance based on the item embedding and the query embedding.”),
generating a loss using the first relevance output and a first label from the first training sample (i.e. computing a loss based on the output and training example) (Xie: ¶¶ [0072] [0073] “For each training example of the training dataset to which the model is applied 315, the online system generates 320 an error term based on a predicted measure of relevance between the item and the query included in the training example and a label associated with the training example. The error term is larger when a difference between the predicted measure of relevance and the label applied to the training example is larger and is smaller when the difference between the predicted measure of relevance and the label applied to the training example is smaller. In various embodiments, the online system generates 320 the error term between the predicted measure of relevance and the label applied to the training example using a loss function. Example loss functions include a mean square error function, a mean absolute error, a hinge loss function, and a cross-entropy loss function…In various embodiments, the online system applies a weight to the difference between the predicted measure of relevance from the loss function and the label applied to the training example when generating the error term. The weight is directly related to a measure of similarity between the item of the training example and the query of the training example. For example, the weight is the measure of similarity between the training example and the query of the training examples. In some embodiments, the measure of similarity is the measure of relevance between the item and the query of the training example based on a current set of parameters comprising the model.”), and
updating parameters of the keyphrase relevance model based on the loss (i.e. updating parameters of model based on loss) (Xie: ¶¶ [0074] “The online concierge system 140 backpropagates 3 25 the error term to update the set of parameters comprising the model and stops 330 backpropagation in response to the error term, or the loss function, satisfying one or more criteria. For example, the online concierge system 140 backpropagates 325 the error term through the model to update parameters of the model until the error term has less than a threshold value. For example, the online system 140 may apply gradient descent to update the set of parameters. The online system stores 335 the set of parameters comprising the model on a non-transitory computer readable storage medium after stopping 330 the backpropagation for subsequent use of the model to determine a measure of relevance between an item and a query. For example, the online system applies the trained model to a received query and items retrieved in response to the received query and ranks the items based on their measures of relevance to the query determined by the trained model. In the preceding example, the model allows the online system to rank items so items having higher measures of relevance to the query are initially displayed.”).
With respect to Claims 21 and 29:
All limitations as recited have been analyzed and rejected to claim 1. Claim 21 recites “A computer-implemented method comprising:” the steps of computer-readable medium claim 1. Claim 29 recites “A system comprising: one or more processors; and one or more computer storage media storing computer-usable instructions that, when used by the one or more processors, cause the system to perform operations comprising:” (Xie: ¶ [0079]) the steps of computer-readable medium claim 1. Claims 21 and 29 do not teach or define any new limitations beyond claim 1. Therefore they are rejected under the same rationale.
With respect to Claim 3:
Xie teaches:
The one or more computer storage media of claim 1, wherein the training dataset comprises a second plurality of positive training samples and a second plurality of negative training samples generated by (i.e. training samples include positive and negative training data) (Xie: ¶ [0056] “For example, the machine learning training module 230 leverages positive training examples for other queries to generate negative training examples for a specific query. For example, items included in positive training examples for other queries are combined with the specific query to generate negative training examples for the specific query, as further described below in conjunction with FIGS. 3 and 4.”):
accessing historical search data identifying a plurality of search queries and one or more item listings for each of the plurality of search queries (i.e. accessing previous queries identifying search queries and items corresponding to queries) (Xie: ¶ [0055] “For example, the machine learning training module 230 retrieves previously received queries and items displayed by the interface module 210 in response to the queries. From the retrieved queries and items displayed in response to the queries, the machine learning training module 230 obtains a training dataset for a model that determines a measure of relevance between a query and an item. The training dataset includes training examples, with each training example including a query and an item. Additionally, each training example includes a label indicating whether a user preformed a specific interaction with the item after the online concierge system 140 received the query. For example, the specific interaction is including the item in an order, while other specific interactions may be identified in other embodiments.”);
generating the second plurality of positive training samples using the historical search data, wherein a first positive training sample comprises: a first search query from the plurality of search queries as a first positive keyphrase, an item listing associated with the first search query in the historical search data, and a relevant label; and generating the second plurality of negative training samples using in-batch random negative sampling on the historical search data, wherein a first negative training sample comprises: a second search query from the plurality of search queries as a first negative keyphrase, an item listing from the historical search data not associated with the search query in the historical search data, and a non-relevant label (i.e. generating positive and negative training samples using previously received search queries, wherein the positive training samples are used to determine a relevant label and negative training samples are used to determine if label is not relevant, wherein the negative training samples are determined via in batch random sampling) (Xie: ¶ [0056] “As the data describing prior queries and items generally includes positive training examples where a label applied to a combination of an item and a query indicates the user performed the specific interaction with the item after the online concierge system 140 received the query, the machine learning training module 230 generates negative training examples to improve the accuracy of the model in various embodiments. A negative training example is a training example where a label applied to a combination of a query and an item indicates the user did not perform the specific interaction with the item after the online concierge system 140 receives the query. As further described below in conjunction with FIGS. 3 and 4, in various embodiments, the machine learning module 230 generates negative training examples from positive training examples in the training dataset. For example, the machine learning training module 230 leverages positive training examples for other queries to generate negative training examples for a specific query. For example, items included in positive training examples for other queries are combined with the specific query to generate negative training examples for the specific query, as further described below in conjunction with FIGS. 3 and 4. Further, in some embodiments, the machine learning training module 230 selects a set of the negative training examples for a query to reduce a number of negative samples used when training the model, as further described below in conjunction with FIGS. 3 and 4.” Furthermore, as cited in ¶ [0065] “Referring back to FIG. 3, while generating 310 negative training examples from positive training examples augments the training dataset, such generation of negative training examples causes the training dataset to include more negative training examples than positive training examples. Such a disparity in a number of positive training examples and a number of negative training examples affects accuracy of a model trained using the training dataset. To reduce the number of negative training examples included in the training dataset, the online system selects a subset of the negative training examples generated using the in-batch negative method further described above and discards negative training examples that are not in the subset. In various embodiments, the online system selects the subset of the negative training examples using uniform sampling or self-adversarial reweighting and sampling. For uniform sampling, the online system randomly selects a number of negative training examples including a query from the negative training examples. Each negative training example for a query has the same probability of being selected when universal sampling is performed. In the example of FIG. 4, using uniform sampling to select one negative training example for query 405A causes each of negative training example 420, negative training example 425, and negative training example 430 to have an equal probability of being selected. Uniform sampling allows the online system to augment a positive training example for a query with a subset of negative training examples for the query generated through the in-batch negative method further described above. Such sampling allows the training dataset to include both positive training examples and negative training examples, while maintaining balance between a number of negative training examples and a number of positive training examples in the training dataset.”).
With respect to Claims 22 and 30:
All limitations as recited have been analyzed and rejected to claim 3. Claims 22 and 30 do not teach or define any new limitations beyond claim 3. Therefore they are rejected under the same rationale.
With respect to Claim 4:
Xie teaches:
The one or more computer storage media of claim 1, wherein the one or more encoders comprise an item listing encoder and a keyphrase encoder; and wherein the one or more embeddings comprises an item listing embedding generated by the item listing encoder using the first item listing data and a keyphrase embedding generated by the keyphrase encoder using the first keyphrase (i.e. item encoder generates item embedding and query encoder generates query or keyphrase embedding) (Xie: ¶ [0068] “The online system trains a model determining a measure of relevance between an item and a query using the training dataset, which includes positive training examples as well as the negative training examples generated 310 as further described above. In various embodiments, the model includes a query encoder, an item encoder, and a fusion layer. The model comprises a set of weights stored on a non-transitory computer readable storage medium in various embodiments. For training, the online system initializes a network of a plurality of layers comprising the model, with each layer including one or more weights. The model is configured to receive a query and an item and to generate a measure of relevance of the item to the query. The weights comprise a set of parameters used by the desirability model to transform input data-an item and a query-received by the model into output data-a measure of relevance between the item and the query. In various embodiments, the model includes a query encoder and an item encoder configured to generate a query embedding for a query and an item embedding for an item, respectively. A fusion layer included in the model receives the item embedding and the query embedding and determines the measure of relevance based on the item embedding and the query embedding.”).
With respect to Claims 23 and 31:
All limitations as recited have been analyzed and rejected to claim 4. Claims 23 and 31 do not teach or define any new limitations beyond claim 4. Therefore they are rejected under the same rationale.
With respect to Claim 5:
Xie teaches:
The one or more computer storage media of claim 1, wherein the one or more encoders comprise a cross-encoder; and wherein the one or more embeddings comprise a combined embedding generated by the cross-encoder using the first item listing data and the first keyphrase (i.e. encoders comprise of a cross-encoder or fusion later, wherein the query embedding and item embedding are combined by the fusion layer or cross-encoder) (Xie: ¶ [0070] “The fusion layer 525 is coupled to the query encoder 505 and to the item encoder 510. The fusion layer 525 receives the query embedding 515 from the query encoder 505 and the item embedding 520 from the item encoder 510 as inputs. The fusion layer 525 determines a measure of relevance 530 between the query embedding 515 and the item embedding 520. In various embodiments the item embedding and the query embedding each have a common number of dimensions in the multidimensional space, and the fusion layer 525 determines a cosine similarity between the query embedding 515 and the item embedding 520 as the measure of relevance 535. As another example, the fusion layer 525 determines the measure of relevance 535 as a dot product between the query embedding 515 and the item embedding 520. While in other embodiments, the fusion layer 525 determines the measure of relevance 535 between the query and the item as another measure of similarity between the corresponding query embedding and item embedding.”).
With respect to Claims 24 and 32:
All limitations as recited have been analyzed and rejected to claim 5. Claims 24 and 32 do not teach or define any new limitations beyond claim 5. Therefore they are rejected under the same rationale.
With respect to Claim 7:
Xie teaches:
The one or more computer storage media of claim 1, wherein the first relevance output from the classifier comprises a relevance score (i.e. score is based on how relevant/related the query term is to the item) (Xie: ¶ [0041] “In some embodiments, the content presentation module 210 scores items based on a search query received from the customer client device 100. A search query is text for a word or set of words that indicate items of interest to the customer. The content presentation module 210 scores items based on a relatedness of the items to the search query. For example, the content presentation module 210 may apply natural language processing (NLP) techniques to the text in the search query to generate a search query representation ( e.g., an embedding) that represents characteristics of the search query. The content presentation module 210 may use the search query representation to score candidate items for presentation to a customer ( e.g., by comparing a search query embedding to an item embedding).”).
With respect to Claim 26:
All limitations as recited have been analyzed and rejected to claim 7. Claim 26 does not teach or define any new limitations beyond claim 7. Therefore it is rejected under the same rationale.
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) 6, 8, 9, 25, 27, and 28 are rejected under 35 U.S.C. 103 as being unpatentable over Xie in view of U.S. Patent 11,004,135 to Sandler.
With respect to Claim 6:
Xie does not explicitly disclose the one or more computer storage media of claim 1, wherein the first relevance output from the classifier comprises a binary indication of relevance.
However, Sandler further discloses wherein the first relevance output from the classifier comprises a binary indication of relevance (i.e. outputs binary indication of relevance corresponding to positive/negative correlations with respect to the relevance of items) (Sandler: Col. 15 Lines 16-29 “It is noted that DPPs have been proposed as a way to select a diverse subset of items given features vectors, or equivalently, a similarity kernel for items. DPPs are probabilistic models of global, negative correlations and are a distribution over subsets of a fixed ground set. For example, a DPP over a ground set of N items can be seen as modeling a binary characteristic vector of length N. A characteristic of a DPP is that these binary variables are negatively correlated in that the inclusion of one item makes the inclusion of other items less likely. The strengths of these correlations can be derived from a kernel matrix that defines a global measure of similarity between pairs of items so that more similar items are less likely to co-occur. As a result, DPPs assign higher probability to sets of items that are diverse.”).
Therefore, it would have been obvious to one of ordinary skill in the art, at the time the invention was made, to add Sandler’s first relevance output from the classifier comprises a binary indication of relevance to Xie’s accessing a training dataset, the training dataset comprising a plurality of training samples where each training sample includes a keyphrase, item listing data for an item listing on a listing platform, and a label indicating relevance of the keyphrase to the item listing based on a search relevance model of a search system for the listing platform. One of ordinary skill in the art would have been motivated to do so because the “relevance model includes a feedforward neural network architecture with an input feature set that includes a variety of features in order to make the network robust to the "cold start" problem ( e.g., where a user has little or no associated behavioral data/purchase data)” and “the neural network architecture alleviates data sparsity and helps the recommendation engine provide recommendations to first time users of the pantry catalog. Thus, the relevance model can be considered to be "regularized." Regularization, in mathematics and statistics and particularly in the fields of machine learning and inverse problems, is a process of introducing additional information in order to solve an ill-posed problem or to prevent overfitting.” (Sandler: Cols. 2-3 Lines 51-7).
With respect to Claim 25:
All limitations as recited have been analyzed and rejected to claim 6. Claim 25 does not teach or define any new limitations beyond claim 6. Therefore it is rejected under the same rationale.
With respect to Claim 8:
Xie does not explicitly disclose the one or more computer storage media of claim 1, wherein the operations further comprise: employing the trained keyphrase relevance model when providing one or more keyphrase recommendations for a target item listing.
However, Sandler further discloses employing the trained keyphrase relevance model when providing one or more keyphrase recommendations for a target item listing (i.e. employing relevance model to keywords and items in order to recommend items) (Sandler: Col. 5 Lines 44-64 “It will be appreciated that, while many of the examples herein focus on the use of purchase history data for identifying relevant items, implementations of the disclosed machine learning system can use one ( or a combination of) the following types of historical user behavioral data: purchase history, click history (e.g., items viewed), keywords entered in search queries, selection of browse nodes ( e.g., catalog categories such as "clothing" or "garden"), text of item reviews submitted by the user, and other data representing interactions between users and items and/or the electronic catalog. Further, the disclosed machine learning system can be used to generate recommendation sets of other types of items (e.g., non-consumable items, promotional items, digital media items). In the digital media context, the disclosed model can be used to generate playlists of music or videos, where the playlists are both relevant (e.g., in-line with the user's interests) and diverse ( e.g., songs of different genres, videos relating to different topics). As well, the disclosed machine learning system can be used to generate recommendations outside of the context of an electronic catalog.” Furthermore, as cited in Col. 6 Lines 19-34 “FIG. 2A illustrates a schematic block diagram of a recommendation engine 200 including the disclosed architecture for a relevance-diversity balancing machine learning system. The machine learning system can be implemented as one or more electronic digital memories and one or more electronic digital processors configured to generate recommendations as described herein. As illustrated, the recommendation engine 200 includes a user profile features data repository 205, item data repository 260, diversified relevant recommendations data repository 230, vector embedding generator 220, relevance model 210, diversity model 225, and diversified relevant recommendations data repository 230. One or both of the relevance model 210 and diversity model 225 can be a machine learning model, and the recommendation engine 200 can be considered as implementing a model ensemble to provide recommendations.”).
Therefore, it would have been obvious to one of ordinary skill in the art, at the time the invention was made, to add Sandler’s employing the trained keyphrase relevance model when providing one or more keyphrase recommendations for a target item listing to Xie’s accessing a training dataset, the training dataset comprising a plurality of training samples where each training sample includes a keyphrase, item listing data for an item listing on a listing platform, and a label indicating relevance of the keyphrase to the item listing based on a search relevance model of a search system for the listing platform. One of ordinary skill in the art would have been motivated to do so because the “relevance model includes a feedforward neural network architecture with an input feature set that includes a variety of features in order to make the network robust to the "cold start" problem ( e.g., where a user has little or no associated behavioral data/purchase data)” and “the neural network architecture alleviates data sparsity and helps the recommendation engine provide recommendations to first time users of the pantry catalog. Thus, the relevance model can be considered to be "regularized." Regularization, in mathematics and statistics and particularly in the fields of machine learning and inverse problems, is a process of introducing additional information in order to solve an ill-posed problem or to prevent overfitting.” (Sandler: Cols. 2-3 Lines 51-7).
With respect to Claim 27:
All limitations as recited have been analyzed and rejected to claim 8. Claim 27 does not teach or define any new limitations beyond claim 8. Therefore it is rejected under the same rationale.
With respect to Claim 9:
Xie does not explicitly disclose the one or more computer storage media of claim 8, wherein employing the trained keyphrase relevance model when providing the one or more keyphrase recommendations for the target item listing comprises: accessing target item listing data for the target item listing; retrieving a plurality of candidate keyphrases using the target item listing data; causing the trained keyphrase relevance model to generate a corresponding relevance output for each candidate keyphrase by providing the target item listing data and each candidate keyphrase as paired inputs to the trained keyphrase relevance model; filtering one or more candidate keyphrases from the plurality of candidate keyphrases based on the corresponding relevance output for each candidate keyphrase to provide a filtered set of keyphrases; ranking the keyphrases from the filtered set of keyphrases; and providing the one or more keyphrase recommendations for the target item listing based on the ranking.
However, Sandler further discloses:
accessing target item listing data for the target item listing (i.e. access item listing data for the target items) (Sandler: Cols. 7-8 Lines 63-13 “During training, the relevance model can learn the internal parameters that produce the best match (objectively or within allowable margins) between the input training data and the known expected output. Once trained, the relevance model can be provided with new input data and used to generate relevance scores for each of a number of items. In the context of the pantry catalog, the output of the relevance model can include a score for each item that qualifies for sale in the pantry catalog, and higher scores indicate higher likelihood of purchase by the user. In other context, higher relevance scores indicate higher likelihood of purchase, rental, streaming, clicking ( e.g., user selecting an advertisement), and the like. The relevance model 210 is depicted with a high-level schematic representation of a neural network, however other machine learning models suitable for generating relevance scores can be used in other embodiments, for example random forests, decision trees, Bayesian probability estimators, and the like.”);
retrieving a plurality of candidate keyphrases using the target item listing data (i.e. retrieving a candidate pool of recommended terms corresponding to items) (Sandler: Col. 13 Lines 29-41 “FIG. 2C illustrates an efficient greedy determinant point process (GDPP) engine 226 usable as the diversity model 225 of FIG. 2A. The efficient GDDP engine 226 includes a candidate recommendations pool data repository 240 that stores data representing the total candidate pool of items and items remaining in the candidate pool for consideration for addition to the recommendations set. This data is passed to a set-selector engine 256 together with vector representations from the vector representations of ranked relevant recommendations data repository 235. The efficient GDDP engine 226 also includes an updated recommendation sets data repository 265 that stores the items currently in the recommendations set, and a size selector engine 245.” Furthermore, as cited in Col. 14 Lines 27-35 “One embodiment of the diversity model 225 can optimize a set-selection objective including terms relating to both item relevance scores (e.g., probabilities provided by the relevance model 210) and item diversity scores (e.g., determined based on volume occupied in feature space). Returning to the example context of the electronic catalog for pantry items, let M be the number of items in the pantry catalog.”);
causing the trained keyphrase relevance model to generate a corresponding relevance output for each candidate keyphrase by providing the target item listing data and each candidate keyphrase as paired inputs to the trained keyphrase relevance model (i.e. relevance model is trained based on pairs of terms relating to items in order to output a relevance score for each term) (Sandler: Col. 14 Lines 27-67 “One embodiment of the diversity model 225 can optimize a set-selection objective including terms relating to both item relevance scores (e.g., probabilities provided by the relevance model 210) and item diversity scores (e.g., determined based on volume occupied in feature space). Returning to the example context of the electronic catalog for pantry items, let M be the number of items in the pantry catalog. Let N be the number of users of the pantry catalog. Let r, be the relevance score of item i, for example predicted by the neural network 211. Let J denote a subset of indices corresponding to a subset of items. Let the identity matrix be denoted by I. Let L denote a matrix of latent vector representations for items ( e.g., FIG. 1A), where the i'th column of L, L,, represents the vector representation of item i. Also, if T is a subset of indices (indexing rows or columns ), let Lr denote the sub-matrix of L whose columns are restricted to the indices in T (i.e. Lr only has ITI columns). Similarly, let Lrr denote the sub-matrix of L whose rows and columns are restricted to the indices in T. Finally, let S denote the similarity matrix of the items, obtained as S=LrL. Thus, Su=( L,, L) , the inner product between L, and LJ"… The objective of equation (3) can be thought of as a relevance-diversity, subset selection problem given an initial set of recommendations for each user. The first term in the objective (r,) quantifies the relevance score of the recommendations and promotes the selection of highly relevant set of items. The values for the first term can be the relevance scores output from the neural network 211 in some implementations. The second term in the objective (log det(S}+yI)) promotes diverse sets to be chosen, with7 being a pre-specified constant that keeps the objective well defined.”);
filtering one or more candidate keyphrases from the plurality of candidate keyphrases based on the corresponding relevance output for each candidate keyphrase to provide a filtered set of keyphrases (i.e. trade-off or filter relevant recommendations based on diversity output for each term) (Sandler: Col. 15 Lines 3-15 “The parameter A is a trade-off parameter that trades-off relevant recommendations with diverse recommendations. The higher A is, the more the objective prioritizes diversifying the recommendations and the objective prioritizes relevance. The opposite is true for smaller values of A. The particular value of A chosen for a given implementation of objective of equation (3) can vary depending upon the desired balance between relevance and diversity, however it can be preferable to make the diversity scores on the same or a similar scale as the relevance terms, and thus the value of A can depend upon the values of the relevance scores. In one example, in various embodiments 0.1 and 0.01 are suitable values for A.” Furthermore, as cited in Col. 17 Lines 11-22 “The efficient GDPP engine 266 draws from a candidate pool of identified relevant recommendations, from which the relevant-diverse subset is created. This candidate pool is stored in the candidate recommendations pool data repository 240. Initially, the candidate pool can include a top number of recommendations ranked by relevance scores determined by a relevance model, for example a top two hundred most relevant items. This pool can be decremented at each stage of the set-building process to remove the item just added to the set, such that the candidate recommendations pool data repository 240 keeps a running index of the items currently in the candidate pool.”);
ranking the keyphrases from the filtered set of keyphrases (i.e. ranking terms from the selected candidate set) (Sandler: Col. 13 Lines 29-57 “FIG. 2C illustrates an efficient greedy determinant point process (GDPP) engine 226 usable as the diversity model 225 of FIG. 2A. The efficient GDDP engine 226 includes a candidate recommendations pool data repository 240 that stores data representing the total candidate pool of items and items remaining in the candidate pool for consideration for addition to the recommendations set. This data is passed to a set-selector engine 256 together with vector representations from the vector representations of ranked relevant recommendations data repository 235. The efficient GDDP engine 226 also includes an updated recommendation sets data repository 265 that stores the items currently in the recommendations set, and a size selector engine 245… Generally, a diversity model 225 according to the present disclosure re-ranks recommendations from the neural network 211 ( or another relevance model 210) to generate a set of both relevant and diverse recommendations. In embodiments of the present disclosure, the diversity model 225 operates on vector embeddings of the recommendations. The vector embedding of recommendation k can be represented as xk, and can be a one-dimensional or high-dimensional vector. Thus, a matrix representing the pool of recommendations, X, is formed from the set of recommendation feature vectors such that X=(x1; ... ; xk) for the set of recommendations 1 through k.”); and
providing the one or more keyphrase recommendations for the target item listing based on the ranking (i.e. providing relevant terms associated with highest score/ranking) (Sandler: Col. 17 Lines 34-64 “Maximizing this initial stage involves finding the item with the highest relevance score. At the next stage, the selection function can iteratively consider pairs-including the first item added to the set and then each of the recommendations remaining in the candidate pool-in light of the objective. The item that maximizes the objective with the first item is then added to the set. At each stage, the relevance term is computed by summing the relevance scores of all the items in the set. The diversity term is computed seeking the maximum volume formed in latent product space by the feature vectors representing the items. The relevance and diversity terms are added together to generate the score for that item set, and the. As illustrated, the set-selection engine 256 seeks to greedily and approximately solve for the set with the maximum objective score with the series of algorithmic steps 252 in the depicted block of code 250. It will be appreciated that variations on the specific depicted code that achieve the same function as the described set-selection engine 256 are also within the scope of the present disclosure. As described above, the reformulated DPP objective includes a (1) first term (the relevance term) that sums the relevance scores of items in a set and a (2) second term (the diversity term) that evaluates diversity of the set as a function of a trade-off parameter set to achieve a desired balance between relevance and diversity, a diversity score computed as the log determinant of the similarity matrix of the items (the square of the volume in latent vector space of the items in the set ( e.g., the disclosed metric used to evaluate the diversity of these items), and a smoothing parameter that introduces Gaussian entropy.”).
Therefore, it would have been obvious to one of ordinary skill in the art, at the time the invention was made, to add Sandler’s accessing target item listing data for the target item listing; retrieving a plurality of candidate keyphrases using the target item listing data; causing the trained keyphrase relevance model to generate a corresponding relevance output for each candidate keyphrase by providing the target item listing data and each candidate keyphrase as paired inputs to the trained keyphrase relevance model; filtering one or more candidate keyphrases from the plurality of candidate keyphrases based on the corresponding relevance output for each candidate keyphrase to provide a filtered set of keyphrases; ranking the keyphrases from the filtered set of keyphrases; and providing the one or more keyphrase recommendations for the target item listing based on the ranking to Xie’s accessing a training dataset, the training dataset comprising a plurality of training samples where each training sample includes a keyphrase, item listing data for an item listing on a listing platform, and a label indicating relevance of the keyphrase to the item listing based on a search relevance model of a search system for the listing platform. One of ordinary skill in the art would have been motivated to do so because the “relevance model includes a feedforward neural network architecture with an input feature set that includes a variety of features in order to make the network robust to the "cold start" problem ( e.g., where a user has little or no associated behavioral data/purchase data)” and “the neural network architecture alleviates data sparsity and helps the recommendation engine provide recommendations to first time users of the pantry catalog. Thus, the relevance model can be considered to be "regularized." Regularization, in mathematics and statistics and particularly in the fields of machine learning and inverse problems, is a process of introducing additional information in order to solve an ill-posed problem or to prevent overfitting.” (Sandler: Cols. 2-3 Lines 51-7).
With respect to Claim 28:
All limitations as recited have been analyzed and rejected to claim 9. Claim 28 does not teach or define any new limitations beyond claim 9. Therefore it is rejected under the same rationale.
Response to Arguments
Applicant’s arguments see page 1 of the Remarks disclosed, filed on 05/20/2026, with respect to the 35 U.S.C. § 101 rejection(s) of claim(s) 10-20 have been considered and are persuasive. The Applicant has states “The Office Action states: "Examiner would also like to note that claims 1-9 are not rejected under 35 U.S.C. 101 for being directed to a judicial exception (i.e. abstract idea) because the claims are directed the training of the machine learning model." Claims 10-20 have been canceled herein, and new claims 21-32 have been added herein, including independent claims 21 and 29. Each of claims 21 and 29 recite elements with the same scope as claim 1. As such, Applicant submits that claims 21-32 are directed to patent eligible subject matter for at the least the reason provided for claim 1.” The Examiner agrees and notes that independent claims 1, 21, and 29 are directed to the training of the machine learning model. Therefore, the rejection(s) of claim(s) 1, 3-9, and 21-32 under 35 U.S.C. § 101 has been withdrawn.
Examiner would like to note that claims 1, 3-9 are not rejected under 35 U.S.C. 101 for claiming signals per se because of ¶¶ [0080] [0081] of Applicant’s specification disclosing “The terms “computer storage media” and “computer storage medium” do not comprise signals per se.”
Applicant’s arguments see pages 1-6 of the Remarks disclosed, filed on 05/20/2026, with respect to the 35 U.S.C. § 102(a)(1) rejection(s) of claim(s) 1-5, 7, 10-13, 15, 19, and 20 as being unpatentable by Xie have been considered but are not persuasive. The Applicant asserts “In contrast, Xie's training data is generated from previously received queries and interactions with results for the queries, where positive examples are query-item pairs with which a user performed a specified interaction (e.g., adding to an order), and negative examples correspond to query-item pairs for which the specific interaction was not performed, with negative examples commonly produced by combining a query from one positive example with an item from a different query's positive example (in-batch negatives), optionally sampling/weighting the negatives using similarity (self-adversarial reweighting). Stated differently, in amended claim 1 the search relevance model itself is the supervision source that decides relevance/non-relevance for each candidate keyphrase-listing pair, whereas in Xie the supervision is derived from user actions (or the absence of user actions) and then negatives are engineered from positives and reweighted, with the labeling premise being user interaction-based. This difference in how training samples are generated matters because it changes what "non-relevant" means and how negatives are obtained. Under amended claim 1, a negative training sample exists because the search relevance model actually output a non-relevance indication for that specific keyphrase-listing pair. As such, "non-relevant" reflects a relevance determination mechanism (i.e., the search relevance model) designed to filter search results, not a lack of downstream engagement as in Xie. In contrast to claim 1, under Xie, a negative training example is fundamentally tied to the proposition that a specific user interaction was not performed with an item after a query, and Xie's in-batch approach further assumes that items that were positives for other queries can serve as negatives for the current query (even if the item might in fact be relevant but simply not selected/seen/available/competitive in ranking). For example, if the query is "wireless mouse" and an item listing for a mid-priced mouse never gets clicked because it's ranked below highly popular brands, Xie's process would treat that query-item pair (or engineered query-item pair) as negative because no qualifying interaction occurred even though the item listing is relevant to the query. By contrast, amended claim 1 would label the pair based on whether the search relevance model deems the mouse listing relevant to "wireless mouse" given its attributes (title/category/description/etc), regardless of whether users happened to click or purchase it in the logged data. Amended claim 1's approach is also an improvement over Xie because it directly addresses the bias problem that arises when no user interactions (e.g., user clicks, sales) are treated as "non-relevant." The application explains that historical search data suffers from missing-not-at- random conditions: "An item listing lacking clicks or sales for a particular search query does not necessarily mean that the search query is non-relevant to the item listing," because user interactions provide only a skewed set of impressions influenced by ranking and visibility, so unpopular or low-ranked items may receive no clicks/sales even if they are relevant. Xie's data generation is rooted in that interaction/no-interaction paradigm where positives come from interactions, and negatives correspond to non-interactions (including negatives synthesized from other queries' positives). This non-interaction approach in Xie is an unreliable proxy for non-relevance. By generating negatives from the search relevance model's explicit non-relevance outputs, amended claim 1 produces negative samples that are better aligned with actual relevance judgments (as determined by the search system) rather than artifacts of exposure, position bias, popularity bias, or auction/promotion effects that suppress clicks. As an example to illustrate the improvement, consider "tail" queries or newly listed items. Suppose a new item listing "Organic Moroccan Argan Oil 50ml" appears and the keyphrase "argan oil hair" has few historical interactions with that listing because the listing is new, has little traffic, or is ranked low. Under Xie, the absence of a "specific interaction" would tend to push the system toward treating many query-item combinations involving that new listing as negative, and in-batch negatives could further reinforce that by pairing the query with items from other queries' positives. Under amended claim 1, the system can still generate training samples for that new listing by pairing candidate keyphrases with the listing data and labeling each pair using the search relevance model's outputs SO the query "argan oil hair" paired with the argan oil listing can be labeled relevant if the search relevance model recognizes, for instance, semantic/attribute alignment, even if user interaction data is sparse or absent. This yields a training dataset with higher-fidelity negative labels (true "non-relevance" as the search system understands it) and avoids teaching the downstream keyphrase relevance model that "low engagement" implies "non- relevance," which is precisely the bias the application identifies and claim 1 avoids. During the interview, the examiner identified paragraphs 58 and 61 as potentially relevant to amended claim 1. Paragraphs 58 and 61 of Xie describe generating training samples from historical query-item interaction data, where positive samples are query-item pairs associated with a user interaction (e.g., purchase or selection), and negative samples are created by (i) treating the absence of such interaction as a negative label and/or (ii) constructing additional negative samples by pairing a given query with items drawn from positive examples of other queries (i.e., in-batch negative sampling). As described above, this approach differs from amended claim 1, which does not infer relevance or non-relevance from user behavior or lack thereof, but instead generates training samples by explicitly providing keyphrase-item listing pairs to a search relevance model and assigning positive or negative labels based directly on that model's relevance output for each pair, such that both positives and negatives reflect an affirmative relevance determination rather than an absence of interaction.” The Examiner respectfully disagrees. Examiner would like to refer the Applicant to ¶¶ [0055] [0056] “In various embodiments, the machine learning training module 230 obtains a training dataset for a model from searches for items previously received by the content presentation module 210. For example, the machine learning training module 230 retrieves previously received queries and items displayed by the interface module 210 in response to the queries. From the retrieved queries and items displayed in response to the queries, the machine learning training module 230 obtains a training dataset for a model that determines a measure of relevance between a query and an item. The training dataset includes training examples, with each training example including a query and an item. Additionally, each training example includes a label indicating whether a user preformed a specific interaction with the item after the online concierge system 140 received the query. For example, the specific interaction is including the item in an order, while other specific interactions may be identified in other embodiments…As the data describing prior queries and items generally includes positive training examples where a label applied to a combination of an item and a query indicates the user performed the specific interaction with the item after the online concierge system 140 received the query, the machine learning training module 230 generates negative training examples to improve the accuracy of the model in various embodiments. A negative training example is a training example where a label applied to a combination of a query and an item indicates the user did not perform the specific interaction with the item after the online concierge system 140 receives the query. As further described below in conjunction with FIGS. 3 and 4, in various embodiments, the machine learning module 230 generates negative training examples from positive training examples in the training dataset. For example, the machine learning training module 230 leverages positive training examples for other queries to generate negative training examples for a specific query. For example, items included in positive training examples for other queries are combined with the specific query to generate negative training examples for the specific query, as further described below in conjunction with FIGS. 3 and 4. Further, in some embodiments, the machine learning training module 230 selects a set of the negative training examples for a query to reduce a number of negative samples used when training the model, as further described below in conjunction with FIGS. 3 and 4.” Furthermore, as cited in ¶ [0068] “The model is configured to receive a query and an item and to generate a measure of relevance of the item to the query. The weights comprise a set of parameters used by the desirability model to transform input data-an item and a query-received by the model into output data-a measure of relevance between the item and the query.” It is clear from the disclosure above that the Xie reference teaches generates training samples by explicitly providing keyphrase-item listing pairs to a search relevance model or search query terms and item data to a machine learning model in order to assign positive or negative labels based directly on that model's relevance output for each pair, such that both positives and negatives reflect an affirmative relevance determination. Therefore, the rejection(s) of claim(s) 1, 3-5, 7, 21-24, 26, and 29-32 under 35 U.S.C. § 102(a)(1) with claims 6, 7, 9, 25, 27, and 28 being rejected under 35 U.S.C. § 103 is provided above with updated citations.
Examiner recommends amending independent claims 1, 21, and 29 to clarify how the classifier generates a binary indication of relevance in order to overcome the Xie reference.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. The following references are cited to further show the state of the art:
U.S. Publication 2022/0004712 to Bahuleyan for disclosing Computer implemented methods and systems are provided for generating diverse key phrases while maintaining competitive output quality. A system for training a sequence to sequence (S2S) machine learning model is proposed where neural unlikelihood objective approaches are used at (1) a target token level to discourage the generation of repeating tokens, and (2) a copy token level to avoid copying repetitive tokens from the source text. K-step ahead token prediction approaches are also proposed as an additional mechanism to augment the approach to further enhance the overall diversity of key phrase outputs.
U.S. Publication 2024/0412060 to Luo for disclosing A computer-implemented method for training a neural network based ranking model includes performing a training data augmentation operation on a set of training data to generate a set of synthesized training data, and training a neural network based ranking model using the set of training data and the set of synthesized training data. The set of training data includes, for each of a plurality of queries, respective query-document data and respective relevance judgement data. The query-document data for a query includes data associated with a plurality of query-documents pairs for the query. The relevance judgement data for a query includes one or more sets of user feedback data associated with the query. The set of training data has an imbalanced training data distribution and the set of synthesized training data is arranged for use to reduce training data distribution imbalance of the set of training data.
U.S. Publication 2023/0099888 to Jaini for disclosing A machine learning model may be trained using annotated communications data. Each communication (e.g., a short messaging system (SMS) message or email) is annotated with a measure of user interaction. The machine learning model is thus trained to predict a measure of user interaction for future communications. Before sending future communications, at least a portion of the communication is provided to the trained machine learning model to predict the expected measure of user interaction with the communication. In response to the prediction, the sender of the communication may alter the communication. The system may automatically send the communication if the predicted measure of user interaction exceeds a predetermined threshold and only prompt the user if the predicted measure of user interaction does not exceed the predetermined threshold.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. 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 extension fee 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 Azam Ansari, whose telephone number is (571) 272-7047. The examiner can normally be reached from Monday to Friday between 8 AM and 4:30 PM.
If any attempt to reach the examiner by telephone is unsuccessful, the examiner's supervisor, Waseem Ashraf, can be reached at (571) 270-3948.
Another resource that is available to applicants is the Patent Application Information Retrieval (PAIR). Information regarding the status of an application can be obtained from the (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAX. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pairdirect.uspto.gov. Should you have questions on access to the Private PAIR system, please feel free to contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free).
Applicants are invited to contact the Office to schedule either an in-person or a telephonic interview to discuss and resolve the issues set forth in this Office Action. Although an interview is not required, the Office believes that an interview can be of use to resolve any issues related to a patent application in an efficient and prompt manner.
/AZAM A ANSARI/
Primary Examiner, Art Unit 3621
August 10, 2026
Prosecution Insights