DETAILED ACTION
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on September 23, 2026. Claims 1, 5, and 9 is/are amended and new claims 10-12 is/are added. Claims 1-12 have been examined in this application.
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 .
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status:
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 of this title, 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.
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The factual inquiries set forth in Graham v. John Deere Co., 383 U.S. 1, 148 USPQ 459 (1966), that are applied for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
Determining the scope and contents of the prior art.
Ascertaining the differences between the prior art and the claims at issue.
Resolving the level of ordinary skill in the pertinent art.
Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1-12 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Pat. 8671120 (“Bicknell”) in view U.S. Pub. 20210304121 (“Lee”) in view U.S. Pub. 20210042330 (“Bremmer”) in further view of NPL “LaTr - Layout-Aware Transformer for Scene-Text Visual Question Answer” (“Biten”).
As per claims 1, 5, and 9, Bicknell discloses, method for identifying a match with a product, comprising (Bicknell discloses a method and system for automatically matching an incoming item description to an authoritative item definition corresponding to a product in a product catalog. Bicknell explains that the matching engine receives an item description and compares that item description to various item definitions, wherein the item definitions correspond to products in the product catalog, and identifies the item definition most similar to the item description as the matching item definition) (col. 2, ll. 54-67):
extracting, by a processor, one or more attributes relating to a product from data indicated in a request for identifying the product, each of the one or more attributes describing the product (Examiner interprets that Bicknell discloses receiving item-description information identifying a product and using attributes of that product to locate a corresponding product definition. Bicknell states that uploaded inventory information may include product attributes such as title, publisher, manufacturer, part number, and price that assist in uniquely identifying the corresponding product in the product catalog. Bicknell further teaches that an item definition may include attribute values, such as title, that uniquely identify an item and that the matching engine compares the received item description with such definitions. Bicknell further discloses that the matching-engine components are implemented as a computer program executing on a computer system and that the computer system includes a central processing unit, thereby teaching processing of the received item description by a processor) (col. 1, ll. 55-63; col. 2, ll. 54-67; col. 5, ll. 31-47);
identifying a match with the product from the plurality of products, the identification based on comparing, by the processor, the one or more attributes with each corresponding entry of the plurality of products (Examiner interprets that Bicknell teaches that the matching engine compares an incoming item description against item definitions corresponding to products and calculates similarity based upon corresponding attributes. Its rules specify how similarity between an item description and an item definition is determined and how a similarity score is calculated based on comparison of their attributes. Bicknell also explains that the product catalog may contain hundreds of thousands of item definitions and that candidate item definitions are identified from this collection for detailed matching analysis. Bicknell expressly teaches rule-defined attribute comparisons, weighted attribute similarity scores, and corresponding threshold similarity scores) (col. 3, ll. 3-38; col. 6, ll. 38-54);
selecting one or more products from the plurality of products in the database based on a rule-based comparison of at least one of the one or more attributes with each corresponding entry in the database (Examiner interprets that Bicknell teaches a rules table containing matching rules for different categories of products. Each rule contains one or more filters specifying criteria for an item description to match an item definition, wherein each filter specifies one or more attributes and a scoring algorithm for quantitatively representing similarity between the item description and the item definition. Bicknell thereafter uses a designated filter of the applicable rule and indexes corresponding to attribute values to select candidate item definitions from the larger item-definition database for detailed analysis. Bicknell's disclosure therefore provides both the rule/attribute comparison and the resulting selection of a candidate subset) (col. 5, l. 57-col. 6, l. 18; col. 6, ll. 38-55);
identifying the match with the product from the selected one or more products, the identification based on comparing, by the processor, the one or more attributes with each corresponding entry of the selected one or more products (Examiner interprets that Bicknell teaches that after candidate item definitions have been identified, the matching engine passes the candidate item definitions to the rule-matching component, which determines which candidates match or potentially match the received item description. Bicknell further teaches calculating a similarity score for each candidate item definition using scoring methods designated by the applicable filters to compare corresponding attributes and generate an indication of similarity) (col. 6, ll. 38-67; col. 7, ll. 1-10);
calculating a score for each of the selected one or more products based on the comparison (Bicknell expressly teaches calculating similarity scores based upon comparisons of corresponding product attributes. The score may be a weighted combination of similarity scores for multiple attributes, with individual weights reflecting the importance of the respective attributes. Bicknell additionally teaches calculating a similarity score for each candidate item definition and using scoring methods to compare attributes and generate the similarity score) (col. 3, ll. 3-38; col. 7, ll. 1-10), wherein the score for each of the selected one or more products is calculated by a machine learning matching model that receives as input a pair comprising the one or more attributes of the product and the corresponding entry of the selected product (Examiner notes that this limitation is disclosed by Lee).
Bicknell does not disclose that the data comprises an image of the product or that the candidate scores are calculated by a machine-learning matching model receiving the claimed product pair, however Lee discloses, wherein the data comprises an image of the product (Lee teaches that product information received with a product request may include a product image URL, along with product name, brand, description, attributes, model number, and other product information, and further teaches obtaining product images for product matching) (0086, 010-0103),
wherein the score for each of the selected one or more products is calculated by a machine learning matching model that receives as input a pair comprising the one or more attributes of the product and the corresponding entry of the selected product (Lee teaches an online matching model trained specifically for product matching. Training data comprises pairs of products whose product information includes product category, name, brand, model number, and other characteristics, wherein the pairs are labeled “match” or “different.” Lee further teaches that the model may be a Naïve Bayes machine-learning model trained to determine the likelihood that a pair of products is identical based on the product information of the pair, and that tagged characteristics such as color, size, and brand can be used to calculate product-pair match scores) (0080-0085),
identifying the match with the product when the score for the match is a highest score of the calculated scores and greater than a threshold value (Lee teaches that the machine-learning model determines a match score between the first product and each candidate product. Lee further teaches determining that the first product is identical to one of the second products when the match score is above a predetermined threshold, including selecting “the second product associated with the highest match score”) (0087-0093),
verifying whether each of the selected one or more products excluding the identified match having a score greater than the threshold value is also a match with the product (Lee teaches identifying the highest-scoring candidate satisfying the predetermined threshold as the identified match at 0093; Bicknell teaches that when more than one candidate item definition has a similarity score above the matching threshold, the result is treated as ambiguous and the candidate item definitions are provided to a manual matching subsystem for verification and resolution, wherein a user determines which candidate item definition, if any, actually matches the item description) (0093; Bicknell, col. 3, ll. 38-59; col. 4, ll. 13-40).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the applicant’s invention to modify Bicknell's candidate-product comparison and scoring mechanism to incorporate Lee's machine-learning product-pair scoring and ranking such that one highest-scoring above-threshold candidate is identified as the primary match while Bicknell's existing ambiguity-resolution procedure is applied to the remaining above-threshold candidates. Such modification would have predictably preserved Bicknell's verification of ambiguous matches while allowing the most strongly matched candidate to be automatically identified.
Bicknell discloses, subjecting additional above-threshold candidate matches to verification as ambiguous matches, but further does not disclose adapting, by the processor, the one or more attributes based on historical data relating to the product into a same format as that of one or more entries in a database, the database comprising one or more entries for each of a plurality of products, however Bremmer discloses, adapting, by the processor, the one or more attributes based on historical data relating to the product into a same format as that of one or more entries in a database, the database comprising one or more entries for each of a plurality of products (Bremer teaches preprocessing record attributes, including transforming their format and mapping values to predefined values; Bremer further teaches a predefined repository structure and attributes that may be determined from historical data, and teaches transforming received records having a different structure into the predefined database structure before processing) (0034, 0068-0070);
and retraining the machine learning matching model based on the verification (Bremer teaches training a machine-learning model on labeled record pairs, applying the model to unlabeled pairs, selecting pairs using model results, obtaining “same” or “different” labels from a data steward, adding the newly labeled pairs to the training data, and repeating the training process) (0079-0089, Fig. 2),
the retraining comprising training the machine learning matching model on training data comprising the pairs for the selected one or more products excluding the identified match having a score greater than the threshold value (Bremer teaches selecting model-classified record pairs, obtaining verification labels for the selected pairs, adding the verified/labeled pairs to the training data, and repeating training of the machine-learning model; Bicknell teaches subjecting additional above-threshold candidate matches to verification as ambiguous matches at col. 3, ll. 38-59, and Lee teaches identifying the highest-scoring above-threshold candidate as the identified match at 0093) (0083, 0088-0089).
each of the pairs being associated with a result of the verification indicating whether the respective selected product is also a match with the product (Bremer teaches that each selected record pair is presented to a data steward/classifier, who compares the records and supplies a label indicating whether they represent the “same” entity or “different” entities, and the newly labeled pairs are incorporated into the labeled training set for subsequent training) (0088-0089).
It would have been obvious to a person of ordinary skill in the art before the effective filling date of the applicant’s invention to modify Bicknell, as modified by Lee, to employ Bremer's active-learning retraining process so that verified ambiguous candidate pairs are retained as labeled training examples and used to retrain the machine-learning matching model. Bicknell already identifies and verifies ambiguous matches, while Bremer teaches that model-selected uncertain pairs may provide informative training examples. Using those verification results as labeled training data would have predictably improved subsequent matching accuracy and reduced repeated manual resolution of similar ambiguous cases.
Bicknell further does not disclose extracting the one or more attributes comprises identifying the one or more attributes from the image based on a visual question and answering (VQA) paradigm, however Biten discloses, and extracting the one or more attributes comprises identifying the one or more attributes from the image based on a visual question and answering (VQA) paradigm (Biten teaches Scene-Text VQA reasoning over vision, language, and scene-text information. The LaTr architecture inputs three modalities—image, OCR information, and a question—to a transformer. Visual features are generated from image patches using a Vision Transformer, OCR information is encoded, the question is encoded, and the combined visual/OCR/question representations are processed by a multimodal transformer encoder-decoder to generate an answer) (“LaTr So far, we explained the building blocks of our method, now we describe how we put it all together, as depicted in Fig. 2 (b). After pre-training the language modality of the model with layout information, we input all three modalities, namely; image, OCR information and question to the transformer encoder. Let V = {V₀, ..., VN } be a set of visual patch features such that V₀ is the [class] embed-ding, Q = {W₁, ..., Wm} be the question tokenized into Wi and O = {O₁, O₂, ..., On} be the OCR tokens. We embed the OCR tokens and questions using Eq. (1) to obtain encoded OCR tokens E and encoded question features E q. For the 2-D spatial embedding of each Wi, we use fixed values (x₀ = y₀ = 0; x₁ = y₁ = 1000). Finally, we concatenate all the inputs [V; E ; E q] to feed to the multimodal trans-former encoder-decoder architecture. Cross entropy loss is used to fine-tune our model”) (Fig. 2(b), §§2-3, “LaTr.”).
It would have been obvious to a person of ordinary skill in the art before the effective filling date of the applicant’s invention to modify Bicknell, as modified by Lee, to employ Biten's known VQA technique for obtaining information from the product-image data taught by Lee because Bicknell relies on descriptive product attributes to identify and match products, Lee teaches that product information can include product-image data, and Biten provides a known technique for deriving responsive textual and visual information from an image based on a question. Using Biten's VQA technique to obtain additional product attributes from an available product image would have predictably increased the product information available to Bicknell's matching engine and improved identification of products whose textual descriptions are incomplete or ambiguous.
As per claims 2 and 6, Bicknell discloses, wherein selecting the one or more products from the plurality of products in the database based on the rule-based comparison further comprises: identifying one or more key attributes from the one or more attributes (Bicknell teaches that each matching rule includes one or more filters, wherein each filter identifies one or more attributes that are used in calculating the similarity score, and further assigns different weights to the selected attributes, thereby indicating the relative importance of those attributes in determining a product match. For example, Bicknell assigns different weights to ISBN, author, title, brand, and product-format attributes depending upon the applicable matching rule) (col. 7, Table 1; col. 8, ll. 20-39);
and selecting the one or more products from the plurality of products in the database based on comparing the one or more key attributes with each corresponding entry in the database (Bicknell further teaches that the candidate-identification component uses the attributes identified in a designated filter to retrieve candidate item definitions. In one example, the applicable filter identifies title and product-type identification attributes, and the component retrieves item definitions whose attribute values correspond to the selected attributes. The system generates a query from those selected attribute values and executes the query against indexes to obtain candidate item definitions. Accordingly, Examiner interprets Bicknell's attributes selected by the applicable filter, including attributes assigned differing weights and used for candidate retrieval, as the claimed one or more key attributes) (col. 10, ll. 60-68; col. 11, ll. 1-30).
As per claims 3 and 7, Bicknell teaches that its product/item records contain product-identifying title/name attributes and that those attributes are compared against corresponding product-catalog entries, and further teaches applying textual comparison techniques such as word matching and edit distance to product-title attributes, but does not expressly disclose adapting, by the processor, the one or more attributes into at least one of a unit of measurement, quantity, language, or item name that is the same as that of one or more entries in the database, however Bremmer discloses, further comprising adapting, by the processor, the one or more attributes into at least one of a unit of measurement, quantity, language or item name that is the same as that of one or more entries in the database (Bremer teaches preprocessing attribute values before record matching, including transforming the format of attribute values, removing noise characters, and performing word mapping of an attribute value to a corresponding predefined value, for example mapping “St.” to “Street” 0034. Bremer further teaches transforming received records having a structure different from the database's predefined data structure into the predefined structure before matching) (0034, 0068-0070).
It would have been obvious to a person of ordinary skill in the art before the effective filling date of the applicant’s invention to modify Bicknell's product-matching system to employ Bremer's known preprocessing and normalization technique to map differently represented product attribute values, including product name/title information, into a predefined form corresponding to the entries used for comparison in Bicknell's product database. Such normalization would have predictably improved the consistency and reliability of Bicknell's attribute comparisons by reducing mismatches caused by differently formatted or differently expressed attribute values.
As per claims 4 and 8, Bicknell discloses, wherein the database further indicates an identifier for each of the plurality of products (Bicknell teaches that merchant inventory information must be mapped to a product in the product catalog and gives ISBN as an example of an identifier associated with the corresponding catalog product. Once the inventory product is mapped to the catalog product, the website can identify the corresponding inventory for that product) (col. 1, ll. 35-50),
the method further comprising mapping the product with the identifier of the identified match (Bicknell further teaches that an authoritative item definition may include an industry-standard unique identifier for the item and that the matching engine identifies the item definition corresponding to the incoming item description. Bicknell additionally teaches that the inventory table may contain “a unique identifier of the matching item definition in the item definition table,” thereby linking the matched product to the identifier of its corresponding authoritative database entry) (col. 2, ll. 54-67, col. 4, ll. 51-61).
As per claims 10, 11, and 12, Bicknell does not disclose wherein the VQA paradigm comprises generating (i) a question embedding, (ii) a plurality of text embeddings derived from optical character recognition (OCR) on the image, and (iii) a plurality of image embeddings generated from image patches, and processing the embeddings using a neural network, however Biten discloses, wherein the VQA paradigm comprises generating (i) a question embedding (Examiner interprets Biten discloses receiving a question as one of three input modalities. The question is tokenized into question tokens Q={W1,…,Wm}, and the question tokens are embedded to produce encoded question features (Eq). The encoded question features are subsequently provided to the multimodal transformer architecture i.e. it then uses those features as one of the three inputs to the multimodal transformer)) (Biten, Fig. 2(b), §3, “LaTr”),
(ii) a plurality of text embeddings derived from optical character recognition (OCR) on the image (Biten obtains OCR tokens from the image and embeds the OCR tokens to produce encoded OCR representations. Figure 2 shows an OCR system and corresponding word embeddings, and Biten explains that the OCR token representations, including associated spatial information, are encoded before being supplied to the transformer encoder) (Fig. 2(b), §3, “Visual Features” and “LaTr”),
and (iii) a plurality of image embeddings generated from image patches, and processing the embeddings using a neural network (Examiner interprets that Biten discloses using a Vision Transformer to generate visual features. The input image is reshaped into a plurality of two-dimensional image patches. The flattened patches are mapped into a multidimensional embedding space using a linear projection layer, and positional embeddings are added to the patch embeddings. The resulting visual output is represented as a sequence of visual patch features (V={VO, …, VN}) (Fig. 2(b), §3, “LaTr”).
It would have been obvious to a person of ordinary skill in the art before the effective filling date of the applicant’s invention to modify Bicknell, as modified by Lee, to employ Biten's multimodal VQA architecture comprising question embeddings, OCR-derived text embeddings, and image-patch-derived visual embeddings processed by a transformer neural network. Lee provides product-image information for product matching, Bicknell relies upon descriptive product attributes to identify corresponding products, and Biten provides a known VQA architecture for jointly processing image, OCR-text, and question information. Employing Biten's multimodal representations would have predictably provided additional image-derived product information for Bicknell's matching process while improving VQA processing where relevant information is distributed between visual and scene-text content.
Response to Arguments
With regards to § 101 rejections:
The arguments filed on September 23, 2026, with respect to the rejection(s) of claims 1-12 under 35 U.S.C 101 have been fully considered and are persuasive in view of the amendments to independent claims 1, 5, and 9. Accordingly, the rejection of claims 1-12 under 35 U.S.C. §101 is withdrawn.
Even assuming, without deciding, that certain comparison, evaluation, ranking, or matching operations recited in the claims constitute an abstract idea under Step 2A, Prong One, the claims as presently amended are considered to integrate any such judicial exception into a practical application under Step 2A, Prong Two. The claims now require: a machine-learning matching model that receives as input a pair comprising the extracted attributes of the product and the corresponding database entry of a selected product; calculation of a score for each such pair; identification of a match where the score is both the highest calculated score and greater than a threshold; verification of the remaining selected products having above-threshold scores; and retraining that machine-learning matching model using training data comprising those verified product pairs, each pair being associated with the result of the verification. Applicant identifies support for this feedback architecture in paragraphs (0073-0074, 0080-0082). Thus, the amended claims specify which model is retrained, which examples are selected for retraining, how those examples arise from operation of the model, and which verification result is associated with each training example. The claimed verification results are therefore not merely collected or displayed; they are returned to the machine-learning matching model as labeled training examples for subsequent operation of the matching process. Applicant specifically characterizes this as a closed feedback loop in which above-threshold non-selected candidates are verified and returned to the model for further training.
This analysis is consistent with MPEP §2106.04(d)(1), which instructs that a claim may integrate a judicial exception into a practical application where the claim reflects an improvement in computer functionality or another technological field. It is also consistent with the precedential decision in Ex parte Desjardins, in which the Appeals Review Panel determined that limitations directed to improving how a machine-learning model itself operates may constitute a technological improvement under Step 2A, Prong Two. The Examiner further notes that the independent claims require obtaining attributes from product-image data using a VQA paradigm, and new dependent claims 10-12 more particularly recite question embeddings, OCR-derived text embeddings, image-patch-derived embeddings, and neural-network processing. These limitations are considered as part of the claims as a whole, rather than as isolated data-gathering steps. Accordingly, on the present record, the claimed combination integrates any recited abstract idea into a practical application, and the claims are eligible at Step 2A, Prong Two. Therefore, it is unnecessary to reach Step 2B, and the previous determination concerning whether the additional elements were well-understood, routine, and conventional is not maintained.
With regards to § 103 rejections:
Applicant's arguments, see pages 10-13, filed September 23, 2026, with respect to the rejection(s) of claims 1-12 under 35 U.S.C 102/103 have been fully considered but are unpersuasive/moots on new ground of rejection. Applicant's arguments concerning the previous rejection under 35 U.S.C. §103 have been fully considered. The previous rejection based on Yoon in view of Krasadakis and Biten is not maintained in the same form. Applicant's arguments regarding Yoon, Krasadakis, and Jain are therefore moot with respect to the present rejection.
Independent claims 1, 5, and 9 were amended to expressly require that the score for each selected product be calculated by a machine-learning matching model receiving a product/candidate pair, and that the matching model be retrained using verified above-threshold candidate pairs excluding the identified match, with each pair associated with its verification result. Applicant identifies these features as the basis for distinguishing the previously applied art. The present rejection applies Bicknell in view of Lee and Bremer, and further in view of Biten. Bicknell teaches rule-based product matching, candidate selection, threshold scoring, and resolution of multiple above-threshold candidate matches; Lee teaches machine-learning product-pair scoring and identifying a highest-scoring candidate satisfying a threshold; Bremer teaches obtaining same/different verification labels for model-selected pairs and using the labeled pairs in subsequent training of the machine-learning model; and Biten is retained for the VQA-based image-processing limitations. Thus, the newly cited combination addresses the limitations added by Applicant's amendment rather than relying on the teachings Applicant challenged in Yoon and Krasadakis. The detailed teachings and reasons for combining the references are set forth in the rejection above. Accordingly, Applicant's arguments are persuasive as to deficiencies alleged in the previously applied combination, but do not overcome the present rejection based on the newly applied prior art.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. US. Pat. 10949907 (“Jain”).
Jain discloses, a method for generating a list of products each matching a reference product are disclosed. A user query is first received, and multi-modal attribute data for the reference product are determined, with each data mode being a type of product characterization having a modality selected from a text data class, categorical data, a pre-compared engineered feature, audio, image, and video. Next, a first list of candidate products is determined based on a product match signature, and a second list of candidate products is generated from the first, wherein for at least one given candidate product, a deep learning multi-modal matching model is selected to determine whether a match is found. Lastly, the second list is filtered to remove outliers and to generate the list of matching products.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GAUTAM UBALE whose telephone number is (571)272-9861. The examiner can normally be reached Mon-Fri. 7:00 AM- 6:30 PM PST.
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, Marissa Thein can be reached at (571) 272-6764. 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.
/GAUTAM UBALE/
Primary Examiner, Art Unit 3689
