DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Examiner’s Note
Applicant’s Patent Application Fee Determination Record (SB06), as filed on 06 February 2025, indicates three (3) independent claims and 15 dependent claims. The listed independent claims appears to correspond to claims 1, 8, and 11 as these are the only claims which are clearly in independent form and contain no statements indicating reliance on another claim. The remaining claims, claims 2-7, 9-10, and 12-15 are understood, as indicated by the applicant, to be asserted as dependent claims.
Applicant is advised regarding the dependency of the claims. The Federal Circuit has indicated that “A claim's status as dependent or independent depends on the substance of the claim in light of the language of § 112, ¶ 4, and not the form alone.” (Monsanto v. Syngenta Seeds, 503 F.3d 1352, 1357 (Fed. Cir. 2007)). The court further explained that, under pre-AIA 35 U.S.C § 112, ¶ 4 (currently 35 U.S.C § 112(d)), “[t]o establish whether a claim is dependent upon another, this court examines if the new claim both refers to an earlier claim and further limits that referent.” (Monsanto at 1357, citing 35 U.S.C § 112, ¶ 4 (2000)). Specifically with reference to dependence from a process claim, the court held that whether a claim is properly read as dependent or independent turns on whether the asserted dependent claim “specifically requires …the performance of the steps” in the referent process claim. (Id. at 1358).
During prosecution, it falls to the examiner to determine status as dependent or independent with reference to 35 U.S.C § 112(d). (See MPEP 2173.05(f)). As such, we perform the same analysis here for claims 2-7, 9-10, and 12-15.
Regarding claims 2-5 and 12-13, these claims both refer to claims 1 and 11, respectively, and provide further limitation to that claim. As well, under the broadest reasonable interpretation, claims 2-5 and 12-13 specifically require the performance of the steps in the referent process claim. As such, claims 2-5 and 12-13 are understood as dependent claims for the purposes of further prosecution.
With reference to the substance of claims 6-7, 9-10, and 14-15, these claims “refer to an earlier claim” but the fail to “further limit that referent,” as the limitations are directed to changing the statutory class of the referent claim. Claims 1, 8, and 11 recite a process which “consist[s] of a series of steps or acts to be performed.” (MPEP 2106). Claims 6-7, 9-10, and 14-15 are asserted as dependent from claims 1, 8, and 11, but act only to convert the previously cited process to another statutory class. Specifically, claims 6, 9, and 14 is directed to a computer readable media, and claims 7, 10, and 15 are directed to an electronic device. Claims 6-7, 9-10, and 14-15 are not understood as a result of, a continuation of, or a further limit to the process of claim 1.
Further, claims 6-7, 9-10, and 14-15 do not specifically require the performance of the steps in the referent process claim. Though claims 1, 8, and 11 recites a series of steps or acts to be performed, the broadest reasonable interpretation of claims 6-7, 9-10, and 14-15 fails to require the actions defined by the process of claims 1, 8, and 11 actually be performed. The process of claims 1, 8, and 11 is incorporated by reference into claims 6-7, 9-10, and 14-15 as a computer code or instruction of some kind, which need not ever be executed. As such, though claims 6-7, 9-10, and 14-15 refer back to claims 1, 8, and 11 in an apparent dependent form, these claims are independent claims.
The application, both at the time of filing and currently, contains nine (9) independent claims and a total of 15 claims. If the applicant wishes for these claims to be treated as dependent, applicant is advised to amend the claims, in light of specification support, such that the claims further limit the referenced independent claim, as required for dependency by 35 U.S.C § 112(d).
Information Disclosure Statement
The information disclosure statement(s) (IDS) submitted on 17 November 2025 is/are being considered by the examiner.
Claim Objections
Claim 12 is objected to because of the following informalities:
Regarding claim 12, the phrase “dimensions of preset instruction” at line 1 should read as “dimensions of preset instruction information”.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 11-15 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claim 11, the limitation “expanding… in dimensions of preset instruction information” lacks clarity. Claim 11 recites “expanding the first prompt sample in dimensions of preset instruction information” at lines 3-4. However, expanding “in dimensions” contains syntactical ambiguity which cannot be clarified based on context of the claim or description in the specification. In data processing, the word “dimensions” is generally understood to refer to separate classes or attributes of information in a data element. However, the claim recites that the dimension is a property of the preset instruction information. Because the first prompt sample is not described as having or corresponding to these dimensions, it is unclear what is being expanded within the first prompt sample or how that expansion operation functionally relates to the “preset instruction information” through the phrase “in dimensions.” Therefore, claim 11 lacks clarity and is rejected.
Regarding claims 12-15, claims 12-15 incorporate all limitations from claim 11, either by reference or by dependency. Therefore, claims 12-15 are rejected at least for the same reasons as described with reference to claim 11.
Appropriate correction is required.
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 (i.e., changing from AIA to pre-AIA ) 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, 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.
Claims 1-4 and 6-7 is/are rejected under 35 U.S.C. 103 as being unpatentable over Goligorsky (U.S. Pat. App. Pub. No. 2024/0265205, hereinafter Goligorsky) in view of De Barros (U.S. Pat. App. Pub. No. 2025/0036695, hereinafter De Barros).
Regarding claim 1, Goligorsky discloses An object material generation method (The systems and methods described with reference to the “example e-commerce platform”; Goligorsky, ¶ [0074]), comprising: performing intent recognition on instruction information indicating the generation of an object material to obtain an intent recognition result, (“the process proceeds to block 730 in which the text is analyzed {intent recognition} to identify feature inputs and prompt instructions {intent recognition results}”; Goligorsky, ¶ [0143]) wherein the intent recognition result includes a material generation scenario and/or key information of the object (The intent recognition results, such as in the context of the “a product description” and/or “a marketing ad” include “feature inputs {key information of the object} and prompt instructions {material generation scenario}”; Goligorsky, ¶ [0143], [0147]); in response to the intent recognition result meeting a preset condition, formatting a preset prompt template based on the intent recognition result to generate a prompt (“Once the text has been analyzed and divided into feature inputs and prompt instructions, the divided text can be provided to a prompt generator at block 740” where “prompts may be provided to a prompt structure analyzer which may, at block 732, analyze the structure of the prompt”, said structure including the minimum requirements for the product descriptions (See FIG. 6) and the maximum character allowance, where the structural analysis results in either “determin[ing] a prompt template from an existing prompt given to the system” or “generat[ing] a prompt template if no existing prompt (or prompt template) is given.” In the template example, as shown in FIGS. 8 and 9, the preset prompt template (FIG. 8) is formatted based on the “feature inputs and prompt instructions {intent recognition result}” to generate the prompt as shown in FIG. 9; Goligorsky, ¶ [0122], [0144], [0146]; FIGS. 6, 8 and 9); in response to the intent recognition result not meeting the preset condition, generating a prompt based on the instruction information using a… prompt generation model (“a prompt generation engine” may be used “to generate prompts to be provided as input to a language model such as a LLM” where “In the case where no existing prompt template is given, an output expectation may be provided instead that describes the type of response a user is expecting from the text generation platform” such as “a product description, a marketing ad” and using “the given output expectation, the prompt structure analyzer may create a simple prompt template of the format: (1) preamble, (2) list of prompt instructions, and (3) list of feature inputs.”; Goligorsky, ¶ [0147]-[0148]); triggering a preset object material generation model based on the generated prompt to produce the object material (“once the prompt is generated, in some embodiments the prompt may be sent to an LLM, as shown at block 750 and response may then be received from the LLM at block 760” and said “response may be presented to a user on a user interface” where the response may be “the product description” as generated by the LLM “for the product outlined in the text input of FIG. 6”; Goligorsky, ¶ [0168]-[0169]; FIG. 6). However, Goligorsky fails to expressly recite using a pre-fine-tuned prompt generation model.
De Barros teaches systems and methods for “providing enhanced output of a generative model.” (De Barros, ¶ [0022]). Regarding claim 1, De Barros teaches using a pre-fine-tuned prompt generation model (As part of the workflow model, the “intent classifier 114 provides its output indicative of a user intent to the workflow model 110 which can then use the output to generate a more accurate and effective prompt for input into the generative model 112” where the “the workflow model is a model (e.g., a generative model) trained using intent-specific data” {pre-fine-tuned prompt generation model} and “responsive to receiving an indication of an intent associated with the input (e.g., from the output of the intent classifier 114)” the “prompt generator 116 is configured to generate a prompt for input into the generative model 112 which causes the generative model 112 to produce an output” and can further incorporate grounding information “wherein the grounding information is associated with the intent” and may be “keywords or other aspects identified explicitly from user input”; De Barros, ¶ [0008], [0034]-[0036]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the product description generation platform of Goligorsky to incorporate the teachings of De Barros to include using a pre-fine-tuned prompt generation model. The product description platform of Goligorsky discloses the use of generated prompts to cause a generative model to produce product descriptions which better reflect the intent of the received input. However. Goligorsky fails to expressly recite a pre-fine-tuned prompt generation model for the generation of said generated prompts. De Barros discloses a broadly applicable workflow model which incorporates a generative model fine-tuned using “intent specific data”. A PHOSITA would be motivated to combine the product description platform of Goligorsky with the workflow model of De Barros, such that the produced product descriptions “have a higher likelihood of responsiveness and relevancy to the input (e.g., by corresponding to an intent of the user)” due to the prompt generated by workflow model having “ground[ing] using data relating to the intent (e.g., based upon the output of the intent classifier which is indicative of an intent of the user input),” where such data is expected to be improved by pre-training using “intent specific data,” as recognized in light of De Barros. (De Barros, ¶ [0040]).
Regarding claim 2, the rejection of claim 1 is incorporated. Goligorsky and De Barros disclose all of the elements of the current invention as stated above. Goligorsky further discloses wherein the key information of the object comprises an original title of the object, an object category, and object attributes (As depicted in FIG. 9, and the accompanying description, the feature inputs can include “product name and product features”, where the product features, as shown in FIG. 6 include “locally sourced” which is an object category, and “10g of protein” which is an object attribute.; Goligorsky, ¶ [0154]; FIG. 5 and 9), and the preset condition comprises: the intent recognition result including the material generation scenario and at least two types of the key information of the object (As shown in FIG. 6, the instructions and at least two product features are required for use of the “autowrite product description” function. As both the instructions and the product features are required prior to the determination of whether a prompt template can be used, the preset conditions comprise the same.; Goligorsky, ¶ [0110]-[0111]; FIG. 6).
Regarding claim 3, the rejection of claim 1 is incorporated. Goligorsky and De Barros disclose all of the elements of the current invention as stated above. Goligorsky further discloses wherein formatting a preset prompt template based on the intent recognition result to generate a prompt comprises: retrieving a pre-established prompt template corresponding to the material generation scenario (“the divided text can be provided to a prompt generator at block 740” which can use a selected prompt template from the “prompt templates (or prompt structures) for the text input field” which “may exist” and a “prompt may be, in some cases, created based on the prompt template”; Goligorsky, ¶ [0144]-[0145]); and formatting the prompt template based on the key information of the object included in the intent recognition result to generate the prompt (“Once the text has been analyzed and divided into feature inputs and prompt instructions, the divided text can be provided to a prompt generator at block 740” where, as shown in FIGS. 6 and 9, the prompt template is formatted to include the feature inputs (see 940 of FIG. 9 incorporating feature inputs 630 of FIG. 6) and the prompt instructions (see 920 of FIG. 9 incorporating the prompt instructions of FIG. 6); Goligorsky, ¶ [0144]-[0145]; FIGS. 6 and 9).
Regarding claim 4, the rejection of claim 1 is incorporated. Goligorsky and De Barros disclose all of the elements of the current invention as stated above. Goligorsky further discloses wherein performing intent recognition on the instruction information indicating the generation of an object material to obtain an intent recognition result comprises: identifying the material generation scenario... and extracting the key information of the object from the instruction information (“the text {the instruction information} is analyzed to identify feature inputs and prompt instructions {material generation scenario}” where the text is “analyzed and divided {...from the instruction information} into feature inputs {extracting key information of the object...} and prompt instructions”; Goligorsky, ¶ [0143]-[0145]). However, Goligorsky fail(s) to expressly recite wherein performing intent recognition on the instruction information indicating the generation of an object material to obtain an intent recognition result comprises: identifying the material generation scenario matched by the instruction information for generating the object material using a pre-fine-tuned intent recognition model.
The relevance of De Barros is described above with relation to claim 1. Regarding claim 4, De Barros teaches wherein performing intent recognition on the instruction information indicating the generation of an object material to obtain an intent recognition result comprises: identifying the material generation scenario matched by the instruction information for generating the object material using a pre-fine-tuned intent recognition model (“The intent classifier 114 receives the input set forth by the user at the client computing device 102. From the input, the intent classifier 114 is configured to produce an output indicative of an intent associated with the input” and “The intent is indicative of an objective related to the input, such as, for example, shopping” which “may be further refined into a sub-intent.” As read in the context of the prompt “Write a short product description for a product with the following features {features} and make sure to follow these instructions when writing: {instructions},” of Goligorsky at [0109], the intent classifier of De Barros determines the intent of generating a product description, as said generating is “indicative of [the] intent associated with the input”, which is a material generation scenario. Further, the “intent classifier 114 may be a classification model trained on intent-specific information such that output of the intent classifier 114 is indicative of an intent associated with the input” thus being fine-tuned for the intent recognition.; De Barros, ¶ [0034]-[0035]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the product description generation platform of Goligorsky to incorporate the teachings of De Barros to include wherein performing intent recognition on the instruction information indicating the generation of an object material to obtain an intent recognition result comprises: identifying the material generation scenario matched by the instruction information for generating the object material using a pre-fine-tuned intent recognition model. The product description platform of Goligorsky discloses the use of generated prompts to cause a generative model to produce product descriptions which better reflect the intent of the received input. However. Goligorsky fails to expressly recite a pre-fine-tuned prompt generation model for the generation of said generated prompts. De Barros discloses a broadly applicable workflow model which incorporates a generative model fine-tuned using “intent specific data”. A PHOSITA would be motivated to combine the product description platform of Goligorsky with the workflow model of De Barros, such that the produced product descriptions “have a higher likelihood of responsiveness and relevancy to the input (e.g., by corresponding to an intent of the user)” due to the prompt generated by workflow model having “ground[ing] using data relating to the intent (e.g., based upon the output of the intent classifier which is indicative of an intent of the user input),” where such data is expected to be improved by pre-training using “intent specific data,” as recognized in light of De Barros. (De Barros, ¶ [0040]).
Regarding claim 6, Goligorsky discloses A non-transitory computer-readable storage medium configured with instructions executable by one or more processors to cause the one or more processors (The systems and methods described with reference to the “example e-commerce platform” as implemented through “example computing system 400” including “memory 404” which is a non-transitory computer readable medium as known in the art, which “may store instructions for execution by the processor 402, to the computing system 400 to carry out examples of the methods, functionalities, systems and modules disclosed herein.”; Goligorsky, ¶ [0069], [0074]) to perform the method of claim 1 (See mapping of claim 1; Goligorsky and De Barros).
Regarding claim 7, Goligorsky discloses An electronic device comprising: one or more processors; and one or more computer-readable memories coupled to the one or more processors and having instructions stored thereon that are executable by the one or more processors (The systems and methods described with reference to the “example e-commerce platform” as implemented through “example computing system 400” including “at least one processing unit, such as a processor 402, and at least one physical memory 404” where “memory 404 may store instructions for execution by the processor 402, to the computing system 400 to carry out examples of the methods, functionalities, systems and modules disclosed herein.”; Goligorsky, ¶ [0069], [0074]) to perform the method of claim 1 (See mapping of claim 1; Goligorsky and De Barros).
Claim 5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Goligorsky and De Barros as applied to claim 1 above, and further in view of Mishra (U.S. Pat. App. Pub. No. 12,579,565, hereinafter Mishra).
Regarding claim 5, the rejection of claim 1 is incorporated. Goligorsky and De Barros disclose all of the elements of the current invention as stated above. Goligorsky further discloses wherein the instruction information comprises user-inputted instruction information (“A user may enter any natural language” as the text for the feature inputs or the prompt instructions; Goligorsky, ¶ [0138]), and after triggering a preset object material generation model based on the generated prompt to produce the object material, the method further comprises: displaying the generated object material to the user (“once the prompt is generated” which is after triggering generation of the prompt at the computing system, “the prompt may be sent to an LLM, as shown at block 750 and response may then be received from the LLM at block 760” and “The response may be presented to a user on a user interface at block 770. For example, the product description for the product outlined in the text input of FIG. 6, as generated by the text generation platform.”; Goligorsky, ¶ [0068], [0168]-[0169]; FIG. 6); and gathering user feedback on the generated object material, (“the product description for the product outlined in the text input of FIG. 6, as generated by the text generation platform, may be presented to the user for approval in some embodiments” where presenting an output to a user for approval is the gathering of user feedback on said output {the generated object material}; Goligorsky, ¶ [0169]; FIG. 6) [wherein the model can be fine-tuned through reinforcement learning] (“a trained ML model may be fine-tuned, meaning that the values of the learned parameters may be adjusted slightly in order for the ML model to better model a specific task” which “typically involves further training the ML model on a number of data samples (which may be smaller in number/cardinality than those used to train the model initially) that closely target the specific task.” and where the training includes inputting into an ML model, the “training data to be processed... processing the training data using the ML model, collecting the output generated by the ML model (e.g. based on the inputted training data), and comparing the output to a desired set of target values” where the target values may be established by “a measure of some target observable effect on the environment” such by “reinforcement learning”; Goligorsky, ¶ [0047], [0050]). However, Goligorsky and De Barros fail to expressly recite wherein information of the feedback is used, in combination with the prompt and the object material, to generate supervised fine-tuning samples, which are used to perform human feedback reinforcement learning on the object material generation model.
Mishra teaches systems and methods for generating content reviews using pretrained generative language models. (Mishra, ¶ col. 2, lines 56-60). Regarding claim 5, Mishra teaches wherein information of the feedback is used, in combination with the prompt and the object material, to generate supervised fine-tuning samples (“the generative language models described herein may be generative pre-trained transformers that are fine-tuned for the specific tasks discussed herein using reinforcement learning” where “the model is presented with a task (e.g., review generation, item question answering, etc.) {the prompt} and the model receives feedback in the form of a reward or penalty (defined by the objective function) {the information of the feedback} based on the actions taken by the model {the object material}”; Mishra, ¶ Col. 6, lines 34-53), which are used to perform human feedback reinforcement learning on the object material generation model (“The model’s parameters are updated based on this cumulative reward calculated mathematically from the objective function” and the available feedback data can be “user response data” which is human feedback reinforcement learning, that, as read in the context of Goligorsky, is performed on the object material generation model; Mishra, ¶ Col. 6, lines 34-53).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the product description generation platform of Goligorsky as modified by the workflow model for prompt generation of De Barros to incorporate the teachings of Mishra to include wherein information of the feedback is used, in combination with the prompt and the object material, to generate supervised fine-tuning samples, which are used to perform human feedback reinforcement learning on the object material generation model. The combination of Goligorsky and De Barros discloses a fine-tuned intent based prompt generation system for generating product descriptions using a generative language model. However, Goligorsky and De Barros, though disclosing the use of reinforcement learning generally, fail to expressly recite the generation of supervised fine-tuning samples from feedback for human feedback reinforcement learning. Mishra discloses the fine tuning of a LLM using “item reviews “ which are ranked based on user indications of helpfulness or usefulness, such that the “fine-tuned generative language model learns to generate high quality reviews that touch upon the most salient aspects of the item.” A PHOSITA would be motivated to combine the product generating system of Goligorsky and De Barros with the human feedback reinforcement learning described in Mishra, to achieve the well-known benefit of improving review generation quality based on intangible aspects of human preference, as evidenced by the received human input described in Mishra. (Mishra, ¶ Col. 8, lines 7-44).
Claims 8-10 is/are rejected under 35 U.S.C. 103 as being unpatentable over De Barros in view of Goligorsky.
Regarding claim 8, De Barros discloses A model fine-tuning method (Systems and methods described with reference to a workflow model.; De Barros, ¶ [0034]), comprising: obtaining a set of first instruction information samples, (“The workflow model 110 comprises an intent classifier 114” and a “prompt generator 116” where “the intent classifier 114 may be... trained on intent-specific information such that output of the intent classifier 114 is indicative of an intent associated with the input” and “grounding information...associated with the intent” comprising “keywords or other aspects identified explicitly from user input”; De Barros, ¶ [0034]-[0036]) wherein each of the first instruction information samples is used to indicate [the objective] (As described above, the intent-specific information includes the “intent” and “grounding information...associated with the intent” comprising “keywords or other aspects identified explicitly from user input” which are “indicative of an objective related to the input”.; De Barros, ¶ [0034]-[0036]); fine-tuning a pre-trained generative model using input-output text pairs… (“The workflow model 110 comprises an intent classifier 114” and a “prompt generator 116” where “the intent classifier 114 may be a classification model trained on intent-specific information such that output of the intent classifier 114 is indicative of an intent associated with the input” and “grounding information...associated with the intent” comprising “keywords or other aspects identified explicitly from user input”.; De Barros, ¶ [0034]-[0036]). However, De Barros fails to expressly recite [wherein the objective is]... generation of an object material; assigning labeling information to each first instruction information sample in the set, wherein the labeling information includes a material generation scenario and key information of the object contained in the first instruction information sample, and wherein the material generation scenario includes any of the following: title generation, selling point generation, object detail generation, and marketing text generation, and the key information of the object includes an original title of the object, an object category, and an object attribute; fine-tuning a pre-trained generative model using input-output text pairs composed of the first instruction information samples and their corresponding labeling information to obtain an intent recognition model.
Goligorsky teaches systems and methods for generation of a product description using a generative language model. (Goligorsky, ¶ [0022]). Regarding claim 8, Goligorsky teaches [wherein the objective is]... generation of an object material (“the process proceeds to block 730 in which the text is analyzed to identify feature inputs and prompt instructions” where the “feature inputs and prompt instructions” are the first instruction information which indicates the generation of object material.; Goligorsky, ¶ [0143]); assigning labeling information to each first instruction information sample in the set, (“Training data may be annotated with ground truth labels (e.g. each data entry in the training dataset may be paired with a label),” where, in supervised learning in the context of fine-tuning a model for identification of “feature inputs and prompt instructions”, the ground truth labels includes labels for the desired output (i.e., labelling of the “feature inputs and prompt instructions”).; Goligorsky, ¶ [0046]) wherein the labeling information includes a material generation scenario and key information of the object contained in the first instruction information sample (As noted above, the training data includes ground truth labels, where the ground truth labels includes labels for the desired output (i.e., labelling of the “feature inputs {key information of the object} and prompt instructions {material generation scenario}”).; Goligorsky, ¶ [0046]), and wherein the material generation scenario includes any of the following: title generation, selling point generation, object detail generation, and marketing text generation (The prompt instructions are directed to, at least object detail generation. The “prompt instructions” may be instructions included as part of a request for “a short product description for a product” where the LLM must “follow these instructions when writing”; Goligorsky, ¶ [0109]), and the key information of the object includes an original title of the object, an object category, and an object attribute (As depicted in FIG. 9, and the accompanying description, the feature inputs can include “product name and product features”, where the product features, as shown in FIG. 6 include “locally sourced” which is an object category, and “10g of protein” which is an object attribute.; Goligorsky, ¶ [0154]; FIG. 5 and 9); fine-tuning a pre-trained generative model using input-output text pairs composed of the first instruction information samples and their corresponding labeling information to obtain an intent recognition model (“a trained ML model may be fine-tuned, meaning that the values of the learned parameters may be adjusted slightly in order for the ML model to better model a specific task” which “involves further training the ML model on a number of data samples (which may be smaller in number/cardinality than those used to train the model initially) that closely target the specific task” where “Training data may be annotated with ground truth labels (e.g. each data entry in the training dataset may be paired with a label)” and “If the training data is labeled, the desired target values may be, e.g., the ground truth labels of the training data.”; Goligorsky, ¶ [0046]-[0047], [0050]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the workflow model of De Barros to incorporate the teachings of Goligorsky to include [wherein the objective is]... generation of an object material; assigning labeling information to each first instruction information sample in the set, wherein the labeling information includes a material generation scenario and key information of the object contained in the first instruction information sample, and wherein the material generation scenario includes any of the following: title generation, selling point generation, object detail generation, and marketing text generation, and the key information of the object includes an original title of the object, an object category, and an object attribute; fine-tuning a pre-trained generative model using input-output text pairs composed of the first instruction information samples and their corresponding labeling information to obtain an intent recognition model. De Barros discloses a broadly applicable workflow model which incorporates a generative model fine-tuned using “intent specific data”. However, De Barros fails to expressly recite the specific steps of supervised fine tuning, as known in the art. The product description platform of Goligorsky discloses the use of generated prompts to cause a generative model to produce product descriptions which better reflect the intent of the received input and further provides for training and fine tuning processes, such as for transformer models, which is a known technique utilized to adapt and align pre-trained language models using domain specific data to domain-specific uses. A PHOSITA would be motivated to combine the workflow model of De Barros with the LLM fine tuning of Goligorsky, because a PHOSITA would recognize that to practically deploy the DE Barros model for end user applications, it requires supervised adaptation. The PHOSITA would naturally look to references in the same art area, such as Goligorsky, which detail the standard supervised fine tuning mechanics (e.g., cross-entropy loss over target tokens, parameter weight updating, learning rate schedules, etc.) to supply the necessary implementation details to make the workflow model of De Barros functional, as recognized in light of Goligorsky. (Goligorsky, ¶ [0043]-[0050]).
Regarding claim 9, De Barros discloses A non-transitory computer-readable storage medium configured with instructions executable by one or more processors to cause the one or more processors (The systems and methods described with reference to the “workflow model” as implemented through a computing system including “computer-readable data storage that is configured with computer-executable instructions that cause certain functionality to be performed when executed by a processor.”; De Barros, ¶ [0024]-[0025]) to perform the method of claim 8 (See mapping of claim 8; De Barros and Goligorsky).
Regarding claim 10, De Barros discloses An electronic device comprising: one or more processors; and one or more computer-readable memories coupled to the one or more processors and having instructions stored thereon that are executable by the one or more processors (The systems and methods described with reference to the “workflow model” as implemented through a computing system including “computer-readable data storage that is configured with computer-executable instructions that cause certain functionality to be performed when executed by a processor.”; De Barros, ¶ [0024]-[0025]) to perform the method of claim 8 (See mapping of claim 8; De Barros and Goligorsky).
Claims 11-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Goligorsky in view of Kobren (U.S. Pat. App. Pub. No. 2023/0032208, hereinafter Kobren).
Regarding claim 11, Goligorsky discloses A model fine-tuning method (The systems and methods described with reference to the “example e-commerce platform”; Goligorsky, ¶ [0074]), comprising: obtaining a first prompt sample and a corresponding first object material sample for the first prompt sample (During supervised training/fine tuning, the system includes using labeled training data, where “the desired target values may be, e.g., the ground truth labels of the training data” and said target values are well understood in the art to be a loss value between the model output and the corresponding ground truth output for a particular input. Thus, the training data for supervised learning, in the context of the generative language model described in Goligorsky, comprises the prompts incorporating the “feature inputs and prompt instructions” and a corresponding ground truth “product description”; Goligorsky, ¶ [0047], [0144], [0168]-[0169]); fine-tuning a pre-trained large language model (“a trained ML model may be fine-tuned, meaning that the values of the learned parameters may be adjusted slightly in order for the ML model to better model a specific task. Fine-tuning of a ML model typically involves further training the ML model on a number of data samples (which may be smaller in number/cardinality than those used to train the model initially) that closely target the specific task.” where a large language model is an exemplary trained ML model.; Goligorsky, ¶ [0050]) based on the first prompt sample and its corresponding first object material sample, as well as the second prompt sample and its corresponding second object material sample, to obtain an object material generation model (training and fine tuning inputting “training data to be processed by the ML model, processing the training data using the ML model, collecting the output generated by the ML model (e.g. based on the inputted training data), and comparing the output to a desired set of target values. If the training data is labeled, the desired target values may be, e.g., the ground truth labels of the training data.” Thus, the training data for supervised learning, in the context of the generative language model described in Goligorsky, comprises the prompts incorporating the “feature inputs and prompt instructions” and a corresponding ground truth “product description”; Goligorsky, ¶ [0047], [0144], [0168]-[0169]). However, Goligorsky fails to expressly recite expanding the first prompt sample in dimensions of preset instruction information based on the first object material sample to obtain a second prompt sample; obtaining a second object material sample corresponding to the second prompt sample using a preset question-answering language model.
Kobren teaches systems and methods for “augmenting data sets used for machine learning models.” (Kobren, ¶ [0001]). Regarding claim 11, Kobren teaches expanding the first prompt sample in dimensions of preset instruction information based on the first object material sample to obtain a second prompt sample (“the method 200 may revise sentences in the set of training sentences to generate another set of training sentences (operation 212). This additional set of training sentences may be based on the content of the initial set of training sentences,” and in revising the sentences “the system extracts a subset of words from the sentence” which “will ultimately be used to generate additional examples for improving the training of one or more machine learning models.” Further “the system may bias selection of tokens based on...type of content word (e.g., based on subject matter associated with the word), or other criteria” and “The execution of the operation 212 and the sub-operations 216-228 causes the system to generate a revised set of clauses that are based on the initial set of sentences received from the dataset generation model.”; Kobren, ¶ [0064], [0068], [0070], [0077]); obtaining a second object material sample corresponding to the second prompt sample using a preset question-answering language model (“The system then uses the generated revised set of clauses as inputs to the dataset generation model to generate a second set of sentences that may be used to fine tune the training of the dataset generation model”; Kobren, ¶ [0078]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the workflow model of Goligorsky to incorporate the teachings of Kobren to include expanding the first prompt sample in dimensions of preset instruction information based on the first object material sample to obtain a second prompt sample; obtaining a second object material sample corresponding to the second prompt sample using a preset question-answering language model. The product description platform of Goligorsky discloses the use of generated prompts to cause a fine-tuned generative model to produce product descriptions which better reflect the intent of the received input. However, Goligorsky relies on a dataset comprising preexisting samples. Kobren discloses systems for expanding an existing training data set. A PHOSITA would be motivated to combine the product description platform of Goligorsky with the dataset expansion of Kobren to address the well-known problem of data insufficiency. The incorporation of Kobren would allow for expansion of limited datasets, as is often seen in domain-specific data sets, resulting in “a greater diversity of examples” which can “improve the precision and accuracy of a machine learning model,” as recognized in light of Kobren. (Kobren, ¶ [0002]).
Regarding claim 12, the rejection of claim 11 is incorporated. Goligorsky and Kobren disclose all of the elements of the current invention as stated above. However, Goligorsky fails to expressly recite wherein the dimensions of preset instruction include one or more of the following dimensions: target generation language, number of languages generated per request, number of content items generated per request, character length of the generated content, and position of key information within the generated content.
The relevance of Kobren is described above with relation to claim 11. Regarding claim 12, Kobren teaches wherein the dimensions of preset instruction include one or more of the following dimensions: target generation language, number of languages generated per request, number of content items generated per request, character length of the generated content, and position of key information within the generated content (“the system may bias selection of tokens based on part of speech, word length, type of content word (e.g., based on subject matter associated with the word), or other criteria” where the selection of tokens is part of the expansion and the bias is the enumerated dimensions for said expansion.; Kobren, ¶ [0077]).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the workflow model of Goligorsky to incorporate the teachings of Kobren to include wherein the dimensions of preset instruction include one or more of the following dimensions: target generation language, number of languages generated per request, number of content items generated per request, character length of the generated content, and position of key information within the generated content. The product description platform of Goligorsky discloses the use of generated prompts to cause a fine-tuned generative model to produce product descriptions which better reflect the intent of the received input. However, Goligorsky relies on a dataset comprising preexisting samples. Kobren discloses systems for expanding an existing training data set. A PHOSITA would be motivated to combine the product description platform of Goligorsky with the dataset expansion of Kobren to address the well-known problem of data insufficiency. The incorporation of Kobren would allow for expansion of limited datasets, as is often seen in domain-specific data sets, resulting in “a greater diversity of examples” which can “improve the precision and accuracy of a machine learning model,” as recognized in light of Kobren. (Kobren, ¶ [0002]).
Regarding claim 13, the rejection of claim 11 is incorporated. Goligorsky and Kobren disclose all of the elements of the current invention as stated above. Goligorsky further discloses further comprising, after fine-tuning the pre-trained large language model to obtain the object material generation model: obtaining supervised fine-tuning samples, (“Training a ML model generally involves inputting... training data to be processed by the ML model, processing the training data using the ML model, collecting the output generated by the ML model (e.g. based on the inputted training data), and comparing the output to a desired set of target values” where “the desired target value... may be a measure of some target observable effect on the environment (e.g., in the case of a reinforcement learning agent)” where reinforcement learning incorporates feedback as the supervision regarding an output from a model {supervised fine-tuning samples}.; Goligorsky, ¶ [0047]) wherein the supervised fine-tuning samples include a third prompt sample, a corresponding third object material sample, and a sample category that matches the third object material sample, (The supervised fine tuning samples are the “generated output value {a corresponding third object material sample}” the corresponding input sample {a third prompt sample}, “and the desired target value,” which can be input received regarding “the product description for the product outlined in the text input of FIG. 6, as generated by the text generation platform” in response to being “presented to the user for approval in some embodiments”, where FIG. 6 establishes a sample category in the instructions (“in the voice of a pirate”); Goligorsky, ¶ [0047], [0169]; FIG. 6) wherein the sample category is determined based on user feedback regarding the third object material sample (The desired target value, which as outlined in FIG. 6, includes a sample category (the voice of a pirate) is determined based on the user response to presentation for approval (approval directly addresses whether the user instructions were followed).; Goligorsky, ¶ [0047], [0169]); performing human feedback reinforcement learning on the object material generation model based on the supervised fine-tuning samples (“The parameters of the ML model are updated based on a difference between the generated output value” as generated based on a corresponding input sample, “and the desired target value.”; Goligorsky, ¶ [0047]).
Regarding claim 14, Goligorsky discloses A non-transitory computer-readable storage medium configured with instructions executable by one or more processors to cause the one or more processors (The systems and methods described with reference to the “example e-commerce platform” as implemented through “example computing system 400” including “memory 404” which is a non-transitory computer readable medium as known in the art, which “may store instructions for execution by the processor 402, to the computing system 400 to carry out examples of the methods, functionalities, systems and modules disclosed herein.”; Goligorsky, ¶ [0069], [0074]) to perform the method of claim 11 (See mapping of claim 11; Goligorsky and Kobren).
Regarding claim 15, Goligorsky discloses An electronic device comprising: one or more processors; and one or more computer-readable memories coupled to the one or more processors and having instructions stored thereon that are executable by the one or more processors (The systems and methods described with reference to the “example e-commerce platform” as implemented through “example computing system 400” including “at least one processing unit, such as a processor 402, and at least one physical memory 404” where “memory 404 may store instructions for execution by the processor 402, to the computing system 400 to carry out examples of the methods, functionalities, systems and modules disclosed herein.”; Goligorsky, ¶ [0069], [0074]) to perform the method of claim 11 (See mapping of claim 11; Goligorsky and Kobren).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Chen (U.S. Pat. App. Pub. No. 2025/0077792) discloses systems and methods for generating training data (including training inputs and associated outputs) to fine-tune a pretrained machine learning model (which may be referred to herein as a base machine learning model or a pretrained machine learning model) using limited domain-specific data.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Sean E. Serraguard whose telephone number is (313)446-6627. The examiner can normally be reached 07:00-17:00 M-F.
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, Daniel C. Washburn can be reached at (571) 272-5551. 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.
/Sean E Serraguard/Primary Examiner, Art Unit 2657
Prosecution Insights