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 .
DETAILED ACTION
Claims 1-15, 17-19, and 21-22 are pending. Claims 1, 7, and 17 are independent.
Claims 2-6 and 21-22 depend from Claim 1.
Claims 8-15 depend from Claim 7.
Claims 18-29 depend from Claim 17.
Claims 16 and 20 are cancelled.
This Application was published as U.S. 2025/0217603.
Response to Amendment
Examiner thanks Applicant for response filed on 21 Jan 2026 which has been correspondingly accepted and considered in this office action. Claims 1-15, 17-19, and 21-22 are pending.
Response to Arguments
Applicant's arguments filed 21 Jan 2026 have been fully considered but they are not persuasive. Each argument of Applicant’s arguments will be addressed in turn.
With regards to objection to drawings:
Applicant has amended the drawings and specification. As a results, the objection to drawings has been withdrawn.
With regards to 35 USC § 103:
Applicant's arguments filed 21 Jan 2026 have been fully considered but they are not persuasive. Applicant arguments will be addressed in turn.
Claim 1:
Applicant argues that reference He et al. (US2024/0419941 hereinafter He), Gandhi et al. (US2024/0378399 hereinafter Gandhi), and Li et al. (US2022/0229832 hereinafter Li) cited in the 21 Oct 2025 office action does not teach the claimed limitations.
MPEP 2145 (I) Argument does not replace evidence where evidence is required states that an argument by the applicant is not evidence unless it is an admission, in which case, an examiner may use the admission in making a rejection.
Here, applicant arguments submitted 21 Jan 2026 pages 13-16 does not cite to any reference’s pages, paragraphs or figures, nor does the applicant cite to any of applicant’s paragraphs or figures. Thus, it is not clear if applicant’s arguments are the applicant’s personal opinion, admission, or what any reference actually teaches. Therefore, all applicants arguments provided will not be given substantial weight.
MPEP 2145 (IV) Arguing against references individually discourages applicants from “attacking references individually where the rejections are based on combinations of references. In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., Inc., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).”
Here, applicant provides arguments against the He, Gandhi, and Li references without taking into consideration all the references together, in particular what the base reference Bhattacharyya et al. (US2025/0119625 hereinafter Bhattacharyya) teaches. As will be discussed below, Bhattacharyya in combination with the other references cited teaches the claimed limitations, and each applicant argument will be addressed in turn.
Applicant argues (21 Jan 2026 arguments page 14) the following with respect to the He reference:
He fails to teach the claimed reason category class because He's fraud inference model determines a score or binary fraud outcome associated with a refund event, rather than categorizing a ''reason'' from LLM output into a reason category based on the natural language text of a transcript. He's use of coded refund reasons is event-driven and outcome-oriented, not a semantic classification of LLM output derived from transcript language.
MPEP 2145 (VI) arguing limitations which are not claimed states that “claims are interpreted in light of the specification, limitations from the specification are not read into the claims. In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993)”
Here, claim 1 does not have the limitation that “semantic classification of classification of LLM output derived from transcript language” because semantic is not mentioned in the claim or specification. Therefore, whether He teaches “semantic classification of classification of LLM output derived from transcript language” is not relevant.
On the other hand, 21 Oct 2025 non-final office action describes how Bhattacharyya teaches accessing a transcript, prompting an LLM from the transcript, and generating, by the LLM (21 Oct 2025 office action, Page 5) Furthermore, combining these limitations plus the reason generation (Bhattacharyya Fig 5 item 550) as well as the reason classifier, reason, and reason category class taught by He (21 Oct 2025 office action, Page 6) teaches all the claimed limitations described in the office action as well as “categorizing a ''reason'' from LLM output into a reason category based on the natural language text of a transcript.”
Applicant argues (21 Jan 2026 arguments pages 14-15) the following with respect to the Gandhi reference:
Gandhi fails to teach the claimed intent category class because Gandhi's intent classifier operates on a user query to drive program execution or DSL generation, not on LLM output generated from a transcript, and does not categorize intent as a semantic class derived from transcript natural language. Gandhi's intent determination is a control mechanism, not a post-LLM semantic classification as required by Claim 1.
Here, as previously discussed, semantic is not mentioned and arguments that Gandhi “does not categorize intent as a semantic class derived from transcript natural language” [or that] Gandhi's intent determination is a control mechanism, not a post-LLM semantic classification as required by Claim 1” are not relevant.
On the other hand, it is relevant that Bhattacharyya teaches the transcript where LLM output is generated (21 Oct 2025 office action, Page 5) Combining this with the intent classifier, intent, and intent category class (id, pages 6-7) teach the claimed limitations. Furthermore, Bhattacharya in view of Gandhi teaches an LLM output and categorizing intent derived from a natural language transcripts because Bhattacharyya teaches accessing a transcript and the natural language text of the transcript as described in 21 Oct 2025 office action, Page 5-9.
Applicant argues (21 Jan 2026 arguments page 15) the following with respect to the Li reference:
Li fails to teach the claimed reason category class because Li determines an action from user intent in a content-generation pipeline and does not disclose categorizing a ''reason'' from LLM output, let alone doing so based on the natural language text of a transcript using a neural-network-based classifier. Li conflates intent and action determination and does not disclose a standalone semantic reason classifier operating on LLM output.
Here, as previously discussed, semantic is not mentioned and arguments that Li “does not disclose a standalone semantic reason classifier operating on LLM output” are not relevant. Furthermore, Li was used to teach action classifier, action, and action category class (21 Oct 2025 office action, Pages 8-9) and reference to Li not teaching “reason category class” or “categorizing a “reason” from LLM output” in applicant arguments above are also not relevant. As discussed above, Gandhi teaches a reason and Bhattacharyya teaches the transcript and natural language text. Furthermore, Bhattacharyya in view of Li teaches the claimed limitations about the transcript, natural language text, intent classifier, intent, and intent category class as described in (21 Oct 2025 office action, Pages 5 and 8-9)
Thus, for all the reasons discussed above Bhattacharyya in view of He, Gandhi, and Li teaches all the limitations of claim 7
Claim 7:
Applicant argues on pages 15-16 of remarks received on 21 Jan 2026 the following:
Bhattacharyya's system does not train the ML models in the manner disclosed and claimed. For instance, the instant specification discloses that the system uses an LLM to label a content (such as a transcript's intent of the caller, a reason for the call, or action resulting from the call) and then using those labels to train a neural network to further classify the transcript. Even if Smith describes using an LLM to generate training data, combining Smith with Bhattacharyya would require repurposing Bhattacharyya's LLM-generated text representation from classifier input to training labels, which is contrary to Bhattacharyya's disclosed use of the text representation. In particular, doing so would change Bhattacharyya's principle of operation from using words in the text representation to classify video to using discrete labels for classifying video.
MPEP 2145 (IV) Arguing against references individually discourages applicants from “attacking references individually where the rejections are based on combinations of references. In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., Inc., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986).”
Here, applicant provides arguments that Bhattacharyya “does not train the ML models in the manner disclosed and claimed.” (21 Jan 2026 arguments pages 15-16) As was discussed in the 21 Oct 2025 non-final office action, Bhattacharyya does not teach the labelling limitation. Thus, it is improper to solely attack the Bhattacharyya reference not teaching training in the manner claimed, when the claimed limitations are taught by Bhattacharyya in view of Smith et al. (US2024/0160900 hereinafter Smith) as explained in the 21 Oct 2025 non-final office action pages 17-19 which shows how Bhattacharyya teaches training limitation, and Smith teaches the labelling limitation.
Applicant further argues that “combining Smith with Bhattacharyya would require repurposing Bhattacharyya's LLM-generated text representation from classifier input to training labels, which is contrary to Bhattacharyya's disclosed use of the text representation.” (21 Jan 2026 arguments pages 15-16)
Here, 21 Oct 2025 non-final office action page 16 states “[Bhattacharyya teaches “classifiers used are trained using a training dataset” (Par [0109]).” Furthermore, Bhattacharyya teaches that classifiers can be trained by annotation by a human annotator and that “labeled videos can then be used to train the persuasion strategy classifier” (Par [0111]) which shows that Bhattacharyya teaches using training labels, which is not contrary to Bhattacharya’s use of text representation. While Bhattacharyya uses training labels, it is not in the same manner as the claimed invention, which is why the Smith reference is used to teach the labelling limitation in the 21 Oct 2025 non-final office action pages 16-19.
Thus, for all the reasons discussed above Bhattacharyya in view of Smith teaches all the limitations of claim 7
Claim 17:
Applicant arguments to claim 17 are the same as claim 7.
MPEP 2145 (I) Argument does not replace evidence where evidence is required states that an argument by the applicant is not evidence unless it is an admission, in which case, an examiner may use the admission in making a rejection.
Here, claim 17 does not cite the Bhattacharyya reference. Arguments that Bhattacharyya does not teach the training limitation are irrelevant, and applicant needs to provide evidence that the Gandhi in view of Smith does not teach the claimed limitations as discussed in 21 Oct 2025 non-final office action pages 28-31. As such, there are no arguments to respond to.
In summary, for all the reasons stated above, applicant arguments are non-persuasive and all rejections are being maintained.
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.
Claims 1-2 are rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) hereinafter Bhattacharyya in view of He et al. (US 2024/0419941) hereinafter He, Gandhi et al. (US2024/0378399) hereinafter Gandhi, and Li et al. (US2022/0229832) hereinafter Li.
With regards to claim 1, Bhattacharyya teaches:
A system, comprising:
a processor programmed to execute one or more neural networks, [Fig 9 item 914]
the one or more neural networks comprising a classifier, and wherein the processor is programmed to: [Bhattacharyya Fig 5 teaches various classifiers (emotion (534) persuasion (538) and topic (542)) classifiers as well as reason generator (550) and action generator (546) where classifiers are neural networks since Bhattacharyya teaches “various forms of machine learning models, such as a classifier(s)” (Par [0100]) and in some embodiments the “machine learning model is a Large Language Model (LLM)” (Par [0085]) which teaches that classifiers are LLMs and that “an LLM is a deep neural network” (Par [0085])]
access a transcript; [Fig 5 item 512 Par [0122]]
prompt a Large Language Model (LLM) from the transcript based on the natural language text of the transcript and the prompt; [Bhattacharyya Fig 5 teaches “transcript 512 can be concatenated into the prompt 530. The prompt 530 is provided to a LLM” (Par [0122])]
generate, by the LLM, responsive to the prompt, an LLM output [Bhattacharyya Fig 5 teaches “prompt 530 is provided to a LLM to generate a text representation 532” (Par [0122])]
With regards to claim 1, Bhattacharyya fails to teach:
a reason classifier trained to classify the reason into one or more reason categories
to identify a reason, and/or
comprising the reason, and/or
determine, by the one or more neural networks, a reason category class that categorizes the reason from the LLM output into a reason category based on the natural language text in the transcript, and/or
With regards to claim 1, He teaches:
a reason classifier trained to classify the reason into one or more reason categories [He Fig 3 teaches “LLM 310 to classify a reason for a refund event associated with the order, one or more coded reasons for the refund event provided by the picker, any other useful information associated with the order, or some combination thereof” (Par [0076])]
to identify a reason, and/or [He Fig 3 teaches identifying “a reason for a refund event associated with the order, one or more coded reasons for the refund event provided by the picker, any other useful information associated with the order, or some combination thereof” (Par [0076])]
comprising the reason, and/or [He Fig 3 Par [0076] teaches LLM output responses (315, 320, and 325) that comprise the reason]
determine, by the one or more neural networks, a reason category class that categorizes the reason from the LLM output into a reason category based on the natural language text in the transcript, and/or [He Par [0037] Table 2 teaches LLM classifies into reason categories “<refund>, <replacement>, <found>.”
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system with the reason generator of Bhattacharyya with the LLM system with the reason classifier as taught by He. The motivation to combine the teachings of Bhattacharyya with He is because “prompt may include a transcript of a chat conversation between a customer and a picker during a picking (i.e., shopping) process” (Par [0035]) which increases the capabilities of the invention of Bhattacharyya to be used in commercial settings with customers and customer service]
With regards to claim 1, Bhattacharyya in view of He fails to teach:
an intent classifier trained to classify an intent into one or more intent categories,
to identify an intent
comprising the intent;
determine, by the one or more neural networks, an intent category class that categorizes the intent from the LLM output into an intent category based on the natural language text in the transcript,
With regards to claim 1, Gandhi teaches:
an intent classifier trained to classify an intent into one or more intent categories, [Gandhi Fig 2 teaches intent classifier (218) and “a classifier model is used to analyze the natural language query to determine the user intent. The classifier model is implemented using a natural language processing (NLP) model that is trained to analyze a textual input and to classify the query into one of a set of categories of request” (Par [0032])]
to identify an intent [Gandhi Fig 2 teaches intent classifier (218) and “a classifier model is used to analyze the natural language query to determine the user intent” (Par [0032])]
comprising the intent; [Gandhi Fig 2 teaches “LLM 222 receives the prompt from the prompt construction unit 220, analyzes the prompt, and outputs DSL program code to implement the user intent expressed in the natural language query” (Par [0036])]
determine, by the one or more neural networks, an intent category class that categorizes the intent from the LLM output into an intent category based on the natural language text in the transcript, [Gandhi Fig 2 teaches intent classifier (218) and the “classifier model is implemented using a natural language processing (NLP) model that is trained to analyze a textual input and to classify the query into one of a set of categories of request” (Par [0032])]
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system of Bhattacharyya and He with the LLM system with the intent classifier as taught by Gandhi. The motivation to combine the teachings of Bhattacharyya and He with Gandhi is because “approach enables the user to communicate their intent to the application in language which is familiar to the user without requiring the user to navigate through the various controls of the user interface of the application” (Par [0020]) which increases the capabilities of the invention of Bhattacharyya and He to be used in settings with customers and customer service]
With regards to claim 1, Bhattacharyya in view of He and Gandhi fails to teach:
an action classifier trained to classify the action into one or more action categories,
to identify an action
comprising the action; and
determine, by the one or more neural networks, an action category class that categorizes the action from the LLM output into an action category based on the natural language text in the transcript.
With regards to claim 1, Li teaches:
an action classifier trained to classify the action into one or more action categories, [Li teaches “The action category may be classified and the user query, action, and/or action category may be provided to a prompt design component (e.g., prompt design component 115)” (Par [0033])]
to identify an action [Li Fig 3 teaches “step 310 the natural language action is determined from an intent of the user query” (Par [0033])]
comprising the action; and [Li Fig 3 teaches “At step 320, the prompt is provided to a natural language generation model (e.g., natural language generation model 125), such as GPT-3. The natural language generation model performs modelling, and at step 325 the output from the model is received” (Par [0033])]
determine, by the one or more neural networks, an action category class that categorizes the action from the LLM output into an action category based on the natural language text in the transcript. [Li teaches “The action category may be classified and the user query, action, and/or action category may be provided to a prompt design component (e.g., prompt design component 115)” (Par [0033]) while Li teaches classifying the action before the prompt and LLM output in Fig 3, it is “understood and appreciated that the methods are not limited by the order of acts, as some acts may, in accordance therewith, occur in a different order and/or concurrently with other acts from that shown and described herein” (Par [0057]) such that Li teaches an action classifier trained to classify the action into one more action categories.
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system and action generator of Bhattacharyya, He, and Gandhi with the LLM system with the action classifier as taught by Li. The motivation to combine the teachings of Bhattacharyya, He, and Gandhi with Li is because the “way to reduce the dependency on user text inputs and provide a way to generate entire texts based on a few inputs from users” (Li Par [0016]) which increases the capabilities of the invention of Bhattacharyya, He, and Gandhi to be used in settings with customers and customer service]
With regards to claim 2, Bhattacharyya, He, Gandhi, and Li teaches:
All the limitations of claim 1
wherein the processor is further programmed to: identify a potential issue that is to be mitigated based on the intent category class, the reason category class, and/or the action category class. [Gandhi teaches “intent classifier unit 218 determines whether the natural language query entered by the user is within scope of the services provided by the application services platform 210” … “determine the user intent” and “classifier model categorizes any natural language queries that do not fall into one of the other categories as an unsupported action that cannot be performed by the application services platform 210” (Par [0032])
It would be obvious to one of ordinary skill in the art to combine the intent category class as taught by Gandhi and combine it with the reason category class as taught by He and the action category class as taught by Li. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, and Li together is because Gandhi teaches “specific categories associated with each application may vary. Furthermore, the classification model can be updated to support additional categories as support for natural language commands are provide” (Par [0032]) which increases the capabilities of the invention of Bhattacharyya, He, and Gandhi to be used in settings with different category classes of data for increased customer service]
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of He et al. (US 2024/0419941), Gandhi et al. (US2024/0378399), and Li et al. (US2022/0229832) and in further view of Khanwalkar et al. (US12393620) hereinafter Khanwalkar.
With regards to claim 3, Bhattacharyya, He, Gandhi, and Li teaches:
All the limitations of claim 1
wherein the LLM is prompted to identify the intent from the transcript and the one or more neural networks comprises an intent classifier that determines the intent category class, and wherein the processor is further programmed to: determine a probability that the intent category class is correct; [Bhattacharyya teaches “classifiers may also output probabilities or likelihoods associated with the predicted classification” (Par [0109]) where reduction of likelihood is a determination of probability]
initiate a discovery process to retrain the intent classifier to identify the category or the sub-category for the intent. [He teaches a discovery process that “receives a response to the prompt from the model serving system 150” and then “use the response to retrain (or more generally “update”) an algorithm executed at the online concierge system 140 that generates the prompt for the LLM” (Par [0038]) where “response can be utilized for both inference and training of the one or more computer models” (Par [0040])]
With regards to claim 3, Bhattacharyya, He, Gandhi, and Li fails to teach:
determine that the probability is below a threshold probability value such that the intent category class is not able to classify the intent into a category or sub-category; and
With regards to claim 3, Khanwalkar teaches:
determine that the probability is below a threshold probability value such that the intent category class is not able to classify the intent into a category or sub-category; and [Khanwalkar teaches “user input query is only passed to the generative question answering engine if the question has a question intent probability above a threshold” (Col 14 lines 15-17)]
It would be obvious to one of ordinary skill in the art to combine the intent category class as taught by Bhattacharyya, He, Gandhi, and Li and combine it with the threshold probability as taught by Khanwalkar. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, and Li together with Khanwalkar is because Khanwalkar teaches “detect a fraudulent behavior associated with an order placed with the online concierge system” (Par [0039]) which increases the capabilities of the invention of Bhattacharyya, He, Gandhi, and Li to be used in settings with customer fraud which increases customer service]
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of He et al. (US 2024/0419941), Gandhi et al. (US2024/0378399), and Li et al. (US2022/0229832) and Khanwalkar et al. (US12393620) in further view of Vakili et al. (US2025/0248667) hereinafter Vakili.
With regards to claim 4, Bhattacharyya, He, Gandhi, Li, and Khanwalkar teaches:
All the limitations of claim 3
wherein the processor is further programmed to: generate a text embedding based on text of the intent; [Gandhi teaches “calculates the embeddings embedding E(q) for the natural language query q input by the user” (Par [0055])]
With regards to claim 4, Bhattacharyya, He, Gandhi, Li, and Khanwalkar fails to teach:
cluster, into a first cluster, the text embedding with other text embeddings of other intents identified from other transcripts in a training dataset, wherein the text embedding is clustered into the first cluster based on a similarity between the text embedding and the other text embeddings, wherein the first cluster represents a sub-category for the intent; and
train the intent classifier based on the first cluster.
With regards to claim 4, Vakili teaches:
cluster, into a first cluster, the text embedding with other text embeddings of other intents identified from other transcripts in a training dataset, wherein the text embedding is clustered into the first cluster based on a similarity between the text embedding and the other text embeddings, wherein the first cluster represents a sub-category for the intent; and [Vakili Fig 8 teaches “classifying the LLM intents may comprise one or more steps or operations for analyzing a plurality of conversational or user-utterance data to create embeddings comprising clusters of semantically similar sentences. In accordance with certain embodiments, the clustered groupings may each comprise an intent, each of which may be classified according to the plurality of agenda prompts” (Par [0091])]
train the intent classifier based on the first cluster. [Vakili Fig 8 teaches “routine 800 may comprise one or more steps or operations for training the LLM according to the configurations from steps 802-810 (Step 814).” (Par [0091])
It would be obvious to one of ordinary skill in the art to combine the intent category class as taught by Bhattacharyya, He, Gandhi, Li, and Khanwalkar and combine it with the cluster as taught by Vakili. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, Li, and Khanwalkar with Vakili is because Vakili teaches “configuring a large language model within the voice-based system for management” (Par [0090]) which increases the capabilities of the invention of Bhattacharyya, He, Gandhi, and Li to be used in better communicate with customers with different problems which increases customer service]
Claims 5-6 are rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of He et al. (US 2024/0419941), Gandhi et al. (US2024/0378399), and Li et al. (US2022/0229832) and Khanwalkar et al. (US12393620) in further view of Vakili et al. (US2025/0248667) in further view of Smith et al.(US2024/0160900) hereinafter Smith.
With regards to claim 5, Bhattacharyya, He, Gandhi, Li, Khanwalkar and Vakili teaches:
All the limitations of claim 4
With regards to claim 5, Bhattacharyya, He, Gandhi, Li, Khanwalkar and Vakili fails to teach:
wherein the processor is further programmed to: prompt the LLM to generate a sub-category name for the first cluster;
label the transcripts in the first cluster with the sub-category name; and
retrain the intent classifier based on the labeled transcripts.
With regards to claim 5, Smith teaches:
wherein the processor is further programmed to: prompt the LLM to generate a sub-category name for the first cluster; [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) where “associated datapoints, forming one or more groups or clusters of datapoints” (Par [0048]) and Fig 1(a) teaches using cluster view to “to explore the clustered data and evaluate possible labeling functions (LFs); a subset of these possible LFs are created” (Par [0083])]
label the transcripts in the first cluster with the sub-category name; and [Smith Fig 1(a) teaches creating labels with the labeling function (Par [0082-83] see Poms et al. (US2023/0419121)]
retrain the intent classifier based on the labeled transcripts. [Smith teaches “LFs are used to train a model; that model is analyzed for errors; and the errors are corrected by using Cluster View to explore for additional data to label.” (Par [0083])
It would be obvious to one of ordinary skill in the art to combine the LLM system with the intent category classifier as taught by Bhattacharyya, He, Gandhi, Li, Khanwalkar, and Vakili and combine it with the labeling function as taught by Smith. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, Li, Khanwalkar, and Vakili with Smith is because Smith teaches “embodiment can be used to implement a form of programmatic labeling to accelerate the development of classifiers and other forms of models. The disclosed methodology is particularly helpful in generating labels or annotations for unstructured data” (Par [0009]) which increases the capabilities of the invention of Bhattacharyya, He, Gandhi, Li, Khanwalkar, and Vakili to better train its classifiers]
With regards to claim 6, Bhattacharyya, He, Gandhi, Li, Khanwalkar, Vakili, and Smith teaches:
All the limitations of claim 5
wherein to re-train the intent classifier, the processor is further programmed to: initiate a review of the sub-category name for use as a label, wherein the intent classifier is retrained after the review of the sub-category name for use as the label. [Smith teaches a “automated or programmatic labeling process” (Par [0080]) which is a “suitable approach” to ensure efficiently generation of labels, and “train a model” (Par [0083]) such as the intent classifier]
Claims 7-8 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of Smith et al.(US2024/0160900).
With regards to claim 7, Bhattacharyya teaches:
A method for training neural networks , comprising: accessing, by a processor, [Fig 9 item 914] a plurality of contents each comprising natural language text; [Fig 5 item 512 Par [0122] were a transcript is a plurality of contents each comprising natural language text]
for each content: prompting, by the processor, a Large Language Model (LLM) to identify a concept from the content based on the prompt and the natural language text; [Bhattacharyya Fig 5 teaches “transcript 512 can be concatenated into the prompt 530. The prompt 530 is provided to a LLM to generate a text representation 532” (Par [0122]) where a text representation identifies a concept]
generating, by the processor executing the LLM, responsive to the prompt, an LLM output comprising words or phrases that specify the concept identified from the content; and [Bhattacharyya Fig 5 item 532 text representation (Par [0122])]
training, by the processor, one or more neural networks to classify input content based on the LLM-mediated training dataset. [Bhattacharyya teaches “classifiers used are trained using a training dataset” (Par [0109]) where “LLM can be used to identify any of such visual insights 536, 540, 544, 548, and 552, among others” (Par [0122]) where Fig 5 shows insights are produced by classifiers and “LLM is a deep neural network” (Par [0085])]
With regards to claim 7, Bhattacharyya fails to teach:
generating, by the processor, a text embedding of the concept based on the words or phrases that specify the concept, wherein an association is stored between each text embedding and the content from which the concept was identified;
generating, by the processor, a plurality of clusters of text embeddings, wherein each cluster comprises at least two text embeddings based on their similarity to one another;
prompting, by the processor, the LLM to generate a label for each cluster of text embeddings;
labeling, by the processor, each content that corresponds to a cluster based on the label generated by the LLM for the cluster to generate an LLM-mediated training dataset; and
With regards to claim 7, Smith teaches:
generating, by the processor, a text embedding [Smith teaches “techniques including one or more of text embeddings” (Par [0016])] of the concept based on the words or phrases that specify the concept, wherein an association is stored between each text embedding and the content from which the concept was identified; [Smith teaches various types of embeddings such as deep learning embeddings that “require the accuracy or adaptability of deep learning to unseen word” (Par [0016])]
generating, by the processor, a plurality of clusters of text embeddings, wherein each cluster comprises at least two text embeddings based on their similarity to one another; [Smith teaches “Attempt to group (cluster) the datapoints in a dataset using techniques that assign datapoints to the same group if they share similarities ... The degree of similarity can be measured by the similarity between two embeddings” (Par [0022])]
prompting, by the processor, the LLM to generate a label for each cluster of text embeddings; [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) where “associated datapoints, forming one or more groups or clusters of datapoints” (Par [0048])]
labeling, by the processor, each content that corresponds to a cluster based on the label generated by the LLM for the cluster to generate an LLM-mediated training dataset; and [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047])
It would be obvious to one of ordinary skill in the art to combine the LLM system as taught by Bhattacharyya and combine it with the system of generating labels of clustering or language model technique as taught by Smith. The motivation to combine the teachings of Bhattacharyya with Smith is because Smith invention teaches “embodiment can be used to implement a form of programmatic labeling to accelerate the development of classifiers and other forms of models. The disclosed methodology is particularly helpful in generating labels or annotations for unstructured data” (Par [0009]) which increases the capabilities of the invention of Bhattacharyya to better train its classifiers]
With regards to claim 8, Bhattacharyya in view of Smith teaches:
All the limitations of claim 7
wherein the content comprises a transcript of multi-media, a communication, a document, or an article. [Bhattacharyya Fig 5 item 512 Par [0122]]
With regards to claim 12, Bhattacharyya in view of Smith teaches:
All the limitations of claim 7
wherein the label corresponds to a sub-category name, the method further comprising: clustering at least some the plurality of clusters into one or more second clusters, wherein each second cluster of the one or more second clusters comprises at least two clusters of text embeddings that are similar to one another beyond a second threshold value and wherein each second cluster represents a category for the sub-categories in the one or more second clusters; [Smith teaches classification of email as HAM or SPAM. (Par [0041-42]) In this example, 10 clusters are created, which have similar embeddings, and clusters were based on “majority or threshold value” (Par [0042])]
for each second cluster: prompt the LLM to generate a category name to be used as a second label for training the one or more neural networks; [Smith teaches “submit a query to a pre-trained foundation or large language model (LLM)” (Par [0045])]
generate the category name based on the words or phrases of the corresponding sub-category names; and identify the corresponding content associated with the second cluster name and label each of the corresponding content based on the category name to generate the LLM- mediated training dataset with both the category name and the label. [Smith teaches creating “weakly supervised labels that may be used downstream to generate annotated training data for a model” (Par [0044]) and to generate a label for individual data points” (Par [0046])]
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of Smith et al.(US2024/0160900) in view of Gandhi et al. (US2024/0378399)
With regards to claim 9, Bhattacharyya in view of Smith teaches:
All the limitations of claim 7
wherein the content comprises a transcript of a call, [Bhattacharyya teaches video transcription 512 (Fig 5 Par [0122]) where a video transcription can be for a call such as video conference for a call center (see Shaffer et al. (US2024/0364770))]
With regards to claim 9, Bhattacharyya in view of Smith fails to teach:
the concept comprises at least an intent of a caller for making the call, and
the one or more neural networks comprise an intent classifier trained to categorize the intent.
With regards to claim 9, Gandhi teaches:
the concept comprises at least an intent of a caller for making the call, and [Gandhi Fig 2 teaches intent classifier (218) and “a classifier model is used to analyze the natural language query to determine the user intent” (Par [0032])]
the one or more neural networks comprise an intent classifier trained to categorize the intent. [Gandhi Fig 2 teaches intent classifier (218) and “a classifier model is used to analyze the natural language query to determine the user intent. The classifier model is implemented using a natural language processing (NLP) model that is trained to analyze a textual input and to classify the query into one of a set of categories of request.” (Par [0032])
It would be obvious to one of ordinary skill in the art to combine the LLM system as taught by Bhattacharyya and Smith and combine it with the LLM system using intent classifiers as taught by Gandhi. The motivation to combine the teachings of Bhattacharyya and Smith with Gandhi is because Gandhi teaches “approach enables the user to communicate their intent to the application in language which is familiar to the user without requiring the user to navigate through the various controls of the user interface of the application” (Par [0020]) which increases the capabilities of the invention of Bhattacharyya and Smith to be used in settings with customers and customer service]
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of Smith et al.(US2024/0160900) and Gandhi et al. (US2024/0378399) in further view of He et al. (US2024/0419441)
With regards to claim 10, Bhattacharyya in view of Smith and Gandhi teaches:
All the limitations of claim 9
With regards to claim 10, Bhattacharyya in view of Smith and Gandhi fails to teach:
wherein the concept further comprises a reason for the call, and
the one or more neural networks comprise a reason classifier trained to categorize the reason, the method further comprising:
determining, by the reason classifier, a category for the reason.
With regards to claim 10, He teaches:
wherein the concept further comprises a reason for the call, and [He Fig 3 Par [0076] teaches LLM output responses (315, 320, and 325) that comprise the reason]
the one or more neural networks comprise a reason classifier trained to categorize the reason, the method further comprising: [He Fig 3 teaches “LLM 310 to classify a reason for a refund event associated with the order, one or more coded reasons for the refund event provided by the picker, any other useful information associated with the order, or some combination thereof” (Par [0076])]
determining, by the reason classifier, a category for the reason. [He Par [0037] Table 2 teaches LLM classifies into reason categories “<refund>, <replacement>, <found>.”
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system with the reason generator of Bhattacharyya in view of Smith and Gandhi with the LLM system with the reason classifier as taught by He. The motivation to combine the teachings of Bhattacharyya in view of Smith and Gandhi with He is because “prompt may include a transcript of a chat conversation between a customer and a picker during a picking (i.e., shopping) process” (Par [0035]) which increases the capabilities of the invention of Bhattacharyya in view of Smith and Gandhi to be used in commercial settings with customers and customer service]
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of Smith et al.(US2024/0160900), Gandhi et al. (US2024/0378399), and He et al. (US2024/0419441) in further view of Li et al (US2022/0229832)
With regards to claim 11, Bhattacharyya in view of Smith and Gandhi teaches:
All the limitations of claim 9
With regards to claim 11, Bhattacharyya in view of Smith and Gandhi fails to teach:
wherein the concept further comprises an action taken in response to the call, and the one or more neural networks comprise an action classifier trained to categorize the action, the method further comprising:
determining, by the action classifier, a category for the action.
With regards to claim 11, Li teaches:
wherein the concept further comprises an action taken in response to the call, and [Li Fig 3 teaches “At step 320, the prompt is provided to a natural language generation model (e.g., natural language generation model 125), such as GPT-3. The natural language generation model performs modelling, and at step 325 the output from the model is received” (Par [0033])] the one or more neural networks comprise an action classifier trained to categorize the action, the method further comprising: [Li teaches “The action category may be classified and the user query, action, and/or action category may be provided to a prompt design component (e.g., prompt design component 115)” (Par [0033])]
determining, by the action classifier, a category for the action. [Li teaches “The action category may be classified and the user query, action, and/or action category may be provided to a prompt design component (e.g., prompt design component 115)” (Par [0033]) while Li teaches classifying the action before the prompt and LLM output in Fig 3, it is “understood and appreciated that the methods are not limited by the order of acts, as some acts may, in accordance therewith, occur in a different order and/or concurrently with other acts from that shown and described herein” (Par [0057]) such that Li teaches an action classifier trained to classify the action into one more action categories.
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system and action generator of Bhattacharyya in view of Smith and Gandhi with the LLM system with the action classifier as taught by Li. The motivation to combine the teachings of Bhattacharyya in view of Smith and Gandhi with Li is because the “way to reduce the dependency on user text inputs and provide a way to generate entire texts based on a few inputs from users” (Li Par [0016]) which increases the capabilities of the invention of Bhattacharyya in view of Smith and Gandhi to be used in settings with customers and customer service]
Claims 13-15 is rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of Smith et al.(US2024/0160900) and He et al. (US2024/0419441).
With regards to claim 13, Bhattacharyya in view of Smith teaches:
All the limitations of claim 7
further comprising:
during an operational phase, determining a probability that a classification for the concept is correct; [Bhattacharyya teaches displaying “probability or likelihood” (Par [00117])]
determining that the probability is below a threshold probability value; and [Smith teaches “for each new datapoint, using the new datapoint as input along with each prompt and threshold and determining the cluster or group” (Par [0050])]
With regards to claim 13, Bhattacharyya in view of Smith fails to teach:
initiating a discovery process to retrain the one or more neural networks.
With regards to claim 13, He teaches:
initiating a discovery process to retrain the one or more neural networks.
[He teaches a discovery process that “receives a response to the prompt from the model serving system 150” and then “use the response to retrain (or more generally “update”) an algorithm executed at the online concierge system 140 that generates the prompt for the LLM” (Par [0038]) where “response can be utilized for both inference and training of the one or more computer models.” (Par [0040])
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system of Bhattacharyya and Smith with the LLM system with the reason classifier as taught by He. The motivation to combine the teachings of Bhattacharyya and Smith with He is because “prompt may include a transcript of a chat conversation between a customer and a picker during a picking (i.e., shopping) process” (Par [0035]) which increases the capabilities of the invention of Bhattacharyya and Smith to be used in commercial settings with customers and customer service]
With regards to claim 14, Bhattacharyya in view of Smith and He teaches:
All the limitations of claim 13
further comprising: generating a text embedding based on text of the concept; and [Smith teaches “techniques including one or more of text embeddings” and various types of embeddings such as deep learning embeddings that “require the accuracy or adaptability of deep learning to unseen word” (Par [0016])]
clustering, into a first cluster, the text embedding with other text embeddings of other concepts identified from other transcripts in a training dataset, wherein the text embedding is clustered into the first cluster based on a similarity between the text embedding and the other text embeddings, wherein the first cluster is used to determine a new category for the concept. [Smith teaches “Attempt to group (cluster) the datapoints in a dataset using techniques that assign datapoints to the same group if they share similarities ... The degree of similarity can be measured by the similarity between two embeddings” (Par [0022]) and determine if the data “is in the cluster or not in the cluster. The threshold value can be set as 0.5 for this task as it is a binary classification problem” (Par [0043])]
With regards to claim 15, Bhattacharyya in view of Smith and He teaches:
All the limitations of claim 14
further comprising: prompting the LLM to generate a sub-category name for the first cluster; [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) which is represented by a “unique identifier” (Par [0050])]
labeling the transcripts in the first cluster with the sub-category name; and [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) which is represented by a “unique identifier” (Par [0050])]
re-training the one or more neural networks based on the labeled transcripts. [He teaches a process that “receives a response to the prompt from the model serving system 150” and then “use the response to retrain (or more generally “update”) an algorithm executed at the online concierge system 140 that generates the prompt for the LLM” (Par [0038]) where “response can be utilized for both inference and training of the one or more computer models.” (Par [0040])
Claims 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over Gandhi et al. (US2024/0378399) in view of Smith et al.(US2024/0160900)
With regards to claim 17, Gandhi teaches:
A non-transitory computer readable medium storing instructions [Gandhi Fig 9 Par [0094]] for training an intent classifier [Fig 2 item 218] comprising a neural network [Gandhi teaches classifier is “implemented using a natural language processing (NLP) mode;” (Par [0032]) where NLP can perform text summarizations using an LLM or deep neural network (see Bhattacharyya (US2025/0119625) Par [0085])]
to categorize an intent of a caller from a transcript of a call, the instructions, when executed by a processor, programs the processor to: [Gandhi Fig 2 teaches “LLM 222 receives the prompt from the prompt construction unit 220, analyzes the prompt, and outputs DSL program code to implement the user intent expressed in the natural language query” (Par [0036]) where intent is based on document context (Fig 2) where document can be transcript for a call (see Bhattacharyya teaches video transcription 512 (Fig 5 Par [0122]) where a video transcription can be for a call such as video conference for a call center (see Shaffer et al. (US2024/0364770) Par [0113]))]
access a plurality of transcripts; [Gandhi Fig 2 teaches intent classifier access a document where a document can be a plurality of transcripts (see Raghavan et al. (US 8761373) Fig 5 where document contains transcripts (515) and transcript tabs (500, 505, 510))]
for each transcript from among a plurality of transcripts: prompt a Large Language Model (LLM) to identify an intent of a caller whose speech is transcribed in the transcript; [Gandhi Fig 2 teaches intent classifier (218) provides info for prompt (220) for an LLM (222) to identify intent of the user]
generate, by the LLM based on the prompt and text of the transcript, an LLM output comprising one or more words or phrases that specify the intent; [Gandhi Fig 2 teaches “LLM 222 receives the prompt from the prompt construction unit 220, analyzes the prompt, and outputs DSL program code to implement the user intent expressed in the natural language query” (Par [0036])]
With regards to claim 17, Gandhi fails to teach:
generate a text embedding of the intent based on the one or more words or phrases;
generate one or more clusters of text embeddings, wherein each cluster of text embeddings comprises at least two text embeddings that are similar to one another beyond a threshold value and wherein each text embedding corresponds to the intent of a corresponding transcript;
for each cluster of text embeddings:
prompt the LLM to generate a cluster name to be used as a label for training the intent classifier;
generate the cluster name based on the words or phrases of the corresponding intents associated with the text embeddings in the cluster;
identify the corresponding transcript associated with the text embedding and label the corresponding transcript based on the cluster name to generate a labeled set of transcripts; and
train the intent classifier based on the labeled set of transcripts to classify an input transcript into a class corresponding to the cluster name.
With regards to claim 17, Smith teaches:
generate a text embedding of the intent based on the one or more words or phrases; [Smith teaches “techniques including one or more of text embeddings” and various types of embeddings such as deep learning embeddings that “require the accuracy or adaptability of deep learning to unseen word” (Par [0016])]
generate one or more clusters of text embeddings, wherein each cluster of text embeddings comprises at least two text embeddings that are similar to one another beyond a threshold value and wherein each text embedding corresponds to the intent of a corresponding transcript; [Smith teaches “Attempt to group (cluster) the datapoints in a dataset using techniques that assign datapoints to the same group if they share similarities ... The degree of similarity can be measured by the similarity between two embeddings” (Par [0022]) and determine if the data “is in the cluster or not in the cluster. The threshold value can be set as 0.5 for this task as it is a binary classification problem” (Par [0043])]
for each cluster of text embeddings: prompt the LLM to generate a cluster name to be used as a label for training the intent classifier; [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) which is represented by a “unique identifier” (Par [0050])]
generate the cluster name based on the words or phrases of the corresponding intents associated with the text embeddings in the cluster; [Smith teaches “unique identifier” (Par [0050]) which is based on the cluster where the cluster is based on the text embeddings and corresponding intents]
identify the corresponding transcript associated with the text embedding and label the corresponding transcript based on the cluster name to generate a labeled set of transcripts; and [Smith teaches creating embeddings for each corresponding datapoint or transcript, (Par [0017]) clustering the datapoints (Par [0022]) and creating labels and naming the cluster by “an ID or attribute may be a randomly generated string of numbers/characters” (Par [0029]) to generate a labeled set of transcripts]
train the intent classifier based on the labeled set of transcripts to classify an input transcript into a class corresponding to the cluster name. [Smith teaches “labels for a set of data used to train a machine learning model” (Par [0016]) or intent classifier.
It would be obvious to one of ordinary skill in the art to combine the LLM system and intent classifier as taught by Gandhi and combine it with the system of generating labels of clustering or language model technique as taught by Smith. The motivation to combine the teachings of Gandhi with Smith is because Smith invention teaches “embodiment can be used to implement a form of programmatic labeling to accelerate the development of classifiers and other forms of models. The disclosed methodology is particularly helpful in generating labels or annotations for unstructured data” (Par [0009]) which increases the capabilities of the invention of to better train its classifiers]
With regards to claim 18, Gandhi in view of Smith teaches:
All the limitations of claim 17
wherein the cluster name corresponds to a sub-category name, [Smith teaches “prompt for the large language model, wherein a system creates corresponding prompts submitted to the language model, and the output of the large language model is a prediction (suggested label) for one or more of the datapoints in the dataset” (Par [0047]) which is represented by a “unique identifier” (Par [0050])]
the one or more clusters of text embeddings comprises a plurality of clusters of text embeddings, [Smith teaches “Attempt to group (cluster) the datapoints in a dataset using techniques that assign datapoints to the same group if they share similarities ... The degree of similarity can be measured by the similarity between two embeddings” (Par [0022])]
and wherein the instructions, when executed by the processor, further programs the processor to: cluster at least some the plurality of clusters into one or more second clusters, wherein each second cluster of the one or more second clusters comprises at least two clusters of text embeddings that are similar to one another beyond a second threshold value and wherein each second cluster represents a category for a corresponding intent; [Smith teaches classification of email as HAM or SPAM. (Par [0041-42]) In this example, 10 clusters are created, which have similar embeddings, and clusters were based on “majority or threshold value” (Par [0042])]
for each second cluster: prompt the LLM to generate a category name to be used as a second label for training the intent classifier; [Smith teaches “submit a query to a pre-trained foundation or large language model (LLM)” (Par [0045]) or intent classifier]
generate the category name based on the words or phrases of the corresponding sub-category names; and identify the corresponding transcripts associated with the second cluster name and label each of the corresponding transcripts based on the category name to generate the labeled set of transcripts with both the category name and the sub-category name, [Smith teaches creating “weakly supervised labels that may be used downstream to generate annotated training data for a model” (Par [0044]) and to generate a label for individual data points” (Par [0046])]
wherein the intent classifier is trained based on the labeled set of transcripts with both the category name and the sub-category name to classify an input transcript in a category class and a sub-category class. [Gandhi Fig 2 teaches intent classifier (218) and “a classifier model is used to analyze the natural language query to determine the user intent. The classifier model is implemented using a natural language processing (NLP) model that is trained to analyze a textual input and to classify the query into one of a set of categories of request.” (Par [0032])
Claim 19 is rejected under 35 U.S.C. 103 as being unpatentable over Gandhi et al. (US2024/0378399) in view of Smith et al.(US2024/0160900), Bhattacharyya et al. (US2025/0119625), He et al. (US 2024/0419941), and Li et al. (US2022/0229832)
With regards to claim 19, Gandhi in view of Smith teaches:
All the limitations of claim 17
wherein the instructions, when executed, further programs the processor to: for each transcript from among the plurality of transcripts: [Gandhi Fig 2 teaches intent classifier access a document where a document can be a plurality of transcripts (see Raghavan et al. (US 8761373) Fig 5 where document contains transcripts (515) and transcript tabs (500, 505, 510))]
With regards to claim 19, Gandhi in view of Smith fails to teach:
prompt the LLM to identify a reason for the call and an action,
responsive to the call, that was taken, wherein the LLM output further comprises one or more second words or phrases that specify the reason and one or more third words or phrases that specify the action;
generate a second text embedding of the reason based on the one or more second words or phrases and a third text embedding of the action based on the one or more third words or phrases;
With regards to claim 19, Bhattacharyya teaches:
prompt the LLM to identify a reason for the call and an action, [Bhattacharyya Fig 5 teaches “ prompt 530 is provided to a LLM to generate a text representation 532 that represents the text associated with the video” (Par [0122]) where text is input to identify the reason generator (550) and action generator (546)]
responsive to the call, that was taken, wherein the LLM output further comprises one or more second words or phrases that specify the reason and one or more third words or phrases that specify the action; [Bhattacharyya Fig 5 teaches LLM output results in reason (552) and action (548)]
generate a second text embedding of the reason based on the one or more second words or phrases and a third text embedding of the action based on the one or more third words or phrases; [Smith teaches various types of embeddings such as deep learning embeddings that “require the accuracy or adaptability of deep learning to unseen word” (Par [0016])
It would be obvious to one of ordinary skill in the art to combine the LLM system and intent classifier as taught by Gandhi and Smith and combine it with the LLM with reason and action generators as taught by Bhattacharyya. The motivation to combine the teachings of Gandhi and Smith with Bhattacharyya is because Bhattacharyya teaches “generate video insights, a text representation that represents the video using a natural language description is generated using a large language model” (Par [0004]) which increases the capabilities of the invention of Gandhi and Smith to better its classifiers on video data]
With regards to claim 19, Gandhi in view of Smith and Bhattacharyya fails to teach:
train a reason classifier based on the second text embedding; and
With regards to claim 19, He teaches:
train a reason classifier based on the second text embedding; and [He Fig 3 teaches “LLM 310 to classify a reason for a refund event associated with the order, one or more coded reasons for the refund event provided by the picker, any other useful information associated with the order, or some combination thereof” (Par [0076])
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system with the reason generator of Gandhi in view of Smith and Bhattacharyya with the LLM system with the reason classifier as taught by He. The motivation to combine the teachings of Gandhi in view of Smith and Bhattacharyya with He is because “prompt may include a transcript of a chat conversation between a customer and a picker during a picking (i.e., shopping) process” (Par [0035]) which increases the capabilities of the invention of Gandhi in view of Smith and Bhattacharyya to be used in commercial settings with customers and customer service]
With regards to claim 19, Gandhi in view of Smith, Bhattacharyya, and He fails to teach:
train an action classifier based on the third text embedding.
With regards to claim 19, Li teaches:
train an action classifier based on the third text embedding. [Li teaches “The action category may be classified and the user query, action, and/or action category may be provided to a prompt design component (e.g., prompt design component 115)” (Par [0033])
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system and action generator of Gandhi in view of Smith, Bhattacharyya, and He with the LLM system with the action classifier as taught by Li. The motivation to combine the teachings of Gandhi in view of Smith, Bhattacharyya, and He with Li is because the “way to reduce the dependency on user text inputs and provide a way to generate entire texts based on a few inputs from users” (Li Par [0016]) which increases the capabilities of the invention of Gandhi in view of Smith, Bhattacharyya, and He to be used in settings with customers and customer service]
Claims 21-22 are rejected under 35 U.S.C. 103 as being unpatentable over Bhattacharyya et al. (US2025/0119625) in view of He et al. (US 2024/0419941), Gandhi et al. (US2024/0378399), and Li et al. (US2022/0229832) in further view of Gado et al. (US12217012 hereinafter Gado)
With regards to claim 21, Bhattacharyya, He, Gandhi, and Li teaches:
All the limitations of claim 1
With regards to claim 21, Bhattacharyya, He, Gandhi, and Li fails to teach:
wherein the one or more neural networks are trained based on one or more labels generated by the LLM from a plurality transcripts.
With regards to claim 21, Gado teaches:
wherein the one or more neural networks are trained based on one or more labels generated by the LLM from a plurality transcripts. [Gado Fig 2b teaches training a language model LLM and classifier model where labels were created by language model LLM (Col 4 lines 34-35) that process the transcripts (step 258), cluster encoder values (step 260) and assign topic labels (step 262) in order to train the classifier model (step 265).
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system and classifiers as taught by Bhattacharyya, He, Gandhi, and Li with the method of training models as taught by Gado. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, and Li with Gado is because Gado teaches “classifier model is a few shot machine learning model that may be sufficiently trained using fewer than millions of examples and may trained use as few as tens to hundreds of examples and have sufficient accuracy (e.g., an accuracy greater than 80%)” (Col 8 lines 46-50) which increases the capabilities of the invention of Bhattacharyya, He, Gandhi, and Li to train neural networks in an efficient and accurate manner]
With regards to claim 22, Bhattacharyya, He, Gandhi, and Li teaches:
All the limitations of claim 1
With regards to claim 22, Bhattacharyya, He, Gandhi, and Li fails to teach:
wherein the processor is further programmed to: determine, by the one or more neural networks, a sub-category class for the intent, the reason, and/or the action based on the natural language text in the transcript; and
determine, by the one or more neural networks, a category class for the intent, the reason, and/or the action based on the determined sub-category class,
wherein the category class represents a higher-level grouping of a plurality of sub- category classes.
With regards to claim 22, Gado teaches:
wherein the processor is further programmed to: determine, by the one or more neural networks, a sub-category class for the intent, the reason, and/or the action based on the natural language text in the transcript; and [Gado Fig 2b teaches language models performs clustering encoder values (step 260) where each cluster represents a sub-category class such as intent, reason, and/or action based on the natural language text in the transcript]
determine, by the one or more neural networks, a category class for the intent, the reason, and/or the action based on the determined sub-category class, [Gado Fig 2b teaches assigning topic labels (step 262) using language model where each topic label that corresponds to the cluster which is a category class for the intent, reason, and action based on the determined sub-category class]
wherein the category class represents a higher-level grouping of a plurality of sub- category classes. [Gado teaches “topics are selected based on similarity between the clusters. For example, clusters with centroids that are less than a threshold distance apart may be combined into a single topic” (Col 8 lines 39-43) where similarity by using distance between centroids is a higher-level grouping.
It would be obvious to one of ordinary skill at the time of applicant’s filing to combine the LLM system and classifiers as taught by Bhattacharyya, He, Gandhi, and Li with the method of training models as taught by Gado. The motivation to combine the teachings of Bhattacharyya, He, Gandhi, and Li with Gado is because Gado teaches “classifier model is a few shot machine learning model that may be sufficiently trained using fewer than millions of examples and may trained use as few as tens to hundreds of examples and have sufficient accuracy (e.g., an accuracy greater than 80%)” (Col 8 lines 46-50) which increases the capabilities of the invention of Bhattacharyya, He, Gandhi, and Li to train neural networks in an efficient and accurate manner]
Conclusion
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 nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Joseph J Yamamoto whose telephone number is (571)272-4020. The examiner can normally be reached M-F 1000-1800 EST.
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, Bhavesh Mehta can be reached at 571-272-7453. 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.
JOSEPH J. YAMAMOTO
Examiner
Art Unit 2656
/BHAVESH M MEHTA/Supervisory Patent Examiner, Art Unit 2656
Prosecution Insights