DETAILED ACTION
This communication is in response to the Applicant Arguments/Remarks filed 10/2/2025. Claims 1-20 are pending in the application.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
Applicant's arguments filed 10/2/2025 have been fully considered.
Regarding the argument I, Mazza teaches at col. 5:5-19: the end user may formulate a query, and transmit the query to the chatbot system as a chat message, text message, social media message, and/or the like. The chatbot system may process the query and determine a user intent. One or more machine learning models may be invoked for predicting the user intent Once the intent is determined, the chatbot may output an answer in response to the query; col. 3, last para: answer questions pertaining to contents of a given data source. The source may include, for example, a company's website (e.g., a Frequently Asked Questions (FAQ) page, help page, etc.), other documents generated for the company (e.g., text documents, image files, sound files, etc.); col.6:39-43: the intent classification system may use the extracted embedding features to predict a user intent. The predicted user intent may be used to identify an answer to the user query, for being returned to the requesting user; col. 9:5-9: the training system executes a language model for generating a response to the candidate question, using the input context. The language model may be the same or different from the language model that is invoked to generate the candidate question; col. 10:27-30: the training system identifies the duplicate questions for separate suggested answers, and identifies an optimal answer from the plurality of separate suggested answers.
Regarding the argument II-III, Mazza et al. teaches at col. 1:56-61: in response to the score being below a threshold: employing a second machine learning model for assigning a classification score to the output; based on the classification score, identifying at least a portion of the second chunk as the answer; and associating the answer to the first candidate question for use as the training data; col. 9:18-41: a string alignment algorithm may be executed to determine how well the strings in the response align with strings in the context. In act 408, the generated score is compared against a threshold alignment value. For example, if the score is less than the threshold alignment value, further evaluation may be conducted for determining the reason for the misalignment; fig. 4: compare response to context and generate score. Thus, the extracted strings in the response/initial text response that is below a threshold is equivalent to the misaligned text portion.
Aggarwal also teaches at para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated; para. 53: the threshold for the conversational alignment score is determined based on data analysis to determine which value results in best performance over a validation dataset. Thus, the trained examples labeled as false are equivalent to the negative example set. The hallucinated content in output of a machine learning model, i.e., content which is contrary to validated retrieved context referred to as hallucination or fabrication or as content that is false, that lacks truth, then detect and/or correct one or more misalignments.
Cui further teaches at para. 33: provide guidance for the model on what to consider, examples of accurate query-answer pairs and examples of incorrect (or hallucinated) query-answer pairs; para. 44-47: generates a closed-ended query that forces the model to respond with a positive or negative (e.g., “YES”/“NO” or “TRUE”/“FALSE”) answer; fig. 4: items 402-420: receive potential answer, generate verification prompt as an hallucination check, if hallucination is not passed, generate an improvement prompt, if passed, output potential answer to the user as a final answer; fig. 5, items 502-514: a negative response by the machine learning model to the verification prompt is indicative of the potential answer being a hallucination. Thus, the combination of references does teach the argued limitations.
Regarding the argument IV, specification, para. 32 teaches “Examples of language machine learning models include BLOOM, Bard AI, ChatGPT (e.g., GPT-3.5, GPT-4, etc.), LaMDA, DialoGPT.”
Mazza teaches at para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated/negative example set. Please also see responses to the arguments II-III above.
In response to applicant’s argument that there is no teaching, suggestion, or motivation to combine the references, the examiner recognizes that obviousness may be established by combining or modifying the teachings of the prior art to produce the claimed invention where there is some teaching, suggestion, or motivation to do so found either in the references themselves or in the knowledge generally available to one of ordinary skill in the art. See In re Fine, 837 F.2d 1071, 5 USPQ2d 1596 (Fed. Cir. 1988), In re Jones, 958 F.2d 347, 21 USPQ2d 1941 (Fed. Cir. 1992), and KSR International Co. v. Teleflex, Inc., 550 U.S. 398, 82 USPQ2d 1385 (2007). In this case, the combination of references is on page 11 of the Office action.
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.
Claim(s) 1-7, 9-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Mazza et al. (US 11676044) in view of Aggarwal et al. (2024/0202284) and further in view of Cui et al. (US 20250077940).
As per claims 1, 9, 15, Mazza et al. teaches
in response to receiving a digital query from a client device, generating a plurality of text responses to the digital query and selecting a text response from the plurality of text responses to transmit to the client device by: selecting one or more supporting digital documents for the digital query from a repository of digital documents (col. 5:5-19: the end user may formulate a query, and transmit the query to the chatbot system as a chat message, text message, social media message, and/or the like. The chatbot system may process the query and determine a user intent. One or more machine learning models may be invoked for predicting the user intent Once the intent is determined, the chatbot may output an answer in response to the query; col. 3, last para: answer questions pertaining to contents of a given data source. The source may include, for example, a company's website (e.g., a Frequently Asked Questions (FAQ) page, help page, etc.), other documents generated for the company (e.g., text documents, image files, sound files, etc.); col.6:39-43: the intent classification system may use the extracted embedding features to predict a user intent. The predicted user intent may be used to identify an answer to the user query, for being returned to the requesting user; col. 9:5-9: the training system executes a language model for generating a response to the candidate question, using the input context. The language model may be the same or different from the language model that is invoked to generate the candidate question; col. 10:27-30: the training system identifies the duplicate questions for separate suggested answers, and identifies an optimal answer from the plurality of separate suggested answers);
generating, utilizing a language machine learning model, an initial text response to the digital query from a first text prompt generated utilizing the digital query (col. 2:19-26: execute a first machine learning model for automatically generating a first candidate question/digital query associated with the first chunk; determine whether the first candidate question satisfies a criterion; col. 8:11-33: the training system may provide a prompt to the language model to generate an output. For example, for company "XYZ," and title "How to Get a Refund," the task provided to the language model may be to generate X number of questions that an XYZ customer may ask about how to get a refund, that can be answered by the identified chunk/initial text response; col. 9:52-55: the training system provides to GPT-3 the candidate question and the input context, and prompt the model to output a label indicative of whether the question can be answered by the response);
extracting, utilizing an alignment score model, a supporting digital document misalignment text portion of the first initial response by comparing the initial text response and the one or more supporting digital documents, wherein the supporting digital document misalignment text portion comprises text content in the initial text response that is hallucinated by the language machine learning model relative to the one or more supporting digital documents (col. 1:56-61: in response to the score being below a threshold: employing a second machine learning model for assigning a classification score to the output; based on the classification score, identifying at least a portion of the second chunk as the answer; and associating the answer to the first candidate question for use as the training data; col. 9:18-41: a string alignment algorithm may be executed to determine how well the strings in the response align with strings in the context. In act 408, the generated score is compared against a threshold alignment value. For example, if the score is less than the threshold alignment value, further evaluation may be conducted for determining the reason for the misalignment; fig. 4: compare response to context and generate score. Thus, the extracted strings in the response/initial text response that is below a threshold is equivalent to the misaligned text portion).
Mazza does not explicitly teach the negative example set.
Aggarwal (US 20240202284) teaches
and the negative example set; wherein the supporting digital document misalignment text portion comprises text content in the initial text response that is hallucinated by the language machine learning model relative to the one or more supporting digital documents (fig. 2, item 216: misalignment detecting module, item 218: alignment score monitoring module, item 211: prompt generating module; para. 6: determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated; para. 53: the threshold for the conversational alignment score is determined based on data analysis to determine which value results in best performance over a validation dataset. Thus, the trained examples labeled as false are equivalent to the negative example set. The hallucinated content in output of a machine learning model, i.e., content which is contrary to validated retrieved context referred to as hallucination or fabrication or as content that is false, that lacks truth, then detect and/or correct one or more misalignments);
generating a negative example set comprising the supporting digital document misalignment text portion of the initial text response (para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated/negative example set);
generating a second text prompt comprising the digital query and the negative example set comprising the supporting digital document misalignment text portion of the initial text response (para. 6: monitoring, using the AI model, the conversation to determine if there is a misalignment in the conversation between the user and the AI chatbot and reduce the conversational alignment score if the misalignment is detected. The method includes determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment. The method includes increasing the conversational alignment score for a third response from the user for the recovery prompt if the conversation is recovered from the misalignment and, using the AI model, the conversational alignment is determined for the third response; para. 11, 22: determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 44: dynamically generates prompts such as a first prompt, a second prompt, an alliance confirmation prompt, and a recovery prompt based on context gathered from the user; para. 69-70);
based on comparing a first alignment score for the initial text response and a second alignment score for the supporting digital document misalignment-informed text response, selecting the supporting digital document misalignment-informed text response to transmit to the client device instead of the initial text response (para. 23: prompts are generated based on a contextual feature that has highest priority among the plurality of contextual features, wherein the contextual feature that has the highest priority is identified by prioritizing the plurality of contextual features in a decreasing order. Thus, as the text with highest score is used for the response; para. 53: the threshold for the conversational alignment score is determined
based on data analysis to determine which value results in best performance over a validation dataset. Thus, comparing the alignment scores; para. 68: the closest semantic and syntactic match among all the representative patterns provided for each intent is determined for each response received from the user; fig. 7, item 720).
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza and the misalignment portion indicates content in the first text response that is hallucinated/false by the language machine learning model of Aggarwal in order to effectively detect user inputs and/or responses generated by the machine learning models are not aligned, recover the setback by addressing the user’s concerns using prompts and/or responses to better reduce the misalignment scores.
Even if Mazza and Aggarwal do not explicitly teach the negative example set.
Cui et al. teaches
the negative example set (para. 33: provide guidance for the model on what to consider, examples of accurate query-answer pairs and examples of incorrect (or hallucinated) query-answer pairs; para. 44-47: generates a closed-ended query that forces the model to respond with a positive or negative (e.g., “YES”/“NO” or “TRUE”/“FALSE”) answer; fig. 4: items 402-420: receive potential answer, generate verification prompt as an hallucination check, if hallucination is not passed, generate an improvement prompt, if passed, output potential answer to the user as a final answer; fig. 5, items 502-514: a negative response by the machine learning model to the verification prompt is indicative of the potential answer being a hallucination);
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza, Aggarwal and negative example set of Cui to effectively detect the hallucination/false answer/response by the model in order to recover the setback by addressing the user’s concerns using prompts and/or responses to provide the user relevant/helpful answers.
The trained examples labeled as false are equivalent to the negative example set. The hallucinated content in output of a machine learning model, i.e., content which is contrary to validated retrieved context referred to as hallucination or fabrication or as content that is false, that lacks truth, then detect and/or correct one or more misalignments in answers to the users.
As per claim 2, Mazza et al. teaches
wherein selecting the one or more supporting digital documents comprises:
generating, utilizing an embedding model, query embeddings from the digital query and document embeddings from the repository of digital documents; and comparing the query embeddings with the document embeddings to identify the one or more supporting digital documents (col. 6:16-67: generate a set of context-aware embeddings (also referred to as features) from the user query. The embeddings may be word and/or sentence embeddings that represent one or more words of the user query as numerical vectors that encode the semantic meaning of the query. In this regard, the embeddings may also be referred to as semantic representations. In one example, the embeddings may be represented as a vector including values representing various characteristics of the word(s) in the query, such as, for example, whether the word(s) is a noun, verb, adverb, adjective, etc., the words that are used before and after each word, and/or the like; col. 10:36-56).
As per claims 3, 16, Mazza et al. teaches
wherein extracting the supporting digital document misalignment text portion of the initial text response utilizing the alignment score model comprises: comparing, utilizing the alignment score model, a first sentence of the initial text response with the one or more supporting digital documents to generate a first alignment score; comparing, utilizing the alignment score model, a second sentence of the initial text response with the one or more supporting digital documents to generate a second alignment score, wherein alignment scores indicate measures of alignment between sentences of the initial text response with the one or more supporting digital documents (col. 6:16-27: one or more of the neural networks may generate a set of context-aware embeddings (also referred to as features) from the user query. The embeddings may be word and/or sentence embeddings that represent one or more words of the user query as numerical vectors that encode the semantic meaning of the query. In this regard, the embeddings may also be referred to as semantic representations. In one example, the embeddings may be represented as a vector including values representing various characteristics of the word(s) in the query, such as, for example, whether the word(s) is a noun, verb, adverb, adjective, etc., the words that are used before and after each word, and/or the like; col. 8:56-67; col. 9:18-41: a string alignment algorithm may be executed to determine how well the strings in the response align with strings in the context. In act 408, the generated score is compared against a threshold alignment value. For example, if the score is less than the threshold alignment value, further evaluation may be conducted for determining the reason for the misalignment).
As per claims 4, 10, Mazza et al. teaches
wherein extracting the supporting digital document misalignment text portion of the initial text response comprises: comparing the first alignment score and the second alignment score to an alignment threshold; extracting the supporting digital document misalignment text portion of the initial text response based on determining that at least one of the first alignment score or the second alignment score fails to satisfy the alignment threshold (col. 1:56-61: in response to the score being below a threshold: employing a second machine learning model for assigning a classification score to the output; based on the classification score, identifying at least a portion of the second chunk as the answer; and associating the answer to the first candidate question for use as the training data; col.6:39-43: the intent classification system may use the extracted embedding features to predict a user intent. The predicted user intent may be used to identify an answer to the user query, for being returned to the requesting user; col. 9:18-41: the generated score is compared against a threshold alignment value. For example, a threshold alignment value of 70% may be used to determine whether the response sufficiently aligns with the input context. If the score is less than the threshold alignment value, further evaluation may be conducted for determining the reason for the misalignment. The misalignment may be because although the substance of the response equals the substance of the input context, the response may be rephrased (e.g., may use different words, may use different synonyms, the order of the words may differ, etc.), causing the response to fail to meet an alignment threshold. Thus, the extracted strings in the response/initial text response that is below a threshold is equivalent to the misaligned text portion).
As per claims 5, 11, Mazza et al. teaches at col. 9:52-55: the training system provides to GPT-3 the candidate question and the input context, and prompt the model to output a label indicative of whether the question can be answered by the response (e.g., 1 or 0).
Mazza et al. does not explicitly teach said claims.
Aggarwal teaches
generating a third text prompt from the digital query and the negative example set
comprising the supporting digital document misalignment text portion of the initial text response and an additional supporting digital document misalignment text portion of the supporting digital document misalignment-informed text response; generating, utilizing the language machine learning model, an additional supporting digital document misalignment-informed text response from the third text prompt (fig. 7; para. 6: determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 8: prompts are generated based on a contextual feature that has highest priority among the plurality of contextual features, wherein the contextual feature that has the highest priority is identified by prioritizing the plurality of contextual features in a decreasing order; para. 17-19, 21: using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 25: the conversational alignment score for a third response from the user for the recovery prompt if the conversation is recovered from the misalignment and the conversational alignment is determined for the third response);
based on comparing the first alignment score for the initial text response, the second alignment score for the supporting digital document misalignment-informed text response, and a third alignment score for the additional supporting
digital document misalignment-informed text response, selecting the supporting
digital document misalignment-informed text response to transmit to the client device instead of the initial text response or the additional supporting
digital document misalignment-informed text response (para. 6: increasing the conversational alignment score for a third response from the user for the recovery prompt if the conversation is recovered from the misalignment and, using the AI model, the conversational alignment is determined for the third response; para. 8: prompts are generated based on a contextual feature that has highest priority among the plurality of contextual features, wherein the contextual feature that has the highest priority is identified by prioritizing the plurality of contextual features in a decreasing order; para. 23-25: prompts are generated based on a contextual feature that has highest priority among the plurality of contextual features, wherein the contextual feature that has the highest priority is identified by prioritizing the plurality of contextual features in a decreasing order. Thus, as the text with highest score is used for the response; ; para.107).
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza and the third prompt and response by the language machine learning model of Aggarwal in order to effectively detect user inputs and responses generated by the machine learning models are not aligned, recover the setback by addressing the user’s concerns using prompts and/or responses to better reduce the misalignment scores.
Cui et al. also teaches
the negative example set (para. 33: provide guidance for the model on what to consider, examples of accurate query-answer pairs and examples of incorrect (or hallucinated) query-answer pairs; para. 44-47: generates a closed-ended query that forces the model to respond with a positive or negative (e.g., “YES”/“NO” or “TRUE”/“FALSE”) answer; fig. 4: items 402-420: receive potential answer, generate verification prompt as an hallucination check, if hallucination is not passed, generate an improvement prompt, if passed, output potential answer to the user as a final answer; fig. 5, items 502-514: a negative response by the machine learning model to the verification prompt is indicative of the potential answer being a hallucination).
As per claim 6, Mazza teaches
wherein generating the negative example set (col. 9:52-55: the training system provides to GPT-3 the candidate question and the input context, and prompt the model to output a label indicative of whether the question can be answered by the response, e.g., 1 or 0. Thus, the model does use the negative sample set for identifying negative/misaligned responses).
further comprises comparing each sentence of the initial text response to the one or more supporting digital documents to generate a plurality of alignment scores; and selecting a subset of sentences from the initial text response to add to the negative example set by comparing the plurality of alignment scores and determining that the subset of sentences of the initial text response are hallucinated by the language machine learning model relative to the one or more supporting digital documents (col. 6:17-21: generate a set of context-aware embeddings (also referred to as features) from the user query. The embeddings may be word and/or sentence embeddings that represent one or more words of the user query as numerical vectors that encode the semantic meaning of the query; col. 10:14-21: if one or more sentences of a paragraph of the input context are selected as the portion that achieved optimal alignment, the training system returns the entire paragraph that contains the one or more sentences as the answer block for the recommended training question; col. 9:52-55: prompt the model to output a label indicative of whether the question can be answered by the response, e.g., 1 or 0. Thus, the model does use the negative sample set for identifying negative/misaligned responses).
Aggarwal (US 20240202284) also teaches
wherein generating the negative example set (fig. 2, item 216: misalignment detecting module, item 218: alignment score monitoring module, item 211: prompt generating module; para. 6: determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated; para. 53: the threshold for the conversational alignment score is determined based on data analysis to determine which value results in best performance over a validation dataset. Thus, the trained examples labeled as false are equivalent to the negative example set. The hallucinated content in output of a machine learning model, i.e., content which is contrary to validated retrieved context referred to as hallucination or fabrication or as content that is false, that lacks truth, then detect and/or correct one or more misalignments);
Even if Mazza and Aggarwal do not explicitly teach negative example set,
Cui et al. teaches
wherein generating the negative example set (para. 33: provide guidance for the model on what to consider, examples of accurate query-answer pairs and examples of incorrect (or hallucinated) query-answer pairs; para. 44-47: generates a closed-ended query that forces the model to respond with a positive or negative (e.g., “YES”/“NO” or “TRUE”/“FALSE”) answer; fig. 4: items 402-420: receive potential answer, generate verification prompt as an hallucination check, if hallucination is not passed, generate an improvement prompt, if passed, output potential answer to the user as a final answer; fig. 5, items 502-514: a negative response by the machine learning model to the verification prompt is indicative of the potential answer being a hallucination).
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza, Cui and negative example set of Cui to effectively detect the hallucination/false answer/response by the model in order to recover the setback by addressing the user’s concerns using prompts and/or responses to provide the user relevant/helpful answers.
As per claim 7, Mazza et al. teaches
identifying a number of responses to generate for the digital query from the client device; and generating the plurality of text responses to the digital query that corresponds to the number of responses (col. 1:62-63: the source data includes questions and answers to the questions; col. 3:30-67: a chatbot builder/administrator manually defines a set of questions and appropriate responses to the questions, and uses the question-answer pairs to train the chatbot; col. 6:59-65: the pretrained
language model may include, for example, a generative language model such as, for example, Generative Pretrained Transformer 3 (GPT-3), that has been trained to
generate new intents/answers based on existing intents/answers; col. 7:8-13: the training system attempts to validate the generated questions for determining whether the input block of source data appropriately answers the question. The question may be discarded, or a different portion of the source data may be identified to answer the question, based on the results of the validation; col. 8:43-44).
As per claim 12, Mazza et al. teaches
extract an additional supporting digital document misalignment text portion of the supporting digital document misalignment-informed text response to add to the negative example set; and generate an additional text prompt comprising the negative example set and the digital text query from the client device (col. 1:56-61: in response to the score being below a threshold: employing a second machine learning model for assigning a classification score to the output; based on the classification score, identifying at least a portion of the second chunk as the answer; and associating the answer to the first candidate question for use as the training data; 10:14-21: if one or more sentences of a paragraph of the input context are selected as the portion that achieved optimal alignment, the training system returns the entire paragraph that contains the one or more sentences as the answer block for the recommended training question).
Mazza does not explicitly teach the negative example set.
Aggarwal (US 20240202284) teaches
the negative example set (fig. 2, item 216: misalignment detecting module, item 218: alignment score monitoring module, item 211: prompt generating module; para. 6: determining, using the AI model, a type of misalignment and generating a recovery prompt to recover from the misalignment; para. 11: the plurality of machine learning models are trained based on training data that comprises examples of user text labeled as true, and the user text labeled as false/hallucinated; para. 53: the threshold for the conversational alignment score is determined based on data analysis to determine which value results in best performance over a validation dataset. Thus, the trained examples labeled as false are equivalent to the negative example set. The hallucinated content in output of a machine learning model, i.e., content which is contrary to validated retrieved context referred to as hallucination or fabrication or as content that is false, that lacks truth, then detect and/or correct one or more misalignments).
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza and the misalignment portion indicates content in the first text response that is hallucinated/false by the language machine learning model of Aggarwal in order to effectively detect user inputs and responses generated by the machine learning models are not aligned, recover the setback by addressing the user’s concerns using prompts and/or responses to better reduce the misalignment scores.
Cui also teaches at fig. 4: items 402-420: receive potential answer, generate verification prompt as an hallucination check, if hallucination is not passed, generate an improvement prompt, if passed, output potential answer to the user as a final answer; fig. 5, items 502-514: a negative response by the machine learning model to the verification prompt is indicative of the potential answer being a hallucination.
As per claim 13, Mazza et al. teaches
wherein the one or more processors are configured to cause the system to generate, utilizing the language machine learning model, an additional supporting digital document misalignment-informed text response to the digital text query from the additional text prompt (col. 8:19-33: the training system may provide a prompt to the language model to generate an output. The language model generates the prompted X
number of questions based on the identified chunk.).
As per claim 14, Mazza et al. teaches
wherein the one or more processors are configured to cause the system to: provide, for display on a graphical user interface of the client device, the supporting digital document misalignment-informed text response with the second alignment score; and provide, for display on the graphical user interface of the client device, a supporting digital documents element, wherein a selection of the supporting digital documents element causes the graphical user interface to display the one or more supporting digital documents with emphasized portions of the one or more supporting digital documents that correspond to the supporting digital
document misalignment-informed text response (figs. 7-8: prompts; col. 12:20-67: an example document (e.g., web page) that may be used for generating training question-answer pairs according to one embodiment. The training system may
analyze the document and partition the document into four chunks. For example, the chunks may be determined based on identifying section headings that have a larger font size than the remaining text; col. 3, last paragraph);
Claim 17 claims similar subject matter as of claims 4-5 and is rejected based on the same ground of rejection.
As per claim 18, Mazza et al. teaches
wherein generating the text responses to the digital query further comprises receiving, via the user interface of the client device, a selection of a number of iterations, the number of iterations indicating a number of text responses to iteratively generate from the digital query (fig. 9: ID change in source data, automatically generate training questions, retrain chatbot, log change; col. 7:36-43: when a change is detected, the training system may provide the updated source data (including context surrounding the source data), to the language model, for generating one or more training questions. The updated question-answer pair may then be used for retraining the inference models; col. 8:27-33; col. 13:31-39: the training system may use the training questions and associated answers for re-training the inference models of the chatbot system, the training system may record the change of the chatbot system in a log. The log may include, for example, a timestamp in which the change was made. Changes made to the chatbot may also be recorded in the log for avoiding redoing a modification that may have already have been performed in a previous iteration).
As per claim 19, Mazza et al. teaches
providing, for display on a graphical user interface of the client device, a comparison element for selection; and in response to a selection of the comparison element, providing, for display on a graphical user interface of the client device, the supporting digital document misalignment-informed text
response with a corresponding alignment score (col. 7:50-67: the GUI displays of the question-answer pairs recommended by the training system; fig. 4, item 408: the training system may determine that the best answer for Q5 is a combination of chunks 804 and 806, and output both of the paragraphs as the answer; col. 6:42-43: the predicted user intent may be used to identify an answer to the user query, for being returned to the requesting user).
Aggarwal also teaches at para. 81-84: The conversation alignment score monitoring module detects a drop in the conversational alignment between the user 102 and the AI chat bot 110 and attempts to recover from this misalignment and bring back the user 102 to the conversation by providing a recovery prompt 418 to the user 102 through the prompt generating module 211, e.g., "I'm trying my best to understand you better. Try venting to me, it will help." In FIG. 4B of a user interface 401, the user 102
may respond with 420 to the recovery prompt 418 provided by the AI chatbot 110, e.g., "Ok, let's keep talking." The AI chatbot 110 recovered from the major misalignment as the user 102 agreed to continue with the conversation.
As per claim 20, Mazza et al. teaches
providing the supporting digital document misalignment-informed response to the client device and further providing, to the client device for display, at least a portion of one or more of the supporting digital documents that emphasized portions of the supporting digital documents that correspond the supporting digital document misalignment-informed text response (col. 7:50-67: the GUI displays of the question-answer pairs recommended by the training system; fig. 4, item 408: the training system may determine that the best answer for Q5 is a combination of chunks 804 and 806, and output both of the paragraphs as the answer; col. 12:20-67).
Claim(s) 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Mazza et al. (US 11676044) in view of Aggarwal et al. (2024/0202284) and further in view of Cui et al. (US 20250077940) and Lee et al. (US 20200372399).
As per claim 8, Mazza et al. teaches
wherein selecting the supporting digital document misalignment-informed text response to transmit to the client device instead of the initial text response further comprise: providing, for display on a graphical user interface of the client device, the supporting digital document misalignment-informed text response with the second alignment score (col. 7:50-67: the GUI displays of the question-answer pairs recommended by the training system; fig. 4, item 408: the training system may determine that the best answer for Q5 is a combination of chunks 804 and 806, and output both of the paragraphs as the answer; col. 12:20-67: an example document (e.g., web page) that may be used for generating training question-answer pairs according to one embodiment. The training system may analyze the document and partition the document into four chunks. For example, the chunks may be determined based on identifying section headings that have a larger font size than the remaining text; col. 3, last paragraph.)
Even if Mazza, Aggarwal, Cui do not explicitly teach transmitting the second alignment score and the second text response to the client device.
Lee teaches
transmitting the second alignment score and the second text response to the client device (para. 127: The misalignment identification system 118 uses the threshold values 1112 to determine misalignments or which misalignments to report to the publisher. The misalignment identification system 118 can determine, via user input, that the threshold value 1112 of 0.5 for the misalignment class 1108 “does the page contain information unrelated with the link?” results in an excessive number of reported misalignments. Based on this determination, the misalignment identification system 118 can adjust the threshold value 1112 to a higher value (e.g., 0.75) to decrease the sensitivity and number of reported misalignments; para. 150).
Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Mazza, Aggarwal, Cui with the present text alignment score in relating to the response to effectively provide users the accuracy or alignments of responses generated by the machine learning models of Lee in order for the users recover the setback by addressing the user’s concerns to better reduce the misalignment scores.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Shamir et al. (US 20250252292) teaches at para. 49: the pretrained model is used to sample sequences, which may then be labeled by humans (or by models) with a preference label. The preference label may designate which sequences/examples satisfy some preference and which do not, or alternatively, may give relative ratings or rankings among sequences preferring one sequence over another to the specific task; para. 207: validation tools include tools that can parse and confirm output(s) of a machine-learned model. Validation tools include engineered heuristics that establish certain thresholds applied to model outputs. For example, validation tools can ground the outputs of machine-learned models to structured data sources (e.g., to mitigate "hallucinations").
Singh et al. (US 20250094462) teaches at para. 96-97: facilitate the generating of the response, prompts may be determined based on the queries, the contextual information, and in-context learning samples. The prompts may relate to instructions that are usable for machine learning models such as, for example, large language models.
Lin (US 20230320642) teaches at para. 90: machine learning model, embedding text documents.
Karpman et al. (US 11995803) teaches at col. 21:38-41: the advanced settings menu can include a header prompting the user to enter a negative prompt (e.g., a description of subject matter that the generated image should not include); col. 21:65-col. 22:2: the software application layer is configured to automatically modify and/or edit the text prompt entered into the interactive text field based on a style selection at the generation interface and/or negative prompts entered by the user in the advanced settings menu; col. 12:29-35: generate a multimodal embedding (e.g., a feature vector) representing image-alternate-text pairs (e.g., each pair in the training corpus).
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 LINH BLACK whose telephone number is (571)272-4106. The examiner can normally be reached 9AM-5PM EST 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, Tony Mahmoudi can be reached on 571-272-4078. 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.
/LINH BLACK/Examiner, Art Unit 2163 1/19/2026
/TONY MAHMOUDI/Supervisory Patent Examiner, Art Unit 2163