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 .
Priority
Acknowledgment is made of applicant’s claim for foreign priority under 35 U.S.C. 119 (a)-(d). The certified copy has been filed in parent Application No. CN 2023113640306, filed on 10/19/2023.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-7, 12-15 are rejected under 35 U.S.C. 103 as being unpatentable by Ahmadidaneshashtiani (US 2021/0249002 A1) in view of Gao (NPL ‘Neural Approaches to Conversational Information Retrieval’ (2022)).
Regarding claim 1, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction method, which is applied at a proxy layer, comprising: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
receiving user interactive information and user preference information transmitted from a smart device side, the user interactive information being information desensitized by the smart device side with respect to original interactive information inputted by a user; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
transmitting, after processing the user interactive information to obtain post-processed interactive information, the post-processed interactive information to a GPT model, so that the GPT model returns task handling information with respect to the post-processed interactive information; and ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
fusing, upon receipt of the task handling information returned from the GPT model, the user preference information and the task handling information to obtain interactive response information, and transmitting the interactive response information to the smart device side, so that the smart device side remediates the interactive response information based on user private data included in the user interactive information and provides a response to the user based on remediated information. ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 2, Ahmadidaneshashtiani in view of Gao teaches claim 1. Ahmadidaneshashtiani further teaches:
matching between respective nouns in the task handling information and respective preference labels in the user preference information to obtain the preference labels matching the respective nouns in the task handling information; and ([0326] “The other approach, tokenization, works but requires context-aware values to be used in place of the sensitive data. A broker's sanitizer cannot, for instance, replace an email address with a string of numbers (112435) because the NLU will no longer adequately be able to discern that the user gave an email address. Instead, the broker system must replace email addresses with something that also looks like an email address.”)
replacing, based on label values of the preference labels matching the respective nouns in the task handling information, corresponding nouns in the task handling information to obtain the interactive response information; wherein the smart device side, upon receipt of the interactive response information, remediates the interactive response information based on the user private data included in the user interactive information to obtain remediated handling information and executes the remediated handling information to implement the response to the user. ([0327] “The broker leverages the second approach for sanitization by allowing for a sanitization “hook” to be added at the front of the conversation. This hook can be run before the NLU sees the query, obfuscating any sensitive data, and then again after the NLU returns, restoring said data before handing off to a fulfillment handler. This retains the sensitive data internally for use while withholding it from external services. The hook also provides a single placeholder for keeping logs of the conversation reducing the attack surface (against any cyber-attacks) and simplifying security controls (hence reducing cost). This may, for example, be superior to alternate approaches where Another superiority—data is stored across different components of the dialog system while in this example embodiment, only the broker memory and downstream API have access to the sensitive data.”)
Regarding claim 3, Ahmadidaneshashtiani in view of Gao teaches claim 1. Gao further teaches:
vectorizing the post-processed interactive information to obtain a post-processed interactive vector; searching, based on the post-processed interactive vector, in knowledge bases of respective domains to determine domain knowledge matching the post-processed interactive vector from within the knowledge bases of the respective domains and a domain knowledge vector of the matching domain knowledge; and; concatenating the post-processed interactive vector and the domain knowledge vector of the matching domain knowledge to obtain a fused query vector, and transmitting the fused query vector to the GPT model. (Section 5.3 page 71 “the encoder, which is similar to what is described in Section 5.2, encodes each 𝑄𝑃𝑛 pair independently, and produces an integrated contextual vector representation for each token, m𝑛 𝑗 for the 𝑗-th token of the 𝑛-th pair. The decoder then performs attention over the concatenation of the representations of all the retrieved passages, and generate the answer string.”)
Regarding claim 4, Ahmadidaneshashtiani in view of Gao teaches claim 3. Ahmadidaneshashtiani further teaches:
transmitting the fused query vector to a plurality of GPT models to obtain task handling information returned from the plurality of GPT models with respect to the post-processed interactive information; and ([0391] “In some embodiments, specific domain-trained NLP and NLU engines are utilized”)
fusing or filtering the task handling information returned from the plurality of GPT models with respect to the post-processed interactive information to obtain optimized handling information, the optimized handling information being fused with the user preference information to obtain the interactive response information. ([0391] “a selection process may be utilized to select a specific NLP and NLU engine that has a high or highest confidence score in relation to generated intent data objects. The intent handlers can link responses to different intents.”)
Regarding claim 5, Ahmadidaneshashtiani in view of Gao teaches claim 1. Gao further teaches:
the proxy layer is constructed by:
building an initial model based on a pre-trained text generation model; (Figure 3.3 shows the model being initialized.)
inputting sample input information to the initial model to obtain test output information outputted by the initial model; and (Figure 3.3 shows the model being pre-trained.)
finely tuning the initial model based on a score of the test output information or based on a difference between the test output information and the sample output information, thereby obtaining the proxy layer. (Figure 3.3 shows the model being fine-tuned with task labels).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine middleware platform of Ahmadidaneshashtiani with the training method of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 6, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction method, which is applied at a smart device side, comprising: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
receiving original interactive information inputted by a user, and desensitizing the original interactive information to obtain user interactive information; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
transmitting the user interactive information and the user's user preference information to a proxy layer, so that the proxy layer transmits, after processing the user interactive information to obtain post-processed interactive information, the post-processed interactive information to a GPT model, whereby the GPT model returns task handling information with respect to the post-processed interactive information; and ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
receiving interactive response information transmitted from the proxy layer, remediating the interactive response information based on user private data included in the user interactive information, and providing a response to the user based on remediated information; wherein the interactive response information is obtained from fusing, by the proxy layer upon receipt of the task handling information returned from the GPT model with respect to the post-processed interactive information, the user preference information and the task handling information. ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 7, Ahmadidaneshashtiani in view of Gao teaches claim 6. Ahmadidaneshashtiani further teaches:
remediating the interactive response information based on the user private data included in the user interactive information to obtain remediated handling information, and executing the remediated handling information to implement a response to the user; ([0326] “The other approach, tokenization, works but requires context-aware values to be used in place of the sensitive data. A broker's sanitizer cannot, for instance, replace an email address with a string of numbers (112435) because the NLU will no longer adequately be able to discern that the user gave an email address. Instead, the broker system must replace email addresses with something that also looks like an email address.”)
wherein the interactive response information is obtained by the proxy layer which, after performing matching between respective nouns in the task handling information and respective preference labels in the user preference information to obtain the preference labels matching the respective nouns in the task handling information, replaces corresponding nouns in the task handling information based on label values of the preference labels matching the respective nouns in the task handling information ([0327] “The broker leverages the second approach for sanitization by allowing for a sanitization “hook” to be added at the front of the conversation. This hook can be run before the NLU sees the query, obfuscating any sensitive data, and then again after the NLU returns, restoring said data before handing off to a fulfillment handler. This retains the sensitive data internally for use while withholding it from external services. The hook also provides a single placeholder for keeping logs of the conversation reducing the attack surface (against any cyber-attacks) and simplifying security controls (hence reducing cost). This may, for example, be superior to alternate approaches where Another superiority—data is stored across different components of the dialog system while in this example embodiment, only the broker memory and downstream API have access to the sensitive data.”)
Regarding claim 12, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction method, comprising: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
receiving, by a smart device side, original interactive information inputted by a user, and desensitizing the original interactive information to obtain user interactive information; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
transmitting, by the smart device side, the user interactive information and the user's user preference information, to a proxy layer; ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
receiving, by the proxy layer, the user interactive information and the user preference information transmitted from the smart device side, processing the user interactive information to obtain post-processed interactive information, and then transmitting the post-processed interactive information to a GPT model, so that the GPT model returns task handling information with respect to the post-processed interactive information; ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
fusing, by the proxy layer, after receiving the task handling information returned from the GPT model with respect to the post-processing interactive information, the user preference information and the task handling information to obtain interactive response information, and transmitting the interactive response information to the smart device side; and ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
receiving, by the smart device side, the interactive response information transmitted from the proxy layer, remediating the interactive response information based on user private data included in the user interactive information, and providing a response to the user based on the remediated information. ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 13, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction apparatus, which is deployed at a proxy layer, comprising: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
an information receiving unit configured to receive user interactive information and user preference information transmitted from a smart device side, the user interactive information being information desensitized by the smart device side with respect to original interactive information inputted by a user; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
a prompt proxy configured to transmit, after processing the user interactive information to obtain post-processed interactive information, the post-processed interactive information to a GPT model, so that the GPT model returns task handling information with respect to the post-processed interactive information; and ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
a response proxy configured to fuse, upon receipt of the task handling information returned from the GPT model, the user preference information and the task handling information to obtain interactive response information, and transmit the interactive response information to the smart device side, so that the smart device side remediates the interactive response information based on user private data included in the user interactive information and provides a response to the user based on remediated information. ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 14, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction apparatus, which is deployed at a smart device side, comprising: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
an input processing unit configured to receive original interactive information inputted by a user, and desensitize the original interactive information to obtain user interactive information; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
an information forwarding unit configured to transmit the user interactive information and the user's user preference information to a proxy layer, so that the proxy layer transmits, after processing the user interactive information to obtain post-processed interactive information, the post-processed interactive information to a GPT model, whereby the GPT model returns task handling information with respect to the post-processed interactive information; and ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
a response information providing unit configured to receive interactive response information transmitted from the proxy layer, remediate the interactive response information based on user private data included in the user interactive information, and provide a response to the user based on remediated information; wherein the interactive response information is obtained from fusing, by the proxy layer upon receipt of the task handling information returned from the GPT model with respect to the post-processed interactive information, the user preference information and the task handling information. ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Regarding claim 15, Ahmadidaneshashtiani teaches:
A GPT model-based smart device interaction system, comprising: a proxy layer; and a smart device side: ([0037] “The computer processor is configured to receive, from a user interface (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”. The conversational middleware platform provided by the system, in some embodiments is adapted for receiving the conversational flows from a plurality of individuals, or to receive conversational flows in a plurality of languages.” And [0097] “The system 100 acts as a conversational agent (intermediary) that interoperates between an application interface, a NLP (either cloud based or any API enabled NLP), and downstream fulfilment APIs as illustrated in FIG. 1A.”)
an information receiving unit configured to receive user interactive information and user preference information transmitted from a smart device side, the user interactive information being information desensitized by the smart device side with respect to original interactive information inputted by a user; ([0080] “The computer processor is configured to receive, from a user interface 102 (e.g., an audio or a visual interface rendered on a display of a mobile device or an automated teller machine (ATM)), input strings representing a conversational request from the user having query tokens. The input strings can include general or specific queries, such as “what is my account balance?”, or “send money to my son”.” And [0081] “The broker processor 108 identifies, from the input strings, sensitive query tokens that need to be redacted or sanitized (e.g., payee names, account numbers).” The system also deals with the users preferences [0127] “An interaction manager may control a number of conversation sessions with various endpoints, and may also establish audit logs for tracking performance, or iteratively updating a user profile following or during a discussion. For each user or type of user, a user profile can be tracked, and as described in various embodiments below, can utilize specific types of data structures that are periodically updated to reflect new biases and weights that impact how the routing and orchestration system of the broker interfaces with the various natural language processing/natural language understanding agents.”)
a prompt proxy configured to transmit, after processing the user interactive information to obtain post-processed interactive information, the post-processed interactive information to a GPT model, so that the GPT model returns task handling information with respect to the post-processed interactive information; and ([0118] “Upon receiving response from the NLU agent (the response is what is referred to as an Intent data object in this document), the broker does a look-up on its registered fulfillment handlers and investigates whether a handler is registered for the given intent.”)
a response proxy configured to fuse, upon receipt of the task handling information returned from the GPT model, the user preference information and the task handling information to obtain interactive response information, and transmit the interactive response information to the smart device side, so that the smart device side remediates the interactive response information based on user private data included in the user interactive information and provides a response to the user based on remediated information; and ([0132] “The utterance string can use placeholders instead of sensitive information, and in some embodiments, the placeholders are unique so that they can be mapped back to the original information when generating responses for the user on an internal conversational flow. In this situation, the modifications are tracked, for example, by maintaining a lookup table of placeholders and sensitive information, which can be used to de-sanitize the outputs (e.g., swapping <TICKER> with the symbol of the stock of relevance in the conversation) prior to presentation through the chatbot interface.” Also see [0327] for more details.)
on the smart device side is deployed the GPT model-based smart device interaction apparatus according to claim 14. (See above)
Ahmadidaneshashtiani does not teach the model being a GPT based model. However, Gao does (Section 6.6.1 SC-GPT on page 97 describes their GPT model).
Ahmadidaneshashtiani and Gao are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO. One would be motivated to do this as Ahmadidaneshashtiani already uses NLP models.
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable by Ahmadidaneshashtiani in view of Gao and Brown (US 2020/0364511 A1).
Regarding claim 8, Ahmadidaneshashtiani in view of Gao teaches claim 6. Neither of them teach the elements of claim 8. However Brown does:
collecting a plurality of pieces of original interactive information inputted by the user and corresponding remediated information, and recording the original interactive information, which corresponds to the remediated information not meeting the user's expectation, as a directed to-be-optimized sample; ([0034] “A feedback loop for a conversational system starts with the user's input about the answer to their question. The system will then store this positive or negative response to be reviewed later by a subject matter expert (SME). If the feedback is negative, the SME will then provide the system enough information to retrain the system.”)
performing intent recognition with respect to the directed to-be-optimized sample to obtain a user intent corresponding to the directed to-be-optimized sample, and setting, based on the user intent, a type label for a concerned text associated with the user intent in the directed to-be-optimized sample; ([0049] “That is, the classification for the user's original question/utterance U1a and/or a substantially similar question/answer as the user's original question/utterance U1a defines the intent of the question/utterance. Thus, for the example of “How can I make Software Product A work for me?” (the user's original question/utterance U1a), this question/utterance is classified as a request for help in making a product work properly.”)
performing word segmentation to the concerned text to obtain a keyword of the concerned text, and searching, based on the keyword of the concerned text and the type label of the concerned text, in knowledge bases of respective domains to obtain domain knowledge matching the concerned text; and ([0074] “a DNN 524 (analogous to DNN 424 shown in FIG. 4) is trained to rank answers to a particular question (e.g., utterance 400 shown in FIG. 4) based on the intent (see intent labels 402 in FIG. 4) and entities (e.g., certain words, phrases, terms, etc.) in that particular question.”)
establishing an association relationship between the concerned text and the domain knowledge matching the concerned text, so that after a same user inputs original interactive information identical to the directed to-be-optimized sample again, searching is performed, based on post-processed interactive vector of the post-processed interactive information corresponding to the original interactive information, in the knowledge bases of respective domains, wherein the domain knowledge matching the concerned text is determined as the domain knowledge matching the corresponding post-processed interactive vector; or, a response to the original interactive information is provided based on the domain knowledge matching the concerned text. ([0084]-[0088] “As described in block 707, the system then reevaluates the user's original question utilizing alternative intent classifications to form alternative answers, such as those shown in FIG. 2. As described in block 709, the system the presents the alternative answers (e.g., utterances U2b-a, U2b-b, U2b-c shown in FIG. 2) to the user. As described in block 711, the system then receives a selection from the alternative answers from the user (e.g., utterance U2b-c from FIG. 2). As described in block 713, the system updates processing rules to generate updated processing rules that are based on the selection (see FIG. 6). As described in block 715, the system then retrains the cognitive interaction information system based on the updated processing rules (see FIGS. 4-6).”)
Ahmadidaneshashtiani, Gao and Brown are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO and the negative feedback retraining of Brown. One would be motivated to do this to be able to update the model based on user feedback.
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable by Ahmadidaneshashtiani in view of Gao and Wang (US 2021/0191506 A1).
Regarding claim 9, Ahmadidaneshashtiani in view of Gao teaches claim 6. Neither of them teach the elements of claim 9. However Wang does:
obtaining feedback information inputted by the user with respect to the response and collecting an image of the user when inputting the feedback information; ([0089] “After starting (1000) an affective interaction session, an affective interaction system (e.g., affective interaction system 100 of FIG. 3A) may receive (1002) an emotion communication from a user and derive (1004) emotion-related data based on the collected emotion communication by using a data collector (e.g., data collector 202 of FIGS. 3A and 3B).”)
performing emotion recognition with respect to the feedback information to obtain an emotion type of the feedback information; ([0091] “At step 1014, the system may recognize (1014) an emotion feature based on the emotion-related data by using an emotion recognizer (e.g., emotion recognizer 204 of FIGS. 3A and 3C).”)
performing expression recognition with respect to the image of the user when inputting the feedback information to obtain an expression type of the user when inputting the feedback information; ([0054] “Collector 202 is configured to capture emotion communication 102 in one or more modalities, such as a text 104, a voice 106, a facial expression 108, a gesture 110, a physiological signal 112, and/or multimodality 114. Collector 202 is used at the beginning stage of an affective interaction session, where it serves as an interface to a user and a communication channel for an affective interaction system to collect data and emotion from a user. “)
determining, in a case that the feedback information includes a negative keyword, the emotion type of the feedback information is a negative emotion, or the expression type of the user when inputting the feedback information is a negative expression, that an interaction barrier occurs currently; and ([0092] “At step 1026, the system may then derive and transmit (1026) a user intention containing the affective intention and interaction intention.”)
obtaining, in a case that the interaction barrier occurs currently, respective pieces of historical prompt-response information corresponding to the user, searching, in the respective pieces of historical prompt-response information, historical prompt-response information matching the original interactive information, and providing a response based on the historical prompt-response information matching the original interactive information. ([0092] “Based on the domain knowledge data, general knowledge data, and the user intention, the system may formulate (1028) an affective strategy for the affective intention and an interaction strategy for the interaction intention by using an affective strategy formulator (e.g., affective strategy formulator 208 of FIGS. 3A and 3C). The affective strategy formulator may then generate (1030) an affective command based on the affective strategy and interaction strategy. Based on the affective command, the system may then provide (1032) an affective expression and convey (1034) the affective expression back to the user, using an affective computing expression generator 210 (e.g., affective computing expression generator 210 of FIGS. 3A and 3B). The system may then go back and start (1000) another affective interaction session or continue the session. The system may perform the above process in whole or in part, in any sequence or order, with or without any additional steps.”)
Ahmadidaneshashtiani, Gao and Wang are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning conversation systems. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO and the emotion recognition of Wang. One would be motivated to do this to be able to update the model based on user feedback.
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable by Ahmadidaneshashtiani in view of Gao and Lien (US 2025/0085986 A1 with provisional filed Sep 12th 2023).
Regarding claim 10, Ahmadidaneshashtiani in view of Gao teaches claim 6. Neither of them teach the elements of claim 10. However Lien does:
determining a current network status type, transmitting, in a case that the network status type is poor network connectivity or network outage, the user interactive information and the user's user preference information to a locally deployed local GPT model to obtain local response information returned from the local GPT model, and providing a response to the user based on the local response information. ([0060] “Such a local LLM could be used when the remote device 102 is offline (e.g., in the event of a network disruption. The local LLM can also process simpler questions that can be resolved without the need for the cloud-based LLM (e.g., “Grandie, open the Call app.”).”)
Ahmadidaneshashtiani, Gao and Lien are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO and offline capabilities of Lien. One would be motivated to do this so models can still be used while offline.
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable by Ahmadidaneshashtiani in view of Gao, Lien and Frantar (NPL ‘GPTQ: Accurate Post-Training Quantization For Generative Pre-Trained Transformers’).
Regarding claim 11, Ahmadidaneshashtiani in view of Gao and Lien teaches claim 10. None of them teach the elements of claim 11. However Frantar does:
the local GPT model is obtained by performing quantized model compression to a sample GPT model. (Section 4 The GPTQ Algorithm teaches their model quantization method.)
Ahmadidaneshashtiani, Gao, Lien and Frantar are considered analogous art to the claimed invention because they are in the same field of endeavor being machine learning. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine privacy preserving aspects of Ahmadidaneshashtiani with the GPT model of GAO and offline capabilities of Lien with the compression of Frantar. One would be motivated to do this so the models take up less storage on the local devices.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DANIEL P GRUSZKA whose telephone number is (571)272-5259. The examiner can normally be reached M-F 9:00 AM - 6:00 PM ET.
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, Li Zhen can be reached at (571) 272-3768. 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.
/DANIEL GRUSZKA/Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights