DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
Applicant's arguments filed 2/9/26 have been fully considered but they are not persuasive.
Regarding the 35 U.S.C, 103 rejection of the claims 1, 19 and 20 with references Kuo and Cohen, Applicant argues that although Cohen is utilized to teach limitations “the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection" and "the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection”, Cohen does not describe the various benefits not recognized in the art of record, such as allowing the customer to train the first classifier as desired and without relying on the service provider to do so, as well as reducing calls to the service provider's classifier, which may cost the customer money, and that the details of the two claimed classifiers are not taught, and that Applicant did not find disclosure in Kuo or Cohen of doing something “in response to a first ANN determining that the user query is not out-of-domain based on the user query corresponding to one of k classes” or “in response to the first ANN determining that the user query is out-of-domain based on no probability scores from a plurality of sigmoid functions corresponding to a respective one of the first plurality of predefined user intents exceeding a predetermined threshold” (Arguments, pg. 11-15).
Examiner respectfully disagrees. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., benefits from allowing the customer to train the first classifier as desired and without relying on the service provider to do so, as well as reducing calls to the service provider's classifier) are not recited in the rejected claims. Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
Nevertheless, Cohen discloses the use of first machine learning models (MLM) 122 trained by a provider as well as a second machine learning models (MLM) 125 trained by a user/customer (para. [0045]; fig. 1, elements 122, 125), where the MLM 125 is configured to identify a speaker's intent in a “user-specific domain” among multiple intents/intentions (para. [0016]; para. [0025]-[0026]), and where the MLM 125 is trained on user-specific data 204 (para. [0029], corresponding to limitation “wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection”. Cohen also discloses where the MLM 122 is configured to identify a topic or a user’s/speaker's intent in a different domain among multiple intents/intentions (para. [0017]; para. [0025]-[0026]), and where the MLM 122 is trained on initial data 204 (para. [0029]), corresponding to limitation “the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection”.
Cohen further discloses that the training of the first machine learning model (MLM) 122 trained by a provider is performed off-site at a training server 162 (fig. 1, element 160, 162; para. [0023]), while the training of the second machine learning model (MLM) 125 trained by a user/customer is performed at a user device 102 (fig. 1, element 102, para. [0019]; para. [0023]), i.e., a decoupling of the training and as a result, limitation “training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer”, where the decoupled training of models 122 and 125 implies facilitation of “independent performance optimization of the first classifier by the customer”. Therefore, Examiner maintains that the combination of Kuo and Cohen discloses the details of the two claimed classifiers.
Regarding claim 21, Kuo discloses its prediction models 206 or 206 (i.e., the claimed first ANN) analyzing an initial user message 432 to determine if the message 432 includes one or more keywords associated with any one of the possible intents and calculating a score for each possible intent, while selecting an intent for the message 432 from the set of possible intents having the highest score (fig. 2; fig. 4; para. [0052]; para. [0067]), i.e., determining that the user query is not out-of-domain based on the probability from the first ANN. Kuo also discloses providing the determined intent to a chat robot 204 (i.e., a virtual assistant), where the chat bot provides dialogue to the user 104 and assists the user based on the intent determined by the prediction model 206, corresponding to “doing something” in response, and as a result, limitation “in response to the first ANN determining that the user query is not out-of-domain based on the probability from the first ANN, control the virtual agent based on a user intent of the first plurality of predefined user intents to which the user query is determined to correspond”.
Kuo further discloses that in response to its system determining that the prediction models 206 and 208 are unable to predict an intent of the user when none of the scores corresponding to the possible intents is above a probability threshold (para. [0064]) i.e., the user query is out-of-domain based on no probability scores exceeding a predetermined threshold, processing the user utterance with a different prediction model 704 (i.e., the claimed second ANN) that was trained using the third-party training data that includes intents not supported by the prediction models 206 and 208, and where the different prediction model 704 is trained to receive user utterance and output/determine an intent associated with the utterance (para. [0072]), corresponding to limitation “in response to the first ANN determining that the user query is out-of-domain based on no probability scores exceeding a predetermined threshold: second process the received data using a second ANN, the second ANN having been trained with a second set of training data that is different from the first set of training data to classify user queries according to a second, different, plurality of predefined user intents, the second ANN trained to receive the user query data as input and output a predicted user intent associated with a class that the second classifier has been trained to recognize and a respective probability”
What Kuo in view of Cohen does not explicitly disclose include the use of “probability scores from a plurality of sigmoid functions corresponding to a respective one of the first plurality of predefined user intents”
Reference Xu discloses this limitation (para. [0037]-[0038]; para. [0047]; para. [0172]; para. [0181]). When combined with the teachings of Kuo and Cohen with the suggestion/motivation of calculating the confidence scores as measured by the probabilities of the classes/Intents being handled by a classifier, among the various activation functions/methods of calculating such scores (Xu, para. [0043]; para. [0164]), the combination of the references discloses the limitations of claim 20.
Regarding the dependent claims rejected with additional references Xu, Li, Wohlwend and Sackett, Applicant argues that the additional references do not address the above argued deficiencies of Kuo and Cohen and as such, the references fail to disclose the limitations recited in the dependent claims (arguments, pg. 14-15). Examiner respectfully disagrees as provided above, and absent any argument as to why the cited portions of the additional references fail to disclose specific limitations in the dependent claim, Examiner maintains that the rejections of the dependent claims are appropriate
Response to Amendment
The prior 35 U.S.C. 112 rejection of as well as the objection to the claims (3/6/26) are hereby withdrawn in light of amendments to the claims
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.
1. Claim 1-4, 6, 8, 11-14 and 16-20 are rejected under 35 U.S.C. 103 as being unpatentable over Kuo et al US 2021/0125025 A1 (“Kuo”) in view of Cohen et al US 2022/0261631 A1 (“Cohen”)
Per claim 1, Kuo discloses a method of processing user queries for a virtual agent, the method comprising:
receiving data representing a user query directed to the virtual agent (a service provider may utilize an online chat system that may include a chat robot (also referred to as a chat bot or simply a bot) to provide automated assistance to a customer in an online chat session …, para. [0015]; The chat interface 402 also includes an input portion 414 that enables the user 140 to input a message (e.g., an utterance that may include text data, audio data …, para. [0049]-[0050]);
determining whether the user query corresponds to any one of a first plurality of predefined user intents, wherein the determining comprises first processing the received data using a first classifier configured to classify user queries according to the first plurality of predefined user intents, the first classifier comprising a first artificial intelligence (AI) model trained with a first set of training data to receive the user query data as input and output a predicted user intent associated with a class that the first classifier has been trained to recognize and a respective probability (para. [0051]; the prediction model 206 may be a natural language model (e.g., a machine learning model) configured to analyze natural language (e.g., a sentence, a phrase, a question, etc.) and to determine (e.g., predict) whether a message obtained from an online chat session is associated with any one of the set of possible intents … The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent …, para. [0052]; para. [0055]; the prediction model 208 may be a bidirectional encoder representations from transformations (BERT) model…., para. [0059]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067]; para. [0071], prediction models 206 or 208 as first classifier, machine learning and BERT model as AI models);
in response to a positive determination based on the probability from the first classifier meeting or exceeding a threshold, generating first output data based on a user intent of the first plurality of predefined user intents to which the user query is determined to correspond (The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent… In some embodiments, the prediction model 206 may select an intent for the message 432 from the set of possible intents having the highest score…, para. [0052]; para. [0053]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067], prediction models 206 or 208 as predicting and outputting intent having highest score exceeding lower scores/threshold);
in response to a negative determination based on the probability from the first classifier falling below the threshold, second processing the received data using a second classifier to generate second output data, the second classifier being configured to classify user queries according to a second plurality of predefined user intents, the second classifier comprising a second AI model trained with a second set of training data that is different from the first set of training data, the second classifier configured to receive the user query data as input and output a predicted user intent associated with a class that the second classifier has been trained to recognize and a respective probability (para. [0021]; In some embodiments, the chat manager 202 may determine that the prediction model 208 is unable to predict an intent of the user 104 based on the utterance (e.g., the utterance 532). For example, the chat manager 202 may determine that the prediction model is unable to predict an intent of the user 104 when none of the scores corresponding to the possible intents is above a probability threshold …, para. [0064]; On the other hand, if the chat manager 202 determines that the prediction model 206 is unable to predict an intent of the user 104 (e.g., the prediction model 206 outputs a null value or otherwise indicates that no intent is determined) …, para. [0066]-[0067]; para. [0071]; Since the prediction model 704 was trained using the third-party training data that includes intents not supported by the prediction models 206 and 208, the prediction model 704 may be able to determine an intent for the utterance 602 when the prediction models 206 and 208 failed to do the same…, para. [0072], prediction models 206 and 208 unable to predict an intent for utterance as negative determination, third party intents not supported by models 206 and 208 as second plurality of predefined user intents, prediction model 704 as second AI model, determined/output intent by prediction model 704 as second output data); and
performing a predetermined action in relation to the virtual agent based on either the first output data or the second output data (the prediction model 206 may output the intent to file a dispute of a transaction to the chat robot 204. The chat robot 204 may provide a dialogue to the user 104 and assist the user 104 based on the intent determined by the prediction model 206 …, para. [0053]; para. [0067], intent predicted by prediction model 206 or 208 as first output data used by chat robot to provided response to user)
Kuo does not explicitly disclose wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection, the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection, and training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer
However, these features are taught by Cohen:
wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection (a language MLM (a language MLM (that identifies a topic, speaker's intent … para. [0016]; Initial training module 210 may also train custom MLMs 125 using custom (user-specific and/or user-provided) data 204…., para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) or custom (trained by a user of the pipeline) MLMs 125 …, para. [0045]),
the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection (a language MLM (that identifies a topic, speaker's intent … para. [0016]; para. [0017]; initial training module 210 may train MLMs 122 using initial data 202.…, para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) …, para. [0045], performing non-user specific intent detection as out-of-domain user intent detection), and
training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer (fig. 1, elements 102, 120, 160, 162; para. [0022]; training engine 160 may perform off-site training of pre-trained MLMs 122 whereas training engine 120 on computing device 102 may perform retraining of pre-trained MLMs 122 as well as training of new (custom) MLMs 125, para. [0023]; para. [0030]; para. [0035]; the user may initiate retraining of some of the MLMs included in the pipeline using training engine 120.…, para. [0039]; The selected MLMs may be pre-trained using a first set of training data, which may be previously supplied by the providers of the pipeline services or by the user…., para. [0045], decoupled training of models 122 and 125 as implying facilitation of independent performance optimization of the first classifier by the customer)
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Cohen with the method of Kuo in arriving at the missing features of Kuo, because such combination would have resulted in improving recognition of speech that may be encountered in a domain environment (Cohen, para. [0022]).
Per claim 2, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the determining comprises performing out-of-domain detection based on the first plurality of predefined user intents (para. [0052]; if the chat manager 202 determines that the prediction model 206 is unable to predict an intent of the user 104 …, para. [0066])
Per claim 3, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the first classifier is independent of the second classifier (fig. 7, elements 206, 208, 704, elements 206 or 204 as independent of element 704).
Per claim 4, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the first plurality of predefined user intents and the second plurality of predefined user intents are mutually exclusive (fig. 7, elements 206, 208, 704; para. [0072]).
Per claim 6, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the first classifier comprises an artificial neural network (the prediction model 208 may be a bidirectional encoder representations from transformations (BERT) model…., para. [0059]; the prediction model 206 may be a natural language model (e.g., a machine learning model) …, para. [0052]).
Per claim 8, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses: wherein the determining comprises, for each of the first plurality of predefined user intents: calculating a probability that the user query corresponds to the predefined user intent (para. [0042]-[0043]); and
comparing the calculated probability with a threshold, the threshold having a value that is dependent on the predefined user intent (an accuracy threshold may be used to determine whether the prediction model 206 is able or unable to determine an intent. Further, the threshold may be adjusted based on different factors, including the type of predicted intent and/or the user 140.…, para. [0043]).
Per claim 11, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the second classifier comprises an artificial neural network (The prediction model 704 may also be a bidirectional encoder representations from transformations (BERT) model …, para. [0071]).
Per claim 12, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the second processing comprises determining that the user query corresponds to a user intent of the second plurality of predefined user intents (para. [0072])
Per claim 13, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses: wherein the method comprises training the first classifier using first training data, the first training data associated with the first plurality of predefined user intents (para. [0071]),
wherein the second classifier has been trained using second, different training data, the second training data associated with the second plurality of predefined user intents, and wherein the first training data comprises a first amount of training data, and the second training data comprises a second, greater amount of training data (For example, the prediction models 206 and 208 may support only 15 intents, but the IVR system may support over 200 intents.…, para. [0071]; For example, the intent generation module 702 may use the prediction model 704 that has been initially trained with the third-party training data to determine (e.g., predict) an intent associated with the utterance 602. Since the prediction model 704 was trained using the third-party training data that includes intents not supported by the prediction models 206 and 208 …, para. [0072]-[0073]).
Per claim 14, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the first plurality of predefined user intents comprises a first number of user intents, and the second plurality of predefined user intents comprises a second, greater number of user intents (For example, the prediction models 206 and 208 may support only 15 intents, but the IVR system may support over 200 intents…., para. [0071]-[0072]).
Per claim 16, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the predetermined action comprises causing the virtual agent to provide an output for a user (para. [0053]; the chat manager 202 may provide the intent predicted by the prediction model 208 to the chat robot 204 and cause the chat robot 204 to provide automated responses to the user 104 via the online chat session 220.…, para. [0067]).
Per claim 17, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the received data comprises text data and/or audio data received via the virtual agent (para. [0049]-[0050])
Per claim 18, Kuo in view of Cohen discloses the method of claim 1,
Kuo discloses wherein the user query comprises a question and/or a command directed to the virtual agent (para. [0049]-[0050])
Per claim 19, Kuo discloses a computer program product comprising a non-transitory computer-readable storage medium having computer-readable instructions stored thereon, the computer-readable storage medium being executable by a computerized device to cause the computerized device to perform a method of processing user queries for a virtual agent, the method comprising:
receiving data representing a user query directed to the virtual agent (a service provider may utilize an online chat system that may include a chat robot (also referred to as a chat bot or simply a bot) to provide automated assistance to a customer in an online chat session …, para. [0015]; The chat interface 402 also includes an input portion 414 that enables the user 140 to input a message (e.g., an utterance that may include text data, audio data …, para. [0049]-[0050]);
determining whether the user query corresponds to any one of a first plurality of predefined user intents, wherein the determining comprises first processing the received data using a first classifier configured to classify user queries according to the first plurality of predefined user intents, the first classifier comprising a first artificial intelligence (AI) model trained with a first set of training data and to receive the user query data as input and output a predicted user intent associated with a class that the first classifier has been trained to recognize and a respective probability (para. [0051]; the prediction model 206 may be a natural language model (e.g., a machine learning model) configured to analyze natural language (e.g., a sentence, a phrase, a question, etc.) and to determine (e.g., predict) whether a message obtained from an online chat session is associated with any one of the set of possible intents … The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent …, para. [0052]; para. [0055]; the prediction model 208 may be a bidirectional encoder representations from transformations (BERT) model…., para. [0059]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067]; para. [0071], prediction models 206 or 208 as first classifier, machine learning and BERT model as AI models);
in response to a positive determination based on the probability from the first classifier meeting or exceeding a threshold, generating first output data based on a user intent of the first plurality of predefined user intents to which the user query is determined to correspond (The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent… In some embodiments, the prediction model 206 may select an intent for the message 432 from the set of possible intents having the highest score…, para. [0052]; para. [0053]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067], prediction models 206 or 208 as predicting and outputting intent having highest score exceeding lower scores/threshold);
in response to a negative determination based on the probability from the first classifier falling below the threshold, second processing the received data using a second classifier to generate second output data, the second classifier being configured to classify user queries according to a second plurality of predefined user intents, the second classifier comprising a second AI model trained with a second set of training data that is different from the first set of training data, the second classifier configured to receive the user query data as input and output a predicted user intent associated with a class that the second classifier has been trained to recognize and a respective probability (para. [0021]; In some embodiments, the chat manager 202 may determine that the prediction model 208 is unable to predict an intent of the user 104 based on the utterance (e.g., the utterance 532). For example, the chat manager 202 may determine that the prediction model is unable to predict an intent of the user 104 when none of the scores corresponding to the possible intents is above a probability threshold …, para. [0064]; On the other hand, if the chat manager 202 determines that the prediction model 206 is unable to predict an intent of the user 104 (e.g., the prediction model 206 outputs a null value or otherwise indicates that no intent is determined) …, para. [0066]-[0067]; para. [0071]; Since the prediction model 704 was trained using the third-party training data that includes intents not supported by the prediction models 206 and 208, the prediction model 704 may be able to determine an intent for the utterance 602 when the prediction models 206 and 208 failed to do the same…, para. [0072], prediction models 206 and 208 unable to predict an intent for utterance as negative determination, third party intents not supported by models 206 and 208 as second plurality of predefined user intents, prediction model 704 as second AI model, determined/output intent by prediction model 704 as second output data); and
performing a predetermined action in relation to the virtual agent based on either the first output data or the second output data (the prediction model 206 may output the intent to file a dispute of a transaction to the chat robot 204. The chat robot 204 may provide a dialogue to the user 104 and assist the user 104 based on the intent determined by the prediction model 206 …, para. [0053]; para. [0067], intent predicted by prediction model 206 or 208 as first output data used by chat robot to provided response to user)
Kuo does not explicitly disclose wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection, the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection, and training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer
However, these features are taught by Cohen:
wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection (a language MLM (that identifies a topic, speaker's intent … para. [0016]; Initial training module 210 may also train custom MLMs 125 using custom (user-specific and/or user-provided) data 204…., para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) or custom (trained by a user of the pipeline) MLMs 125 …, para. [0045]),
the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection (a language MLM (that identifies a topic, speaker's intent … para. [0016]; para. [0017]; initial training module 210 may train MLMs 122 using initial data 202.…, para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) …, para. [0045], performing non-user specific intent detection as out-of-domain user intent detection), and
training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer (fig. 1, elements 102, 120, 160, 162; para. [0022]; training engine 160 may perform off-site training of pre-trained MLMs 122 whereas training engine 120 on computing device 102 may perform retraining of pre-trained MLMs 122 as well as training of new (custom) MLMs 125, para. [0023]; para. [0030]; para. [0035]; the user may initiate retraining of some of the MLMs included in the pipeline using training engine 120.…, para. [0039]; The selected MLMs may be pre-trained using a first set of training data, which may be previously supplied by the providers of the pipeline services or by the user…., para. [0045], decoupled training of models 122 and 125 as implying facilitation of independent performance optimization of the first classifier by the customer)
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Cohen with the product of Kuo in arriving at the missing features of Kuo, because such combination would have resulted in improving recognition of speech that may be encountered in a domain environment (Cohen, para. [0022]).
2. Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Kuo in view of Cohen and Xu US 2022/0171947 A1 (“Xu”)
Per claim 20, Kuo discloses an apparatus comprising:
at least one processor (para. [0080]); and
at least one memory including computer program code, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the apparatus (para. [0080]) to:
receive data representing a user query directed to a virtual agent (a service provider may utilize an online chat system that may include a chat robot (also referred to as a chat bot or simply a bot) to provide automated assistance to a customer in an online chat session …, para. [0015]; The chat interface 402 also includes an input portion 414 that enables the user 140 to input a message (e.g., an utterance that may include text data, audio data …, para. [0049]-[0050]);
first process the received data using a first artificial neural network, ANN, the first ANN having been trained with a first set of training data and to classify user queries according to a first plurality of predefined user intents, wherein the first processing comprises performing, at the first ANN, out-of-domain detection based on the first plurality of predefined user intents, the first ANN trained to receive the user query data as input and output a prediction of whether the user query is out-of-domain and a respective probability (para. [0051]; the prediction model 206 may be a natural language model (e.g., a machine learning model) configured to analyze natural language (e.g., a sentence, a phrase, a question, etc.) and to determine (e.g., predict) whether a message obtained from an online chat session is associated with any one of the set of possible intents … The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent …, para. [0052]; para. [0055]; the prediction model 208 may be a bidirectional encoder representations from transformations (BERT) model…., para. [0059]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067]; para. [0071], prediction models 206 or 208 as first ANN, machine learning or BERT model as ANN models);
in response to the first ANN determining that the user query is not out-of-domain based on the probability from the first ANN, control the virtual agent based on a user intent of the first plurality of predefined user intents to which the user query is determined to correspond (The prediction model 206 may analyze the message 432 to determine if the message 432 includes one or more keywords (or keyword combinations such as “dispute” and “transaction”) associated with any one of the possible intents and may calculate a score for each possible intent… In some embodiments, the prediction model 206 may select an intent for the message 432 from the set of possible intents having the highest score…, para. [0052]; para. [0053]; para. [0064]; If the prediction model 208 is able to predict an intent (e.g., the prediction model 208 outputs an intent from the list of possible intents, the prediction model calculated a highest score for one of the intents …, para. [0067], prediction models 206 and 208 able to predict an intent for utterance as implying user query is not out-of-domain, predicted highest scores from models 206 and 209 as exceeding lower scores/threshold, intents with highest scores as output); and
in response to the first ANN determining that the user query is out-of-domain based on no probability scores exceeding a predetermined threshold: second process the received data using a second ANN, the second ANN having been trained with a second set of training data that is different from the first set of training data to classify user queries according to a second, different, plurality of predefined user intents, the second ANN trained to receive the user query data as input and output a predicted user intent associated with a class that the second classifier has been trained to recognize and a respective probability (para. [0021]; In some embodiments, the chat manager 202 may determine that the prediction model 208 is unable to predict an intent of the user 104 based on the utterance (e.g., the utterance 532). For example, the chat manager 202 may determine that the prediction model is unable to predict an intent of the user 104 when none of the scores corresponding to the possible intents is above a probability threshold (e.g., 60%, 70%, etc.).…, para. [0064]; para. [0066]-[0067]; If the chat manager 202 determines that the prediction model 208 is unable to predict an intent of the user 104 (e.g., the prediction model 208 outputs a null value, the highest score outputted by the prediction model 208 is below a threshold …, para. [0068]; para. [0071]; Since the prediction model 704 was trained using the third-party training data that includes intents not supported by the prediction models 206 and 208, the prediction model 704 may be able to determine an intent for the utterance 602 when the prediction models 206 and 208 failed to do the same…, para. [0072], prediction models 206 or 208 as first ANN, prediction models 206 and 208 unable to predict an intent for utterance (i.e., no score exceeding threshold) as implying user query is out-of-domain, prediction model 704 as second ANN, prediction model 704/second ANN as trained using different training data that includes intents not supported by prediction models 206 or 208/first ANN); and
control the virtual agent based on a user intent of the second plurality of predefined user intents to which the user query is determined to correspond (fig. 2, elements 204, 206, 208; para. [0017]; The chat manager 202 and/or the chat robot 204 may use the prediction model 206 to analyze the utterances and predict an intent of the user 140 …, para. [0042]; After discovering the new sub-intents, the process 800 then labels (at step 825) the utterances with the discovered sub-intents and retrains (at step 830) the first prediction model and/or the second prediction model using the labeled utterances. …, para. [0074], chat robot making predictions using re-trained prediction models as being controlled based on predicted output of prediction model 704)
Kuo does not explicitly disclose wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection, the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection, and training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer
However, these features are taught by Cohen:
wherein the first classifier is a k-class classifier trained by a customer and configured to perform in-domain, customer-specific user intent domain detection (a language MLM (that identifies a topic, speaker's intent … para. [0016]; Initial training module 210 may also train custom MLMs 125 using custom (user-specific and/or user-provided) data 204…., para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) or custom (trained by a user of the pipeline) MLMs 125 …, para. [0045]),
the second classifier is an n-class classifier trained by a service provider and configured to perform out-of-domain user intent detection (a language MLM (that identifies a topic, speaker's intent … para. [0016]; para. [0017]; initial training module 210 may train MLMs 122 using initial data 202.…, para. [0029]; a plurality of trained MLMs, which may be pre-trained MLMs 122 (e.g., by the providers of the pipeline services) …, para. [0045], performing non-user specific intent detection as out-of-domain user intent detection), and
training of the first classifier and the second classifier is decoupled, facilitating independent performance optimization of the first classifier by the customer (fig. 1, elements 102, 120, 160, 162; para. [0017]; para. [0022]; training engine 160 may perform off-site training of pre-trained MLMs 122 whereas training engine 120 on computing device 102 may perform retraining of pre-trained MLMs 122 as well as training of new (custom) MLMs 125, para. [0023]; para. [0030]; para. [0035]; the user may initiate retraining of some of the MLMs included in the pipeline using training engine 120.…, para. [0039]; The selected MLMs may be pre-trained using a first set of training data, which may be previously supplied by the providers of the pipeline services or by the user…., para. [0045], decoupled training of models 122 and 125 as implying facilitation of independent performance optimization of the first classifier by the customer)
Kuo in view of Cohen does not explicitly disclose the use of probability scores from a plurality of sigmoid functions corresponding to a respective one of the first plurality of predefined user intents
However, this feature is taught by Xu (para. [0037]-[0038]; The method includes determining, by the chatbot system, an intent for the one or more utterances or messages, using a set of binary classifiers (e.g., an n-binary classifier). Each binary classifier of the a set of binary classifiers includes a modified logit function for outputting a corresponding distance-based logit value. For example, the distance-based logit value may be calculated by using the modified logit function for a given class Ci …, para. [0047]; An activation function (e.g., a sigmoid function) can be applied to the distance-based logit value outputted by the modified logit function to generate a predicted output, in which the expected output identifies a probability predictive of whether the utterance corresponds to a respective class of the set of classes, para. [0172]; The logit value can be processed through an activation function to generate the probability value for the particular class. The probability value ranges between [0, 1]. The initialization of the machine-learning model can include defining a number of layers, a type of each layer (e.g., fully-connected, convolutional neural network), and a type of an activation function for each layer (e.g., sigmoid), para. [0181])
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Cohen with the apparatus of Kuo in arriving at the missing features of Kuo, as well as to combine the teachings of Xu with the apparatus of Kuo in view of Cohen in arriving at the missing features of Kuo in view of Cohen because such combination would have resulted in improving recognition of speech that may be encountered in a domain environment (Cohen, para. [0022]) as well as in calculating the confidence scores as measured by the probabilities of the classes/Intents being handled by a classifier, among the various function/methods of calculating such scores (Xu, para. [0043]; para. [0164]).
3. Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Kuo in view of Cohen as applied to claim 1 above, and further in view of Li et al US 2014/0101119 A1 (“Li”)
Per claim 5, Kuo discloses the method of claim 1,
Kuo does not explicitly disclose wherein the first plurality of predefined user intents overlaps with the second plurality of predefined user intents
However, this feature is taught by Li (The training set for the meta-classifier can have no overlap with the training set for domain classifiers … Alternatively, relative to the total number of items in the training set, the number of different items in the training set for the meta-classifier relative to the domain classifiers and/or corresponding subject matter domain classifier can be at least about 90%, or at least about 80%, or at least about 60%, or at least about 50%, para. [0031]; para. [0049])
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Li with the method of Kuo in view of Cohen in arriving at the missing features of Kuo in view of Cohen because such combination would have resulted in alternate methods of classifying data as a matter of design choice as well as in allowing separate domain/class searches based on particular queries (Li, para. [0013]; para. [0031]).
4. Claim 7 is rejected under 35 U.S.C. 103 as being unpatentable over Kuo in view of Cohen as applied to claim 1 above, and further in view of Xu
Per claim 7, Kuo discloses the method of claim 1,
Kuo does not explicitly disclose wherein the first classifier comprises a plurality of sigmoid functions, each of the plurality of sigmoid functions corresponding to a respective one of the first plurality of predefined user intents
However, this feature is taught by Xu (para. [0095]; The initialization of the machine-learning model can include defining a number of layers, a type of each layer (e.g., fully-connected, convolutional neural network), and a type of an activation function for each layer (e.g., sigmoid), para. [0196])
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Xu with the method of Kuo in view of Cohen in arriving at the missing features of Kuo in view of Cohen, because such combination would have resulted in calculating the confidence scores as measured by the probabilities of the classes/Intents being handled by a classifier, among the various function/methods of calculating such scores (Xu, para. [0043]; para. [0164]).
5. Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Kuo in view of Cohen as applied to claim 1 above, and further in view of Wohlwend US 2020/0151254 A1 (“Wohlwend”)
Per claim 9, Kuo discloses the method of claim 1,
Kuo does not explicitly disclose training the first classifier by minimizing a cumulative negative log-likelihood over the first plurality of predefined user intents.
However, this feature is taught by Wohlwend (para. [0054]-[0055])
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Wohlwend with the method of Kuo in view of Cohen in arriving at the missing features of Kuo in view of Cohen, because such combination would have resulted in ensuring the classifier/model is updated so that message training data in the same class/intent are moved closer to each other and training data in different classes are moved further apart (Wohlwend, para. [0054]-[0055]).
6. Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Kuo in view of Cohen as applied to claim 1 above, and further in view of Sackett et al US 2023/0147359 A1 (“Sackett”)
Per claim 10, Kuo discloses the method of claim 1,
Kuo does not explicitly disclose wherein the second classifier is configured to perform out-of-domain detection based on the second plurality of predefined user intents
However, this feature is taught by Sackett (para. [0006]; For example, the fastest classifier (e.g., the pattern matching classifier) may process the input phrase first. If a positive classification is made (e.g., one or more topics are identified, optionally with a confidence score above a threshold value), the process can simply move to using the identified one or more topics to generate a response. If a negative classification is made (e.g., no topics are identified and/or the confidence scores for any identified topics are below a threshold value), the process may proceed to the next classifier in line. If the next classifier renders a positive classification, it can be used to generate a response. If the next classifier renders a negative classification, the process can continue to the third classifier. The results of the third classifier can then be used to generate a response …, para. [0038], topics as intents)
It would have been obvious to one of ordinary skill in the art before the effective filing of the invention to combine the teachings of Sackett with the method of Kuo in view of Cohen in arriving at the missing features of Kuo in view of Cohen, because such combination would have resulted in providing conformation information as well as information used in the generation of a response (Sackett, para. [0077]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. See PTO 892 form.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 OLUJIMI A ADESANYA whose telephone number is (571)270-3307. The examiner can normally be reached Monday-Friday 8:30-5:00pm.
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, Richemond Dorvil can be reached at 571-272-7602. 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.
/OLUJIMI A ADESANYA/Primary Examiner, Art Unit 2658
Prosecution Insights