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 .
Acknowledgement
Acknowledgement is made of applicant’s amendment made on 07/08/2026. Applicant’s submission filed has been entered and made of record.
Status of the Claims
Claims 1-20 are pending.
Response to Applicant’s Arguments
In response to “The Applicant has amended claim 1 (and similarly, claim 19) to recite: "reducing memory requirements and improving storage input/output latency involved in performing multiple tasks by employing a single embedding model for determining embeddings of a set of user- specific embeddings and a set of content embeddings", such that the claims are directed to a concrete computer-functioning improvement (e.g., model structure that provides a reduced memory footprint with improved input/output latency characteristics), similar to the manner in which the patent eligible claims in Enfish L.L.C. v. Microsoft Corp. feature a structure that improves how a computer stores and retrieves data, with efficiency gains using a specific implementation. As such, the amended claim 1 is eligible under step one of the Alice test”.
Under Prong (2) of Step 2A, the goal is to determine whether the claim is directed to the recited exception by evaluating whether the claim as a whole integrates the recited judicial exception into a practical application of the exception. See MPEP 2106.04II(A).
In particular, evaluating integration into a practical application requires identifying whether there are any additional elements recited in the claim beyond the judicial exception and evaluating those additional elements, individually and in combination, to determine whether they integrate the exception into a practical application, using one or more of the considerations laid out by the Supreme Court and the Federal Circuit (“CAFC”). See MPEP 2106.04(d).
According to the Supreme Court, a patent may issue for the means or method of producing a certain result, or effect, and not for the result or effect produced. Diamond v. Diehr, 450 U.S. 175, 182 n. 7 (1981). Therefore, the focus is on whether the claim “focus on a specific means or method that improves the relevant technology or are instead directed to a result or effect that itself is the abstract idea and merely invoke generic processes and machinery”. Enfish, L.L.C. v. Microsoft Corp., 822 F.3d 1327, 1336 (Fed. Cir. 2016).
For example, in Enfish, the CAFC found it relevant to ask whether claims were directed to an improvement to computer functionality versus being directed to an abstract idea. Enfish, 822 F.3d at 1335. To that extent, the CAFC found that the claims were specifically directed to a self-referential table for a computer database. Id. at 1337. In particular, the claim language required a four step algorithm specifically directed to a self-referential table for a computer database that improved upon prior art information search and retrieval systems by employing a flexible, self-referential table to store data. Id. at 1336-37.
Therefore, the focus of the claims was on a specific asserted improvement in computer capabilities (i.e., the self-referential table for a computer database), not on economic or other tasks for which a computer was used in its ordinary capacity. Id. at 1336. See also MPEP 2106.04(d)I (“an improvement in the functioning of a computer or an improvement to other technology or technical field, as discussed in MPEP 2106.04(d)(1) and 2106.05(a)”).
Exemplary claim 1 recites a method comprising:
reducing memory requirements and improving storage input/output latency involved in performing multiple tasks by employing a single embedding model for determining embeddings of a set of user-specific embeddings and a set of content embeddings, wherein employing the single embedding model comprises:
receiving a plurality of user-specific text strings associated with a user;
using the single embedding model comprising a natural language processing (NLP) model, generating the set of user-specific embeddings, comprising, for each user-specific text string of the plurality: generating a respective user-specific embedding based on the user-specific text string, wherein the respective user-specific embedding preserves semantic language information from the user-specific text string;
generating a first set of content embeddings of the set of content embeddings using the single embedding model, comprising, for each content block of a set of content blocks, generating a respective content embedding associated with the content block;
based on a set of semantic similarity metrics determined between the set of user- specific embeddings and the set of content embeddings, determining a ranked list of content blocks selected from the set of content blocks; and
based on the ranked list, providing a first content block to the user within a digital mental health platform, in response to a mental health state of the user.
According to the specification US 2024/0411996 A1 at ¶26: “Third, some such variants can additionally or alternatively function to reduce memory requirements by employing a single embedding model for determining embeddings of multiple different inputs, such as for determining both user-specific embeddings (e.g., embeddings of user-specific inputs such as text strings) and content embeddings (e.g., embeddings associated with content blocks, such as embeddings of content titles, descriptions, and/or excerpts); additionally or alternatively, a single embedding model (e.g., multi-lingual embedding model, such as a language-agnostic embedding model) may be used for determining embeddings of inputs in multiple languages. In using only a single embedding model for multiple tasks (e.g., determining embeddings of different types of inputs and/or of different language inputs), memory requirements can be reduced as compared with using two (or more) different embedding models (e.g., wherein both/all embedding models would need to be kept in memory to avoid undesirable storage I/O latency from switching between use of the different models). Further, a person of skill in the art will recognize that the use of a particular type of model (e.g., transformer model, embedding model, semantic embedding model, multi-lingual embedding model, etc.) confers specific structural elements to the model used”.
At a minimum, using a single embedding model for embedding user specific text strings and content and improves search and retrieval of content semantically similar to the user specific text strings by avoiding switching between use of different embedding models, resulting in reduced memory requirements.
Therefore, like the specifically asserted self-referential table for improving database search and retrieval in Enfish, the instant claims specifically asserted the application of a single embedding model to generate user-specific embedding and content embedding that reduces memory requirement when performing search and retrieval of contents semantically similar to the user specific text strings.
Therefore, Claims 1-20 are patent eligible.
In response to “The Applicant has amended independent claim 1, as indicated above, and respectfully asserts that Devesa fails to teach or anticipate the "single embedding model" architecture recited in claims 1 (and similarly, claim 19). Instead, Devesa teaches multiple structurally distinct encoding pipelines (i.e., not a shared model), including 1) a user-query pipeline (see, for example, Devesa paragraphs [0047]-[0048]) and 2) a content/corpus pipeline (see, for example, Devesa paragraphs [0060]-[0063], where separated RNNs are highlighted (e.g., FIG. 7 and paragraph[oo21] of Devesa). In more detail, Devesa requires and explicitly teaches architecture where the content side relies on pre-stored word-level graph representations of the medical corpus, which are structurally and operationally separate from the RNN-based user embedding pipeline”.
Devesa teaches a medical assistant system 206 creating word embedding of sentences present in a health profile of a user 102 to select one or more relevant answers from a corpus of medical triage conversation (¶60); i.e., the medical assistant system 206 creates / generates a first set of content embeddings.
Further, the medical assistant system 206 converts user enquiry into word embedding by converting tokenized sentence of user enquiry into one-hot vector representation (¶48); i.e., the medical assistant system 206 creates / generates a set of user-specific embeddings based on user specific text strings of the user enquiry.
Here, the medical assistant system 206 is based on an RNN architecture (Fig. 3, RNN model comprising a recurrent neural network 306 and a fully connected -> TANH activation layer to generate user enquiry embedding 310 per ¶49 and to generate sentence embedding 310 per ¶107) that encodes input sequence into a state vector that compresses semantic information and a second RNN decodes the semantic information to produce one or more relevant answers (¶73).
In other words, Devesa teaches using a single (i.e., the same) RNN-encoder based embedding model to encode input sequences in order to create (1) the first set of content embeddings and (2) the set of user-specific embeddings.
While the medical assistant system 206 does map the graph of user enquiry with graph present in a corpus of medical triage conversation (¶51), the medical assistant system 206 created the graph of the user enquiry based on the conversion of the user enquiry into the word embedding to show relation between the words used by the user in the user enquiry present in the word embedding (¶50).
In other words, by using the RNN architecture / single embedding model to convert the set of user-specific text strings into the set of user specific embedding, the single embedding model thereby created the graph showing relation between user enquiry words present in the word embedding.
Nonstatutory Double Patent Rejections
The nonstatutory double patenting rejection is based on a judicially created doctrine grounded in public policy (a policy reflected in the statute) so as to prevent the unjustified or improper timewise extension of the "right to exclude" granted by a patent and to prevent possible harassment by multiple assignees. See In re Goodman, 11 F.3d 1046, 29 USPQ2d 2010 (Fed. Cir. 1993); In re Longi, 759 F.2d 887, 225 USPQ 645 (Fed. Cir. 1985); In re Van Ornum, 686 F.2d 937, 214 USPQ 761 (CCPA 1982); In re Vogel, 422 F.2d 438, 164 USPQ 619 (CCPA 1970); and In re Thorington, 418 F.2d 528, 163 USPQ 644 (CCPA 1969).
A timely filed terminal disclaimer in compliance with 37 CFR 1.321(c) may be used to overcome an actual or provisional rejection based on a nonstatutory double patenting ground provided the conflicting application or patent is shown to be commonly owned with this application. See 37 CFR 1.130(b).
Effective January 1, 1994, a registered attorney or agent of record may sign a terminal disclaimer. A terminal disclaimer signed by the assignee must fully comply with 37 CFR 3.73(b).
The USPTO Internet website contains terminal disclaimer forms which may be used. Please visit www.uspto.gov/forms/. The filing date of the application will determine what form should be used. A web-based eTerminal Disclaimer may be filled out completely online using web-screens. An eTerminal Disclaimer that meets all requirements is auto-processed and approved immediately upon submission. For more information about eTerminal Disclaimers, refer to www.uspto.gov/patents/process/file/efs/guidance/eTD-info-I.jsp.
Claims 1, 4, 10-13, and 16-18 are rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 in view of Maitra et al. (US 11854540 B2).
The limitations of claim 1 of US 12099808 B2 (“model for natural language processing, the model comprising a trained semantic embedding model” being the single embedding model that reduces memory requirements and improving storage input / output latency involved in performing multiple tasks) fully encompasses the limitations of claims 1 and 10 in the instant application except based on the ranked list, providing a first content block to the user in a digital health platform, in response to a health state of the user.
Maitra teaches a digital mental health platform (Col 4, Rows 14-18, machine learning models based wellness system to improve mental wellness by providing empathetic advisement) using an encoder-decoder bidirectional LSTM model to process user specific text strings and to provide response to a mental health state of the user (Col 11, Rows 9-19, Rows 35-42, and Fig. 1E, using bidirectional LSTM model to process text, audio, and video of the user to classify an act a speaker is performing (e.g., Col 4, Row 65 – Col 5, Row 4, identify an affect on the user including pain, haste, engagement level, despair, longing, fondness etc.) in order to generate contextual conversation data; Col 11, Rows 43-50, wellness system performs one or more actions to provide an empathetic, context-aware, multimodal conversational agent to assist mental wellness of the user based on the contextual conversation data).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to implement the digital health platform as a digital mental health platform to provide response to a mental health state of the user in order to improve mental wellness by providing empathetic advisement in a proactive, personalized, contextual, and guided manner (Devesa, Col 4, Rows 14-18).
The limitations of claim 4 of US 12099808 B2 fully encompasses the limitations of claim 4 in the instant application.
The limitations of claim 5 of US 12099808 B2 fully encompasses the limitations of claim 17 in the instant application.
The limitations of claim 8 of US 12099808 B2 fully encompasses the limitations of claim 11 in the instant application.
The limitations of claim 9 of US 12099808 B2 fully encompasses the limitations of claim 12 in the instant application.
The limitations of claim 10 of US 12099808 B2 fully encompasses the limitations of claim 13 in the instant application.
The limitations of claim 11 of US 12099808 B2 fully encompasses the limitations of claim 18 in the instant application.
The limitations of claim 13 of US 12099808 B2 fully encompasses the limitations of claim 16 in the instant application.
Claim 2 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Chang et al. (US 2020/0152304 A1).
The limitations of claims in US 12099808 B2 do not disclose wherein the first content block is structured to provide a meditation exercise.
Chang discloses generating personalized and effective mental health therapies and recommendations (¶2) by structing content block to provide a meditation exercise (¶77).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a meditation exercise in order to generate personalized and effective mental health therapies (Chang, ¶2).
Claim 3 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of in view of Misrilall et al. (US 2022/0093253 A1).
The limitations of claims in US 12099808 B2 do not disclose wherein the first content block is structured to provide a breathing exercise.
Misrilall discloses performs natural language processing to analyze inputs from a target patient to generate a content block (Abstract) that is structured to provide a breathing exercise (¶80).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a breathing exercise in order to make specific type of content available to the user (Misrilall, ¶98).
Claim 5 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 4, in view of Bouyarmane (US 11797530 B1).
The limitations of claims in US 12099808 B2 do not disclose wherein a set of embeddings of the multi-lingual embedding model is language agnostic.
Bouyarmane discloses using language agnostic multi-lingual embedding architecture-based NLP model to generate embeddings for entity records expressed in different languages (Col 8, Rows 45-54, entity records expressed in multiple languages; Col 10, Rows 1-12, use hierarchical embedding models to generate embeddings at word level and entity record level; in view of Col 6, Rows 18-20 and Col 10, Rows 15-16, hierarchical embedding model comprises neural network based models at individual layers of the hierarchy such as bidirectional long short term memory units “BiLSTMs”).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to use a NLP model comprising language agnostic multi-lingual embedding architecture (compare Devesa, ¶48, using RNN-LSTM architecture for generating embeddings with Bouyarmane, Col 6, Rows 18-20, HEM comprising bidirectional LSTMs) to generate the set of user specific embeddings (Devesa, ¶62) in order to generate respective semantic similarity metrics between the user specific embeddings and set of content embeddings (i.e., entity records) expressed in different languages (Bouyarmane, Col 5, Rows 20-24).
Claim 6 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2), Bouyarmane (US 11797530 B1) as applied to claim 5, in view of Aikawa et al. (US 2019/0354589 A1).
The limitations of claims in US 12099808 B2 do not disclose wherein the subset of the set of embeddings encode the shared semantic meaning and are characterized by the cosine similarity greater than 0.8.
Aikawa discloses collecting word embedding corresponding to word / phrase and to determine words / phrases of similar expression (Abstract) based on a cosine similarity greater than 0.8 (¶213).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to characterize embeddings encode shared semantic meaning to have a cosine similarity greater than 0.8 to ensure that similarity measure of two matching embedding is maximal (Devesa, ¶63).
Claims 7-8 are rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Devesa (US 2020/0211709 A1).
Regarding Claims 7-8 in the instant application, the limitations of claims in US 12099808 B2 do not disclose wherein the NLP model is trained on monolingual word embeddings and wherein the NLP model is trained on a pseudo-cross-lingual corpus including mixed contents of different languages.
Devesa discloses wherein the NLP model is trained on monolingual word embeddings (¶63, the recurrent neural network is trained in such a manner that similarity measure of two matching embedding is maximal where a single training input includes a question, its correct answer, and a randomly chosen incorrect answer; ¶69, the medical assistant 206 trained in any one of the one or more languages; i.e., the RNN is trained using training input comprising matching embeddings in at least one language) and wherein the NLP model is trained on a pseudo-cross-lingual corpus including mixed contents of different languages (¶63, the recurrent neural network is trained in such a manner that similarity measure of two matching embedding is maximal where a single training input includes a question, its correct answer, and a randomly chosen incorrect answer; ¶69, the medical assistant 206 trained in any one of the one or more languages; i.e., the RNN is trained using training input comprising matching embeddings in one or more languages).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to training the mono-lingual word embeddings and on a pseudo-cross-lingual corpus including mixed contents of different languages in order to enable dialogue with user in one or more languages (Devesa, ¶69).
Claim 9 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Kalns et al. (US 2014/0310001 A1).
The limitations of claims in US 12099808 B2 do not disclose minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with the ranked list of content blocks in response to completion of a conversation with the user, caching the recommendation in a fast retrieval database, and refreshing a user application with the recommendation when the user next opens the user application.
Kalns discloses a virtual personal assistant (Abstract) implementing machine learning model (¶18, analyzing user’s past interaction with VPA based on historical record of natural language dialog of the user and the VPA application; ¶32 and ¶49, implement a VPA model to provide dialog models) minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with a list of content blocks (¶47, automatically search a list of products as gift for another person based on understanding the likely intent of user’s gaze or current time / date as “SearchProduct (recipient = other)”) in response to completion of a conversation with the user (¶108, determine a previously concluded conversation), caching the recommendation in a fast retrieval database (¶64, dialog context 212 implements a data structure to keep track of intent history for a current dialog session; ¶70, intents from dialog context 212 can be retained temporarily in a cache memory), and refreshing a user application with the recommendation when the user next opens the user application (¶108, when resuming a previously concluded conversation, ask whether the user would like to return to looking for a gift for his or her daughter).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with the ranked list of content blocks in response to completion of a conversation with the user, caching the recommendation in a fast retrieval database, and refreshing a user application with the recommendation when the user next opens the user application in order to make recommendations based on multi-modal inputs other than user natural language input such as user’s gaze or current time / date (Kalns, ¶47) and to make recommendations based on inferences and contextual information derived from dialog context (Kalns, ¶108) being tracked in a computerized data structure (Kalns, ¶64) extracted into a cache memory (Kalns, ¶70).
Claim 14 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 13, in view of Kurowski et al. (US 2016/0071432 A1).
The limitations of claims in US 12099808 B2 do not disclose wherein the first conversational session comprises a first conversation between a coach and the user.
Kurowski teaches a health and wellness management system generating recommendations to a user (Abstract) in a conversational session comprising a conversation between a coach and the user (¶296, primary care provider being a health coach).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to provide a health coach in order to provide virtual primary care providers as medical practitioner (Kurowski, ¶296).
Claim 15 is rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 and Maitra et al. (US 11854540 B2) as applied to claim 13, in view of Kurowski et al. (US 2016/0071432 A1).
The limitations of claims in US 12099808 B2 do not disclose wherein the first conversational session satisfies a minimum conversation length criterion.
Ladkat discloses classifying high quality conversational session based on a conversational session satisfying a minimum conversation length criterion (¶48).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to classify first conversational session satisfying a minimum conversation length criterion as a high quality conversation (Ladkat, ¶48).
Claims 19-20 are rejected under the judicially created doctrine of obviousness-type double patenting as being unpatentable over the claims of US 12099808 B2 in view of Bouyarmane (US 11797530 B1) and Chang et al. (US 2020/0152304 A1).
The limitations of claim 1 of US 12099808 B2 (“model for natural language processing, the model comprising a trained semantic embedding model” being the single embedding model that reduces memory requirements and improving storage input / output latency involved in performing multiple tasks) encompasses the limitations of claim 19 in the instant application except trained semantic embedding model does not include a language agnostic multilingual embedding architecture and wherein the first content block is structured to provide a meditation exercise.
Bouyarmane discloses using language agnostic multi-lingual embedding architecture-based NLP model to generate embeddings for entity records expressed in different languages (Col 8, Rows 45-54, entity records expressed in multiple languages; Col 10, Rows 1-12, use hierarchical embedding models to generate embeddings at word level and entity record level; in view of Col 6, Rows 18-20 and Col 10, Rows 15-16, hierarchical embedding model comprises neural network based models at individual layers of the hierarchy such as bidirectional long short term memory units “BiLSTMs”).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to use a semantic embedding NLP model comprising language agnostic multi-lingual embedding architecture (compare Devesa, ¶48, using RNN-LSTM architecture for generating embeddings with Bouyarmane, Col 6, Rows 18-20, HEM comprising bidirectional LSTMs) to generate the set of user specific embeddings (Devesa, ¶62) in order to generate respective semantic similarity metrics between the user specific embeddings and set of content embeddings (i.e., entity records) expressed in different languages (Bouyarmane, Col 5, Rows 20-24).
Further, Chang discloses generating personalized and effective mental health therapies and recommendations (¶2) by structing content block to provide a meditation exercise (¶77).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a meditation exercise in order to generate personalized and effective mental health therapies (Chang, ¶2).
The limitations of claim 12 of US 12099808 B2 fully encompasses the limitations of claim 20 in the instant application.
Claim Rejections - 35 USC § 103
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 103 that form the basis for the rejections under this section made in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 4, 7-8, 10-13 and 17 are rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) in view of Maitra et al. (US 11854540 B2).
Regarding Claim 1, Devesa discloses a method comprising:
reducing memory requirements and improving storage input/output latency involved in performing multiple tasks by employing a single embedding model (Fig. 3 shows an RNN architecture comprising a recurrent neural network 306 and a fully connected TANH activation layer 308; ¶73, RNN encodes input sequence into a state vector that compresses semantic information) for determining embeddings of a set of user-specific embeddings (¶¶48-49, medical assistant system 206 has a RNN based architecture to create user enquiry word embedding 310 in accordance to the architecture of Fig. 3) and a set of content embeddings (¶60, medical assistant system 206 uses the same RNN based architecture to create word embeddings of sentences present in the health profile of the user 102 to create sentence embedding 310 per ¶107), wherein employing the single embedding model comprises:
receiving a plurality of user-specific text strings associated with a user (¶45 and ¶47, receiving user enquiry comprising text strings and converting text strings into tokens);
using the single embedding model comprising a natural language processing (NLP) model (Fig. 3 and ¶49, RNN architecture generates user enquiry word embedding 310), generating the set of user-specific embeddings (¶47, convert text string of user enquiry into tokens; ¶48, converting tokens into word embedding using RNN-LSTM based model), comprising, for each user-specific text string of the plurality: generating a respective user-specific embedding based on the user-specific text string, wherein the respective user-specific embedding preserves semantic language information from the user-specific text string (¶73, RNN reads input sequence (user enquiry or the user symptoms) one word at a time, encodes the input sequence into a state vector that compresses semantic information);
generating a first set of content embeddings of the set of content embeddings using the single embedding model (Fig. 3 and ¶107, using the same RNN architecture to generate sentence embedding 310), comprising, for each content block of a set of content blocks, generating a respective content embedding associated with the content block (¶60, creating word embedding of words presented in a health profile of the user / corpus of medical triage conversation);
based on a set of semantic similarity metrics determined between the set of user-specific embeddings and the set of content embeddings, determining a ranked list of content blocks selected from the set of content blocks (¶63 in view of ¶59, using cosine-similarity / rank loss function to match word embedding of user enquiry with answers in user health profile to select one or more relevant answers (i.e., relevant answers with respective cosine-similarity measure)); and
based on the ranked list, providing a first content block to the user (¶64 in view of ¶59, display one or more relevant answers for the user enquiry based on the selection of the one or more relevant answers) in a digital health platform, in response to a health state of the user (¶41, a bi-directional conversation between medical assistant system 206 and user 102 corresponding to a medical knowledge based dialogue based on a health profile of user 102 per ¶42: create a corpus of medical triage conversation by creating word embeddings of words present in user health profile per ¶60 and map user enquiry word embedding to word embeddings present in the corpus of medical triage conversation to generate relevant answer to the user enquiry per ¶53).
Devesa does not teach the digital health platform is a digital mental health platform that responds to a mental health state of the user.
Maitra teaches a digital mental health platform (Col 4, Rows 14-18, machine learning models based wellness system to improve mental wellness by providing empathetic advisement) using an encoder-decoder bidirectional LSTM model to process user specific text strings and to provide response to a mental health state of the user (Col 11, Rows 9-19, Rows 35-42, and Fig. 1E, using bidirectional LSTM model to process text, audio, and video of the user to classify an act a speaker is performing (e.g., Col 4, Row 65 – Col 5, Row 4, identify an affect on the user including pain, haste, engagement level, despair, longing, fondness etc.) in order to generate contextual conversation data; Col 11, Rows 43-50, wellness system performs one or more actions to provide an empathetic, context-aware, multimodal conversational agent to assist mental wellness of the user based on the contextual conversation data).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to implement the digital health platform as a digital mental health platform to provide response to a mental health state of the user in order to improve mental wellness by providing empathetic advisement in a proactive, personalized, contextual, and guided manner (Devesa, Col 4, Rows 14-18).
Regarding Claim 4, Devesa discloses wherein the NLP model comprises a multi-lingual embedding model trained using user-specific text data in multiple languages (¶62, medical assistant system 206 creates sentence embedding of entire sentences in the plurality of dialogue conversations and use the sentence embedding to produce end to end trainable neural translation models with decoder that uses global embedding to generate natural language from different language embedding; ¶69, medical assistant system 206 trained in one or more languages of the dialogue with the user).
Regarding Claim 7, Devesa discloses wherein the NLP model is trained on monolingual word embeddings (¶63, the recurrent neural network is trained in such a manner that similarity measure of two matching embedding is maximal where a single training input includes a question, its correct answer, and a randomly chosen incorrect answer; ¶69, the medical assistant 206 trained in any one of the one or more languages; i.e., the RNN is trained using training input comprising matching embeddings in at least one language).
Regarding Claim 8, Devesa discloses wherein the NLP model is trained on a pseudo-cross-lingual corpus including mixed contents of different languages (¶63, the recurrent neural network is trained in such a manner that similarity measure of two matching embedding is maximal where a single training input includes a question, its correct answer, and a randomly chosen incorrect answer; ¶69, the medical assistant 206 trained in any one of the one or more languages; i.e., the RNN is trained using training input comprising matching embeddings in one or more languages).
Regarding Claim 10, Devesa discloses after providing the first content block to the user:
receiving a second plurality of user-specific text strings associated with the user (¶65, updates user enquiry after receiving updated user enquiry from the user based on selection from the one or more relevant answers displayed to the user 102; i.e., a new set of user enquiry);
using the NLP model, generating a second set of user-specific embeddings, comprising, for each user-specific text string of the second plurality: generating a respective user-specific embedding based on the user-specific text string (¶47, convert text string of user enquiry into tokens; ¶48, converting tokens into word embedding using RNN-LSTM based model; ¶62, using recurrent neural networks to generate sentence embedding);
based on the second set of user-specific embeddings and a second set of content embeddings ((¶60, creating word embedding of words presented in a health profile of the user / corpus of medical triage conversation)), determining a second ranked list of content blocks selected from the set of content blocks (¶51, map word bedding of user enquiry with word embedding of question words in the medical triage conversation corpus; per ¶63, each question-answer pair was prepared using cosine similarity ranking loss function); and
selecting a second content block based on the second ranked list (¶59, select one or more relevant answers for user enquiry based on mapping and health profile of the user with confidence level / probability of one or more relevant answers being the correct answer to the user enquiry); and
providing the second content block to the user (¶64 in view of ¶59, display one or more relevant answers for the user enquiry based on the selection of the one or more relevant answers).
Regarding Claim 11, Devesa discloses generating a second set of content embeddings, comprising, for each content block of a second set of content blocks, generating a respective content embedding associated with the content block (¶51, identify graph with same structure of the user enquiry in the corpus of medical training dataset by mapping word embedding of the user enquiry with word embedding of question words present in the corpus of medical triage conversation).
Regarding Claim 12, Devesa discloses wherein the second set of content embeddings is equivalent to the first set of content embeddings (¶59, select one or more relevant answers to the user based on one or more relevant answers being the correct answer to the user enquiry; i.e., a first answer and a second answer are equivalent to the extent that they are correct answers to the user enquiry; see e.g., ¶¶88-89 and ¶91, answer rank 1.0 and answer rank 3.0 corresponding to amoxicillin are equivalent / correct answers to the enquiry “What are the symptoms of the flu?”).
Regarding Claim 13, Devesa discloses wherein receiving the plurality of user-specific text strings comprises receiving a set of messages of a first conversational session with the user (¶41, user 102 interacts with medical assistant system 206 in a bi-directional conversation; e.g., medical assistant 206 enquires the user “How are you feeling today”? and user initially enquires a question from the medical assistant system 206), and selecting the plurality of user-specific text strings from the set of messages (¶47, insert delineation tokens between the context and utterances to distinguish between responses of the medical assistant system 206 and the user 102; i.e., distinguish and select tokens of the user enquiry for conversion into word embedding per ¶48).
Regarding Claim 17, Devesa discloses wherein the first content block is provided to the user upon generating similarity metrics from the set of messages, and identifying the first content block based upon the similarity metrics corresponding to the set of messages (¶51, map word bedding of user enquiry with word embedding of question words in the medical triage conversation corpus; ¶59, select one or more relevant answers for user enquiry based on mapping and health profile of the user with confidence level / probability of one or more relevant answers being the correct answer to the user enquiry; per ¶63, each question-answer pair was prepared using cosine similarity ranking loss function).
Claim 2 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Chang et al. (US 2020/0152304 A1).
Regarding Claim 2, Devesa does not disclose wherein the first content block is structured to provide a meditation exercise.
Chang discloses generating personalized and effective mental health therapies and recommendations (¶2) by structing content block to provide a meditation exercise (¶77).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a meditation exercise in order to generate personalized and effective mental health therapies (Chang, ¶2).
Claim 3 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Misrilall et al. (US 2022/0093253 A1).
Regarding Claim 3, Devesa does not disclose wherein the first content block is structured to provide a breathing exercise.
Misrilall discloses performs natural language processing to analyze inputs from a target patient to generate a content block (Abstract) that is structured to provide a breathing exercise (¶80).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a breathing exercise in order to make specific type of content available to the user (Misrilall, ¶98).
Claim 5 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Bouyarmane (US 11797530 B1).
Regarding Claim 5, Devesa does not disclose wherein a set of embeddings of the multi-lingual embedding model is language agnostic.
Bouyarmane discloses using language agnostic multi-lingual embedding architecture-based NLP model to generate embeddings for entity records expressed in different languages (Col 8, Rows 45-54, entity records expressed in multiple languages; Col 10, Rows 1-12, use hierarchical embedding models to generate embeddings at word level and entity record level; in view of Col 6, Rows 18-20 and Col 10, Rows 15-16, hierarchical embedding model comprises neural network based models at individual layers of the hierarchy such as bidirectional long short term memory units “BiLSTMs”).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to use a NLP model comprising language agnostic multi-lingual embedding architecture (compare Devesa, ¶48, using RNN-LSTM architecture for generating embeddings with Bouyarmane, Col 6, Rows 18-20, HEM comprising bidirectional LSTMs) to generate the set of user specific embeddings (Devesa, ¶62) in order to generate respective semantic similarity metrics between the user specific embeddings and set of content embeddings (i.e., entity records) expressed in different languages (Bouyarmane, Col 5, Rows 20-24).
Claim 6 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) in view of Maitra et al. (US 11854540 B2) and Bouyarmane (US 11797530 B1) as applied to claim 5, in further view of Aikawa et al. (US 2019/0354589 A1).
Regarding Claim 6, Devesa discloses wherein a subset of the set of embeddings encode a shared semantic meaning and are characterized by a cosine similarity (¶63, medical assistant system 206 utilizes embedding functions for embedding context and utterance pairs using ranking loss function including cosine similarity measure for preparing the one or more relevant answers where the RNN is trained to ensure similarity measure of two matching embedding is maximal)
Devesa does not disclose wherein the subset of the set of embeddings encode the shared semantic meaning and are characterized by the cosine similarity greater than 0.8.
Aikawa discloses collecting word embedding corresponding to word / phrase and to determine words / phrases of similar expression (Abstract) based on a cosine similarity greater than 0.8 (¶213).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to characterize embeddings encode shared semantic meaning to have a cosine similarity greater than 0.8 to ensure that similarity measure of two matching embedding is maximal (Devesa, ¶63).
Claim 9 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Kalns et al. (US 2014/0310001 A1).
Regarding Claim 9, Devesa as modified by Maitra discloses minimizing a computational cost and reducing latency for the user by automatically generating (Maitra, Col 13, Rows 21-41, wellness system utilizes machine learning models to generate automated empathetic conversations that, in turn, conserves computing resources, networking resources, and human resources) a recommendation associated with the ranked list of content blocks (Maitra, Col 11, Rows 27-43, bidirectional LSTM RNN model processes text, audio, video, context, and response from the user to determine dialogue act classification data for contextual conversation data; e.g., Col 13, Rows 12-20, act as a friend or companion to keep the user in good spirits like recommending yoga or enroll in a flying class).
Devesa as modified by Maitra does not disclose minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with the ranked list of content blocks in response to completion of a conversation with the user, caching the recommendation in a fast retrieval database, and refreshing a user application with the recommendation when the user next opens the user application.
Kalns discloses a virtual personal assistant (Abstract) implementing machine learning model (¶18, analyzing user’s past interaction with VPA based on historical record of natural language dialog of the user and the VPA application; ¶32 and ¶49, implement a VPA model to provide dialog models) minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with a list of content blocks (¶47, automatically search a list of products as gift for another person based on understanding the likely intent of user’s gaze or current time / date as “SearchProduct (recipient = other)”) in response to completion of a conversation with the user (¶108, determine a previously concluded conversation), caching the recommendation in a fast retrieval database (¶64, dialog context 212 implements a data structure to keep track of intent history for a current dialog session; ¶70, intents from dialog context 212 can be retained temporarily in a cache memory), and refreshing a user application with the recommendation when the user next opens the user application (¶108, when resuming a previously concluded conversation, ask whether the user would like to return to looking for a gift for his or her daughter).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to minimizing a computational cost and reducing latency for the user by automatically generating a recommendation associated with the ranked list of content blocks in response to completion of a conversation with the user, caching the recommendation in a fast retrieval database, and refreshing a user application with the recommendation when the user next opens the user application in order to make recommendations based on multi-modal inputs other than user natural language input such as user’s gaze or current time / date (Kalns, ¶47) and to make recommendations based on inferences and contextual information derived from dialog context (Kalns, ¶108; compare Devesa, ¶40, provide relevant answers to the user based on user context) being tracked in a computerized data structure (Kalns, ¶64) extracted into a cache memory (Kalns, ¶70).
Claim 15 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 1, in view of Ladkat et al. (US 2022/0172220 A1).
Regarding Claim 15, Devesa does not disclose wherein the first conversational session satisfies a minimum conversation length criterion.
Ladkat discloses classifying high quality conversational session based on a conversational session satisfying a minimum conversation length criterion (¶48).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to classify first conversational session satisfying a minimum conversation length criterion as a high quality conversation (Ladkat, ¶48).
Claims 16 and 18 are rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claims 1 and 13, in view of Cai et al. (US 2019/0349321 A1).
Regarding Claim 16, Devesa does not teach identifying an important subset of messages from the set of messages upon processing the set of messages with a classifier.
Cai discloses a system (Fig. 1), comprising: receiving, from a client device, a user query (¶49, processor 104 can receive a user query and parse the user query) and using a classifier to identify an important subset of user query to process the user query (¶123, processor 104 can remove stop words, count frequency of stemmed words, sort the frequencies, rank each sentence based on keywords it contains, and select sentences based on the ranked sentences).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to implement a classifier to process the set of messages to identify an important subset of the messages from the set of messages in order to token text data and remove stop words (Cai, ¶123).
Upon such processing, the established function of Devesa generates the set of user-specific embeddings from the important subset of messages (Devesa ¶47, convert text string of user enquiry into tokens; ¶48, converting tokens into word embedding using RNN-LSTM based model; ¶62, using recurrent neural networks to generate sentence embedding) without stop words.
Regarding Claim 18, Devesa does not teach wherein receiving the plurality of user-specific text strings comprises receiving a summary of a first conversational session with the user, and selecting the plurality of user-specific text strings from the summary.
Cai discloses a system (Fig. 1), comprising: receiving, from a client device, a user query / user specific text strings (¶49, processor 104 can receive a user query and parse the user query) and generating a summary of user query (¶49 and ¶51, processor 104 can receive a tuple or sequence of elements based on a parsed user query; ¶¶15-16, tokenizing data representation of user inputted text to output a summary of user inputted text).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to generate and transmit a summary of a first conversational session with a user and select the user specific text strings from the summary in order to limit the report summary to a number of sentences or words (Cai, ¶121).
Claim 14 is rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) and Maitra et al. (US 11854540 B2) as applied to claim 13, in view of Kurowski et al. (US 2016/0071432 A1).
Regarding Claim 14, Devesa discloses wherein the first conversational session comprises a first conversation between a medical practitioner and the user (¶51 and ¶70, past interaction history of the user includes conversations between a user and a professional medical practitioner).
Devesa does not disclose the medical practitioner is a coach.
Kurowski teaches a health and wellness management system generating recommendations to a user (Abstract) in a conversational session comprising a conversation between a coach and the user (¶296, primary care provider being a health coach).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to provide a health coach as medical practitioner in order to provide virtual primary care providers as medical practitioner (Kurowski, ¶296).
Claims 19-20 are rejected under 35 USC 103(a) as being unpatentable over Devesa (US 2020/0211709 A1) in view of Bouyarmane (US 11797530 B1) and Chang et al. (US 2020/0152304 A1).
Regarding Claim 19, Devesa discloses a method comprising:
reducing memory requirements and improving storage input/output latency involved in performing multiple tasks by employing a single embedding model (Fig. 3 shows an RNN architecture comprising a recurrent neural network 306 and a fully connected TANH activation layer 308; ¶73, RNN encodes input sequence into a state vector that compresses semantic information) for determining embeddings of a set of user-specific embeddings (¶¶48-49, medical assistant system 206 has a RNN based architecture to create user enquiry word embedding 310 in accordance to the architecture of Fig. 3) and a set of content embeddings (¶60, medical assistant system 206 uses the same RNN based architecture to create word embeddings of sentences present in the health profile of the user 102 to create sentence embedding 310 per ¶107), wherein employing the single embedding model comprises:
receiving a plurality of user-specific text strings associated with a user (¶45 and ¶47, receiving user enquiry comprising text strings and converting text strings into tokens);
using the single embedding model comprising a natural language processing (NLP) model (Fig. 3 and ¶49, RNN architecture generates user enquiry word embedding 310) comprising natural language embedding architecture (¶62, use recurrent neural networks to create sentence embeddings; ¶73, RNN architecture for modeling natural language), generating a set of user-specific embeddings (¶47, convert text string of user enquiry into tokens; ¶48, converting tokens into word embedding using RNN-LSTM based model), comprising, for each user-specific text string of the plurality: generating a respective user-specific embedding based on the user-specific text string, wherein the respective user-specific embedding preserves semantic language information from the user-specific text string (¶73, RNN reads input sequence (user enquiry or the user symptoms) one word at a time, encodes the input sequence into a state vector that compresses semantic information);
using the single embedding model, generating a first set of content embeddings (Fig. 3 and ¶107, using the same RNN architecture to generate sentence embedding 310), comprising, for each content block of a set of content blocks, generating a respective content embedding associated with the content block (¶60, creating word embedding of words presented in a health profile of the user / corpus of medical triage conversation);
based on a set of semantic similarity metrics determined between the set of user-specific embeddings and the set of content embeddings, determining a ranked list of content blocks selected from the set of content blocks (¶63 in view of ¶59, using cosine-similarity / rank loss function to match word embedding of user enquiry with answers in user health profile to select one or more relevant answers (i.e., relevant answers with respective cosine-similarity measure)); and
based on the ranked list, providing a first content block to the user (¶64 in view of ¶59, display one or more relevant answers for the user enquiry based on the selection of the one or more relevant answers).
Devesa does not disclose the natural language processing (NLP) model comprising language-agnostic multi-lingual embedding architecture.
Bouyarmane discloses using language agnostic multi-lingual embedding architecture-based NLP model to generate embeddings for entity records expressed in different languages (Col 8, Rows 45-54, entity records expressed in multiple languages; Col 10, Rows 1-12, use hierarchical embedding models to generate embeddings at word level and entity record level; in view of Col 6, Rows 18-20 and Col 10, Rows 15-16, hierarchical embedding model comprises neural network based models at individual layers of the hierarchy such as bidirectional long short term memory units “BiLSTMs”).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to use a NLP model comprising language agnostic multi-lingual embedding architecture (compare Devesa, ¶48, using RNN-LSTM architecture for generating embeddings with Bouyarmane, Col 6, Rows 18-20, HEM comprising bidirectional LSTMs) to generate the set of user specific embeddings (Devesa, ¶62) in order to generate respective semantic similarity metrics between the user specific embeddings and set of content embeddings (i.e., entity records) expressed in different languages (Bouyarmane, Col 5, Rows 20-24).
Devesa does not disclose wherein the first content block is structured to provide a meditation exercise.
Chang discloses generating personalized and effective mental health therapies and recommendations (¶2) by structing content block to provide a meditation exercise (¶77).
It would’ve been obvious to one ordinarily skilled in the art before the effective filing date of the invention to structure the first content block to provide a meditation exercise in order to generate personalized and effective mental health therapies (Chang, ¶2).
Regarding Claim 20, Devesa discloses initiating a first conversational session with the user using a mobile device of the user (¶41, user 102 with communication device 104 interacts with medical assistant system 206 in a bi-directional conversation; e.g., medical assistant 206 enquires the user “How are you feeling today”? and user initially enquires a question from the medical assistant system 206), wherein receiving the plurality of user-specific text strings comprises receiving a set of messages of the first conversational session with the user (¶47, tokenize user enquiry into tokens), and selecting the plurality of user-specific text strings from the set of messages (¶47, insert delineation tokens between the context and utterances to distinguish between responses of the medical assistant system 206 and the user 102; i.e., distinguish and select tokens of the user enquiry for conversion into word embedding per ¶48).
Conclusion
Applicant's amendment necessitated the new grounds 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 extension fee 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 examiner Richard Z. Zhu whose telephone number is 571-270-1587 or examiner’s supervisor Hai Phan whose telephone number is 571-272-6338. Examiner Richard Zhu can normally be reached on M-Th, 0730:1700.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/RICHARD Z ZHU/Primary Examiner, Art Unit 2654 09/17/2026
Prosecution Insights