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 .
All objections/rejections not mentioned in this Office Action have been withdrawn by the Examiner.
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 20 April 2026 has been entered.
Status of the Claims
Prior to entry of the amendment(s) and/or consideration of the argument(s), the status of the claims is as follows.
Claim(s) 1-20 is/are pending.
Claim(s) 3 is/are objected to because of informalities.
Claims 1-20 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement.
Claim(s) 2, and 6-7 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite.
Claims 1-3, 9-16, and 18-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Non-Patent Literature to Salemi (Salemi, A., Mysore, S., Bendersky, M. and Zamani, H., 2023. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406 (2023), hereinafter Salemi) in view of Dai (WO 2024064249 A1, hereinafter Dai).
Response to Amendments
Applicant’s amendment filed on 20 April 2026 has been entered.
In view of the amendment to the claim(s), the amendment of claim(s) 1-3, 6-7, 12, 16, and 18-20, and the cancellation of claim(s) 9 and 13 have been acknowledged and entered.
In view of the amendment to claim(s) 3, the objection to claim(s) 3 is withdrawn.
In view of the amendment to claim(s) 1-2, 6-7, 12, and 18 and the cancellation of claim(s) 9 and 13, the rejection of claim(s) 1-20 under 35 U.S.C. §112 is withdrawn.
In view of the amendment to claim(s) 1, 12, and 18 and the cancellation of claim(s) 9 and 13, the rejection(s) of claims 1-20 under 35 U.S.C. § 103 is withdrawn.
In light of the amended claims, new grounds for rejection under 35 U.S.C. §103 are provided in the action below.
Response to Arguments
Applicant’s arguments regarding the prior art rejections under 35 U.S.C. §103, see pages 8-14 of the Response to Final Office Action dated 20 February 2026, which was received on 20 April 2026 (hereinafter Response and Office Action, respectively), have been fully considered.
Applicant’s arguments as directed to the claims as previously presented and the prior art were addressed in the Advisory Action dated 30 April 2026. Said arguments and explanation provided therein are incorporated by reference.
With respect to the rejection(s) of amended claim(s) 1, 12, and 18 under 35 U.S.C. §103 as being obvious in light of Salemi and Dai, applicant asserts that Salemi and Dai fail to teach or suggest all limitations of the claims as amended. Applicant’s arguments are persuasive. As such, the rejections of claims 1, 12, and 18 under 35 U.S.C. §103 are withdrawn.
Claims 9 and 13 are cancelled in this response. Therefore, the rejections of claims 9 and 13 are withdrawn.
Applicant further argues that the rejection(s) of dependent claims 2-8, 10-11, 14-17, and 19-20 should be withdrawn for at least the same reasons as independent claims 1, 12, and 18. Applicant’s arguments in light of the amended claims are persuasive. As such, the rejections of claims 2-11, 13-17, and 19-20 under 35 U.S.C. §102 and 35 U.S.C. §103 are withdrawn.
However, upon further consideration, new ground(s) of rejection under 35 U.S.C. §103 are made in light of combinations of Salemi and newly cited references Non-patent literature to Shi (Shi, W., Min, S., Yasunaga, M., Seo, M., James, R., Lewis, M., Zettlemoyer, L. and Yih, W.T., 2023. REPLUG: Retrieval-Augmented Black-Box Language Models. arXiv preprint arXiv:2301.12652. (Year: 2023), hereinafter Shi) and Non-patent literature to Li (Li, X., Yu, P., Zhou, C., Schick, T., Levy, O., Zettlemoyer, L., Weston, J.E. and Lewis, M., 2023. Self-Alignment with Instruction Backtranslation. arXiv preprint arXiv:2308.06259. (Year: 2023), hereinafter Li).
The Applicant has not provided any further statement and therefore, the Examiner directs the Applicant to the below rationale.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim(s) 3-8 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Regarding claim 3, claim 3 is rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being incomplete for omitting essential steps, such omission amounting to a gap between the steps. See MPEP § 2172.01. The omitted steps are “altering the previously trained retriever model.” Claim 3 recites “wherein altering the retriever model includes…” which, as originally dependent from claim 2, modified the limitation “altering the previously trained retriever model based on the first score.” By the amendment changing the dependency of claim 3 to claim 1, the above limitations from claim 2 is no longer incorporated in claim 3. Amended claim 3 now contains a wherein clause purporting to modify a limitation which no longer exists in the claim. Therefore, claim 3 lacks an essential method step and is rejected.
Regarding claims 4-8, claims 4-8 depend from claim 3 and incorporate all limitations therefrom. Therefore, claims 4-8 are rejected under 35 U.S.C. 112(b) for at least the same reasons as claim 3.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1-3, 10-12, 14-16, and 18-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Salemi in view of Shi, with further evidence from Lee (U.S. Pat. No. 11,003,865, hereinafter Lee) and Non-patent literature to Guu (Guu, K., Lee, K., Tung, Z., Pasupat, P. and Chang, M., 2020, November. Retrieval augmented language model pre-training. In International conference on machine learning (pp. 3929-3938). PMLR.(2020), hereinafter Guu).
Regarding claim 1, Salemi discloses A personally-stylized content generation method comprising (Systems and methods described with reference to “a retrieval augmentation approach that retrieves personalized items from user profiles to construct personalized prompts for large language models”; Salemi, ¶ Abstract): receiving a first request for first content to be stylized in a style of written prose previously produced by a user (The system receives “input xi” from “user u” for generating text “while aligning with the user’s writing style” based on “user u’s profile” where said profile consists of a “collection of data points pertaining to the user” and where the “profile of each user encompasses all the papers they have authored.”; Salemi, ¶ pg. 2, col. 2, para. 2; pg. 3, col. 2, para. 2, pg. 5, col. 2, para. 3; Figure. 1); applying a previously trained retriever model to the first request to obtain second content previously produced by the user resulting in obtained content (“To achieve personalization for a given sample (xi, yi) associated with user u” the system uses “(1) a query generation function φq that transforms the input xi into a query q for retrieving from the user u’s profile” and “(2) a retrieval model R(q, Pu, k) that accepts a query q, a user profile Pu and retrieves k most pertinent entries from the user profile”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); populating a prompt with the obtained content and the first request resulting in an augmented prompt (The system further includes “a prompt construction function φp that assembles a prompt for user u based on input xi and the retrieved entries.”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); providing the augmented prompt to the LLM (As shown in Figure 1, and described throughout, the modified prompt based on input xi and the retrieved entries, is received by a large language model, with examples including FlanT5-XXL and GPT-3..; Salemi, ¶ Pg. 8, col. 1, para. 2; Figure 1); receiving personally-stylized content from the LLM (“for a given textual input x, the goal is to develop a model M that generates personalized output y for user u. One can model this task as arg maxy p(y|x, u). For each user u, the model M can take advantage of Pu = {(xu1, yu1),(xu2, yu2), (xumu, yumu) where each (xui, yui) denotes a pair of input and personalized output for user u in the same format as x and y. “; Salemi, ¶ pg. 2, col. 1, para. 2), the personally- stylized content including elements of the style of the written prose of the user (Discloses that the output is “personalized... for user u in the same format as x and y” and aligns with the user’s writing style.; Salemi, ¶ pg. 2, col. 1, para. 2; pg. 5, col. 2, para. 3); and providing the personally-stylized content to the user (As the “personalized output” is intended and output for “user u”, the personalized output is also understood as being provided to user u.; Salemi, ¶ pg. 2, col. 1, para. 2). However, Salemi fails to expressly recite the previously trained retriever model trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content.
Shi teaches “a retrieval-augmented language modeling framework including a tunable retrieval model.” (Shi, ¶ Abstract). Regarding claim 1, Shi teaches the previously trained retriever model trained based on a first score (“our approach can be seen as adjusting the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model,” where the retriever model is trained based on the perplexity scores generated by the language model.; Shi, ¶ pg. 4, col. 1, lines 9-21) generated by a language model (LM) based on (i) target content that is stylized in a style of the user (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a style of the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), (ii) content previously generated by the user (“training algorithm” includes “(1) retrieving documents… [and], (2) scoring the retrieved documents by the language model,” where d in the perplexity score is the retrieved document, where the retrieved document, as read in the context of Salemi, is content previously generated by the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), and (iii) a second request for the target content (As indicated above, the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). The input context x is the input prompt, where the input context x, as read in light of the personalized content that aligns with the user’s writing style” described in Salemi is the second request for the target content (that is stylized in a style of the user).; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content (The described “training algorithm” adjusts “the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model” where the perplexity score is the “LM probability of the ground truth output y given the input context x and a document d,” which is an indication of how much document d {previously generated content} will help the language model {a large language model (LLM)} in generating the ground truth output y {target content}; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-38).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include the previously trained retriever model trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which “significantly improves” overall performance for a designated generative task, as shown in the context of GPT-3 language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 2, the rejection of claim 1 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite further comprising: further training the previously trained retriever model by: generating, by a language model (LM) and based on target content that is stylized in the style of the user, content previously generated by the user, and a second request for the target content, the first score indicating how much the previously generated content will help the LLM in generating the target content; and altering the previously trained retriever model based on the first score.
The relevance of Shi is described above with relation to claim 1. Regarding claim 2, Shi teaches further comprising: further training the previously trained retriever model by (As indicated above, the training algorithm trains “the retriever to find documents that result in lower perplexity scores”; Shi, ¶ pg. 4, col. 1, lines 9-21): generating, by a language model (LM) and based on target content that is stylized in the style of the user, content previously generated by the user, and a second request for the target content (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a style of the user, document d is content previously generated by the user, and input context x corresponds to the second request for target content.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), the first score indicating how much the previously generated content will help the LLM in generating the target content (The described “training algorithm” adjusts “the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model” where the perplexity score is the “LM probability of the ground truth output y given the input context x and a document d,” which is an indication of how much document d {previously generated content} will help the language model {a large language model (LLM)} in generating the ground truth output y {target content}; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-38); and altering the previously trained retriever model based on the first score (The above perplexity score is used to “compute the LM likelihood of each document d” in the top k documents and then “Given the input context x and the corresponding ground truth continuation y, we compute the retrieval likelihood and the language model likelihood. The dense retriever is trained by minimizing the KL divergence between these two distributions”; Shi, ¶ pg. 4, col. 1, line 32 - col. 2, line 9).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include further comprising: further training the previously trained retriever model by: generating, by a language model (LM) and based on target content that is stylized in the style of the user, content previously generated by the user, and a second request for the target content, the first score indicating how much the previously generated content will help the LLM in generating the target content; and altering the previously trained retriever model based on the first score. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 3, the rejection of claim 2 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite wherein the previously trained retriever model is trained by: generating, by the LM and based on the target content that is stylized in the writing style of the user and the second request for the target content, a second score indicating how well the LLM can generate the target content based on just the second request; and wherein altering the retriever model includes altering the retriever model based on the first score and the second score.
The relevance of Shi is described above with relation to claim 1. Regarding claim 3, Shi teaches wherein the previously trained retriever model is trained by: generating, by the LM and based on the target content that is stylized in the writing style of the user and the second request for the target content, a second score (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y), and the retrieved documents correspond to “the top 20 documents [retrieved] from the FAISS index” to “compute the retrieval likelihood and the LM likelihood with a temperature of 0”. As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a style of the user, document d is content previously generated by the user, and input context x corresponds to the second request for target content based on the top k documents, where k=20.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36) [the second score] indicating how well the LLM can generate the target content based on just the second request (Shi discloses the training algorithm including the top k documents, where each of the top k documents receives a score. Though not expressly describing the use of a null document, the use of a null document as part of the top k documents was well known to a person of ordinary skill in the art before the effective filing date of the claimed invention. Further, a null document is used for the express purpose of “indicating how well the model can generate the target content based on just the query.” As evidenced, for example, by Lee, it was well known to artisans to “add an additional null document to the top k documents” where “the null document has no content, and thus can be used to identify situations in which the language model would be able to correctly predict an answer without referencing any document.” (Lee, Col. 9, lines 10-23). Lee demonstrates that it was standard practice to use a null document to determine if the second request “is informative enough by itself for the language model to predict the answer without consulting additional documents {indicating how well the LLM can generate the target content based on just the second request}.” (Id.) Further support for the use of a null document for modeling when no retrieval is necessary, can be found in Guu, at pg. 5, col. 2, lines 17-21. As such, the teachings of Shi, as applied to the top k documents including the null document, results in the first score, as described above, and the second score which indicates how well the LLM can generate the target content based on just the second request.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36); and wherein altering the retriever model includes altering the retriever model based on the first score and the second score (The above perplexity score is used to “compute the LM likelihood of each document d” in the top k documents and then “Given the input context x and the corresponding ground truth continuation y, we compute the retrieval likelihood and the language model likelihood. The dense retriever is trained by minimizing the KL divergence between these two distributions”; Shi, ¶ pg. 4, col. 1, line 32 - col. 2, line 9).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include wherein the previously trained retriever model is trained by: generating, by the LM and based on the target content that is stylized in the writing style of the user and the second request for the target content, a second score indicating how well the LLM can generate the target content based on just the second request; and wherein altering the retriever model includes altering the retriever model based on the first score and the second score. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract). Further, the use of a null document as part of the top k documents used in training the tunable retrieval model of Shi was a well-known technique in the art which yields the predictable result of generating a second score indicative of the ability of the language model to predict the answer without consulting additional documents.
Regarding claim 10, the rejection of claim 1 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite wherein the retriever model is an encoder model.
The relevance of Shi is described above with relation to claim 1. Regarding claim 10, Shi teaches wherein the retriever model is an encoder model (“we use a dense retriever based on the dual encoder architecture, where an encoder is used to encode both the input context x and the document d.”; Shi, ¶ pg. 3, col. 1, lines 15-19).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include wherein the retriever model is an encoder model. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 11, the rejection of claim 1 is incorporated. Salemi discloses all of the elements of the current invention as stated above. Salemi further discloses wherein the personally-stylized content includes emulation of a writing style or communication style in the second content (Discloses that the output is “personalized... for user u in the same format as x and y” and aligns with the user’s writing style based on “a collection of user records associated with each sample for all tasks”; Salemi, ¶ pg. 2, col. 1, para. 2; pg. 5, col. 2, para. 3).
Regarding claim 12, Salemi discloses A personally-stylized content generation system comprising (Systems and methods described with reference to “a retrieval augmentation approach that retrieves personalized items from user profiles to construct personalized prompts for large language models”; Salemi, ¶ Abstract): a database storing content items previously generated by users (The system includes a user profile, where said profile consists of a “collection of data points pertaining to the user” and where the “the profile of each user encompasses all the papers they have authored.”; Salemi, ¶ pg. 3, col. 2, para. 2, pg. 5, col. 2, para. 3; Figure. 1); a pre-trained retriever model (Describes a model for generating text “while aligning with the user’s writing style” based on “user u’s profile”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); configured to: receive, from a user, a query for personally-stylized content (The system receives “input xi” from “user u” for generating text “while aligning with the user’s writing style” based on “user u’s profile”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); score each item of content generated by the user in the database (The system describes comparing Best Match (BM) 25 and Contriever as the retrieval models, both of which provide a score for the candidate documents (Contriever generates a relevance score for each of the documents based on the dot product between the query embedding and each document embedding; BM25 ranks documents based on TF-IDF).; Salemi, ¶ pg. 8, col. 1, para. 1); and provide a specified number of content items that were previously generated by the user and associated with highest scores (Discloses retrieving k items from a user profile, based on “the retrieval model (BM25 vs. Contriever)”; Salemi, ¶ pg. 8, col. 1, para. 1); a prompt augmenter configured to: receive the specified number of content items (The system receives the retrieved entries/items which are integrated into the assembled prompt by a language model “with the inputs corresponding to individual tasks” and the system then “assess[es] their performance based on the generated outputs.”; Salemi, ¶ pg. 8, col. 1-2, para. 2); augment a prompt to include content of the specified number of content items resulting in an augmented prompt (The system further includes “a prompt construction function φp that assembles a prompt for user u based on input xi and the retrieved entries.”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); and provide the augmented prompt to a generative model (As shown in Figure 1, and described throughout, the modified prompt based on input xi and the retrieved entries, is received by a large language model, with examples including FlanT5-XXL and GPT-3..; Salemi, ¶ Pg. 8, col. 1, para. 2; Figure 1); and an application configured to provide, to the user, personally-stylized content from the generative model responsive to the augmented prompt (“for a given textual input x, the goal is to develop a model M that generates personalized output y for user u. One can model this task as arg maxy p(y|x, u). For each user u, the model M can take advantage of Pu = {(xu1, yu1),(xu2, yu2), (xumu, yumu) where each (xui, yui) denotes a pair of input and personalized output for user u in the same format as x and y.” As the “personalized output” is intended and output for “user u”, the personalized output is also understood as being provided to user u, and where the output is produced by an application.; Salemi, ¶ pg. 2, col. 1, para. 2). However, Salemi fails to expressly recite trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content.
The relevance of Shi is described above with relation to claim 1. Regarding claim 12, Shi teaches the previously trained retriever model trained based on a first score (“our approach can be seen as adjusting the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model,” where the retriever model is trained based on the perplexity scores generated by the language model.; Shi, ¶ pg. 4, col. 1, lines 9-21) generated by a language model (LM) based on (i) target content that is stylized in a writing style of the user (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a writing style of the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), (ii) content previously generated by the user (“training algorithm” includes “(1) retrieving documents… [and], (2) scoring the retrieved documents by the language model,” where d in the perplexity score is the retrieved document, where the retrieved document, as read in the context of Salemi, is content previously generated by the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), and (iii) a second request for the target content (As indicated above, the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). The input context x is the input prompt, where the input context x, as read in light of the personalized content that aligns with the user’s writing style” described in Salemi is the second request for the target content (that is stylized in a style of the user).; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content (The described “training algorithm” adjusts “the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model” where the perplexity score is the “LM probability of the ground truth output y given the input context x and a document d,” which is an indication of how much document d {previously generated content} will help the language model {a large language model (LLM)} in generating the ground truth output y {target content}; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-38).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 14, the rejection of claim 12 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite wherein the retriever model is an encoder model.
The relevance of Shi is described above with relation to claim 1. Regarding claim 14, Shi teaches wherein the retriever model is an encoder model (“we use a dense retriever based on the dual encoder architecture, where an encoder is used to encode both the input context x and the document d.”; Shi, ¶ pg. 3, col. 1, lines 15-19).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include wherein the retriever model is an encoder model. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 15, the rejection of claim 12 is incorporated. Salemi discloses all of the elements of the current invention as stated above. Salemi further discloses wherein the personally-stylized content includes emulation of a writing style or communication style in the second content (Discloses that the output is “personalized... for user u in the same format as x and y” and aligns with the user’s writing style based on “a collection of user records associated with each sample for all tasks”; Salemi, ¶ pg. 2, col. 1, para. 2; pg. 5, col. 2, para. 3).
Regarding claim 16, the rejection of claim 12 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite further comprising: a language model configured to: generate, based on target content that is stylized in the writing style of the user, content previously generated by the user, and a second request for the target content, a first score indicating how much the previously generated content will help the LLM in generating the target content; and generate, based on the target content that is stylized in the voice of the user and the second request for the target content, a second score indicating how well the LLM can generate the target content based on just the second request.
The relevance of Shi is described above with relation to claim 1. Regarding claim 16, Shi teaches further comprising: a language model configured to: generate, based on target content that is stylized in the writing style of the user, content previously generated by the user, and a second request for the target content (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a style of the user, document d is content previously generated by the user, and input context x corresponds to the second request for target content.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), a first score indicating how much the previously generated content will help the LLM in generating the target content (The described “training algorithm” adjusts “the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model” where the perplexity score is the “LM probability of the ground truth output y given the input context x and a document d,” which is an indication of how much document d {previously generated content} will help the language model {a large language model (LLM)} in generating the ground truth output y {target content}; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-38); and generate, based on the target content that is stylized in the voice of the user and the second request for the target content, a second score (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y), and the retrieved documents correspond to “the top 20 documents [retrieved] from the FAISS index” to “compute the retrieval likelihood and the LM likelihood with a temperature of 0..”. As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a style of the user, document d is content previously generated by the user, and input context x corresponds to the second request for target content based on the top k documents, where k=20.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36) [the second score] indicating how well the LLM can generate the target content based on just the second request (Shi discloses the training algorithm including the top k documents, where each of the top k documents receives a score. Though not expressly describing the use of a null document, the use of a null document as part of the top k documents was well known to a person of ordinary skill in the art before the effective filing date of the claimed invention. Further, a null document is used for the express purpose of “indicating how well the model can generate the target content based on just the query.” As evidenced, for example, by Lee, it was well known to artisans to “add an additional null document to the top k documents” where “the null document has no content, and thus can be used to identify situations in which the language model would be able to correctly predict an answer without referencing any document.” (Lee, Col. 9, lines 10-23). Lee demonstrates that it was standard practice to use a null document to determine if the second request “is informative enough by itself for the language model to predict the answer without consulting additional documents {indicating how well the LLM can generate the target content based on just the second request}.” (Id.) Further support for the use of a null document for modeling when no retrieval is necessary, can be found in Guu, at pg. 5, col. 2, lines 17-21. As such, the teachings of Shi, as applied to the top k documents including the null document, results in the first score, as described above, and the second score which indicates how well the LLM can generate the target content based on just the second request.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include further comprising: a language model configured to: generate, based on target content that is stylized in the writing style of the user, content previously generated by the user, and a second request for the target content, a first score indicating how much the previously generated content will help the LLM in generating the target content; and generate, based on the target content that is stylized in the voice of the user and the second request for the target content, a second score indicating how well the LLM can generate the target content based on just the second request. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which significantly improves overall performance for a designated generative task, as shown in the context of GPT-3 on language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract). Further, the use of a null document as part of the top k documents used in training the tunable retrieval model of Shi was a well-known technique in the art which yields the predictable result of generating a second score indicative of the ability of the language model to predict the answer without consulting additional documents.
Regarding claim 18, Salemi discloses A non-transitory machine-readable medium including instructions that, when executed by a machine, cause the machine to perform operations for generating personally-stylized content, the operations comprising (Systems and methods described with reference to “a retrieval augmentation approach that retrieves personalized items from user profiles to construct personalized prompts for large language models”; Salemi, ¶ Abstract): receiving a first request for personally-stylized content from a user (The system receives “input xi” from “user u” for generating text “while aligning with the user’s writing style” based on “user u’s profile” where said profile consists of a “collection of data points pertaining to the user” and where the “profile of each user encompasses all the papers they have authored.”; Salemi, ¶ pg. 2, col. 2, para. 2; pg. 3, col. 2, para. 2, pg. 5, col. 2, para. 3; Figure. 1); providing a retriever model with the first request (“To achieve personalization for a given sample (xi, yi) associated with user u” the system uses “(1) a query generation function φq that transforms the input xi into a query q for retrieving from the user u’s profile” and “(2) a retrieval model R(q, Pu, k) that accepts a query q”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1); receiving, from the retriever model and responsive to the first request, content previously generated by the user resulting in obtained content (The retrieval model “accepts a query q, a user profile Pu and retrieves k most pertinent entries from the user profile” which are received/used by the prompt construction function; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1), populating a prompt with the obtained content resulting in an augmented prompt, (The system further includes “a prompt construction function φp that assembles a prompt for user u based on input xi and the retrieved entries.”; Salemi, ¶ pg. 2, col. 2, para. 2; Figure. 1) providing the augmented prompt to a generative model (As shown in Figure 1, and described throughout, the modified prompt based on input xi and the retrieved entries, is received by a large language model, with examples including FlanT5-XXL and GPT-3..; Salemi, ¶ Pg. 8, col. 1, para. 2; Figure 1); receiving the personally-stylized content from the generative model (“for a given textual input x, the goal is to develop a model M that generates personalized output y for user u. One can model this task as arg maxy p(y|x, u). For each user u, the model M can take advantage of Pu = {(xu1, yu1),(xu2, yu2), (xumu, yumu) where each (xui, yui) denotes a pair of input and personalized output for user u in the same format as x and y” where the output is “personalized... for user u in the same format as x and y” and aligns with the user’s writing style.; Salemi, ¶ pg. 2, col. 1, para. 2; pg. 5, col. 2, para. 3); and providing the personally-stylized content to the user (As the “personalized output” is intended and output for “user u”, the personalized output is also understood as being provided to user u.; Salemi, ¶ pg. 2, col. 1, para. 2). However, Salemi fails to expressly recite the retriever model trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a generative model in generating the target content.
The relevance of Shi is described above with relation to claim 1. Regarding claim 18, Shi teaches the retriever model trained based on a first score (“our approach can be seen as adjusting the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model,” where the retriever model is trained based on the perplexity scores generated by the language model.; Shi, ¶ pg. 4, col. 1, lines 9-21) generated by a language model (LM) based on (i) target content that is stylized in a writing or communication style of the user (The described “training algorithm” for the retriever model includes “(2) scoring the retrieved documents by the language model (§4.) {generated by a language model(LM) based on...}” where the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). As read in the context of the personalized retrieval augmentation of Salemi, the target tokens (y) are the target content that is stylized in a writing or communication style of the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), (ii) content previously generated by the user (“training algorithm” includes “(1) retrieving documents… [and], (2) scoring the retrieved documents by the language model,” where d in the perplexity score is the retrieved document, where the retrieved document, as read in the context of Salemi, is content previously generated by the user.; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), and (iii) a second request for the target content (As indicated above, the score is the sequence probability (P_LM(y|d,x)) calculated across the target tokens (y). The input context x is the input prompt, where the input context x, as read in light of the personalized content that aligns with the user’s writing style” described in Salemi is the second request for the target content (that is stylized in a style of the user).; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-36), the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content (The described “training algorithm” adjusts “the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model” where the perplexity score is the “LM probability of the ground truth output y given the input context x and a document d,” which is an indication of how much document d {previously generated content} will help the language model {a large language model (LLM)} in generating the ground truth output y {target content}; Shi, ¶ pg. 4, col. 1, lines 9-21 and 32-38).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include the previously trained retriever model trained based on a first score generated by a language model (LM) based on (i) target content that is stylized in the writing style of the user, (ii) content previously generated by the user, and (iii) a second request for the target content, the first score indicating how much the previously generated content will help a large language model (LLM) in generating the target content. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which “significantly improves” overall performance for a designated generative task, as shown in the context of GPT-3 language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Regarding claim 19, Salemi discloses wherein the personally-stylized content includes emulation of a writing or communication style in the previously generated content (Discloses that the output is “personalized... for user u in the same format as x and y” and aligns with the user’s writing style based on “a collection of user records associated with each sample for all tasks”, and where a writing style is also a communication style; Salemi, ¶ pg. 2, col. 1, para. 2; pg. 5, col. 2, para. 3).
Regarding claim 20, the rejection of claim 18 is incorporated. Salemi discloses all of the elements of the current invention as stated above. However, Salemi fails to expressly recite wherein the retriever model is trained to obtain previously generated content that is most likely to help the LLM emulate the previously generated content.
The relevance of Dai is described above with relation to claim 1. Regarding claim 20, Dai teaches wherein the retriever model is trained to obtain previously generated content that is most likely to help the generative model emulate the previously generated content (The above perplexity score is used to “compute the LM likelihood of each document d” in the top k documents and then “Given the input context x and the corresponding ground truth continuation y, we compute the retrieval likelihood and the language model likelihood. The dense retriever is trained by minimizing the KL divergence between these twodistributions” which results in “adjusting the probabilities of the retrieved documents to match the probabilities of the output sequence perplexities of the language model.”; Shi, ¶ pg. 4, col. 1, lines 9-12, and col. 1, line 32 - col. 2, line 9).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi to incorporate the teachings of Shi to include wherein the retriever model is trained to obtain previously generated content that is most likely to help the LLM emulate the previously generated content. The tunable retrieval model of Shi, which is explicitly described as a “simple design” which “can be easily applied to any existing retrieval and language models,” provides a demonstrable retrieval benefit which “significantly improves” overall performance for a designated generative task, as shown in the context of GPT-3 language modeling providing a performance improvement of 6.3% over prior art models without tuning. (Shi, Abstract).
Claim(s) 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Salemi and Shi as applied to claim 3 above, and further in view of Li.
Regarding claim 8, the rejection of claim 3 is incorporated. Salemi and Shi disclose all of the elements of the current invention as stated above. However, Salemi fails to expressly recite wherein the second request is reverse engineered based on the target content resulting in training or testing data.
Li teaches “instruction backtranslation…to construct training examples by generating instruction prompts for web documents.” (Li, Abstract). Regarding claim 8, Li teaches wherein the second request is reverse engineered based on the target content resulting in training or testing data ("Our overall process, which we call instruction backtranslation, thus performs two core steps: 1. Self-augment: Generate instructions for unlabeled data, i.e. the web corpus, to produce candidate training data of (instruction, output) pairs for instruction tuning {reverse engineered based on the target content...}. 2. Self-curate: Self-select high quality demonstration examples as training data to finetune the base model to follow instructions. This approach is done iteratively where a better intermediate instruction-following model can improve on selecting data for finetuning in the next iteration {resulting in training or testing data}."; Li, ¶ pg. 2, lines 8-14).
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the personalized text generation systems of Salemi, as modified by the tunable retrieval models of Shi, to incorporate the teachings of Li to include wherein the second request is reverse engineered based on the target content resulting in training or testing data. The self-augmentation and self-curation described in Li allows for the generation of training data from otherwise unusable text which “scal[es] up data quantity” and selects the “high quality portion of the augmented data” for training which “leads to increasing instruction following performance” for the underlying model, as recognized by Li. (Li, pg. 5, lines 28-36).
Allowable Subject Matter
Claim(s) 4-7 would be allowable if rewritten to overcome the rejection(s) under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), 2nd paragraph, set forth in this Office action and to include all of the limitations of the base claim and any intervening claims.
Claim 17 is objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
The following is an examiner’s statement of reasons for indicating allowable subject matter:
Regarding claim 4, the closest prior art of record Salemi discloses the limitations of claim 1 as presented in the rejection above. However, Salemi does not specifically teach wherein training the retriever model further includes: generating, by the retriever model and based on the second request and the content previously generated by the user, a third score; and altering the retriever model based on a loss that considers (i) the third score and (ii) a fourth score that is a difference between the first score and the second score.
Shi teaches the remaining limitations of claim 1 as mapped in the rejection above. Shi further teaches the limitations of claim 3 as presented above.
However, none of the prior art references of record, either alone or in combination, teaches, suggests, or makes obvious the combination of limitations as recited in claim 4.
Regarding claim 17, the closest prior art of record Salemi discloses the limitations of claim 12 as presented in the rejection above. However, Salemi does not specifically teach wherein the retriever model is further configured to generate, based on the second request and the content previously generated by the user, a third score and the system further comprises: a compute device configured to alter the retriever model based on a difference between (i) the third score and (ii) a difference between the first score and the second score.
Shi does teach the remaining limitations of claim 12 as presented above. Further, Shi does teach the limitations of intervening claim 16, as described in the rejection above.
However, none of the prior art references of record, either alone or in combination, teaches, suggests, or makes obvious the combination of limitations as recited in claim 17.
Regarding claim 4, and mutatis mutandis claim 17, the limitation of “generating, by the retriever model and based on the second request and the content previously generated by the user, a third score; and altering the retriever model based on a loss that considers (i) the third score and (ii) a fourth score that is a difference between the first score and the second score,” in light of all remaining claim limitations, is not taught by the prior art of record.
Though the combination of Salemi and Shi discloses perplexity score in the context of retrieved documents {first score} indicating how much the previously generated content will help the LLM in generating the target content and a perplexity score in the context of a null document {second score} indicating how well the LLM can generate the target content based on just the second request, these scores are generated by the language model, not the retriever model. Shi does not further consider the generation of hypothetical or actual combinations of the first score and the second score. As such, Salemi and Shi, alone or in any combination, fail to describe a third score and a fourth score, either simultaneously or sequentially, for altering the retriever model. Further, no art of record is known to cure this deficiency.
Regarding claims 5-8, claims 5-8 depend from claim 4 and would be allowable for at least the same reasons as described above with reference to claim 4.
Any comments considered necessary by applicant must be submitted no later than the payment of the issue fee and, to avoid processing delays, should preferably accompany the issue fee. Such submissions should be clearly labeled “Comments on Statement of Reasons for Allowance.”
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Sean E. Serraguard whose telephone number is (313)446-6627. The examiner can normally be reached 07:00-17:00 M-F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Daniel C. Washburn can be reached at (571) 272-5551. 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.
/Sean E Serraguard/Primary Examiner, Art Unit 2657
Prosecution Insights