DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Objections
Claims 5 to 20 are objected to because of the following informalities:
Independent claims 5 and 14 set forth a limitations of “instructions to change the summarization task”, but “the summarization task” does not have antecedent basis. These independent claims provide a prior recitation of “a text analysis task”, but not “a summarization task”. Generally, any initial occurrence of a claimed element should be accompanied by an indefinite article of “a” or “an”, and not a definite article of “the” or “said”.
Appropriate correction is required.
Claim Rejections - 35 USC § 103
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 to 2, 5, 7, 12, 14, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Gray et al. (U.S. Patent Publication 2025/0005305) in view of Qin (U.S. Patent Publication 2024/0346256).
Concerning independent claims 1, 5, and 14, Gray et al. discloses a method, system, and computer program product for generative summaries of search results, comprising:
“one or more computing devices, respectively comprising at least one processor and a memory; wherein the one or more computing devices store program instructions that when executed by the one or more computing devices:” – computing device 810 typically includes at least one processor 814 which communicates with a number of peripheral devices that may include a storage subsystem 824 (¶[0143]: Figure 8); storage subsystem 824 stores programming and data constructs that provide the functionality of some or all of the modules described herein; storage subsystem 824 may include the logic to perform selected aspects of the methods disclosed; software modules are generally executed by processor 814 alone or in combination with other processors (¶[0146] - ¶[0147]: Figure 8);
“receive an input text for performing [a text analysis task or] a request to perform a summarization task on a natural language text” – a large language model (LLM) generates a natural language (NL) based summary that is rendered in response to a query (Abstract); implementations disclosed herein are directed to at least selectively utilizing an LLM in generating an NL based summary to be rendered in response to a query (¶[0004]); LLM input engine 134 can, in response to receiving a query, generate LLM input that is to be processed using an LLM in generating an NL based response to the query; LLM response generation engine 136 can process LLM input that is generated by the LLM input engine 134 using an LLM to generate an NL based summary (¶[0043] - ¶[0044]: Figure 2); here, a query is “an input text” that serves as “a request to perform a summarization task on a natural language text”; broadly, a summarization task is a specific variety of “a text analysis task”;
“input the generated instructions to the pre-trained large language model fine-tuned to the domain to prompt the pre-trained large language model fine-tuned to the domain to perform the summarization task [or text analysis task] on the natural language text using the generated instructions” – an LLM that is utilized in generating the NL based summary can be one that is fine-tuned to a summary prompt (¶[0066]); system generates the revised NL based summary using an additional LLM, relative to the one used in block 454, that is fine-tuned based on a prompt that reflects familiarity with content of the search result documents (SRDs) interacted with; a fine-tuned LLM model can be trained to receive known content that the user already knows, followed by additional content to be summarized in view of the known content (¶[0108]: Figure 4); a system generates the NL based summary using an LLM fine-tuned based on a prompt that reflects familiarity with the content; a fine-tuned LLM model can be trained to receive known content that the user is already familiar with, followed by additional content to be summarized in view of the known content (¶[0128]: Figure 6);
“provide the result of the summarization task [or the text analysis task] performed on the natural language text” – system generates an NL based summary based on processing, using an LLM, corresponding content from each of the search result documents (¶[0065]: Figure 2); the system causes the revised NL based summary, generated to be rendered, i.e., at the client device that submitted the query and/or a related client device (¶[0109]: Figure 4); system causes the generated NL based summary to be rendered (¶[0129]: Figure 6).
Concerning independent claims 1, 5, and 14, Gray et al. discloses augmenting a prompt with additional content from search results that is provided to a large language model (LLM). (¶[0008]) Additional information to information of the query is leveraged to provide a more targeted prompt without the need for additional interaction with a user. (¶[0107]) Gray et al., then, discloses to “insert . . . instructions to change the summarization task and the instructions to include further instructions to be specific . . . as part of generating the instructions to perform the summarization task using a pre-trained large language model fine-tuned to the domain”, but does not expressly provide “extract one or more domain entities from the natural language text using a machine learning model” and “insert the one or more domain entities into instructions to change the summarization task and the instructions to include further instructions to be specific to the domain”. That is, Gray et al. is not clearly directed to extracting domain entities and inserting domain entities into a prompt.
Concerning independent claims 1, 5, and 14, Qin teaches response generation using a retrieval augmented artificial intelligence model. (Abstract) Embodiments are disclosed herein that improve the scope and accuracy of responses generated by an LLM. An LLM may be augmented with augmentation information, e.g., domain-specific information and entity-specific information. A retrieval augmented generation (RAG) approach is disclosed herein that adds an information retrieval component to create augmented prompts to feed into the generative language model for generating the final answer/prediction. (¶[0018]) Response generator 110 generates a response to the query based on augmentation information from datasets 112. Datasets 112 may include one or more databases storing augmentation information that is used to respond to the query from GUI 114. Augmentation information stored in datasets 112 may include domain-specific information, e.g., information related to specific topics or fields and entity-specific information, e.g., internal or proprietary corporate information. (¶[0035] - ¶[0036]: Figure 1) Pre-processor 202 may also receive augmentation information 218 from datasets 112. Augmentation information 218 may include, but is not limited to, domain-specific information and entity-specific information. (¶[0040]: Figure 2) Prompt generator 212 may generate a prompt for LLM 214 based on one or more of query 216 and augmentation information 232. Prompt generator 212 may generate an augmented prompt 236 that includes the original query, contextual information and content information, e.g., retrieved augmentation information 232, and a request to answer the original query based on the provided contextual information using the included content information. (¶[0046]: Figure 2) Here, entity-specific information and domain-specific information are extracted from datasets 112 (“extract one or more domain entities from the natural language text”). Qin, then, teaches “extracting one or more domain entities from natural language text” and “inserting the one or more domain entities into instructions to change . . . the instructions to include further instructions to be specific to the domain”. An objective is to improve the scope and accuracy of responses generated by an LLM. (¶[0018]) It would have been obvious to one having ordinary skill in the art to augment instructions to a large language model with domain entities to provide instructions that are specific to a domain as taught by Qin to provide additional content in a prompt to a large language model of Gray et al. for a purpose of improving a scope and accuracy of responses generated by an LLM.
Concerning claims 2, 7, and 16, Gray et al. discloses a fine-tuned LLM model can be trained to receive known content that the user already knows, followed by additional content to be summarized in view of the known content (¶[0108]: Figure 4); a system generates the NL based summary using an LLM fine-tuned based on a prompt that reflects familiarity with the content; a fine-tuned LLM model can be trained to receive known content that the user is already familiar with, followed by additional content to be summarized in view of the known content (¶[0128]: Figure 6). Qin teaches augmentation information that is additional content of entity-specific information and domain-specific information extracted from datasets. (¶[0035] - ¶[0036]: Figure 1) Gray et al., then, discloses “instructions to perform the summarization task” and Qin teaches “one or more domain entities are to be included” in a prompt to a large language model (LLM).
Concerning claim 12, Gray et al. discloses a summarization task (“wherein the text analysis task is a summarization task”).
Claims 3, 4, 6, 11, 13, 15, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Gray et al. (U.S. Patent Publication 2025/0005305) in view of Qin (U.S. Patent Publication 2024/0346256) as applied to claims 1, 5, and 14 above, and further in view of Strader et al. (U.S. Patent Publication 2019/0122766).
Concerning claims 3, 6, and 15, Gray et al. discloses that client device 110 can be equipped with one or more microphones that capture audio data, such as audio data corresponding to spoken utterances of the user. (¶[0034]: Figure 1) Additionally, if a query is a voice query the system can perform automatic speech recognition (ASR) to convert the query to textual format. (¶[0051]: Figure 2) However, Gray et al. does not expressly disclose “generating the input text as a transcript from obtained audio data using an automatic speech recognition system.” Still, Strader et al. teaches an interface for a patient/provider conversation and auto-generation of a note or summary that extracts words and phrases with a trained machine learning model. (Abstract) Specifically, audio input is provided to speech to text conversion model 110, and text is generated by speech-to-text conversion model 110 from audio input. (¶[0034] - ¶[0035]: Figure 1) A transcript is generated from an audio recording using a speech-to-text engine 110 (“generate natural language text as a transcript from obtained audio data using an automatic speech recognition system” or “generate the input text as a transcript from obtained audio data using an automatic speech recognition system”). (¶[0042]: Figure 1) An objective is to improve the efficiency of generating notes of patient visits that takes less time of physicians, allowing them to see more patients and spend more time on patient care. (¶[0003]) It would have been obvious to one having ordinary skill in the art to generate input text as a transcript obtained from audio data using automatic speech recognition as taught by Strader et al. to perform a summarization task with a large language model of Gray et al. for a purpose of improving an efficiency of generating summaries that takes less time of physicians,
Concerning claims 4, 13, and 20, Strader et al. discloses automatically generating a note summarizing a conversation between a patient and healthcare provider from an audio recording (“a medical audio summarization service”) (Abstract); a computing environment of a clinic, hospital, or medical office 200 is a location of a visit between a healthcare provider 202 and patient 204; a recording device 206 at workstation 210 captures and records speech, and workstation 210 is used by provider 202 to view the transcript and note; audio input may be sent over a network 208 to a service provider which implements machine learning models (“offered as part of a provider network”); an entity or service provider generates the transcript and note and transmits them over an application programming interface to software resident on workstation 210 (“and wherein the request is received via an interface of the medical audio summarization service”) (¶[0037] - ¶[0038]: Figure 2).
Concerning claim 11, Strader et al. discloses a computing environment of a clinic, hospital, or medical office 200 is a location of a visit between a healthcare provider 202 and patient 204; audio input may be sent over a network 208 to a service provider which implements machine learning models (“sending one or more requests to a remote host for the machine learning model to perform recognition on the input text”); an entity or service provider generates the transcript and note and transmits them over an application programming interface to software resident on workstation 210 (¶[0037] - ¶[0038]: Figure 2).
Claims 8 to 10 and 17 to 19 are rejected under 35 U.S.C. 103 as being unpatentable over Gray et al. (U.S. Patent Publication 2025/0005305) in view of Qin (U.S. Patent Publication 2024/0346256) as applied to claims 5 and 14 above, and further in view of Mahapatra et al. (U.S. Patent Publication 2023/0153533).
Concerning claims 8 and 17, Gray et al. discloses that a fine-tuned LLM model can be trained to receive known content that the user already knows, followed by, e.g., after a delimiter, additional content to be summarized in view of the known content. (¶[0108]: Figure 4; ¶[0128]: Figure 6) Gray et al., then, discloses that a large language model is pre-trained. However, Gray et al. does not expressly disclose training by “a selection of the domain of a plurality of domains supported by the text analysis system, wherein the machine learning model and the pre-trained large language model correspond to the selected domain and are respectively selected for performing the text analysis task out of respective pluralities of machine learning models that recognize entities out of different ones of the plurality of domains and pre-trained large language models fine-tuned to the different ones of the plurality of domains.”
Concerning claims 8 and 17, Mahapatra et al. teaches pre-training techniques for entity extraction models to facilitate domain adaptation. One or more selection models are used to select a subset of an available source domain corpus. (¶[0002]) One or more selection models are used to select a subset of available source domain training data for an entity extraction model based on similarity to a particular target domain. (¶[0016]) Application 110 and entity extraction tool 150 coordinate to train or adapt entity extraction model 160 to an appropriate domain. Application 110 is designed specifically for documents in a particular domain, e.g., contracts or financial documents. Document dashboard 115 may prompt a user to identify an applicable domain for documents 185, e.g., contracts, financial documents, biomedical documents, artificial intelligence research papers, scientific publications, political publications, musical publications, and literary publications. Dashboard 115 prompts and/or accepts inputs from the user identifying ground truth entities (text and corresponding classifications) from a subset of target corpus 192. (¶[0028]: Figure 1) Predicted classes are compared with ground truth and updated to update entity extraction model 160. (¶[0035]) Mahapatra et al., then, teaches “a selection of the domain out of a plurality of domains supported by the text analysis system” by accepting user input to identify a domain for an application being trained to perform entity extraction. An objective is to perform entity extraction in a circumstance that models do not exist for every domain and do not provide sufficient accuracy in certain domains that diverge from training datasets. (¶[0012]) It would have been obvious to one having ordinary skill in the art to select a domain out of a plurality of domains supported by a text analysis system to recognize entities as taught by Mahapatra et al. to fine-tune large language models of Gray et al. for a purpose of performing entity extraction with sufficient accuracy for models that do not exist for every domain and domains that diverge from training datasets.
Concerning claims 9 and 18, Gray et al. discloses that a large language model is pre-trained, but does not expressly provide for “receiving a request to fine-tune the pre-trained large language models for one or more additional domain entities, wherein the request identifies further training data for fine-tuning that includes one or more additional domain entities in ground truth data”, and “performing further fine-tuning on the pre-trained large language models for the domain using the further training data annotated with the one or more additional domain entities extracted from the ground truth data.” However, Mahapatra et al. teaches that one or more selection models are used to select a subset of available source domain training data for an entity extraction model. (¶[0016]) After pre-training a sentence selection model on an unlabeled target domain corpus, the sentence selection model is fine-tuned on the labeled target domain corpus and used to select sentences from a source domain corpus. (¶[0018]) Document dashboard 115 prompts input from a user identifying ground truth entities (text and corresponding classifications) from a subset of target domain corpus 192. Entity extraction model 160 is pre-trained on training sentences 198 selected from document domain corpus 196 and/or fine-tuned on labeled target domain corpus 194. (¶[0028] - ¶[0029]: Figure 1) Predicted classes are compared with ground truth and updated to update entity extraction model 160. (¶[0035]) Mahapatra et al., then, receives a request from a user to fine-tune a language model to a domain for one or more additional domain ground truth entities.
Concerning claims 10 and 19, Gray et al. discloses that a large language model is pre-trained, but does not expressly provide for “using domain entities from ground truth data included in the training data set.” However, Mahapatra et al. teaches that document dashboard 115 prompts input from a user identifying ground truth entities (text and corresponding classifications) from a subset of target domain corpus 192. Entity extraction model 160 is pre-trained on training sentences 198 selected from document domain corpus 196 and/or fine-tuned on labeled target domain corpus 194. (¶[0028] - ¶[0029]: Figure 1) Predicted classes are compared with ground truth and updated to update entity extraction model 160. (¶[0035]) Mahapatra et al., then, teaches fine-tuning a language model for a domain with ground truth entities.
Response to Arguments
Applicants’ arguments filed 26 May 2026 have been considered but are moot in view of new grounds of rejection necessitated by amendment.
Applicants amend independent claims 1, 5, and 14 to set forth a new limitation of instructions “to change the summarization task and the instructions include further instructions to be specific to the domain”, and present arguments traversing the prior rejection of these independent claims as being obvious under 35 U.S.C. §103 over Strader et al. (U.S. Patent Publication 2019/0122766) in view of Mathur et al. (U.S. Patent Publication 2024/0412226). Applicants delete a limitation of these independent claims directed to “trained to recognize entities of a domain in a given text”. Specifically, Applicants argue that Strader et al. does not disclose these new limitations with domain entities being inserted to change the summarization task and instructions specific to a domain.
Applicants’ amendments introduce a new limitation of “the summarization task” into independent claims 5 and 14 which does not have proper antecedent basis. New claim objections are set forth due to this informality.
Generally, new grounds of rejection are now set forth as directed to the independent claims being obvious under 35 U.S.C. §103 over Gray et al. (U.S. Patent Publication 2025/0005305) in view of Qin (U.S. Patent Publication 2024/0346256). The rejection of some dependent claims continues to rely upon Strader et al. (U.S. Patent Publication 2019/0122766) and Mahapatra et al. (U.S. Patent Publication 2023/0153533). The rejection no longer relies upon Mathur et al. (U.S. Patent Publication 2024/0412226). All of the new grounds of rejection are maintained to be necessitated by amendment. Specifically, even if these amendments can be characterized as relatively minor, they do provide new limitations which appear to imply that ‘instructions’ should be construed as a prompt to a large language model, and that these ‘instructions’ cannot reasonably be construed as generic computer instructions.
Mainly, Gray et al. discloses including additional content in a prompt to a large language to generate a more targeted summary. (¶[0058]) A system generates the revised NL based summary using the same LLM as used in block 454, but using a revised input with a revised prompt that reflects familiarity with the content of search result documents interacted with. A prompt used in block 454 could be ‘create a summary of the following’, whereas the prompt used in sub-block 460A could be ‘create a summary of the following and assuming the user already knows X’, where ‘X’ is a description of the content of the SRD(s) interacted with. (¶[0107]) Figure 4, then, illustrates “instructions to change the summarization task and the instructions to include further instructions to be specific to the domain”. That is, Gray et al. modifies a prompt (“the instructions”) to a large language model so as to change a summary to generate a revised summary by reflecting something the user already knows in a revised prompt. Gray et al. modifies a prompt but does not modify the prompt by inserting domain entities from natural language text.
However, Qin teaches that domain-specific and entity-specific augmentation information is added to a prompt to generate an augmented prompt. (¶[0018] - ¶[0019]) Augmentation information may be generated from datasets 112 that include domain-specific and entity-specific augmentation information. (¶[0035] - ¶[0036] and ¶[0040]) This augmentation information is ‘extracted’ from datasets as text using an encoder 204 and a comparator 208 to determine a similarity between feature vectors that implicitly operates by machine learning.
A rationale for a rejection can be based upon KSR International Co. v. Teleflex Inc. (KSR), 550 U.S. 398, 82 USPQ2d 1385 (2007): (B) Simple substitution of one known element for another to obtain predictable results. See MPEP §2141. Here, Gray et al. discloses including additional content in a prompt to a large language to generate a more targeted summary, but does not clearly disclose that this additional content is domain entities extracted from natural language text. However, it is a simple substitution of domain-specific and entity-specific augmentation information of Qin for additional content in Gray et al. Substitution of domain-specific and entity-specific augmentation information of Qin for additional information of Gray et al. would provide predictable results to improve the scope and accuracy of responses generated by an LLM.
Applicants’ amendments necessitate these new grounds of rejection. All of the new grounds of rejection are necessitated by amendment. Applicants’ arguments are moot. Accordingly, this rejection is properly FINAL.
Conclusion
The prior art made of record and not relied upon is considered pertinent to Applicants’ disclosure.
Kotaru discloses related prior art.
Applicants’ amendment necessitated the new grounds of rejection presented in this Office Action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP §706.07(a). Applicants are reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MARTIN LERNER whose telephone number is (571) 272-7608. The examiner can normally be reached Monday-Thursday 8:30 AM-6:00 PM.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Richemond Dorvil can be reached at (571) 272-7602. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/MARTIN LERNER/Primary Examiner
Art Unit 2658
June 26, 2026
Prosecution Insights