DETAILED ACTION
This communication is in response to the Amendments and Arguments filed on 12/23/2025.
Claims 1, 3-6, 8, 10-13, 15, and 17-20, are pending and have been examined.
All previous objections/rejections not mentioned in this Office Action have been withdrawn by the examiner.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
Applicant's arguments filed 12/23/2025 have been fully considered but they are not persuasive and/or are moot.
Regarding the 101 rejections, Applicant asserts on pgs 7-9 that the claims recite a method for training a model using a specific method, which provides a tangible benefit including higher factuality of outputs by models and/or the ability of a model to accurately detect the factuality of a summary, and thus would integrate any alleged abstract idea into a practical application of improved training of a neural network model. The Examiner respectfully disagrees with this assertion. While the claims recite the generation of a second summary from a factually consistent summary by performing one of three edits to the factually consistent summary, it is not clear from the claim language how the specific use of the second summary in combination with the third summary, which is generated from the source document by the language model, implements the technological improvement disclosed in the specification as discussed by the Applicant. Further, there is no recited claim language that would read to utilizing the neural network based language model after it has been trained, demonstrating that the model does, in fact, output summaries with higher factuality and/or more accurately detect the factuality of a summary. As the claims recite an abstract idea as described in the 101 rejection, and do not recite additional elements that integrate the abstract idea into a practical application, the claims are not patent eligible.
Regarding the 103 rejections, please refer to the updated mappings below, where the new ground of rejection relies on the combination of Kryscinski, Sridhar, and the newly-cited art of Du to teach the amended claim language.
Hence, Applicant’s arguments are not persuasive and/or are moot.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 3-6, 8, 10-13, 15, and 17-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim(s) 1, 8, and 15, the limitation(s) of receiving a source document, displaying the source document and first summary, receiving a first indication, generating a second summary, replacing a word, negating an assertion, swapping words, receiving a second indication, storing the second summary, generating a third summary, computing a loss objective, and training the neural network based language model, as drafted, are processes that, under broadest reasonable interpretation, covers performance of the limitation in the mind and/or with pen and paper but for the recitation of generic computer components. More specifically, the mental process of a human reading a written document, reading a summary of the document and a note about whether it is factual, both of which are written out on paper, recognizing that the summary is factual and editing the summary by making specific text alterations to write down a new summary, recognizing that the new summary is factual or not and writing down the determination, using an understanding of human language to write down a completely new summary of the document, comparing the two most recently written summaries using a specific calculation, and using the results of the comparison in a specific manner to improve the human’s understanding of how to write a factual summary. The neural network based language model reads on rules for understanding human language in order to summarize text. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind and/or with pen and paper but for the recitation of generic computer components, then it falls within the --Mental Processes-- grouping of abstract ideas. Accordingly, the claim(s) recite(s) an abstract idea.
This judicial exception is not integrated into a practical application because the recitation of a data interface and a user interface of claim 1, a system, memory, communication interface, processors, memory, and user interface of claim 8, and a machine-readable medium, processors, data interface, and user interface in claim 15, reads to generalized computer components, based upon the claim interpretation wherein the structure is interpreted using [0032-6] in the specification. Accordingly, these additional elements do not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea. The claim(s) is/are directed to an abstract idea.
The claim(s) do(es) not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to the integration of the abstract idea into a practical application, the additional element of using generalized computer components to receive, display, receive, generate, replace, negate, swap, receive, store, generate, compute, and train, amounts to no more than mere instructions to apply the exception using a generic computer component. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept. The claim(s) is/are not patent eligible.
With respect to claim(s) 3, 10, and 17, the claim(s) recite(s) computing the loss objective, which reads on a human performing a specific analysis when comparing two summaries in order to obtain a specific value. No additional limitations are present.
With respect to claim(s) 4, 11, and 18, the claim(s) recite(s) inputting the third summary to a factuality detector model, which reads on a human using rules regarding determining factuality to determine if the newest summary is factual and using the determination when comparing the two newest summaries to obtain a specific value. No additional limitations are present.
With respect to claim(s) 5, 12, and 19, the claim(s) recite(s) predicting a factuality and comparing, which reads on a human using a comparison of a known factuality of a summary with their predicted factuality to learn how to improve their ability to recognize factual summaries. No additional limitations are present.
With respect to claim(s) 6, 13, and 20, the claim(s) recite(s) receiving a third indication, which reads on a human reading whether a summary is well-written, and then being told whether the summary is also factual. No additional limitations are present.
These claims further do not remedy the judicial exception being integrated into a practical application and further fail to include additional elements that are sufficient to amount to significantly more than the judicial exception.
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.
Claim(s) 1, 3-5, 8, 10-12, 15, and 17-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kryscinski et al. (U.S. PG Pub No. 2021/0124876), hereinafter Kryscinski, in view of Sridhar et al. (U.S. PG Pub No. 2024/0184988), hereinafter Sridhar, and further in view of Du et al. (“Constructing Contrastive Samples via Summarization for Text Classification with Limited Annotations”, arXiv:2104.05094v3, 29 Nov 2021), hereinafter Du.
Regarding claims 1, 8, and 15, Kryscinski teaches
(claim 1) A method of training a neural network based language model (method [0024],[0032],[0034]), the method comprising:
(claim 8) A system for training a neural network based language model (a computing device, i.e. system [0030]), the system comprising:
(claim 8) a memory that stores the neural network based language model and a plurality of processor executable instructions (a memory may be used to store software executed by the computing device and one or more data structures used during operation of the computing device [0030-1],[0034]);
(claim 8) one or more hardware processors that read and execute the plurality of processor-executable instructions from the memory to perform operations comprising (one or more processors run the executable code on the memory to perform the methods [0030-2]):
(claim 15) A non-transitory machine-readable medium comprising a plurality of machine-executable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform operations comprising (one or more processors run the executable code on the memory to perform the methods [0030-2]):
receiving, via a data interface, a source document and a first summary associated with the source document (the training data generation module receives an unannotated collection or set of source documents, i.e. receiving…a source document, and abstractive text summaries for the source text documents, i.e. a first summary associated with the source document, which are received as input data by the computing device that has different components coupled to each other, i.e. via a data interface [0029-32],[0039],[0045]);
displaying the source document and the first summary to a user via a user interface (an interface is provided by which human users can receive, retrieve, and view data, i.e. displaying…to a user via a user interface, such as datasets of textual documents, including the source document and summaries, i.e. displaying the source document and the first summary [0059-60]);
receiving a first indication, from a user via the user interface, of factual consistency of the first summary (text samples are extracted from the source documents, where each sample is a single sentence, and text summarization can be in the form of a claim sentence, i.e. first summary, and a user can use an interface to label a summary, i.e. from a user via the user interface, such as an extractive summary, where a label can be consistent or correct, i.e. receiving a first indication…of factual consistency [0045-6],[0059-60],[0065]);
generating, by an edits generator, …, a second summary of the source document by editing a portion of the first summary by at least one of (each sampled sentence can be changed through one or more text transformations, i.e. editing a portion of the first summary, by a transform module, i.e. an edits generator, to create a novel claim sentence, i.e. generating…a second summary of the source document [0045-7],[0059-60],[0065]):
replacing a single word of the first summary (a text transformation can include entity and number swapping, as well as pronoun swapping [0047],[0049-50]),
negating an assertion of the first summary (a text transformation can include sentence negation [0047],[0051]), or
swapping multiple words of the first summary;
receiving a second indication, from the user via the user interface, of the factual consistency of the second summary (each novel claim sentence is labeled either correct/consistent or incorrect/inconsistent, i.e. receiving a second indication…of the factual consistency of the second summary, where an annotation can be provided by a user through an interface, i.e. from the user via the user interface [0055],[0059-60]);
storing the second summary in a memory associated with a label based on the second indication (memory can be used to store data structures used during operation of the device, i.e. storing…in a memory, where the training data generation module generates an artificial training set of claim sentences and label them as correct or incorrect, and the generated training data can be provided as output, i.e. storing the second summary…associated with a label based on the second indication [0031],[0035-6],[0039]);
generating, by the neural network based language model, a third summary using an input of the source document (the data set includes a text summarization generated, i.e. generating…a third summary, by one or more summarization models for respective source documents, i.e. by the neural network based language model…using an input of the source document [0018-9],[0059-60],[0063],[0065]).
While Kryscinski provides evaluating accuracy of generated text summaries from a summarization model, Kryscinski does not specifically teach computing a loss objecting to train the language model, and thus does not teach
generating,…in response to the first indication being positive, a second summary of the source document by editing a portion of the first summary…;
computing a … loss objective by computing a distance in a representation space between the second summary and the third summary, …according to the label; and
training the neural network based language model based on the computed … loss objective.
Sridhar, however, teaches generating,…in response to the first indication being positive, a second summary of the source document by editing a portion of the first summary… (a factual consistency checking model can be fine-tuned on synthetically hallucinated summaries using semantically variant/invariant transformations, i.e. generating…a second summary of the source document by editing a portion of the first summary, where the hallucinated summary is in reference to a gold summary with an NLI score suggesting factual accuracy relative to the input content, i.e. in response to the first indication being positive [0061],[0091],[0096],[0104-5]);
computing a … loss objective by computing a distance in a representation space between the second summary and the third summary, … according to the label (an NLG system receives input text and has a decoder that is trained to generate output text as a summary, can use NLI scoring to measure the factual consistency of the output text, i.e. third summary, and can perform evaluations through a comparison of gold summaries, i.e. second summary and the label, and generated summaries, i.e. third summary [0048],[0060-1],[0065],[0081-3],[0090],[0096], where the training of the neural network is performed using data that has the data and a label, such as the NLI score, and based on a loss function between the ground-truth output and the predicted output, i.e. computing a loss objective, such as a mean squared error, which is based on the target/ground truth output minus the predicted output, i.e. computing a distance in a representation space [0104-5],[0120],[0124],[0127-8]); and
training the neural network based language model based on the computed … loss objective (the neural network system is trained, i.e. training the neural network based language model, using backpropagation including a loss function, i.e. based on the computed loss objective [0042],[0060],[0120],[0124-5],[0127-8]).
Kryscinski and Sridhar are analogous art because they are from a similar field of endeavor in improving text summarization models. Thus, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the evaluating accuracy of generated text summaries from a summarization model teachings of Kryscinski with training the model using a comparison of different summaries with a loss function and backpropagation as taught by Sridhar. It would have been obvious to combine the references to detect and prevent hallucinations in summary generation while outperforming vanilla beam decoding (Sridhar [0054]).
While Kryscinski in view of Sridhar provides computing a loss objective by comparing the summaries, Kryscinski in view of Sridhar does not specifically teach computing a contrastive loss where one summary is either a positive or negative sample, and thus does not teach
computing a contrastive loss objective by computing a distance in a representation space between the second summary and the third summary, with the third summary used as a positive or negative sample according to the label; and
training the neural network based language model based on the computed contrastive loss objective.
Du, however, teaches computing a contrastive loss objective by computing a distance in a representation space between the second summary and the third summary, with the third summary used as a positive or negative sample according to the label (text summarization is used to construct positive and negative sample pairs for contrastive learning, where positive samples are the augmented anchor sample and negative samples are set to all other samples, and some samples may be the concatenation of two summaries, i.e. with the third summary used as a positive or negative sample according to the label, and contrastive learning is performed by minimizing the vector distance between anchor examples and positive examples while maximizing the distance between anchor examples and negative examples, i.e. computing a contrastive loss objective by computing a distance in a representation space between the second summary and the third summary (Sec. 2.1, 3.2-3.4)); and
training the neural network based language model based on the computed contrastive loss objective (the pre-trained transformer is fine-tuned on the data using contrastive learning (Sec. 2.1,3.3)).
Where Sridhar teaches that the loss is calculated between two summaries (see [0048],[0060-1],[0065],[0081-3],[0090],[0096])
Kryscinski, Sridhar, and Du are analogous art because they are from a similar field of endeavor in utilizing text summarization models. Thus, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the computing a loss objective by comparing the summaries teachings of Kryscinski, as modified by Sridhar, with the use of contrastive learning as taught by Du. It would have been obvious to combine the references to improve the supervised contrastive learning by mixing up the samples in different categories utilizing text summarization as a data augmentation method (Du Introduction).
Regarding claims 3, 10, and 17, Kryscinski in view of Sridhar and Du teaches claims 1, 8, and 15, and Du further teaches
computing the loss objective includes computing a value proportional to the distance (contrastive learning is minimizing the vector distance between anchor examples and positive examples, while maximizing the vector distance between anchor examples and negative examples, i.e. computing the loss objective includes computing a value proportional to the distance (Sec. 2.1,3.3)).
Where the motivation to combine is the same as previously presented.
Regarding claims 4, 11, and 18, Kryscinski in view of Sridhar and Du teaches claims 1, 8, and 15, and Sridhar further teaches
computing the loss objective includes inputting the third summary to a factuality detector model, wherein the loss objective is based on an output of the factuality detector model (the NLG system includes a decoder that generates output text, i.e. third summary, that was ranked highest in factual accuracy by an NLI scoring system, where the NLI score of the final output is computed, i.e. inputting the third summary to a factuality detector model…based on an output of the factuality detector model [0049-50],[0058-9],[0072],[0096], where the training of the neural network is performed using data that has the data and a label, such as the NLI score, and based on a loss function between the ground-truth output and the predicted output, i.e. computing a loss objective, such as a mean squared error, which is based on a calculation involving the target/ground truth output minus the predicted output, i.e. loss objective is based on [0104-5],[0120],[0124],[0127-8]).
Where the motivation to combine is the same as previously presented.
Regarding claims 5, 12, and 19, Kryscinski in view of Sridhar and Du teaches claims 4, 11, and 18, and Kryscinski further teaches
the factuality detector model is trained by predicting a factuality of the second summary and comparing the predicted factuality with the label associated with the second summary (each novel claim sentence, i.e. second summary, is labeled correct or incorrect respective to the original sentence, i.e. the label associated with the second summary, where the labeled novel claim sentences are provided as a training data set to a neural network language model for factual consistency verification, i.e. factuality detector model is trained by predicting the factuality of the second summary [0055-6],[0063]).
Where Sridhar further teaches trained by predicting a factuality of the second summary and comparing the predicted factuality with the label (a neural network is trained using training data with labels, where the difference between the target and the predicted output, i.e. predicting a factuality of the second summary, is used to determine a loss, and training is performed to minimize the amount of loss so that the predicted output is the same as the training label, i.e. comparing the predicted factuality with the label [0061-2],[0124-5],[0127-8]).
And where the motivation to combine is the same as previously presented.
Claim(s) 6, 13, and 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kryscinski, in view of Sridhar, in view of Du, and further in view of Boni et al. (U.S. Patent No. 10,019,525), hereinafter Boni.
Regarding claims 6, 13, and 20, Kryscinski in view of Sridhar and Du teaches claims 1, 8, and 15, and Kryscinski further teaches
…a third indication…of linguistic quality of the first summary associated with the source document, wherein the receiving the first indication of factual consistency is in response to the third indication (the summaries can be manually annotated by the user through the user interface, i.e. the receiving the first indication of factual consistency, where unreadable sentences caused by poor generation are not labeled, i.e. a third indication of linguistic quality of the first summary associated with the source document…in response to the third indication [0055-6],[0059-60]).
While Kryscinski in view of Sridhar and Du provides only annotating sentences that are not the result of poor generation, Kryscinski in view of Sridhar and Du does not specifically teach receiving an indication of linguistic quality via the user interface, and thus does not teach
receiving a third indication, via the user interface, of linguistic quality of the first summary associated with the source document.
Boni, however, teaches receiving a third indication, via the user interface, of linguistic quality of the first summary associated with the source document (the highest quality score summary is presented to a user, i.e. receiving a third indication…of linguistic quality of the first summary associated with the source document, where the user can review the selected summary on a graphical user interface, i.e. via the user interface (5:56-67),(7:63-8:19),(9:37-46)).
Kryscinski, Sridhar, Du, and Boni, are analogous art because they are from a similar field of endeavor in improving summary generation. Thus, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the only annotating sentences that are not the result of poor generation teachings of Kryscinski, as modified by Sridhar and Du, with the interface specifically presenting high quality summaries to the user as taught by Boni. It would have been obvious to combine the references to generate high-quality document summaries that satisfy user-defined goals and a provided summary constraint without requiring domain knowledge (Boni (2:9-20)).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NICOLE A K SCHMIEDER whose telephone number is (571)270-1474. The examiner can normally be reached 8:00 - 5:00 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, Pierre-Louis Desir can be reached at (571) 272-7799. 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.
/NICOLE A K SCHMIEDER/Primary Examiner, Art Unit 2659