Notice of Pre-AIA or AIA Status
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 103
2. 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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
3. Claims 1-2, 9-13 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Vu (US 2023/0153687) in view of Hoch (US 2025/0021862).
Regarding Claim 1:
Vu discloses a system, comprising:
a memory that stores computer executable components (Vu: ¶14 discloses a computer system with a memory with executable instructions); and
a processor that executes the computer executable components stored in the memory components (Vu: ¶14 discloses a computer system with a memory with executable instructions operated by a processor), wherein the computer executable components comprise:
an identification component that identifies one or more entities in one or more utterances components (Vu: ¶46 teaches identifying named entities in text/utterances using natural language processing / named entity recognition (NLP/NER). The utterance context is also explicit because Vu defines user requests as utterances and discusses chatbot/user utterances. The identification component is taught through the natural language utterances (NLU) and NER functions);
a simulation component that generates simulated utterances based on the one or more entities components (Vu: ¶226 each named entity is replaced by a corresponding entity type tag to generate a template, examples include <LOC>, <PER>, and <ORF>. ¶227 generates an invariance group by iteratively replacing the entity type tag with a named entity selected from the list of named entities, this is analogous to simulated utterances);
a model component that executes the simulated utterances on nodes of a natural language processing (NLP) model to identify bias components (Vu: ¶¶114-115 teach a model component that executes the simulated utterances on nodes of the NLP model to identify bias. Specifically, teaching a model system including a sentiment analysis model that may comprise a deep neural network and a transformer based NLP model. ¶159 teaches using an NER model to identify named entities in utterances and replaces the entities with corresponding entity type tags to create templates. ¶164 then selects multiple named entities of the same type and replaces the entity type tag with each selected entity to generate an invariance groups of simulated utterances that differ only in the substituted entity. ¶217 evaluates the sentiment analysis model against the invariance group, therefore executing each simulated utterance through the disclosed neural NLP model and its hidden computational units, weights and biases. ¶164-168 discloses comparing the predicted sentiment probabilities for the simulated utterances and determines the model fails the named entity behavioral test when the change in probabilities exceeds a threshold, thereby identifying named entity bias in the NLP model); and
a selection component that mitigates the bias (Vu: ¶234-239 discloses different named entity debiasing methods including tag replacement, tag replacement + K1, tag replacement + sum of squared differences and tag replacement + max absolute differences)
Vu does not disclose a selection component that mitigates the bias by selecting one or more debiasing methods using quadratic algorithm selection components.
However, Hoch discloses a selection component that mitigates the bias by selecting one or more debiasing methods using quadratic algorithm selection components (Hoch: ¶68 teaches a processor executed feature selection process configured to select a best subset from a plurality of candidate features and to store the selected subset. ¶80- 81 discloses selecting a target subset from the plurality of target candidate features by defining a target function according to the candidate features and their correlation with an expected outcome of a machine learning model and finding a minimum or maximum of the target function representing the selected subset. ¶83 discloses automatically converting or defining the feature selection process as a quadratic unconstrained binary optimization or QUBO, wherein the minimum or maximum indicates the selected candidates).
Vu and Hoch disclose pertinent disclosures. Vu teaches multiple named entity debiasing approaches. Hoch describes the selectable candidates as features. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose applying the quadratic subset selection process to Vu’s plurality of candidate debiasing methods because each debiasing method constitutes a selectable candidate whose expected effect on bias and model performance can be represented in the target function. The resulting system would select the debiasing method or combination of methods represented by the minimum or maximum of the quadratic binary objective. The motivation for doing so is directly taught in ¶48 of Hoch “ improve upon the aforementioned technical field, and/or improve upon prior approaches, by extracting features from a current dataset (which is being fed into the ML model) and a historical dataset. The features are analyzed to determine whether the features extracted from the current dataset are sufficiently different from features extracted from historical dataset.”
Regarding Claim 2:
The proposed combination of Vu in view of Hoch further discloses the system of claim 1, wherein generating the simulated utterances comprises:
identifying respective matching equivalents of the one or more entities (Vu: ¶164 teaches selecting for each template a subset of named entities having the same entity type or class such as LOCATION);
forming one or more stub sentences by replacing the one or more entities with the respective matching equivalents (Vu: ¶159 and 164 discloses replacing each original entity with its corresponding entity type tag therefore creating an utterance template and then replaces the entity type tag with each selected same type entity. Vu explicitly states that the resulting examples differ only in the replaced value of the entity tag); and
executing the one or more stub sentences on the nodes of the NLP model (Vu: ¶164 teaches evaluating the sentiment analysis model against the generated invariance group test).
Regarding Claim 9:
The proposed combination of Vu in view of Hoch further discloses the system of claim 1, wherein the simulated utterances comprise inter-utterances or intra-utterances (Vu: ¶164 teaches generating an invariance group containing multiple separate utterances by replacing an entity type tag in a template with each of multiple type entities, such that the generated utterances differ from one another only in the replaced entity value. The separately generated and compared utterances teach inter utterances. ¶180 teaches also teaches generating an utterance from a template containing multiple entity tags within the same sentence such as replacing two <LOC> tags with “Portland and Geelong: or “Fremantle and Kimberley” thereby teaching intra utterances having multiple entity replacements within a single utterance).
Regarding Claim 10:
The proposed combination of Vu in view of Hoch further discloses the system of claim 1, further comprising: a display component that displays, via a user interface, corresponding bias factors for the one or more entities (Vu: ¶144 and ¶156-157 discloses determining and presenting for each named entity corresponding sentiment probabilities, predicted sentiment labels and failed template percentages, a high failed template percentage confirms bias against the corresponding named entity. ¶245 and ¶283 further discloses a client device interface configured to output information to a user and a user interface display subsystem configured to visually convey textual and graphical information).
Regarding Claim 11:
The proposed combination of Vu in view of Hoch further discloses the system of claim 1, wherein the NLP model employs a large language model (Vu: teaches that the NLP model employs a large language model. Specifically ¶115 discloses the NLP model may be a transformer model such as Bidirectional Encoder Representations from Transformers (BERT) trained for natural language processing, understanding or generation).
Regarding Claim 12:
Claim 12 has been analyzed with regard to claim 1 (see rejection above) and is rejected for the same reasons of obviousness as used above.
Regarding Claim 13:
Claim 13 has been analyzed with regard to claim 2 (see rejection above) and is rejected for the same reasons of obviousness as used above.
Regarding Claim 19:
Claim 19 has been analyzed with regard to claim 1 (see rejection above) and is rejected for the same reasons of obviousness as used above.
Regarding Claim 20:
Claim 20 has been analyzed with regard to claim 2 (see rejection above) and is rejected for the same reasons of obviousness as used above.
4. Claims 3 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch and further in view of Cao (US 2021/0034998).
Regarding Claim 3:
The proposed combination of Vu in view of Hoch further discloses the system of claim 1, except wherein the quadratic algorithm selection uses a binary selector to select the one or more debiasing methods based on a quadratic function.
However, (Cao: ¶75 expressly teaches the binary vector encodes the choice for the subset of features).
Vu and Hoch in view of Cao are combinable because Cao discloses pertinent teachings to the combination. Vu and Hoch in combination teach selecting debiasing algorithms using a quadratic selection algorithm. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to implement Hoch’s quadratic selection process using Cao’s explicitly defined binary vector encoding. The motivation for doing so is explained by Cao in ¶71 “ Currently this challenge typically is addressed manually using human intuition, relying on domain knowledge related to the nature of the data at hand. The success of this approach often is limited in cases in which a vast number of features are concerned.” In other words, manually selecting relevant candidates using human intuition has limited success when the number of candidates is large because the number of possible subsets quickly becomes intractable.
Regarding Claim 15:
Claim 15 has been analyzed with regard to claim 3 (see rejection above) and is rejected for the same reasons of obviousness as used above.
5. Claims 4 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch and further in view of Bellamy “AI FAIRNESS 360: AN EXTENSIBLE TOOLKIT FOR DETECTING, UNDERSTANDING, AND MITIGATING UNWANTED ALGORITHMIC BIAS”.
Regarding Claim 4:
The proposed combination of Vu in view of Hoch further discloses the method system of claim 1, except wherein the one or more debiasing methods comprise fair pre- processing, fair in-processing or fair post-processing.
However, Bellamy discloses wherein the one or more debiasing methods comprise fair pre- processing, fair in-processing or fair post-processing (Bellamy: Section 3 teaches that the one or more debiasing methods comprise fair preprocessing, fair in processing or fair post processing and explains that these categories respectively improve fairness by modifying the training data, modifying the learning procedure of the machine learning model or modifying the predictions produced by the trained model).
Vu, Hoch and Bellamy is analogous because each concerns machine learning models and techniques for mitigating undesirable or biased model behavior. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to implement Vu’s bias mitigation system using Bellamy’s fair pre-processing, fair in processing and fair post processing categories. Bellamy explains that Bias handling algorithms operate at different portions of the model life cycle and that understanding how, when and why to use the different techniques is challenging even for experts in the Introduction.
Regarding Claim 14:
Claim 14 has been analyzed with regard to claim 4 (see rejection above) and is rejected for the same reasons of obviousness as used above.
6. Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch, further in view of Cao and further in view of Moreira-Matias (US 2019/0318248).
Regarding Claim 5:
The proposed combination of Vu, Hoch and Cao further discloses the system of claim 3, wherein the selection component minimizes the quadratic function (Hoch: ¶81 defines a target function, finds its minimum and states that the minimum represents the selected subset)
The proposed combination of Vu, Hoch and Cao do not explicitly disclose to select a minimum number of the one or more debiasing methods.
However, Moreira-Matias discloses to select a minimum number of the one or more debiasing methods (Moreira-Matias: ¶27 discloses evaluating selected features, determines whether the number of selected features is less than a set number, remove the two least important features and retrains and reevaluates the model, ¶28 discloses the method continues until the stopping criteria is met (minimum number of features OR maximum error allowed) ¶29 discloses a minimum set of relevant features for the prediction task at hand).
Vu, Hoch and Cao in view of Moreira-Matias are combinable because each concerns machine learning system and automated selection among candidate techniques or features affecting model performance. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to configure Hoch and Cao’s quadratic selection process according to Moreira-Matias’ teaching of progressively removing the least important candidates until a minimum number of relevant candidates remains. Moreira-Matias explicitly discloses that its feature space reduction heuristic can produce “faster results with similar predictive power” and that reducing the effective dimensionality of the selection space considerably speeds up an otherwise computationally intensive optimization procedure in ¶23.
7. Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch, further in view of Cao and further in view of Liang “Towards Understanding and Mitigating Social Biases in Language Models.”
Regarding Claim 6:
The proposed combination of Vu, Hoch and Cao further discloses the system of claim 3, except wherein the selection component minimizes the quadratic function based on a liveness metric (Liang: Section 5.3 explanation of Table 7 and 8 discloses clarity defined as sentence coherence, grammar and spelling, content, defined as whether the sentence preserves realistic semantic meaning and language model performance measured using an language model score, and discloses that its debiased model retains clarity and content while improving fairness. Liang also values different debiasing settings using a fairness performance plot in which performance is measured by LM score and identifies settings that maintain the original model’s LM score while improving fairness).
Vu, Hoch and Cao in view of Liang are combinable because each concerns machine learning models and the selection or application of techniques affecting model bias and performance. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify Cao’s quadratic selection function to account for Liang’s language quality metric when selecting among Vu’s debiasing methods. Liang explicitly explains that controlling the level of debiasing can produce an initial improvement in fairness at only a small expense in model performance just under Table 5 of Section 5.3.
8. Claim 16 is rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch, further in view of Cao, further in view of Moreira-Matias and further in view of Liang.
Regarding Claim 16:
The proposed combination of Vu, Hoch and Cao further discloses the computer-implemented method of claim 15, further comprising:
minimizing, by the system, the quadratic function (Hoch: ¶81 defines a target function, finds its minimum and states that the minimum represents the selected subset)
The proposed combination of Vu, Hoch and Cao do not explicitly disclose:
to select a minimum number of the one or more debiasing methods
However, Moreira-Matias discloses to select a minimum number of the one or more debiasing methods (Moreira-Matias: ¶27 discloses evaluating selected features, determines whether the number of selected features is less than a set number, remove the two least important features and retrains and reevaluates the model, ¶28 discloses the method continues until the stopping criteria is met (minimum number of features OR maximum error allowed) ¶29 discloses a minimum set of relevant features for the prediction task at hand).
Vu, Hoch and Cao in view of Moreira-Matias are combinable because each concerns machine learning system and automated selection among candidate techniques or features affecting model performance. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to configure Hoch and Cao’s quadratic selection process according to Moreira-Matias’ teaching of progressively removing the least important candidates until a minimum number of relevant candidates remains. Moreira-Matias explicitly discloses that its feature space reduction heuristic can produce “faster results with similar predictive power” and that reducing the effective dimensionality of the selection space considerably speeds up an otherwise computationally intensive optimization procedure in ¶23.
The proposed combination of Bu, Hoch, Cao and Moreira-Matias does not explicitly disclose
minimizing, by the system, the quadratic function based on a liveness metric
However, Liang discloses minimizing, by the system, the quadratic function based on a liveness metric (Liang: Section 5.3 explanation of Table 7 and 8 discloses clarity defined as sentence coherence, grammar and spelling, content, defined as whether the sentence preserves realistic semantic meaning and language model performance measured using an language model score, and discloses that its debiased model retains clarity and content while improving fairness. Liang also values different debiasing settings using a fairness performance plot in which performance is measured by LM score and identifies settings that maintain the original model’s LM score while improving fairness).
Vu, Hoch, Cao, Moreira-Matias in view of Liang are combinable because each concerns machine learning models and the selection or application of techniques affecting model bias and performance. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify Cao’s quadratic selection function to account for Liang’s language quality metric when selecting among Vu’s debiasing methods. Liang explicitly explains that controlling the level of debiasing can produce an initial improvement in fairness at only a small expense in model performance just under Table 5 of Section 5.3.
9. Claims 7-8 and 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over Vu in view of Hoch, further in view of Cao, further in view of Liang and further in view of Xu “Fairness GAN: Generating datasets with fairness properties using a generative adversarial network.”
Regarding Claim 7:
The proposed combination of Vu, Hoch, Cao and Liang further discloses the system of claim 6, except wherein the selection component measures the liveness metric using Generative Adversarial Networks (GANs), wherein a generator produces debiased results, and wherein a discriminator determines if the debiased results are synthetic data.
However, Xu discloses wherein the selection component measures the liveness metric using Generative Adversarial Networks (GANs), wherein a generator produces debiased results, and wherein a discriminator determines if the debiased results are synthetic data (Xu: Section 3.2 teaches that a GAN comprises a generator G that generates fake samples and a discriminator D that operates as a binary classifier predicting whether an input is real or fake data generated by G. The discriminator outputs a probability that the input is real rather than generated fake data. Section 4.2 further teaches that the generated produces fake samples conditioned on a protected attribute and that discriminator D1 distinguishes real data from generated fake data and that discriminator D2 imposes a fairness constraint on the generated samples. Therefore, the generator produces fair or debiased synthetic results, while D1 determines whether those results constitute generated synthetic data).
Vu, Hoch, Cao, Liang and Xu are combinable because each concerns machine learning techniques for reducing bias while preserving the useful quality or performance of model results. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose a generator and discriminator architecture to measure the liveness metric considered by the quadratic selection process when selecting among Vu’s debiasing methods. Xu’s introduction explains that GANs demonstrate impressive performance in modeling real data distribution and generating high quality synthetic data similar to real data and that the system generates discrimination free data while preserving good data utility.
Regarding Claim 8:
The proposed combination of Vu, Hoch, Cao, Liang and Xu further discloses the system of claim 7, wherein the liveness metric equals an inverse of an accuracy of the discriminator (Xu: Section 4.3 discloses that the discriminator is a binary classifier that predicts real or fake data and outputs the probability that the input is real. The generator is trained to fool the discriminator so that the discriminator is unable to distinguish generated data from real data. Therefore, the discriminator accuracy and the realism or liveness of the generated results have an inverse relationship, as the generated results become more realistic, the discriminator becomes less accurate).
Regarding Claim 17:
Claim 17 has been analyzed with regard to claim 7 (see rejection above) and is rejected for the same reasons of obviousness as used above.
Regarding Claim 18:
Claim 18 has been analyzed with regard to claim 8 (see rejection above) and is rejected for the same reasons of obviousness as used above.
Conclusion
The following prior art made of record but not relied upon in the rejection is considered pertinent to Applicant’s disclosure:
Shrawgi et al. US 2026/0073192 discloses a framework for removing or mitigating inconsistency bias in generative artificial intelligence (AI) model responses, which inherently provide generative outputs that may include biases for certain groups. Specifically, this disclosure describes a model bias removal system (e.g., a model inconsistency bias mitigation system) that influences a generative AI model to respond to user prompts without inconsistency biases while not influencing or affecting other aspects of the model's ability to generate user responses. By doing so, the model bias removal system improves the accuracy and efficiency of generative AI models. Additionally, the model bias removal system enhances the fairness, consistency, and impartiality of generative AI model responses.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to IAN SCOTT MCLEAN whose telephone number is (703)756-4599. The examiner can normally be reached "Monday - Friday 8:00-5:00 EST, off Every 2nd Friday".
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, Hai Phan can be reached at (571) 272-6338. 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.
/IAN SCOTT MCLEAN/Examiner, Art Unit 2654
/HAI PHAN/Supervisory Patent Examiner, Art Unit 2654
Prosecution Insights