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 .
Claims 1-13 and 15-21 are pending. This Office Action is responsive to the amendment filed on 06/12/2026, which has been entered in the above identified application.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 3 and 21 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. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Regarding Claim 3, the specification fails to recite or suggest “the training data includes at least one exemplar from the first set of training data”. At best, the specification recites that “the synthetic data 228 is combined with a sample (e.g., portion) of the training data 222 to generating training data used to train the event detection model 226” [0041], but makes no reference to a “exemplar”. As such, this limitation is considered new matter and fails to comply with the written description requirement.
Regarding Claim 21, the specification fails to recite or suggest “the pre-training task includes providing a seed to the generative model”. At best, the specification recites that “values can be used as seeds to the generative model 224 to generate 204 the synthetic data 228” [0042], presumably as part of the “iterative process… during joint training” [0042], but makes no reference to utilizing a seed as part of the pre-training task. As such, this limitation is considered new matter and fails to comply with the written description requirement.
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.
Claims 3 is 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.
Claim 3 recites the limitation "and the training data includes at least one exemplar from the first set of training data". It is unclear what “the training data” refers to, whether it be the first, second, or third sets of training data (i.e. ground truth, synthetic, or test sets) recited in Claim 1, the training data used in the pre-training task in Claim 1, or the “human labeled training data” recited in Claim 3. For the purposes of examination, “the training data” will be interpreted as the combination of the second set of training data and a sample of the first set of training data for training the event detection model, according to paragraph [0041] of the specification (i.e., “the synthetic data 228 is combined with a sample (e.g., portion) of the training data 222 to generating training data used to train the event detection model 226”).
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 1-4, 6-13 and 15-20 are rejected under 35 U.S.C. 103 as being unpatentable over Hataya et al. (“Meta Approach to Data Augmentation Optimization”, published 06/14/2020), hereinafter Hataya; in view of Liu et al. (“MRCAug: Data Augmentation via Machine Reading Comprehension for Document-Level Event Argument Extraction”, published 10/14/2022), hereinafter Liu. Hataya and Liu were cited in a previous Office Action.
Regarding Claim 1, Hataya teaches a method comprising:
obtaining a first set of training data for training a model, the first set training data including labeled data (Hataya: “Let us define a set of input images X and a set of operations S consisting of data augmentation operations such as rotation and color inversion.” [Section 2.1. Designing Data Augmentation Space]);
causing a generative model to generate a second set of training data generated by the generative model (Hataya: “In the AutoAugment family, each image x ∈ X ⊂ [0, 1]D is augmented by an operation O : X → X with a probability of pO ∈ [0, 1] and a magnitude of µO ∈ [0, 1] as illustrated in Figure 1.” [2.1. Designinig Data Augmentation Space]);
training the model based on the first set of training data and the second set of training data (Hataya: “the inner process optimizes parameters of a CNN on training data using a given combination of operations” [Section 1. Introduction]; See [Algorithm 1], where “input = policy(train_data[i]); criterion = cnn.train(input)”; “Our proposed method, MADAO, can optimize a CNN and its data augmentation policy simultaneously by gradient descent in an online manner. Namely, the parameters of the CNN θ is updated to minimize the training loss Ltrain (also written as f), and the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]);
determining a reward value based on performance of the model to detect events based on a third set of training data and a gradient of a loss function based on the second set of training data (Hataya: “we set 10 % of the original training data aside as validation data DV and report error rates on test data.” [Section 4. Experiments and Results]; “Our proposed method, MADAO, can optimize a CNN and its data augmentation policy simultaneously by gradient descent in an online manner. Namely, the parameters of the CNN θ is updated to minimize the training loss Ltrain (also written as f), and the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]; See [Algorithm 1], where “vcriterion = cnn.val(val_data)”; In light of Paragraph [0033] of the specification, which states “the performance of the event detection model 126 Mϴ on the development training data Odev (e.g., measured by F1 scores and/or loss function values) is used as the reward for the synthetic data BG generated by the generative model124 to update Mψ with reinforcement learning”, BRI would support that “determining a reward value based on performance” constitutes calculating loss values with a loss function); and
updating a parameter of the generative model to reduce loss values associated with the loss function based on the reward value (Hataya: “the outer process optimizes the combination of operations to maximize the validation performance.” [Section 1. Introduction]; “Our proposed method, MADAO, can optimize a CNN and its data augmentation policy simultaneously by gradient descent in an online manner. Namely, the parameters of the CNN θ is updated to minimize the training loss Ltrain (also written as f), and the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]).
However, Hataya fails to expressly disclose an event detection model; and causing a generative model, as a result of a pre-training task including training data that is augmented with labels indicating event type information, to generate a second set of training data, not including training data from the first set of training data, the second set of training data including labeled synthetic data.
In the same field of endeavor, Liu teaches an event detection model (Liu: “we devise two data augmentation regimes via MRC, including an implicit knowledge transfer method, which enables knowledge transfer from other tasks to the document-level [event argument extraction] task, and an explicit data generation method, which can explicitly generate new training examples by treating a pre-trained MRC model as an annotator.” [Abstract]); and
causing a generative model, as a result of a pre-training task including training data that is augmented with labels indicating event type information, to generate a second set of training data, not including training data from the first set of training data, the second set of training data including labeled synthetic data (Liu: “we devise two data augmentation regimes via MRC, including an implicit knowledge transfer method, which enables knowledge transfer from other tasks to the document-level EAE task, and an explicit data generation method, which can explicitly generate new training examples by treating a pre-trained MRC model as an annotator.” [Abstract]; “Despite its effectiveness, one disadvantage of implicit knowledge transfer is that it cannot create explicit training data, hence it can only benefit a model in an MRC formulation but not in other formulations [2], [4]. To overcome this issue, we propose another data augmentation approach named as explicit data generation, which can generate new training examples explicitly to enlarge the training set and hence can benefit any model proposed for document-level EAE. As shown in Fig. 3, the method consists of two major stages: a) New Training Data Annotation: The fundamental idea behind explicit data generation is to use the pre-trained MRC model as an annotator to label new instances from unlabeled documents. Let an unlabeled document be D’, and then in our method the following steps are performed: 1) We first identify all event triggers in D’, using an event detector pre-trained on the in-domain labeled triggers.” [Section III.C. Explicit Data Generation via MRC]; “Particularly, in our method, we first create: (i) a clean set, which contains all of the labeled data for the document-level EAE task, and (ii) a wild set, which contains the labeled data from other tasks for the implicit knowledge transfer method, or the automatically generated data for the explicit data generation method.” [Section III.D. Noise Filtering via a Self-Training Regime]; “b) A Joint Training Stage: We devise the following loss function to combine the originally labeled data with the automatically generated data for model training: [Equation 5, 6] where δ is a weight used to balance the contributions of two different forms of data.” [Section III.C. Explicit Data Generation via MRC]; In light of Paragraph [0013] of the specification, which states “the generative model generates a batch of synthetic data (e.g., new labeled training data)”, BRI of “synthetic data” would encompass labeled training data that is generated by a generative model).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated an event detection model; and causing a generative model, as a result of a pre-training task including training data that is augmented with labels indicating event type information, to generate a second set of training data, not including training data from the first set of training data, the second set of training data including labeled synthetic data, as taught by Liu to the method of Hataya because both of these methods are directed towards augmentation of training data by a generative model for joint optimization with a separate neural network. Image classification and event detection are both fields in which a large amount of labeled data is required for model training, but might not exist due to the cost and inadequacy of manual annotation. In making this combination and applying the method of Hataya to the environment of event detection, as well as generating synthetic data for training using a pre-trained generative model, it would allow for the mitigation of the data sparsity problem faced by document level event argument extraction (Liu: [Abstract]) by “generat[ing] new training examples explicitly to enlarge the training set and hence can benefit any model proposed for document-level EAE” (Liu: [Section III.C. Explicit Data Generation via MRC]).
Regarding Claims 8 and 16, they are non-transitory computer readable medium and system claims that correspond to Claim 1. Therefore, they are rejected for the same reasons as Claim 1 above.
Regarding Claim 2, Hataya and Liu teach the method of Claim 1, wherein performance of the event detection model is determined based on the event detection model detecting events within the third set of training data (Hataya: “the outer process optimizes the combination of operations to maximize the validation performance.” [Section 1. Introduction]; “We propose Meta Approach to Data Augmentation Optimization (MADAO), which optimizes CNNs and augmentation policies simultaneously by using gradient based optimization. Here, policies are updated so that they directly increase CNNs’ validation performance.” [Section 1. Introduction]).
Regarding Claim 3, Hataya and Liu teach the method of Claim 1, wherein the first set of training data is sampled from human labeled training data and the training data includes at least one exemplar from the first set of training data (Hataya: “We used CIFAR-10, CIFAR-100, SVHN and ImageNet (ILSVRC-2012). In addition, we also used four fine-grained classification datasets: Oxford 102 Flowers, Oxford-IITT Pets, FGVC Aircraft, and Stanford Cars.” [Section 1. Introduction]; Liu: “we run the following steps: 1) Train a model for document-level EAE on the clean set. 2) Apply the model on the wild set and choose a set of examples with a predictive probability falling in a range [β1, β2] (these examples are the most compatible with the current model). Note that we do not remove them from the wild set to enable a dynamic training paradigm. 3) Retrain the model on a set that combines examples in the clean set and those chosen from the step 2). 4) Repeat step 2) and 3) until convergence.” [Section III.D. Noise Filtering via a Self-Training Regime]; The specification does not explicitly recite an “exemplar”, therefore in light of Paragraph [0041], which recites “the synthetic data 228 is combined with a sample (e.g., portion) of the training data 222 to generating training data used to train the event detection model 226”, BRI of “exemplar” is interpreted to encompass examples (i.e., samples) of the human-labeled training data).
Regarding Claim 4, Hataya and Liu teach the method of Claim 1, wherein the reward value indicates a similarity between the gradient of the loss function and a second gradient of the loss function based on the third set of training data (Hataya: “the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]; “Gradient-based optimization of Equation (1) requires ∇ϕg for iterative updating. Since the data augmentation implicitly affects the validation criterion, in other words, data augmentation is not used for validation, we obtain [Equation 3]. Because of the requirement of g, ∇θg can be obtained.” [Section 3.1. Optimizing Policies by Gradient Descent]).
Regarding Claim 6, Hataya and Liu teach the method of Claim 1, wherein the method further comprises causing the event detection model to perform an event detection task (Liu: “Document-level event argument extraction (EAE) is such a task requiring a model to extract arguments (i.e., participants) of an event at the document level” [Section I. Introduction]).
Regarding Claim 7, Hataya and Liu teach the method of Claim 1, wherein the event detection model is included in an information extraction pipeline (Hataya: “Data augmentation is an effective way to improve the performance of CNN models for image recognition tasks, particularly when its policy is optimized for the target model and dataset.” [Section 1. Introduction]; BRI of information extraction pipeline is that it involves tasks in which information is extracted from data).
Regarding Claim 9, Hataya and Liu teach the medium of Claim 8, wherein the result of the updated event detection model is generated based on a third set of labeled sequences (Hataya: “Optimization of data augmentation policy in AutoAugment family methods can be generalized as [Equation 1] that is, optimizing CNNs on training data with policies that minimize validation criteria on validation data.” [Section 2.2. Generalizing AutoAugment Family]).
Regarding Claim 10, Hataya and Liu teach the medium of Claim 8, wherein updating the parameters of the generative model based on the set of reward values further includes determining a third gradient of a second loss function based on the set of reward values and a set of labels of the second set of labeled sequences (Hataya: “Our proposed method, MADAO, can optimize a CNN and its data augmentation policy simultaneously by gradient descent in an online manner. Namely, the parameters of the CNN θ is updated to minimize the training loss Ltrain (also written as f), and the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]; See [Section 3.2. Approximating Gradients of Policy and Inverse Hessian]),
labels of the set of labels generated by the generative model and indicate an event trigger within the labeled sequences of the second set of labeled sequences (Liu: “we adopt an MRC viewpoint to the document-level EAE task for data augmentation. Let D be a document containing a set of event instances E(D) = {ei}ni=1, each represented by an event trigger.” [Section III. Approach]; “[explicit data generation] uses an MRC model as an annotator to label new training examples for explicitly expanding the training set.” [Figure 3]).
Regarding Claim 11, Hataya and Liu teach the medium of Claim 8, wherein the result of the updated event detection model is generated based on a third set of labeled sequences (Hataya: “Optimization of data augmentation policy in AutoAugment family methods can be generalized as [Equation 1] that is, optimizing CNNs on training data with policies that minimize validation criteria on validation data.” [Section 2.2. Generalizing AutoAugment Family]).
Regarding Claim 12, Hataya and Liu teach the medium of Claim 11, wherein the result indicate performance of the updated event detection model to detect events within the third set of labeled sequences (Hataya: “Optimization of data augmentation policy in AutoAugment family methods can be generalized as [Equation 1] that is, optimizing CNNs on training data with policies that minimize validation criteria on validation data.” [Section 2.2. Generalizing AutoAugment Family]; Liu: “we adopt an MRC viewpoint to the document-level EAE task for data augmentation. Let D be a document containing a set of event instances E(D) = {ei}ni=1, each represented by an event trigger.” [Section III. Approach]).
Regarding Claim 13, Hataya and Liu teach the medium of Claim 8, wherein the first set of labeled sequences are sampled from a set of labeled training data (Hataya: “We empirically demonstrate that MADAO learns effective data augmentation policies and achieves performance comparable or even superior to existing methods on benchmark datasets for image classification: CIFAR-10, CIFAR-100, SVHN, and ImageNet, as well as fine-grained datasets.” [Section 1. Introduction]).
Regarding Claim 15, Hataya and Liu teach the medium of Claim 8, wherein a labeled sequence of the first set of labeled sequences includes a first vector indicating words in the labeled sequence and a second vector indication labels associated with the words (Liu: “In particular, we use a pre-trained MRC model as an annotator to label new training examples in unlabeled documents. For example, we may use a question Who is the attacker in the bombarding event? to query each document, and treat those with answers as new training examples annotated with an attacker role. In contrast to implicit knowledge transfer, explicit data generation can produce tangible training examples, which is shown to benefit a wide range of existing models for the task (e.g., those based on sequence labeling).” [Section I. Introduction]).
Regarding Claim 17, Hataya and Liu teach the system of Claim 16, wherein the processing device to perform the operations comprising pre-training the generative model based on the annotated dataset (Liu: “we use a pre-trained MRC model as an annotator to label new training examples in unlabeled documents.” [Section I. Introduction]).
Regarding Claim 18, Hataya and Liu teach the system of Claim 16, wherein the result of the updated event detection model includes a second gradient of the loss function based on the third set of labeled sequences (Hataya: “Optimization of data augmentation policy in AutoAugment family methods can be generalized as [Equation 1] that is, optimizing CNNs on training data with policies that minimize validation criteria on validation data.” [Section 2.2. Generalizing AutoAugment Family]).
Regarding Claim 19, Hataya and Liu teach the system of Claim 18, wherein the second gradient indicates a performance of the updated event detection model to detect a set of event triggers within the third set of labeled sequences (Hataya: “Optimization of data augmentation policy in AutoAugment family methods can be generalized as [Equation 1] that is, optimizing CNNs on training data with policies that minimize validation criteria on validation data.” [Section 2.2. Generalizing AutoAugment Family]; “Gradient-based optimization of Equation (1) requires ∇ϕg for iterative updating. Since the data augmentation implicitly affects the validation criterion, in other words, data augmentation is not used for validation, we obtain [Equation 3]. Because of the requirement of g, ∇θg can be obtained.” [Section 3.1. Optimizing Policies by Gradient Descent]; Liu: “we adopt an MRC viewpoint to the document-level EAE task for data augmentation. Let D be a document containing a set of event instances E(D) = {ei}ni=1, each represented by an event trigger.” [Section III. Approach]).
Regarding Claim 20, Hataya and Liu teach the system of Claim 19, wherein the reward value indicates a similarity between the gradient of the loss function based on the second set of labeled sequences and the second gradient of the loss function based on the third set of labeled sequences (Hataya: “the parameters of the policy ϕ = {p, µ, π} is updated to minimize the validation loss Lval (also written as g).” [Figure 1]; “Gradient-based optimization of Equation (1) requires ∇ϕg for iterative updating. Since the data augmentation implicitly affects the validation criterion, in other words, data augmentation is not used for validation, we obtain [Equation 3]. Because of the requirement of g, ∇θg can be obtained.” [Section 3.1. Optimizing Policies by Gradient Descent]).
Claim 5 is rejected under 35 U.S.C. 103 as being unpatentable over Hataya in view of Liu, as applied to Claim 4, in further view of Zheng et al. (“Deep AutoAugment”, published 03/15/2022), hereinafter Zheng. Zheng was cited in a previous Office Action.
Regarding Claim 5, Hataya and Liu teach the method of Claim 4. However, they fail to expressly disclose wherein the loss function further comprises a cosine similarity.
In the same field of endeavor, Zheng teaches wherein the loss function further comprises a cosine similarity (Zheng: “the policy is optimized to maximize the cosine similarity between the gradients of the original and augmented data along the direction with low variance.” [Abstract]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated wherein the loss function further comprises a cosine similarity, as taught by Zheng to the method of Hataya and Liu because both of these methods are directed towards augmentation of training data by a generative model for joint optimization with a separate neural network through gradient calculations. Cosine similarity and cross entropy are both commonly used as loss functions when updating parameters of neural networks. In making this combination and utilizing a cosine similarity-based loss function, it would allow for “detect[ing] when an auxiliary loss is helpful to the main loss” to optimize the data augmentation policy (Zheng: [Section 2. Related Work]).
Claim 21 is rejected under 35 U.S.C. 103 as being unpatentable over Hataya in view of Liu, as applied to Claim 11, in further view of Edwards et al. (“Guiding Generative Language Models for Data Augmentation in Few-Shot Text Classification”, published 11/17/2021), hereinafter Edwards. Regarding Claim 21, Hataya and Liu teach the medium of Claim 11. However, they fail to expressly disclose wherein the pre-training task includes providing a seed to the generative model.
In the same field of endeavor, Edwards teaches wherein the pre-training task includes providing a seed to the generative model (Edwards: “In the first step, Seed Selection, we select samples (i.e., seeds) from the original labeled data based on four different strategies (Section 3.1). In the second step, Text Generation, we generate additional artificial training data using a generative language model using three strategies to fine-tune it on different input texts (Section 3.2). Finally, the augmented training data is used in combination with the original data to train a Text Classifier (Section 3.3).” [Section 3. Methodology]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated wherein the pre-training task includes providing a seed to the generative model, as taught by Edwards to the medium of Hataya and Liu because both of these systems are directed towards generating augmented training data using a generative model for training a text-based classifier. In making this combination and providing a seed to the generative model, it would allow the system of Hataya and Liu to “select[[ing]] the most informative samples (or seeds) from the original data”, thereby “prevent[ing] an unnecessary waste of resources and time of generating unused generated documents” (Edwards: [Section 2. Data Augmentation: Related Work]).
Response to Arguments
Examiner acknowledges the Applicant’s amendments to Claims 1, 3, 8, 16, and 21.
Applicant's arguments, filed 06/12/2026, regarding the rejection of Claims 1-13 and 15-21 under 35 U.S.C. § 112(a) with respect to the new matter involving a “prompt” have been fully considered and are persuasive. The rejection has been withdrawn.
The Applicant has not made amendments nor substantive arguments regarding the rejection of Claim 3 under 35 U.S.C. § 112(a) with respect to the new matter involving an “exemplar”. As such, the rejection stands.
Applicant's arguments, filed 06/12/2026, regarding the rejection of Claims 1-13 and 15-21 under 35 U.S.C. § 101 have been fully considered and are persuasive. The rejection has been withdrawn.
Applicant’s arguments, filed 06/12/2026, regarding the rejection of Claims 1-13 and 15-21 under 35 U.S.C. § 103 have been fully considered but are not persuasive.
Applicant alleges, on Pages 7-8 of the Remarks, that the combination of Hataya and Liu fails to teach the limitations of the amended independent Claim 1 because:
A) The cited art does not disclose a generative model that produces, during a pre-training task, new labeled synthetic data distinct from the original training set, as Hataya carries forward existing labels and does not synthesize new labeled sequences or a disjoint “second set” generated by a generative model and Liu fails to cure the deficiencies of Hataya as it uses a pre-trained model as an annotator to mine labels from unlabeled real documents and does not generate synthetic sequences with model-created labels.
B) None of the references recite a reinforcement-learning update policy for the generative model and therefore do not disclose the reward function as claimed, as Hataya uses objective-based policy updates for image augmentation and Liu employs pre-training, fine-tuning, and self-training noise reduction without any reward-driven parameter updates of a generative model.
Examiner respectfully disagrees.
Regarding point A, the written description describes "synthetic training data" as being "e.g., annotated and/or labeled data generated by the generative model" [0004], which Liu explicitly recites ("explicitly generate new training examples by treating a pre-trained MRC model as an annotator" [Abstract]). The Applicant appears to be asserting that Liu does not teach the claim limitations because the new training examples originate from unlabeled documents, implying that "generating synthetic sequences with model-created labels" requires the model to literally create the new training samples from nothing, but this is supported by neither the claims nor the specification. In fact, the written description is largely silent on how the generative model generates synthetic data, merely reciting that synthetic data is new labeled and/or annotated data generated by the generative model, as opposed to being ground truth data labeled by humans. As such, BRI of the claimed "labeled synthetic data generated by the generative model" in light of the specification would encompass Liu's explicit data generation method, in which new training examples are generated by annotating/labeling unlabeled data samples using the pre-trained MRC model. Additionally, the written description does not recite the production of synthetic data during a pre-training task as argued, but rather “one or more pre-training tasks are used to prepare the generative model 124 to generate labeled synthetic data” [0030], in other words, pre-training for the purpose of generating synthetic data during the joint training.
Regarding point B, it is important to note that "reading a claim in light of the specification, to thereby interpret limitations explicitly recited in the claim, is a quite different thing from ‘reading limitations of the specification into a claim,’ to thereby narrow the scope of the claim by implicitly adding disclosed limitations which have no express basis in the claim." See In re Morris, 127 F.3d 1048, 1054-55, 44 USPQ2d 1023, 1027-28 (Fed. Cir. 1997). As such, in this particular instance, a reinforcement learning paradigm is not established in the claim language. The applicant appears to be relying on the "reward" terminology for making this assertion. While a reward is certainly a hallmark of reinforcement learning, it does not on its own exclusively presuppose reinforcement learning, especially when it is being broadly claimed as a performance indicator used to update model parameters, which is a widely-used concept in broader machine learning. While the specification recites the reward as being "for reinforcement learning", it is improper to import those limitations from the specification into the claim language when a BRI analysis of the claimed "reward" comfortably encompasses the evaluation metrics recited in both Hataya and Liu.
As such, the Examiner asserts that the combination of Hataya and Liu teach all of the recited limitations of the amended independent claim 1. Claims 8 and 16 are unallowable for the same reasons discussed above in connection with Claim 1, and Claims 2-7, 9-23, 15, and 17-21 are unallowable for their dependence on an unallowable independent claim, as well as for their own deficiencies described in the 103 prior art rejection (see rejection above).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Cao et al. (“HateGAN: Adversarial Generative-Based Data Augmentation for Hate Speech Detection”) discusses generating an augmented dataset using a deep generative reinforcement learning model for adversarial joint-training with a binary classifier to detect hate speech in social media posts.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MEGAN E HWANG whose telephone number is (703)756-1377. The examiner can normally be reached Monday-Thursday 10:00AM-7:30PM ET.
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, Jennifer Welch can be reached at (571) 272-7212. 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.
/M.E.H./Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights