DETAILED ACTION
This communication is in response to the Application filed on 02/04/2026. Claims 1-20 are pending and have been examined.
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 02/04/2026 has been entered.
Response to Arguments
The reply filed on 02/04/2026 has been entered. Applicant’s arguments with respect to claims 1-20 have been considered but are not persuasive/moot in view of new ground(s) of rejection caused by the amendments.
With respect to the applicant’s arguments to claim rejections under 35 U.S.C § 112(a), Applicant has amended each of the independent claims and asserts that “the specification clearly demonstrates possession of each and every limitation of the pending claims, including claims 1, 8, and 15, through detailed architectural descriptions, algorithmic workflows, and illustrative embodiments. In particular, the specification expressly describes (i) pre-trained natural language models trained on unstructured text corpora, including encoder-decoder neural network architectures, (ii) the use of sentence-level semantic embeddings to represent unstructured text reports in a fixed-dimensional embedding space, (iii) deep learning autoencoders trained in an unsupervised manner on the same training corpus used for the natural language model, (iv) reconstruction-error computation based on distance or similarity metrics, and (v) the use of reconstruction-error behavior to detect semantic drift and control model lifecycle actions such as alerting, inference suspension, and retraining. These features are not introduced for the first time in the claims, but rather are drawn directly from and consistently reflected throughout the specification and accompanying drawings.” The examiner respectfully disagrees with these assertions. The applicant states that the specification demonstrates possession, but then proceeds to merely state features derived from claim limitations. In particular, the applicant fails to cite any part of the specification to support their assertions.
Applicant further asserts that “while the claims recite distributions of reconstruction errors, monitoring intervals, and adaptive thresholds, the specification describes reconstruction-error behavior over time, statistical comparison of current versus training-time representations, and dynamic drift detection mechanisms that inherently encompass the claimed formulations. A person of ordinary skill in the art would readily recognize these disclosures as supporting the full scope of the claims.” The examiner respectfully disagrees with these assertions. First, the applicant asserts that the specification describes the above features, but fails to cite specific text within the specification to support such assertions. Second, the examiner disagrees that a person of ordinary skill in the art would readily recognize these disclosures as supporting the full scope of the claims. As described in further detail below with respect to claim rejections under 35 USC § 112(a), the specification lacks a written description of a dynamic threshold.
With respect to the applicant’s arguments to claim rejections under 35 U.S.C § 103, the applicant’s arguments with respect to claims 1-20 have been considered but are moot in view of new ground(s) of rejection caused by the amendments.
The applicant’s remarks fail to respond to claim rejections under 35 USC § 112(b).
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.
Claims 1-20 are rejected under 35 U.S.C. 112(a) 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, had possession of the claimed invention.
Independent claim 1 recites:
“an access component that (i) accesses a pre-trained natural language model comprising an encoder-decoder neural network” (emphasis added).
Applicant fails to identify paragraphs or figures of the instant application that allegedly support these amendments.
The specification fails to provide an adequate written description of a pre-trained natural language model comprising a “decoder.” Applicant failed to show adequate support in their instant specification for these amendments in direct contradiction to the requirements of MPEP 2163(II)(A) and 2163.04. Furthermore, the support for these limitations is not apparent. The closest specification to supporting these amendments is paragraph [0030], “In various aspects, the pre-trained natural language model can exhibit any suitable deep learning internal architecture. For example, the pre-trained natural language model can include any suitable numbers of any suitable types of layers (e.g., input layer, one or more hidden layers, output layer, any of which can be convolutional layers, dense layers, non-linearity layers, pooling layers, batch normalization layers, or padding layers).” Paragraph [0030] of the specification supports the pre-trained natural language model comprising a neural network, but there is no language in the specification that supports the pre-trained natural language model to also comprise a decoder. Thus, the amended limitations relating to the “decoder” and their use in the pre-trained natural language model are not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventors, at the time the application was filed, had possession of the claimed invention.
Independent claims 8 and 15 are similarly rejected due to parallel claim language.
Further, independent claim 8 recites:
“wherein the threshold margin is dynamically adapted based on a rolling average of recent reconstruction errors computed across successive monitoring intervals and a variance term representing temporal instability in reconstruction error behavior” (emphasis added).
Applicant fails to identify paragraphs or figures of the instant application that allegedly support these amendments.
The specification fails to provide an adequate written description of “a rolling average of recent reconstruction errors and a drift variance parameter” or “a variance term representing temporal instability in reconstruction error behavior” and their association with a dynamically adapted threshold margin. Applicant failed to show adequate support in their instant specification for these amendments in direct contradiction to the requirements of MPEP 2163(II)(A) and 2163.04. Furthermore, the support for these limitations is not apparent. The closest specification to supporting these amendments is paragraphs [0046]-[0047], “example, the drift component can compute a first aggregated reconstruction error (e.g., average reconstruction error, mean square reconstruction error, root mean square reconstruction error) based on the first set of reconstruction errors, and the drift component can likewise compute a second aggregated reconstruction error based on the second set of reconstruction errors. If the first aggregated reconstruction error is within any suitable threshold margin of the second aggregated reconstruction error, the drift component can determine that the set of inferencing unstructured text reports is sufficiently similar to, and thus has not drifted too far from, the set of training unstructured text reports.” Paragraphs [0046]-[0047] of the specification supports aggregating reconstruction errors, but there is no language in the specification that supports dynamically adapting the threshold margin using this aggregated reconstruction error. Further, there is no language in the specification regarding a variance term that represents temporal instability. Thus, the amended limitations relating to a “rolling average” and a “variance term” as well as their use in dynamically adapting a threshold margin, are not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventors, at the time the application was filed, had possession of the claimed invention.
Claims 2-7, 9-14, and 16-20 are rejected due to their dependency upon claims 1, 8, and 15.
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.
Claim 1 is rejected under 35 U.S.C. 112(b), 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 1, line 19 recites the limitation "original collections". It is unclear whether the term “original collections” refers to the collection of sentence embeddings converted by the autoencoder from text reports, or the collection of sentence embeddings received from executing the pre-trained natural language model on text reports. There is insufficient antecedent basis for this limitation in the claim.
Claims 2-7 are rejected due to their dependency upon claims 1.
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, 3-7, 8 and 10-14 are rejected under 35 U.S.C. 103 as obvious over “TCR-M: A Topic Change Recognition-based Method for Data Stream Learning”, (Wang et al.) in view of US Patent Publication 20230316045 A1 (Rama et al.) in view of “Autoencoder-based Anomaly Detection in Streaming Data with Incremental Learning and Concept Drift Adaptation” (Li et al.) in view of "BED, Bi-Encoder-Based Detectors for Out-of-Distribution Detection" (Owen et al.).
Claim 1
Regarding claim 1, Wang et al. disclose a system comprising:
a processor that executes computer-executable components stored in a non-transitory computer-readable memory (Wang et al. pg. 3007, Section 4.3, Paragraph 2, "The experiment is implemented by Python 3.7, and the computational environment is: Red Hat Enterprise Linux Workstation release 7.9 (Maipo), Intel(R) Xeon(R) Gold 6238R CPU @ 2.20GHz. The Learn++.NSE [6] and LeverageBagging[3] were implemented by the ScikitMultiflow meta module [22], with default parameter settings." A computation environment as described above implies the use of non-transitory computer-readable memory), wherein the computer-executable components comprise:
an access component that (i) accesses a pre-trained natural language model (Wang et al. pg. 3005, Section 3.3, Paragraph 2, "we separately train a base learner
f
to help correct the model results. This base learner is independent and updated based on the topic change recognition results. Unlike the ensemble learning model, which is retrained and updated at every moment, this learner is separately retrained only when the data changes." Model
f
is considered analogous to a pre-trained natural language model) [comprising an encoder-decoder neural network] trained on a plurality of unstructured training text reports (Wang et al. pg. 3007, Section 4.2, Paragraph 1, "The datasets chosen to be tested in the experiment is online user reviews collected from e-commerce websites. ... The data were organized and pre-processed to eliminate some invalid reviews, and initially organized into a continuous stream of data to help the model conduct test experiments.") to perform an inferencing task, wherein the inferencing task comprises generating a [structured] analytical output from unstructured text (Wang et al. pg. 3004, Section 3.2, Paragraph 1, "we use the LDA model [4] to help extract topics from data instances. Here, assuming there are
K
topics extracted, ... the probability of each topic
P
β
k
is calculated, where
k
is the topic index."), and (ii) receives a set of unstructured text reports on which the pre-trained natural language model is to be executed (Wang et al. pg. 3006, Algorithm 1, "Input: Data chunks
D
of text stream data.");
a drift component (Wang et al. pg. 3005, Section 3.3, Paragraph 1, "we first learn a Bagging model [5]
F
(
x
)
with K base learners, and retrain it at each time point." Model
F
is considered analogous to a drift component) that determines, prior to execution of the pre-trained natural language model on the set of unstructured text reports (Wang et al. pg. 3006, Algorithm 1 line 11 executes pre-trained natural language model
f
on chunk
t
. Thus, lines 5-9 of Algorithm 1 are considered analogous to determining drift prior to executing the pre-trained natural language model), whether a statistical distribution of semantic representations (Wang et al. pg. 3007, Section 4.2, Paragraph 1, "The data were organized and pre-processed to ... initially organized into a continuous stream of data to help the model conduct test experiments.") associated with the set of unstructured text reports deviates from a statistical distribution learned during training of the pre-trained natural language model (Wang et al. pg. 3005, Section 3.2, Paragraphs 1-3, "assuming there are K topics extracted, ... the probability of each topic can be expressed as [Equation 3] ... we simply calculate the topic change severity of two consecutive time steps by [see equation 5]. Thus, the topic change severity
∆
t
and
∆
t
+
1
at time
t
and time
t
+
1
can be got. Each of them is an array that contains
K
elements. ... To identify whether the topic change is significant, we use the statistical test to calculate the significant level of the topic change severity of two consecutive time points.") via execution of a [deep learning autoencoder having an encoder portion and a decoder portion] model trained [in an unsupervised manner] on the same plurality of training [unstructured text reports] used to train the pre-trained [natural language] model (Wang et al. pg. 3006, Algorithm 1, lines 2-3, "train a Bagging model
F
K
with
K
base learners on chunk 1; Train a single decision tree model
f
with
K
base learners on chunk 1")…; and
a result component that generates, in response to a determination that [the first distribution of reconstruction errors differs from the second distribution of reconstruction errors by more than a threshold margin indicative of semantic drift, a first alert indicating that] data drift has occurred and that the pre-trained natural language model is thereby unable to confidently analyze the set of unstructured text reports (Wang et al. pg. 3002, Section 1, Paragraph 2, "A data stream learning process triggered by topic change recognition results is given. The learning model will self-update when a change is identified."), and that execution of the pre-trained natural language model on the set of unstructured text reports would produce unreliable inferencing results (Wang et al. pg. 3002, Section 2.1, Paragraph 1, "Concept drift in the data stream will affect the learning efficiency that the machine learning model may not adapt to the change in time, leading to learning decay and performance degradation."), and automatically prevent execution of the pre-trained natural language model on the set of unstructured text reports until a retraining operation is initiated using the plurality of training unstructured text reports (Wang et al. pg. 3005, Section 3.3, Paragraph 2, "we separately train a base learner
f
to help correct the model results. This base learner is independent and updated based on the topic change recognition results. Unlike the ensemble learning model, which is retrained and updated at every moment, this learner is separately retrained only when the data changes.").
Wang et al. do not explicitly disclose all of an autoencoder for drift detection.
However, Rama et al. disclose a drift component that determines, prior to execution of the pre-trained [natural language] model [on the set of unstructured text reports], whether a statistical distribution of semantic representations associated with the set of unstructured text reports deviates from a statistical distribution learned during training of the pre-trained [natural language] model (Rama et al. ¶ [0034], "re-construction loss determiner 106 may be configured to determine the mean squared error between each piece of data provided to autoencoder 105 (i.e., input feature vector(s) 108) and the reconstructed version of that data (i.e., reconstructed data 112).") via execution of a deep learning autoencoder (Rama et al. ¶ [0028]-[0033], "Data drift determiner 102 is configured to receive one or more input feature vector(s) 108, each comprising a plurality of feature values utilized to train and/or build a machine learning model (e.g., machine learning model 104) … Autoencoders, such as autoencoder 200 are utilized for deep learning techniques; in particular, autoencoders are a type of a neural network. ") having an encoder portion and a decoder portion (Rama et al. ¶ [0032], "Autoencoder 200 generally comprises three parts: an encoder, a bottleneck, and a decoder") trained in an unsupervised manner (Rama et al. ¶ [0030], "Autoencoder 105 may comprise a self-supervised neural network." Self-supervised learning is considered a subset of unsupervised learning) [on the same plurality of training unstructured text reports used to train the pre-trained natural language model], wherein the drift component determines how different a first distribution of reconstruction errors [associated with the set of unstructured text reports] is from a second distribution of reconstruction errors [associated with the plurality of training unstructured text reports] (Rama et al. ¶ [0034], "re-construction loss determiner 106 may be configured to determine the mean squared error between each piece of data provided to autoencoder 105 (i.e., input feature vector(s) 108) and the reconstructed version of that data (i.e., reconstructed data 112).") by
[(a) converting each report into a collection of sentence-level semantic embeddings using a universal sentence encoding technique that maps sentences into a fixed-dimensional embedding space],
(b) generating, by the encoder portion, a dimensionally-reduced latent vector representation of the collection of [sentence-level semantic] embeddings (Rama et al. ¶ [0032], "The encoder (or encoder network) encodes the input data (i.e., input feature vector(s) 108) into increasingly lower dimensions. That is, the encoder is configured to compress the input data (i.e., input feature vector(s) 108) into an encoded representation that is typically several orders of magnitude smaller than the input data." A compressed encoded representation is considered analogous to a dimensionally-reduced latent vector representation),
(c) reconstructing, by the decoder portion, a reconstructed collection of sentence-level semantic embeddings from the latent vector representation (Rama et al. ¶ [0032], " the decoder is configured to decompress the knowledge representations and reconstruct input feature vector(s) 108 back from their encoded form."), and
(d) computing, by the drift component, for respective reports, a numerical reconstruction error that quantifies divergence between the reconstructed collection and the original collection, based on at least one of a Euclidean-distance metric or cosine-similarity metric (Rama et al. ¶ [0039], "the Euclidean distance between the re-constructed input and the original input, is the loss, and is denoted by Equation 6."); and
a result component that generates, in response to a determination that the first distribution of reconstruction errors differs from the second distribution of reconstruction errors by more than a threshold margin indicative of [semantic] drift (Rama et al. ¶ [0034]-[0040], "re-construction loss determiner 106 may be configured to determine the mean squared error between each piece of data provided to autoencoder 105 (i.e., input feature vector(s) 108) and the reconstructed version of that data (i.e., reconstructed data 112). ... Weighted re-construction loss value 114 is provided to threshold condition analyzer 116. Threshold condition analyzer 116 is configured to determine whether weighted re-construction loss value 114 meets a threshold condition"), a first alert indicating that data drift has occurred and that the [pre-trained natural language] model is thereby unable to confidently analyze the set of unstructured text reports (Rama et al. ¶ [0042], "In response to detecting that data drift has occurred with respect to the more important, threshold condition analyzer 116 may cause an action to be performed. For example, threshold condition analyzer 116 may issue a notification 118 (e.g., to an administrator) that indicates that the data drift has been detected and that indicates that machine learning model 104 should be de-activated"), and that execution of the [pre-trained natural language] model on the set of unstructured text reports would produce unreliable inferencing results (Rama et al. ¶ [0019], "When the models are in production and degrade or suffer from under performance, it is referred to as drift." Detecting drift is considered analogous to detecting whether a model would produce unreliable inference results), and automatically prevent execution of the pre-trained natural language model on the set of unstructured text reports until a retraining operation is initiated using the plurality of training unstructured text reports (Rama et al. ¶ [0042], "threshold condition analyzer 116 may cause machine learning model 104 to be automatically de-activated and/or re-trained by sending a command 120 to an application and/or service that manages machine learning model 104.").
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify Wang et al.’s drift detection system to include Rama et al.’s deep learning autoencoder because such a modification is the result of simple substitution of one known element for another producing a predictable result. More specifically, Wang et al.’s drift component and Rama et al.’s deep learning autoencoder perform the same general and predictable function, the predictable function being calculating data drift from reconstruction loss. Since each individual element and its function are shown in the prior art, albeit shown in separate references, the difference between the claimed subject matter and the prior art rests not on any individual element or function but in the very combination itself - that is in the substitution of Wang et al.’s drift component by replacing it with Rama et al.’s deep learning autoencoder. Thus, the simple substitution of one known element for another producing a predictable result renders the claim obvious.
Wang et al. in view of Rama et al. do not explicitly disclose all of a pre-trained model comprising an encoder-decoder neural network.
However, Li et al. disclose a pre-trained [natural language] model (Li et al. pg. 4, Algorithm 1, line 2, "
h
=
p
r
e
t
r
a
i
n
_
m
o
d
e
l
(
D
)
") comprising an encoder-decoder neural network (Li et al. pg. 3, Section (IV)(A), Paragraph 1, "We consider an AE, which is a special type of a neural network that attempts to copy its input
h
:
X
→
X
[31]. it consists of an encoder followed by a decoder").
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify Wang et al. in view of Rama et al. to include Li et al.’s pretrained encoder-decoder neural network because such a modification is the result of combining prior art elements according to known methods to yield predictable results. More specifically, Wang et al.’s pretrained language model as modified by Li et al.’s pretrained encoder-decoder neural network can yield a predictable result of increasing system applicability since an encoder-decoder neural network would allow for an application to generate text instead of only classifying it. Thus, a person of ordinary skill would have appreciated including in Wang et al.’s pretrained language model the ability to do Li et al.’s pretrained encoding and decoding since the claimed invention is merely a combination of old elements, and in the combination each element merely would have performed the same function as it did separately, and one of ordinary skill in the art would have recognized that the results of the combination were predictable.
Wang et al. in view of Rama et al. in view of Li et al. do not explicitly disclose all of a universal sentence encoding technique.
However, Owen et al. disclose converting each report into a collection of sentence-level semantic embeddings using a universal sentence encoding technique that maps sentences into a fixed-dimensional embedding space (Owen et al. pg. 2, Section 3A, Paragraph 1, "We utilize several feature extraction methods to capture the semantic information from the textual data [such as] Universal Sentence Encoder (USE) [2]: This is a pre-trained model designed to generate high-quality fixed-length vector representations, or embeddings, for sentences and short texts.").
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify Wang et al. in view of Li et al. to incorporate Owen et al.’s universal sentence encoding technique because such a modification is the result of simple substitution of one known element for another producing a predictable result. More specifically, Wang et al.’s data pre-processing and Owen et al.’s universal sentence encoding technique perform the same general and predictable function, the predictable function being transforming input data into an appropriate format for a drift detection model. Since each individual element and its function are shown in the prior art, albeit shown in separate references, the difference between the claimed subject matter and the prior art rests not on any individual element or function but in the very combination itself - that is in the substitution of Wang et al.’s data pre-processing by replacing it with Owen et al.’s universal sentence encoding technique. Thus, the simple substitution of one known element for another producing a predictable result renders the claim obvious.
Claim 3
Regarding claim 3, the rejection of claim 1 is incorporated.
Wang et al. further disclose a training component that trains the [deep learning autoencoder] model (Wang et al. pg. 3005, Section 3.3, Paragraph 1, "we first learn a Bagging model [5]
F
(
x
)
with
K
base learners, and retrain it at each time point") on the set of training unstructured text reports (Wang et al. pg. 3007, Section 4.2, Paragraph 1, "The datasets chosen to be tested in the experiment is online user reviews collected from e-commerce websites. ... The data were organized and pre-processed to eliminate some invalid reviews, and initially organized into a continuous stream of data to help the model conduct test experiments.").
Rama et al. further disclose a training component that trains the deep learning autoencoder on the set of training [unstructured text reports] (Rama et al. ¶ [0030], "Autoencoder 200 is configured to learn data encodings representative of the feature values of input feature vector(s) 108, for example, in a semi-supervised manner").
Claim 4
Regarding claim 4, the rejection of claim 3 is incorporated. Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. disclose all the elements of the claimed invention as stated above.
Rama et al. further disclose wherein a training component:
randomly initializes trainable internal parameters of the deep learning autoencoder (Rama et al. ¶ [0030]-[0031], "As shown in FIG. 2, autoencoder comprises a plurality of nodes 202-244. Each of nodes 202-244 are associated with a weight, which emphasizes the importance of a particular node (also referred to as a neuron). … The weights of a neural network are initialized randomly and are learned through training on a training data set through a process of stochastic gradient descent to reduce the loss as described below.");
[converts the first training unstructured text report into a collection of sentence embeddings;]
executes the deep learning autoencoder on the collection of [sentence] embeddings (Rama et al. ¶ [0032], "Autoencoder 200 generally comprises three parts: an encoder, a bottleneck, and a decoder, each of which comprising one or more nodes. The encoder may be represented by nodes 202-220. Nodes 202, 204, 206, 208, 210, and 212 may represent an input layer by which input data (e.g., input feature vector(s) 108, as shown in FIG. 1) are received by autoencoder 200"), wherein an encoder portion of the deep learning autoencoder receives the collection of [sentence] embeddings and produces a dimensionally-reduced latent vector (Rama et al. ¶ [0032], "The encoder (or encoder network) encodes the input data (i.e., input feature vector(s) 108) into increasingly lower dimensions. That is, the encoder is configured to compress the input data (i.e., input feature vector(s) 108) into an encoded representation that is typically several orders of magnitude smaller than the input data."), and wherein a decoder portion of the deep learning autoencoder receives the dimensionally-reduced latent vector and produces a collection of reconstructed [sentence] embeddings (Rama et al. ¶ [0032], "The decoder (or decoder network) is configured to decode input feature vector(s) 108 into higher increasingly higher dimensions. That is, the decoder is configured to decompress the knowledge representations and reconstruct input feature vector(s) 108 back from their encoded form. ");
computes an error between the collection of [sentence] embeddings and the collection of reconstructed [sentence] embeddings (Rama et al. ¶ [0034], "re-construction loss determiner 106 may be configured to determine the mean squared error between each piece of data provided to autoencoder 105 (i.e., input feature vector(s) 108) and the reconstructed version of that data (i.e., reconstructed data 112). "); and
updates, via backpropagation, the trainable internal parameters of the deep learning autoencoder based on the error (Rama et al. ¶ [0031], "The neural network executes multiple times, changing its weights through backpropagation with respect to a loss function.").
Owen et al. further disclose converting the first training unstructured text report into a collection of sentence embeddings (Owen et al. pg. 2, Section 3A, Paragraph 1, "We utilize several feature extraction methods to capture the semantic information from the textual data [such as] Universal Sentence Encoder (USE) [2]: This is a pre-trained model designed to generate high-quality fixed-length vector representations, or embeddings, for sentences and short texts.").
Claim 5
Regarding claim 5, the rejection of claim 4 is incorporated.
Wang et al. further disclose wherein the training component converts the first training unstructured text report into the collection of sentence embeddings (Wang et al. pg. 3007, Section 4.2, Paragraph 1, "The data were organized and pre-processed to ... initially organized into a continuous stream of data to help the model conduct test experiments."), [via a universal sentence encoding technique].
Owen et al. further disclose wherein the training component converts the first training unstructured text report into the collection of sentence embeddings (Owen et al. pg. 2, Section 3A, Paragraph 1, "We utilize several feature extraction methods to capture the semantic information from the textual data [such as] Universal Sentence Encoder (USE) [2]: This is a pre-trained model designed to generate high-quality fixed-length vector representations, or embeddings, for sentences and short texts."), via a universal sentence encoding technique (Owen et al. pg. 2, Section 3A, Paragraph 1, "We utilize several feature extraction methods to capture the semantic information from the textual data [such as] Universal Sentence Encoder (USE) [2]: This is a pre-trained model designed to generate high-quality fixed-length vector representations, or embeddings, for sentences and short texts.").
Claim 6
Regarding claim 6, the rejection of claim 1 is incorporated. Rama et al. further disclose wherein the first set of reconstruction errors and the second set of reconstruction errors are based on Euclidean distance computations (Rama et al. ¶ [0039], "the Euclidean distance between the re-constructed input and the original input, is the loss, and is denoted by Equation 6.").
Claim 7
Regarding claim 7, the rejection of claim 1 is incorporated.
Owen et al. further disclose wherein a deep learning autoencoder comprises a long short-term memory architecture (Owen et al. pg. 2, Section 2, Paragraph 6, "Unsupervised methods have also been proposed for out-of-domain detection. One such approach is the LSTM-AutoEncoder [18], which uses only in-domain examples to train an autoencoder for out-of-domain detection.").
Claim 8
Regarding claim 8, Wang et al. disclose a device operatively coupled to a processor (Wang et al. pg. 3007, Section 4.3, Paragraph 2, "The experiment is implemented by Python 3.7, and the computational environment is: Red Hat Enterprise Linux Workstation release 7.9 (Maipo), Intel(R) Xeon(R) Gold 6238R CPU @ 2.20GHz.”).
Li et al. disclose training a deep learning autoencoder (Li et al. pg. 3, Section IV, Paragraph 1, "the AE is incrementally updated") comprising an encoder portion and a decoder portion (Li et al. pg. 3, Section (IV)(A), Paragraph 1, "We consider an [autoencoder (AE)], ... it consists of an encoder followed by a decoder") on a set of [unstructured] training (Li et al. pg. 5, Section (V)(A), Paragraph 1, "Our experimental study considers: (i) synthetic (Sea, Circle) and real-world (MNIST-01, MNIST-23) datasets") [text reports that were used to train the pre-trained natural language model], wherein the deep learning autoencoder is trained in an unsupervised manner (Li et al. pg. 3, Section IV, Paragraph 1, "the AE is incrementally updated ... The proposed method is completely unsupervised") to minimize a reconstruction-error loss between original and reconstructed [sentence-]embedding collections (Li et al. pg. 4, Section (IV)(A), Paragraph 1, "[The AE] is trained to minimise the reconstruction loss between an input ... and its decoded version");
determining, by the device and via execution of a deep learning autoencoder, over a monitoring interval (Li et al. pg. 4, Section (IV)(B), Paragraph 1-2, "This method uses a drift reference window
r
e
f
d
r
i
f
t
x
and a drift moving window
m
o
v
d
r
i
f
t
x
of size
W
d
r
i
f
t
. ... These two windows are used for statistical comparison to detect concept drift.), how different a first distribution of reconstruction errors associated with the set [of unstructured text reports] generated during the monitoring interval is from a second distribution of reconstruction errors associated with a set (Li et al. pg. 4, Section (IV)(B), Paragraph 1-2, "The detection mechanism is realised by comparing the reconstruction loss of the reference window
r
e
f
d
r
i
f
t
l
and moving window
m
o
v
d
r
i
f
t
l
using the Mann–Whitney U Test.") [of training unstructured text reports on which the pre-trained natural language model was trained]…; and
generating, by the device and in response to a determination that the first distribution of reconstruction errors differ from the second distribution of reconstruction errors by more than a threshold margin, a first alert indicating that data drift has occurred and that the pre-trained [natural language] model is thereby unable to confidently analyze (Li et al. pg. 4, Section (IV)(B), Paragraph 13, "We will use this test to raise two flags
f
l
a
g
w
a
r
n
and
f
l
a
g
a
l
a
r
m
as shown below. The warning flag indicates a warning for a potential concept drift, while the other flag triggers an alarm for an actual concept drift.") [the set of unstructured text reports], wherein the threshold margin is dynamically adapted based on a rolling average of recent reconstruction errors computed across successive monitoring intervals and a variance term representing temporal instability in reconstruction error behavior (Li et al. pg. 4, Section (IV)(B), Paragraph 11, "The P-value is defined in Eq. (11), where
μ
is the mean and
σ
2
is the variance of the observation
Z
(
t
)
.
P
w
a
r
n
and
P
a
l
a
r
m
are threshold values set for warning and alarm. The P-value for
f
l
a
g
w
a
r
n
should be larger than the p-value for
f
l
a
g
a
l
a
r
m
, i.e.,
P
w
a
r
n
>
P
a
l
a
r
m
. Once there is a flag alarm, then H0 (null hypothesis) is rejected and H1 (alternative hypothesis) is satisfied. The above description is shown in Lines 24 - 28 in Algorithm 1." See Equations 7-11.
μ
is considered analogous to a rolling average of recent reconsturction errors.
σ
2
is considered analogous to a variance term representing temporal instability), and wherein the method further comprises [temporarily suspending inferencing operations of the pre-trained natural language model and] automatically initiating a retraining cycle of the pre-trained [natural language] model using at least a subset of the [unstructured] training [text reports] upon generation of the first alert (Li et al. pg. 4, Section (IV)(B), Paragraph 13, "When a
f
l
a
g
a
l
a
r
m
is raised, we create a new autoencoder to replace the old one and train this new model with
m
o
v
w
a
r
n
."
m
o
v
w
a
r
n
is considered analogous to a subset of training data).
The remaining limitations of claim 8 are similar in scope to the limitations of claim 1 and therefore are rejected for similar reasons as described above.
Claim 10
Regarding claim 10, the rejection of claim 8 is incorporated. The limitations of claim 10 are similar in scope to that of claim 3 and therefore are rejected for similar reasons as described above.
Claim 11
Regarding claim 11, the rejection of claim 10 is incorporated. The limitations of claim 11 are similar in scope to that of claim 4 and therefore are rejected for similar reasons as described above.
Claim 12
Regarding claim 12, the rejection of claim 11 is incorporated. The limitations of claim 12 are similar in scope to that of claim 5 and therefore are rejected for similar reasons as described above.
Claim 13
Regarding claim 13, the rejection of claim 8 is incorporated. The limitations of claim 13 are similar in scope to that of claim 6 and therefore are rejected for similar reasons as described above.
Claim 14
Regarding claim 14, the rejection of claim 8 is incorporated. The limitations of claim 14 are similar in scope to that of claim 7 and therefore are rejected for similar reasons as described above.
Claims 2 and 9 are rejected under 35 U.S.C. 103 as obvious over Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. as applied to claim 1 above, and further in view of US Patent Publication 20240144662 A1 (Subhash Dhok et al.).
Claim 2
Regarding claim 2, the rejection of claim 1 is incorporated. Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. disclose all the elements of the claimed invention as stated above.
Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. do not explicitly disclose all of indicating that data drift has not occurred.
However, Subhash Dhok et al. disclose wherein the result component generates, in response to a determination that the first set of reconstruction errors differ from the second set of reconstruction errors by less than the threshold margin, a second alert indicating that data drift has not occurred and that the pre-trained natural language model is thereby able to confidently analyze the set of unstructured text reports (Subhash Dhok et al. ¶ [0039], "The data drift monitor 122 may be configured to monitor the data drift, and to provide notifications of the data drift related to the threshold(s) to the classification engine 124, periodically, continually, or based on one or more trigger conditions ... the data drift monitor 122 may be configured to monitor the data drift continually, or as continually as possible, and either provide continual notifications to the classification engine 124 or only provide notifications to the classification engine 124 when the data drift exceeds or falls below one of the one or more thresholds stored at the memory 106.").
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify Wang et al.’s data drift detection system to include Subhash Dhok et al.’s continuous alerts because such a modification is the result of combining prior art elements according to known methods to yield predictable results. More specifically, Wang et al.’s data drift detection system as modified by Subhash Dhok et al.’s continuous alerts can yield a predictable result of keeping a system administrator informed since providing continuous notifications regarding the difference in reconstruction errors compared to a threshold value would allow a system administrator to gain valuable insight into the system. Thus, a person of ordinary skill would have appreciated including in Wang et al.’s data drift detection system the ability to do Subhash Dhok et al.’s continuous alerts since the claimed invention is merely a combination of old elements, and in the combination each element merely would have performed the same function as it did separately, and one of ordinary skill in the art would have recognized that the results of the combination were predictable.
Claim 9
Regarding claim 9, the rejection of claim 8 is incorporated. The limitations of claim 9 are similar in scope to that of claim 2 and therefore are rejected for similar reasons as described above.
Claims 15 and 17-20 are rejected under 35 U.S.C. 103 as obvious over Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. in view of "Analysis of semantic drifting in diagnostic texts for sleep disorders" (Deng et al.).
Claim 15
Regarding claim 15, Wang et al. disclose a computer program product for facilitating data drift detection for unstructured texts (Wang et al. pg. 3002, Section 1, Paragraph 2, "this paper considers a method that can simply calculate the severity of topic change and find the time point of significant topic change, and based on this, help adjust the machine learning model to further support text data stream learning.") [via deep learning autoencoders], the computer program product comprising a non-transitory computer-readable memory having program instructions embodied therewith (Wang et al. pg. 3007, Section 4.3, Paragraph 2, "The experiment is implemented by Python 3.7, and the computational environment is: Red Hat Enterprise Linux Workstation release 7.9 (Maipo), Intel(R) Xeon(R) Gold 6238R CPU @ 2.20GHz. The Learn++.NSE [6] and LeverageBagging[3] were implemented by the ScikitMultiflow meta module [22], with default parameter settings." A computation environment as described above implies the use of non-transitory computer-readable memory).
Rama et al. disclose facilitating data drift detection [for unstructured texts] via deep learning autoencoders (Rama et al. ¶ [0034], "re-construction loss determiner 106 may be configured to determine the mean squared error between each piece of data provided to autoencoder 105 (i.e., input feature vector(s) 108) and the reconstructed version of that data (i.e., reconstructed data 112).")…; and
temporarily suspending [diagnostic or prognostic] inferencing operations of a [pre-trained clinical natural language] model (Rama et al. ¶ [0042], "threshold condition analyzer 116 may cause machine learning model 104 to be automatically de-activated and/or re-trained by sending a command 120 to an application and/or service that manages machine learning model 104.").
The remaining limitations of claim 15 are similar in scope to the limitations of claim 8 and therefore are rejected for similar reasons as described above.
Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. do not disclose all of converting clinical reports into a collection of sentence embeddings or performing diagnostic or prognostic inferencing.
However, Deng et al. disclose converting respective clinical text reports into a collection of sentence embeddings (Deng et al. pg. 2, Section 3A, Paragraph 1, “The following types of text documents are extracted from this database: 1) Corpus1-First contact protocol (format: free text) … 2) Corpus2-Data from assessments (format: templated)… 3) Corpus3-Hospital specific diagnostic nomenclature (format: semi-structured) … 4) Corpus4-ICSD-III classification (format: textual in hierarchical structure)” pg. 2, Section 3B, Paragraph 2, “FastText embeddings using all free text (no ICSD labels) from the four aforementioned subsets were learned to simulate four different language spaces” Figure 1 illustrates these four datasets being converted into an embedding space)…; and
diagnostic or prognostic inferencing operations of a pre-trained clinical natural language model (Deng et al. pg. 2, Section 3(B), Paragraph 3, “We have therefore used a text classification model using Convolutional Neural Network (CNN) [6] and Hierarchical Attentive Networks (HAN) [7] to conduct multilabel classification on top level categories (8 classes) of the ICSDIII. Aforementioned four embeddings were used as input to conduct the task.” See Figure 2, which illustrates examples of diagnostic classifications.).
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. to incorporate Deng et al.’s application of drift detection to the medical field.
The suggestion/motivation for doing so would have been that, “textual records generated through patient-physician interaction may contain different types of biases caused by the heterogeneous nature of medical specialties and the used vocabulary in the documentation,” as noted by Deng et al. in the abstract.
Claim 17
Regarding claim 17, the rejection of claim 15 is incorporated. The limitations of claim 17 are similar in scope to that of claim 10 and therefore are rejected for similar reasons as described above.
Claim 18
Regarding claim 18, the rejection of claim 17 is incorporated. The limitations of claim 18 are similar in scope to that of claim 11 and therefore are rejected for similar reasons as described above.
Claim 19
Regarding claim 19, the rejection of claim 18 is incorporated. The limitations of claim 19 are similar in scope to that of claim 12 and therefore are rejected for similar reasons as described above.
Claim 20
Regarding claim 20, the rejection of claim 15 is incorporated. The limitations of claim 20 are similar in scope to that of claim 13 and therefore are rejected for similar reasons as described above.
Claim 16 is rejected under 35 U.S.C. 103 as obvious over Wang et al. in view of Rama et al. in view of Li et al. in view of Owen et al. in view of Deng et al. as applied to claim 15 above, and further in view of US Patent Publication 20240144662 A1 (Subhash Dhok et al.).
Claim 16
Regarding claim 16, the rejection of claim 15 is incorporated. The limitations of claim 16 are similar in scope to that of claim 9 and therefore are rejected for similar reasons as described above.
References Cited
“Concept Drift Detection in Phishing Using Autoencoders” to Menon et al. discloses batch drift detection using an autoencoder. A base classifier is trained parallel and dependent upon the autoencoder, similar to the pre-trained natural language model of this application.
“Reliable and Interpretable Drift Detection in Streams of Short Texts” to Rabinovich et al. discloses an end-to-end framework for model-agnostic change-point detection using autoencoders, reconstruction errors, and sliding windows. Importantly, this art applies drift detection to the field of natural language processing and more specifically, in the context of natural language user requests to a task-oriented system.
“Concept Drift Adaptation in Text Stream Mining Settings: A Systematic Review” to Garcia et al. is a 2025 survey disclosing various papers that tackle drift detection in the field of natural language processing. The survey serves as a descriptive collection of the most influential papers in drift detection. Notably, many of the papers cited in the survey were published before the effective filing date of the instant application.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JACOB B VOGT whose telephone number is (571)272-7028. The examiner can normally be reached Monday - Friday 9:30am - 7pm EST.
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, Paras D Shah can be reached at (571)270-1650. 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.
/JACOB B VOGT/ Examiner, Art Unit 2653
/Paras D Shah/ Supervisory Patent Examiner, Art Unit 2653
04/15/2026
Prosecution Insights