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 .
Examiner notes the entry of the following papers:
Claims filed 4/30/2025.
Applicant’s arguments/remarks filed 4/30/2025.
Claim 26 is new. Claims 1-26 are pending.
Response to Arguments
Applicant presents arguments. Each is addressed.
Applicant argues “Applicant respectfully disagrees. The Office characterizes the features of current claim 1 as mental processes by abstracting them to a high level (e.g., ‘determining relationships’ or ‘generating values’). However, this characterization is inconsistent with revisions to M.P.E.P. §§ 2106.04(d) and 2106.05(a), which expressly caution examiners against oversimplifying claims and ignoring meaningful technical limitations.” (Remarks, page 11, paragraph 2, line 9.) However, applicant has not identified any limitation where the characterization is an oversimplification, or why it is an oversimplification. Therefore, the rejection is proper and maintained.
Applicant argues that “For example, current claim 1 recites the additional elements that integrate the judicial exception into a practical application. Specifically, current claim 1 recites ‘the one or more neural networks are to: compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase; propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities.’ Thus, as recited in claim 1, by computing conditional probabilities for words in a query relative to a target phrase, the system enhances its understanding of context and relevance, leading to more accurate information retrieval and enhancing the search experience with more relevant results.” However, “compute one or more conditional probabilities” and “generate one or more values” have been identified as abstract ideas, not additional elements. The MPEP 2106.05(a) recites “It is important to note, the judicial exception alone cannot provide the improvement. The improvement can be provided by one or more additional elements.” Applicant has not provided objective evidence of a technical improvement. Instead, applicant is relying entirely on the judicial exception to provide a technical improvement. Therefore, the rejection is proper and maintained.
Applicant argues “The above recitations bear certain similarities to the claims at issue in Ex Parte Desjardins, Appeal No. 2024-000567, Dec. on Req. for Rehearing (Sep. 26, 2026) (precedential) (hereinafter “Desjardins”). In Desjardins, the Appeals Review Panel vacated the P.T.A.B.’s new ground of rejection under 35 U.S.C. § 101 of a claim directed to "computing ..., an approximation of a posterior distribution over possible values of the plurality of parameters" on the basis of the Step 2A, Prong Two, relying in part on the reasoning set forth in Enfish, LLC v. Microsoft Corp., 822 F.3d 1327, 118 USPQ2d 1684 (Fed. Cir. 2016): "Enfish recognized that '[m]uch of the advancement made in computer technology consists of improvements to software that, by their very nature, may not be defined by particular physical features but rather by logical structures and processes."' Desjardins, at 8. As further stated by M.P.E.P. § 2106.04(d)(1), "[o]ne way to demonstrate ... integration [of a judicial exception into a practical application] is when the claimed invention improves the functioning of a computer or improves another technology or technical field." The current claim 1 recites how conditional probabilities are computed, how they are propagated across multiple layers, and how relationship values are generated at a subsequent layer. This is a technology based solution to a technological problem, not a black box invocation of machine learning.” (Remarks, page 12, paragraph 2, line 1.) However, applicant has not identified the limitations that are similar to the limitations of Desjardins, nor argued how they have been incorrectly classified. In addition, “computing conditional probabilities” is identified as an abstract idea. As mentioned above in paragraph b, the improvement must come from the additional elements, not the abstract idea. Applicant is relying entirely on the judicial exception to provide a technical improvement. Therefore, the rejection is proper and maintained.
Applicant argues “Further, the claims recite patent eligible subject matter under Step 2B because, as explained above, they recite improvements to the functioning of a computer or other technology or technical field, which have been ‘found to qualify as 'significantly more' when recited in a claim with a judicial exception.’ M.P.E.P. § 2106.05(I)(A).” (Remarks, page 13, paragraph 1, line 2.) However, applicant has not provided objective evidence of a technical improvement. Nor has applicant identified exactly what limitations have provided the improvement.
Applicant remarks “The Office has already acknowledged that dependent claims 3, 4, 7, 9, 11-14, 17-18, 22, and 24-25 are patent eligible.” (Remarks, page 13, paragraph 1, line 6.) Examiner agrees. Examiner suggests incorporating the limitations of the patent eligible claims into the independent claims.
Applicant argues “…the Office has not demonstrated how the implementation of BERT is mapped to features in current claim 1.” (Remarks, page 14, paragraph 2, line 9.) However, since Devlin is clearly mapped to the limitation “compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase” it is unclear exactly what has not been demonstrated. Therefore, the rejection is proper and maintained.
Applicant argues “Even assuming arguendo, that cited portion of Devlin discloses the use of softmax function, which Applicant does not concede…” (Remarks, page 14, paragraph 2, line 11.) However, the mapping recites “...a dot product between Ti and S followed by a softmax over all the words in the paragraph.” Therefore, examiner is interpreting that Devlin discloses use of a softmax function.
Applicant argues “As such, the probabilities in Devlin are output scores for answer span positions, and not conditional probabilities that quantify relationships between query phrase words and target-phrase words as recited in current claim 1.” (Remarks, page 14, paragraph 2, line 16.) However, the limitation recites “compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase”. There is no mention of quantifying relationships in the claim. The specification recites “In at least one embodiment, a conditional probability is computed as a softmax function that normalizes an output of one or more transformer-based language neural networks.” (Specification, paragraph [0073], line 5.) There is no further detailed description of the function in either the specification or the claims. Therefore, examiner is interpreting the softmax function broadly, as simply a softmax function. Devlin recites “the probability of word i being the start of the answer span is computed as a dot product between Ti and S followed by a softmax over all of the word in the paragraph:
PNG
media_image1.png
42
128
media_image1.png
Greyscale
This specifically calculates the probability of word i of the query phrase as being conditioned on the values of S and Ti. Further, though not specifically mentioned in the claims or the mappings, Devlin quantifies relationships between query phrase words and target phrase words “Many important downstream tasks such as Question Answering (QA) and Natural Language Inference (NLI) are based on understanding the relationship between two sentences, which is not directly captured by language modeling. In order to train a model that understands sentence relationships, we pre-train for a binarized next sentence prediction task that can be trivially generated from any monolingual corpus.” (Devlin, page 4, column 2, paragraph 3, line 1.) Therefore, rejection is proper and maintained.
Applicant argues “Furthermore, neither Wang or Devlin disclose ‘propagat[ing] the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generat[ing] one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities.” (Remarks, page 14, paragraph 3, line 1.) However, this is merely an assertion and does not address any particular failing of a specific mapping. Therefore, the rejection is proper and maintained.
Applicant argues “It appears that the Office mistakenly maps ‘embeddings’ that are fed to ‘bunch of layers of transformers’ to ‘propagat[ing] the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks.’” However, it is known in the art that “forward propagation” in a neural network refers to the process by which values are passed through the layers of a neural network. Similarly, it is known that “back propagation” is passing values back through the layers of the neural network. The claims do not describe something different. Merely mentioning “computing conditional probabilities” does not change the meaning of propagating. Furthermore “computing conditional probabilities” is addressed in paragraph h. above. Therefore, the rejection is proper and maintained.
Applicant argues “Applicant respectfully submits that claims 10, 15, and 23 are allowable at least for reasons including some of those discussed above in connection with claim 1.” (Remarks, page 15, paragraph 3, line 1.) However, claim 1 remains rejected. Claims 10, 15, and 23, which recite the same relevant claims remain rejected as well.
Applicant argues “Applicant respectfully submits that claims 2-9, 11-14, 16-22, 24, and 25 are allowable at least for depending from an allowable independent claim.” (Remarks, page 16, paragraph 1, line 2.) However, the independent claims remain rejected. The dependent claims remain rejected, at least for depending from rejected base claims.
Applicant argues “New claim 26 depends from claim 1 described above. Accordingly, Applicant respectfully submits that claim 26 is allowable at least for depending from an allowable independent claim.” (Remarks, page 16, paragraph 6, line 1.) However, claim 1 remans rejected. The rejection for claim 26 is detailed below.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 2, 8, 10, 15-16, 21, and 23 are rejected under 35 U.S.C. § 101 because the claim is directed to an abstract idea without significantly more.
Step 1: Is the claim to a process, machine, manufacture, or composition of matter?
Claims 1-14 are directed to one or more processors (i.e., machine/apparatus), claims 15-25 are directed to a system (i.e., machine/apparatus); therefore, all pending claims are directed to one of the four statutory categories of invention.
Step 2A, Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Claim 1 recites the limitations of:
generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase – mental process (observation, evaluation, judgment). In this case, examiner is interpreting generating values as the mental process of determining values to indicate that one or more relationships exist, rather than a calculation.
the one or more neural networks are to: compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase – mathematical concepts (relationships, formulas or equations, calculations). See MPEP 2106.04(a)(2). Using one or more generic machine learning models to perform the calculation does not change the fact that the limitation is directed to a mathematical calculation. See MPEP 2106.05(f)(1).
generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities – mental process (observation, evaluation, judgment). In this case, examiner is interpreting generating values as the mental process of determining values to indicate one or more relationships exist, rather than a calculation.
Step 2A Prong 2: Does the claim recite additional elements that integrate the judicial exception into a practical application?
The claim recites the additional elements of:
one or more processors comprising: circuitry […] – computer components recited at a high level are construed as generic computer components used to implement the abstract idea. See MPEP 2106.05(f)(2).
to cause one or more neural networks to identify information from one or more datasets– machine learning models recited at a high level are construed as generic. “to cause” is mere instructions to apply. See MPEP 2106.05(f)(1).
propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks – propagating values from one layer to another is a recitation of how a generic neural network functions. As such, this is merely processing input. See MPEP 2106.05(g).
The additional elements do not integrate the judicial exception into a practical application.
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
The additional elements of:
one or more processors comprising: circuitry […] – computer components recited at a high level are construed as generic computer components used to implement the abstract idea. See MPEP 2106.05(f)(2).
to cause one or more neural networks to identify information from one or more datasets– machine learning models recited at a high level are construed as generic. “to cause” is mere instructions to apply. See MPEP 2106.05(f)(1).
propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks – propagating values from one layer to another is a recitation of how a generic neural network functions. As such, this is merely processing input. See MPEP 2106.05(g). Transmitting data is well-understood, routine and conventional. See MPEP 2106.05(d)II)(i).
The limitations do not amount to significantly more than the abstract idea.
Therefore, claim 1 is not patent eligible.
Independent claim 10 recites the same substantive limitations as claim 1 and is similarly not patent eligible.
Independent claim 15 recites the additional elements of “a system comprising: one or more processors” -computer components recited at a high level of generality are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2); and “one or more memories to store parameters associated with one or more neural networks” -computer components recited at a high level of generality are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2).
Independent claim 23 recites the additional elements of “a system comprising: one or more processors”- computer components recited at a high level of generality are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2); and “one or more memories to store parameters associated with the neural networks” - computer components recited at a high level of generality are construed as generic components used to implement the abstract idea. See MPEP 2106.05(f)(2). .
The additional limitations do not integrate the abstract idea into a practical application. Nor do they amount to significantly more than the abstract idea. Therefore, the independent claims are not patent eligible.
The above analysis similarly applies to the dependent claims.
Claims 2 and 16 recite the additional elements of “one or more values are one or more scores” – description of the result of a mental process, as such, a mental process; and “the one or more scores are a positive number or a negative number”- further description of the result of the mental process, as such, a mental process.
Claims 8 and 21 recite the additional elements of “the one or more neural networks are trained with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa)”- reciting training a known neural network model without a description of the training besides identifying a domain, is mere instructions to apply. See MPEP 2106.05(f)(1).
Therefore, claims 1, 2, 8, 10, 15-16, 21, and 23 are not patent eligible.
Claims 3, 4, 7, 9, 11-14, 17-18, 22, and 24-26 are not rejected under § 101 because they each recite functionality that cannot be practically performed by a human mind.
Dependent claims 5-6, and 19-20 depend from claims 4 and 18, respectively, and inherit the limitations of the claims from which they depend. Therefore, claims 5-6 and 19-20 are not rejected under § 101.
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, 3-4, 7, 10-12, 15, 17-18, and 23-24 are rejected as being unpatentable over Wang et al (Contextualized Relevance Feedback for Precision Medicine, herein Wang), Devlin et al (BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, herein Devlin), and Lu, et al (Assessing Semantic Similarity Using A Dual-Encoder Neural Network, herein Lu).
Regarding claim 1,
Wang teaches [one or more processors comprising: circuitry] to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase (Wang, Fig. 1, and page 1673, column 2, paragraph 3, line 1 “For tackling the insufficiency of the previous methods used in PM (precision medicine), we integrate pseudo relevance feedbacks with BERT (Bidirectional Encoder Representations from Transformers) to propose a BERTNPRF model, which explore to obtain the contextual representation understanding of the interaction between a target document and pseudo relevant documents and improve the performance of reranking, so that more precise medical articles can be retrieved and conveyed to clinicians for their decisions.” And, page 1673, column 1, paragraph 1, line 1”Therefore an approach that can employ pseudo relevance feedback to obtain richer relevance information and capture a better representation describing the relations between a query and document to improve retrieval effect in PM task is the goal of our paper.” And, page 1673, column 2, paragraph 3, line 1 “For tackling the insufficiency of the previous methods used in PM, we integrate pseudo relevance feedbacks with BERT to propose a BERTNPRF model, which explore to obtain the contextual representation understanding of the interaction between a target document and pseudo relevant documents and improve the performance of reranking, so that more precise medical articles can be retrieved and conveyed to clinicians for their decisions.”
PNG
media_image2.png
600
797
media_image2.png
Greyscale
Examiner notes it is implicit that a BERT model requires at least one or more processors comprising circuitry in order to execute. In other words, BERT requires one or more processors comprising circuitry, BERTnprf is one or more neural networks, relevance feedbacks is identifying information, medical articles is one or more datasets, query is one or more first words of a query phrase, pseudo relevant document is one or more second words in the target phrase, relations between a query and a document is one or more values that indicate one or more relationships, and, obtain contextual representation understanding of the interaction between a target document and pseudo relevant documents is indicate one or more relationships based on one or more properties of the one or more items.)
Examiner notes that using a BERT model is disclosed in Wang. However, though Wang uses BERT, it does not provide the details of how to implement BERT. Devlin provides those details. It would be obvious to combine Devlin into Wang for the fact that Wang makes use of the developments of Devlin. (Devlin, Abstract, line 1 “We introduce a new language representation model called BERT, which stand for Bidirectional Encoder Representations from Transformers… As a result, the pre-trained BERT model can be finetuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task specific architecture modifications.”). For ease of reference, Figure 1 of Devlin is included below.
PNG
media_image3.png
490
876
media_image3.png
Greyscale
In other words, layers are layers in a neural network, and BERT is one or more neural networks that generate values that indicate relationships between questions and answers, among other things.), wherein the one or more neural networks are to:
compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase (Devlin, Figure 1, and page 6, column 2, paragraph 2, line 7 “The probability of word i being the start of the answer span is computed as a dot product between Ti and S followed by a softmax over all of the words in the paragraph:
PNG
media_image4.png
54
153
media_image4.png
Greyscale
And, page 7, column 2, paragraph 4, line 4 “The only task-specific parameters introduced is a vector whose dot product with the [CLS] token representation C denotes a score for each choice which is normalized with a softmax layer.” The specification of the instant application recites “In at least one embodiment, a conditional probability is computed as a softmax function that normalizes an output of one or more transformer-based language neural networks.” (Specification, paragraph [0073, line 5.) Therefore, examiner is interpreting “compute one or more conditional probabilities” as using a softmax function to normalize the output of the one or more transformer-base language neural networks. In other words, using a softmax function to normalize the output is compute one or more conditional probabilities. Examiner notes query phrase and target phrase are previously mapped.)
propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities (Wang, Figure 1. And page 1675, column 2, paragraph 2, line 5 “Furthermore, the other two embeddings, namely type embeddings, separating query type (PRF type) and document type, and position embeddings, capturing the whole context associated with word order, are added to the tokens. Then these tokens are fed into a bunch of layers of transformers, in which each token gains a updated contextualized embeddings through summing up all other tokens with weights.” In other words, fed into a bunch of layers is propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks. Examiner notes that one or more neural networks, generate values, and indicate one or more relationships are previously mapped.)
The combination of Wang and Devlin implicitly teaches one or more processors comprising circuitry because in order to run a BERT model at least one or more processors comprising one or more circuits is required.
However, for clarity, Lu explicitly teaches one or more processors comprising circuitry (Lu, column 2, line 37 “As a preliminary matter, the term "hardware logic circuitry" corresponds to one or more hardware processors (e.g., CPUs, GPUs, etc.) that execute machine-readable instructions stored in a memory, and/or one or more other hardware logic units (e.g., FPGAs) that perform operations using a task-specific collection of fixed and/or programmable logic gates.” And, column 5, line 29 “As will be described in connection with FIG. 2, the first encoder 120 corresponds to a multi-layer neural network that includes at least one self-attention mechanism and at least one feed-forward neural network. These are components also found in neural networks referred to in the technical literature as "transformers,” e.g., as described in Devlin, et al., "BERT: Pre-training 35 of Deep Bidirectional Transformers for Language Understanding," arXiv:1810.04805v2 [cs.CL], May 24, 2019, 16 pages, and VASWANI, et al., "Attention Is All You Need," arXiv:1706.03762v5 [cs.CL], Dec. 6, 2017, 15 pages.” In other words, one or more hardware processors is one or more processors, and hardware logic circuitry is circuitry.).
Both Lu and the combination of Wang and Devlin are directed to transformer models, e.g. BERT, among other things. The combination of Wang and Devlin teaches one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase; but does not explicitly teach one or more processors comprising circuitry. Lu teaches one or more processors comprising circuitry.
In view of the teaching of the combination of Wang and Devlin, it would be obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teaching of Lu into the combination of Wang and Devlin. This would result in one or more processors comprising circuitry to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase.
One of ordinary skill in the art would be motivated to do this because complex neural networks require a lot of processing power to execute and using one or more processors would increase processing power. (Lu, column 1, line 9 “For instance, in the case of neural networks, the models may include a relatively large number of layers and a relatively large number of machine-trained parameter values. This complexity introduces various challenges. For example, a developer may conclude that a complex model is too slow to deliver output results in a required amount of time.”)
Regarding claim 3,
The combination of Wang, Devlin and Lu teaches the one or more processors of claim 1, wherein the one or more neural networks comprise:
a query-target conditioning layer, that is to use a softmax function, to compute the one or more conditional probabilities (Wang, abstract, line 12 “ By considering the multi-aspect word relations, we propose a BERTNPRF model to integrate PRF with the finetuned BERT model for contextualized interaction of document-document pairs. Experimental results on the standard Text REtrieval Conference (TREC) PM track benchmark show that our proposed method with interpolation can improve the performance in PM.” In other words, BERT, which has a query-target conditioning layer, uses softmax, and computes a conditional probability for each target word.); and
a summation layer to sum the one or more conditional probabilities for each target word in the target phrase to generate the one or more values (Wang, Fig. 1, page 1676, column 1, paragraph 2, line 1 “When all scores of every pair of di from an initial ranked list of [d1d2, · · · , dn] and a target document d are obtained, we merge the scores by summing them up to get the final relevance score, namely rel(d, q, Dq), where i means 1 to n pseudo relevant document’s index.” In other words, we merge the scores by summing them up is a summation layer to sum the conditional probabilities of each target word, and scores is one or more values.).
Regarding claim 4,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 1, wherein the one or more neural networks comprise a layer to:
compute a first masked language prediction for a query word; compute a second masked language prediction for each target word in a target phrase (Wang, page 1675, column 1, paragraph 3, line 34 “BERTNPRF takes as input the concatenated sequence
and generates the output embeddings which represent the interaction between sentence A and sentence B. Through MLP consuming the output embeddings, a ranked score can
be predicted.” And, page 1675, column 2, paragraph 4, line 9 “Our model predicts the relevance of every passage in every feedback document independently and adopts its score based on four schemes: best passage, first passage, passage consisted of head part and tail part, and aggregated passage.” Examiner notes that BERT uses masked language model for prediction “BERT alleviates the previously mentioned unidirectionality constraint by using a ‘masked language model’ (MLM) pre-training objective, inspired by the Cloze task (Taylor, 1953).” (Devlin, page 1, column 2, paragraph 3.) In other words, masked language is masked language, and, predicts every passage is predict every target word in every target phrase.) ;
perform a dot product multiplication of the first masked language prediction and the second masked language prediction to obtain query-target predictions for the target phrase (Devlin, page 6, column 2, paragraph 2, line 7 “The probability of word i being the start of the answer span is computed as a dot product between Ti and S followed by a softmax over
all of the words in the paragraph:
PNG
media_image5.png
45
132
media_image5.png
Greyscale
In other words, computed as a dot product is perform a dot product multiplication. Examiner notes that both Wang and Lu disclose this because they both inherently disclose Devlin, which is a description of BERT.); and
sum the query-target predictions for the target phrase to obtain the one or more values (Wang, Figure 1, page 1676, column 1, paragraph 2, line 1 “When all scores of every pair of di from an initial ranked list of [d1d2, · · · , dn] and a target document d are obtained, we merge the scores by summing them up to get the final relevance score, namely rel(d, q, Dq), where i means 1 to n pseudo relevant document’s index.” In other words, scores is one or more values, merge the scores by summing them up is sum the query-target predictions, and to get the final relevance score is to obtain the one or more values.)
14. Regarding claim 7,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 1, wherein to generate the one or more values the one or more neural networks are to use:
a scoring function to sum and normalize one or more values (Wang, Figure 1, and page 1676, column 1, column 1, paragraph 2, line 1 “When all scores of every pair of di from an initial ranked list of [d1d2, · · · , dn] and a target document d are obtained, we merge the scores by summing them up to get the final relevance score, namely rel (d, q, Dq), where i means 1 to n pseudo relevant document’s index.” And, Devlin, page 6, column 2, paragraph 4, line 4 “The only task-specific parameters introduced is a vector whose dot product with the [CLS] token representation C denotes a score for each choice which is normalized with a softmax layer.” In other words, rel(d, q, Dq) is scoring function, q is query, target is target, sum is sum and normalize is normalize.); and
a ranking function to rank an item of interest in the query phrase for a desirable property in the target phrase (Wang, Figure 1, In other words, unsupervised ranking method is a ranking function.).
Claim 10 is a one or more processors comprising circuitry claim that corresponds to one or more processors comprising circuitry claim 1. They have the same substantive limitations. Therefore, claim 10 is rejected for the same reasons as claim 1.
Regarding claim 11,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 10, wherein circuitry is to use the one or more neural networks to
rank one or more drug candidates from a clinical trials dataset based, at least in part, on the one or more conditional probabilities (Wang, Figure 1, page 1673, column 2, paragraph 3, line 24 “The experimental results on the standard Text REtrieval Conference (TREC) 2018 PM track dataset[10] show that the performance of ranking in PM collection can be improved by the proposed BERTNPRF model with interpolation and demonstrate the effectiveness of the proposed BERTNPRF model.” In other words, ranking in PM collection is rank one or more drug candidates from the dataset, and BERTNPRF is using conditional probabilities for association of the candidates and at least one target property.).
Regarding claim 12,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 10, wherein circuitry is to use the one or more neural networks
to determine a drug candidate for drug approval (Wang, Figure 1, In other words, relevance score highest ranked candidate is determine a candidate for approval.) by
ranking drug candidates from a clinical trials dataset based, at least in part, on query-target conditioning predictions of the drug candidates as query words in the clinical trials dataset and an efficacy property as a target in the clinical trials dataset (Wang, Figure 1, and page 1675, column 1, paragraph 3, line 1 “In this section, we describe the proposed BERTNPRF as summarized in Figure 1. Traditionally, given a query q, a score s of how relevant a target document d is to this q is estimated by a reranker model. However, pseudo relevance feedbacks, which can be viewed as augmented information for queries, especially short queries, have seldom been utilized and combined to form a deeper contextual representation for better capturing the relations between queries and documents in PM task. The queries, namely patients’ records in PM dataset, usually only containing several words describing patients’ information of “disease”, “gene” and “demographic”. Dai et al. [20]” In other words, reranker is ranking candidates, PM is drug candidates, PM dataset is dataset, query and target is query-target, and disease, gene, etc. is an efficacy property.).
Claim 15 is a system claim corresponding to one or more processors comprising circuitry claim 1 with the added limitation of one or more memories to store parameters associated with the one or more neural networks. Otherwise, they are same. The combination of Wang, Devlin, and Lu teaches one or more memories. (Lu, column 14, line 58 “ The computing device 802 can utilize any instance of the computer-readable storage media 806 in different ways. For
example, any instance of the computer-readable storage media 806 may represent a hardware memory unit (such as Random Access Memory (RAM)) for storing transient information during execution of a program by the computing device 802, and/or a hardware storage unit (such as a hard disk) for retaining/archiving information on a more permanent basis.” In other words, random access memory is one or more memories to store parameters with the one or more neural networks.) Therefore, claim 15 is rejected for the same reasons as claim 1. Claim 17 is system claim corresponding to one or more processors claim 3. Otherwise, they are the same. Therefore, claim 17 is rejected for the same reasons as claim 3.
Claim 18 is a system claim corresponding to one or more processors claim 4. Otherwise, they are the same. Therefore, claim 18 is rejected for the same reasons as claim 4.
Claim 23 has the same substantive limitations as claim 15. Therefore, claim 23 is rejected for the same reasons as claim 15.
Claim 24 is a system comprising one or more processors claim corresponding to one or more processor claim 11. Otherwise, they are the same. Therefore, claims 24 is rejected for the same reasons as claim 11.
Regarding claim 26,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 1, wherein
the one or more conditional probabilities are computed prior to normalization performed by the at least one subsequent layer of the one or more neural networks ( Figure 1, and page 6, column 2, paragraph 2, line 7 “The probability of word i being the start of the answer span is computed as a dot product between Ti and S followed by a softmax over all of the words in the paragraph:
PNG
media_image4.png
54
153
media_image4.png
Greyscale
And, page 7, column 2, paragraph 4, line 4 “The only task-specific parameters introduced is a vector whose dot product with the [CLS] token representation C denotes a score for each choice which is normalized with a softmax layer.” And, page 7, column 2, paragraph 4, line 4 “The only task-specific parameters introduced is a vector whose dot product with the [CLS] token representation C denotes a score for each choice which is normalized with a softmax layer.” The specification of the instant application recites “In at least one embodiment, a conditional probability is computed as a softmax function that normalizes an output of one or more transformer-based language neural networks.” (Specification, paragraph [0073, line 5.) Therefore, examiner is interpreting “compute one or more conditional probabilities” as using a softmax function to normalize the output of the one or more transformer-base language neural networks. Further, based on the same citation of the specification, examiner is interpreting that the softmax function performs the normalization. In other words, using a softmax function to normalize the output is compute one or more conditional probabilities, and normalized with a softmax layer is normalization is performed.)
Claims 2, 5-6, 16, and 19-20 are rejected under 35 U.S.C. § 103 as being unpatentable over Wang, Devlin, Lu, and Montague et al (Relevance Score Normalization for Metasearch*, herein Montague).
Regarding claim 2,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 1, wherein
the one or more values are one or more scores (Wang, Fig. 1. And, page 1675, column 1, paragraph 3, line 1 “In this section, we describe the proposed BERTNPRF as summarized in Figure 1. Traditionally, given a query q, a score s of how relevant a target document d is to this q is estimated by a reranker model.” In other words, from Fig. 1, relevance score is one or more values, where the value is a score.), wherein
Thus far, the combination of Wang, Devlin, and Lu, does not explicitly teach the one or more scores is a positive number or a negative number.
Montague teaches the one or more scores is a positive number or a negative number (Montague, page 431, column 1, paragraph 2, line 2 “ZMUV performs very poorly with CombMNZ, but this is easily understood: CombMNZ assumes all relevance scores are positive. But roughly half of the scores produced by ZMUV are negative since it shifts the mean to zero.” In other words, roughly half the scores are negative and the remaining scores are positive is the score is a positive number or a negative number.).
Both Montague and the combination of Wang, Devlin, and Lu are directed to information retrieval based on queries, among other things. The combination of Wang, Devlin, and Lu teaches one or more processors comprising circuitry to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase, wherein the one or more neural networks are to compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase; propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities; but does not explicitly teach the one or more scores is a positive or negative number. Montague teaches the one or more prediction scores is a positive or negative number.
In view of the teaching of the combination of Wang, Devlin, and Lu, it would be obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teaching of Montague into the combination of Wang, and Devlin. This would result one or more processors comprising circuitry to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase, wherein the one or more neural networks are to compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase; propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities; where the one or more scores is a positive or negative number.
One of ordinary skill in the art would be motivated to do this because better techniques for normalizing and estimating relevance score can have a significant impact on computational performance. (Montague, page 1, paragraph 2, line 1 “Research on the problem of metasearch has historically concentrated on algorithms for combining (normalized) scores. In this paper, we show that the techniques used for normalizing relevance scores and estimating the relevance scores of unretrieved documents can have a significant effect on the overall performance of metasearch. We propose two new normalization/estimation techniques and demonstrate empirically that the performance of well-known metasearch algorithms can be significantly improved through their use.”)
Regarding claim 5,
The combination of Wang, Devlin, and Montague teaches the one or more processors of claim 4, wherein the query-target predictions comprise
a first query-target prediction that is a positive number indicating a positive relationship between the query word and a corresponding target word in the target phrase (Montague, See mapping of claim 2. In other words, relevance score is positive if there is a positive relationship between the query and the target.).
Regarding claim 6,
The combination of Wang, Devlin, and Montague teach the one or more processors of claim 4, wherein the query-target predictions comprises
a first query-target prediction that is a negative number indicating a negative relationship between the query word and a corresponding target word in the target phrase (Montague, See mapping of claim 2. And, page 428, column 2, paragraph 8, line 6 “Let us denote the score given by S to a document a as scrS(a). By convention, if scrS(a) > scrS(b), then we know that S is asserting that a is more relevant, or more likely relevant than b.” In other words, a score of lesser numeric value is less relevant than a score of higher numeric value is a negative number indicating a negative relationship between the query and the target.).
Claims 16 and 19-20 are system claims corresponding to one or more processor claims 2 and 5-6, respectively. Otherwise, they are the same. Therefore, claims 16 and 19-20 are rejected for the same reasons as claims 2 and 5-6, respectively.
Claims 8-9, 13-14, 21-22, and 25 are rejected under 35 U.S.C. § 103 as being unpatentable over Wang, Devlin, Lu, and Liu et al (RoBERTa: A Robustly Optimized BERT Pretraining Approach, herein Liu).
Regarding claim 8,
The combination of Wang, Devlin, and Lu teaches the one or more processors of claim 1, wherein the one or more neural networks are
Thus far, the combination of Wang, Devlin, and Lu does not explicitly teach trained with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa).
Liu teaches trained with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa) (Liu, page 3, column 1, paragraph 5, line 1 “We consider five English-language corpora of varying sizes and domains, totaling over 160GB of uncompressed text.” And, page 6, column 2, paragraph 3, line 1 “In the previous section we propose modifications to the BERT pretraining procedure that improve end-task performance. We now aggregate these improvements and evaluate their combined impact. We call this configuration RoBERTa for Robustly optimized BERT approach. Specifically, RoBERTa is trained with dynamic masking (Section 4.1), FULL-SENTENCES without NSP loss (Section 4.2), large mini-batches (Section 4.3) and a larger byte-level BPE (Section 4.4).” In other words, domains is domain specific data and RoBERTa is trained using RoBERTa.)
Both Liu and the combination of Wang, Devlin, and Lu are directed to language processing with transformers, among other things. The combination of Wang, Devlin, and Lu teaches one or more processors comprising circuitry to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase, wherein the one or more neural networks are to compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase; propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities; but does not explicitly teach training with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa). Liu teaches training with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa).
In view of the teaching of the combination of Wang, Devlin, and Lu, it would be obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the teaching of Liu into the combination of Wang, Devlin, and Lu. This would result in one or more processors comprising circuitry to cause one or more neural networks to identify information from one or more datasets by generating one or more values that indicate one or more relationships between one or more words of a query phrase and one or more words of a target phrase, wherein the one or more neural networks are to compute one or more conditional probabilities for the one or more words in the query phrase given the one or more words in the target phrase; propagate the one or more conditional probabilities to at least one subsequent layer of the one or more neural networks; and generate one or more values, at the subsequent layer, that indicate one or more relationships based on the propagated conditional probabilities; and training with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa).
One of ordinary skill in the art would be motivated to do this because training transformer models is computationally expensive and is often performed with an insufficient amount data to fully train the model. (Liu, Abstract, line 4 “Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it.”)
Regarding claim 9,
The combination of Wang, Devlin, Lu, and Liu teaches the one or more processors of claim 1, wherein the one or more neural networks comprise: an input layer to:
receive additional domain-specific data during an inference phase (Wang, Figure 1, In other words, we augment PM data for retraining BERT to explore if retrained BERT brings different effect or not is receive additional domain-specific data during an inference phase.);
receive a query phrase and encode the query phrase into a first vector of tokens using byte-pair encoding (BPE) (Liu, page 6, column 2, paragraph 3, line 1 “In the previous section we propose modifications to the BERT pretraining procedure that improve end-task performance. We now aggregate these improvements and evaluate their combined impact. We call this configuration RoBERTa for Robustly optimized BERT approach. Specifically, RoBERTa is trained with dynamic masking (Section 4.1), FULL-SENTENCES without NSP loss (Section 4.2), large mini-batches (Section 4.3) and a larger byte-level BPE (Section 4.4).” In other words, byte-level BPE is encode the query phrase into a vector of tokens using byte-pair encoding.);
receive a target phrase and encode the target phrase into a second vector of tokens using BPE (See above mapping. Examiner notes that receive a target phrase and encode are previously mapped.);
a Bidirectional Encoder Representations from Transformers (BERT) layer that is trained using a Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa) (Liu, See above mapping. In other words, RoBERTa is trained is trained using a Robustly Optimize Bidirectional Encoder Representations from Transformers Approach (RoBERTa).) and comprises:
a first attention head to receive the first vector of tokens and compute a statistical prediction for each token in the first vector of tokens (Wang, page 1673, column 2, paragraph 3, line 11 “And we extract top-m ones from these ranked articles and use each of them to interact with a candidate article by attention mechanism in BERT, where token embeddings, type embeddings (differentiating query or document) and position embeddings (earning word order information) are added, and multi-head attention is used to earn various type of word relations, such as exact match and synonyms.” In other words, attention heads is attention head, token embedding is first vector of tokens, and BERT architecture means receive a first vector of tokens and compute a statistical prediction for each token.);
a second attention head to receive the second vector of tokens and compute a statistical prediction for each token in the second vector of tokens (Wang, See above mapping. In other words, multi-head attention is a second attention head, token embeddings is a second vector of tokens, and BERT architecture means compute a statistical prediction for each token.); and
an output layer to determine a query-target score by performing a dot product multiplication on the statistical predictions of the first vector of tokens and the statistical prediction of the second vector of tokens (Wang, Figure 1, and page 1673, column 2, paragraph 3, line 19 “With going through a bunch of layers of transformers, it generates an output embeddings.” In other words, output embeddings is output layer, dot product has previously been mapped, and from Figure 1, Relevance Score is query-target score. ).
Regarding claim 13,
The combination of Wang, Devlin, Lu, and Liu teaches the one or more processors of claim 10, wherein the circuitry is to identify the one or more drugs, are further to: for a set of drug candidate from a clinical trials dataset,
compute the one or more conditional probabilities (Devlin, Figure 1, and page 6, column 2, paragraph 2, line 7. See mapping of claim 1, office action page 12.), and
sum the one or more conditional probabilities for the target phrase given the query word to generate one or more values for the respective drug candidate (Wang, Fig. 1, page 1676, column 1, paragraph 2, line 1. See mapping of claim 3, office action page 15.); and
rank the drug candidates according to the one or more values ( Wang, Figure 1, page 1673, column 2, paragraph 3, line 24. See mapping of claim 11, office action page 18.), wherein the one or more neural networks are
trained with domain-specific data using a Robustly Optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa) (Liu, page 3, column 1, paragraph 5, line 1. See mapping of claim 8, office action page 24.).
Regarding claim 14,
The combination of Wang, Devlin, Lu, and Liu teaches the one or more processors of claim 10, wherein the one or more neural networks comprises: an input layer to:
receive a clinical trials dataset for a set of drugs (Wang, Figure 1, and page 1675, column 1, paragraph 3, line 1. See mapping of claim 12, office action page 19.) ; for each drug of the set of drugs,
receive a query word corresponding to the respective drug and encode the query word into a first vector using byte-pair encoding (BPE) (Liu, page 6, column 2, paragraph 3, line 1. See mapping of claim 9, office action page 26.);
receive a target phrase and encode the target phrase into a second vector of tokens using BPE (Liu, See mapping of claim 9, office action page 26.) , wherein
the target phrase comprises a target property of efficacy (Wang, page 1674, column 2, paragraph 2, line 21 “Ran et al. construct a document based neural relevance model, DNRM, which employs PRF to interact with a target document to extract document-to-document
features for improving the performance of Clinical Decision Support (CDS) system. NPRF [7] is also a document based neural IR framework, integrating with two state-of-the-art neural IR models to improve reranking effectiveness in IR task.” Examiner notes that effectiveness is efficacy. In other words, the success of the retrieval is based on the effectiveness of the result.);
a Bidirectional Encoder Representations from Transformers (BERT) layer that is trained using a Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa) and determines a drug score for each drug of the set of drugs; and an output layer to rank the set of drugs according to drug values ( Liu, page 3, column 1, paragraph 5, line 1. See mapping of claim 8, office action page 24 and claim 9, office action page 26.)
Claims 21-22, and 25 are system claims corresponding to one or more processors claims 8-9, and 14, respectively. Otherwise, they are the same. Therefore, claims 21-22, and 25 are rejected for the same reasons as claim 8-9, and 14, respectively.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BART RYLANDER whose telephone number is (571)272-8359. The examiner can normally be reached Monday - Thursday 8:00 to 5:30.
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, Miranda Huang can be reached at 571-270-7092. 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.
/B.I.R./Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124