DETAILED ACTION
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 .
Status of Application
The following is a Final Office Action. In response to Examiner's communication on 10/02/2025, Applicant on 01/02/2026, amended Claims 1, 11, and 20. Claims 1-20 are now pending in this application and have been rejected below.
Response to Amendment
Applicants’ amendments are insufficient to overcome the 35 USC 101 rejections set forth in the previous action, except as it pertains to amended Claims 11-19. The rejections as they pertain to Claims 1-10, 20 are maintained below.
Applicants’ amendments render moot the 35 USC 103 rejections set forth in the previous action. Therefore, these rejections have been withdrawn in favor of new grounds of rejection necessitated by Applicant’s amendments as outlined below.
Response to Arguments – 35 USC § 101
Applicant's arguments with respect to the 35 USC 101 rejections have been fully considered but they are not persuasive with respect to Claims 1-10, 20.
Firstly addressing Claims 11-19, Applicant’s amendments to Claim 11 have been found to surmount the 35 USC 101 rejection by virtue of integrating the claim into a practical application, as indicated by the specificity of the claimed system and the operations it performs. In light of the focus of the claim being directed towards an arrangement of specific components that serve to perform discrete machine learning functionalities, the claimed subject matter can be considered to be outside the bounds of a mental process applied with generic computing components. In light of this, the rejection of Claims 11-19 solely as it pertains to 35 USC 101 has been withdrawn.
However, Claims 1 and 20 depart from this by virtue of the generality of the link of the claimed subject matter to the “processor-based machine learning system” – in Claim 11, there is a clear and present link between the implementation of the “processor-based machine learning system” and the claimed subject matter, namely the electronic device, its processor, and the system that actually effectuates the claimed operations. Put simply, the “processor-based machine learning system” that serves as the focus of Claim 11 is subject matter eligible as what is claimed is a system capable of processing the output of a neural network and adjusting characteristics with a learning agent, whereas the language of using a system “configured to” perform certain operations without actually claiming the performance of those operations in the method claim of Claim 1 cannot be said to be subject matter eligible. Similar logic applies to Claim 20; leaving the performance of subject matter eligible operations as the result of “machine-executable instructions” means that the substrate by which operations are effectuated is not a specific hardware system that must be capable of performing certain functionalities but rather broadly recited “machine-executable instructions”; if the machine-executable instructions are to be subject matter eligible, the actual performance of the steps they cause must be incorporated into Claim 20 instead of the system with a computer-readable medium merely containing them.
Response to Arguments – 35 USC § 103
Applicant' s arguments with respect to the rejection of Claims 1-20 under 35 USC 103 have been considered but are moot in light of new grounds of rejections necessitated by applicant’s amendments.
Applicant’s arguments regarding the applicability of Wang and Hu, as well as their teachings failing to disclose limitations of claims as amended, have been rendered moot in light of new grounds of rejection necessitated by Applicant’s amendments. The rejections have been updated as outlined 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.
Claim 1-10, 20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
101 Analysis – Step 1
The claims are directed to a method and apparatus. Therefore, the claims are directed to at least one of the four statutory categories.
101 Analysis – Step 2A
Regarding Prong 1 of the Step 2A analysis in the MPEP, the claims are to be analyzed to determine whether they recite subject matter that is directed to a judicial expectation, namely a law of nature, a natural phenomenon, or one of the follow groups of abstract ideas: a) mathematical concepts, b) certain methods of organizing human activity, and/or c) mental processes.
Independent Claim 1 includes limitations that recite an abstract idea and will henceforth be used as a representative claim for the 101 rejection until otherwise noted. Claim 1 recites:
A method for generating a report, comprising: acquiring, in a processor-based machine learning system, object data associated with a user's evaluation of an object; generating, in the processor-based machine learning system, first text of the object by at least one language model according to the object data; and generating, in the processor-based machine learning system, the report according to the first text and a graph neural network, wherein the graph neural network is associated with a plurality of objects; wherein the processor-based machine learning system in generating the report is configured to extract feature vectors from the first text to obtain text tokens, to extract additional features from nodes of the graph neural network as graph tokens, and to process the text tokens and the graph tokens, utilizing at least a portion of the at least one language model, to generate enhanced content; wherein the processor-based machine learning system in generating the report is further configured to combine the enhanced content with at least one template; and wherein the processor-based machine learning system further comprises a reinforcement learning model configured to adjust one or more characteristics of the at least one template via a reinforcement learning agent in accordance with a computed reward value.
The examiner submits that the foregoing bolded limitations constitute an abstract idea because under its broadest reasonable interpretation, the claim covers a mental process.
“generating a report…”, “acquiring…data…”, “generating…text” recite abstract ideas - namely, mental processes that could be
performed by a human with a pen and paper, per the MPEP, merely
adapting them into the context of a technological environment with computing parts does not preclude them from being abstract.
Accordingly, the claim recites at least one abstract idea.
Independent Claim 20 recite at least one abstract ideas by presenting substantially similar limitations.
Dependent Claims 2-10 recite at least one abstract idea by virtue of their dependency on independent Claims 1 and 11 respectively.
101 Analysis – Step 2A, Prong II
Regarding Prong II of the Step 2A analysis in the MPEP, the claims are to be analyzed to determine whether the claim, as a whole, integrates the abstract into practical application. As noted in the MPEP, it must be determined whether any additional elements in the claim beyond the judicial exception integrate the exception into a practical application in a manner that imposes a meaningful limit on the judicial exception. The courts have indicated that additional elements, such as merely using a computer to implement an abstract idea, adding insignificant extra solution activity, or generally linking use of a judicial exception to a particular technological environment or field of use do not integrate a judicial exception into a “practical application.
In the present case, the additional limitations beyond the above-noted abstract idea are as follows (where the underlined portions are the “additional limitations” while the bolded portions continue to represent the “abstract idea”):
A method for generating a report, comprising: acquiring, in a processor-based machine learning system, object data associated with a user's evaluation of an object; generating, in the processor-based machine learning system, first text of the object by at least one language model according to the object data; and generating, in the processor-based machine learning system, the report according to the first text and a graph neural network, wherein the graph neural network is associated with a plurality of objects; wherein the processor-based machine learning system in generating the report is configured to extract feature vectors from the first text to obtain text tokens, to extract additional features from nodes of the graph neural network as graph tokens, and to process the text tokens and the graph tokens, utilizing at least a portion of the at least one language model, to generate enhanced content; wherein the processor-based machine learning system in generating the report is further configured to combine the enhanced content with at least one template; and wherein the processor-based machine learning system further comprises a reinforcement learning model configured to adjust one or more characteristics of the at least one template via a reinforcement learning agent in accordance with a computed reward value.
For the following reason(s), the examiner submits that the above identified additional limitations do not integrate the above-noted abstract idea into a practical application.
As it pertains to Claim 1, the additional elements in the claims include “a language model”, “a graph neural network…”, “a reinforcement learning model”. When considered in view of the claim as a whole, the additional elements do not integrate the abstract idea into a practical application because the additional elements are generic computing components that are merely used as a tool to perform the recited abstract idea and/or do no more than generally link the use of the recited abstract idea to a particular technological environment or field of use under Step 2A Prong Two.
Thus, taken alone, the additional elements do not integrate the abstract idea into a practical application. Further, looking at the additional limitation(s) as an ordered combination or as a whole, the limitation(s) add nothing that is not already present when looking at the elements taken individually. For instance, there is no indication that the additional elements, when considered as a whole, reflect an improvement in the functioning of a computer or an improvement to another technology or technical field, apply or use the above-noted judicial exception to effect a particular treatment or prophylaxis for a disease or medical condition, implement/use the above-noted judicial exception with a particular machine or manufacture that is integral to the claim, effect a transformation or reduction of a particular article to a different state or thing, or apply or use the judicial exception in some other meaningful way beyond generally linking the use of the judicial exception to a particular technological environment, such that the claim as a whole is not more than a drafting effort designed to monopolize the exception (MPEP § 2106.05). Accordingly, the additional limitation(s) does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing an abstract idea.
Claims 8 recites “training the language model”.
Claim 20 recites “A computer program product…a non-transitory computer-readable medium and comprising machine-executable instructions”.
These do not integrate the abstract idea into a practical application by analogous reasoning as above.
Claims 2-7, 9-10 do not recite additional limitations beyond those found in claims from which they depend, and therefore do not serve to integrate the recited abstract ideas into a practical application.
101 Analysis – Step 2B
Regarding Step 2B of the MPEP, representative independent claim 1 does not include additional elements (considered both individually and as an ordered combination) that are sufficient to amount to significantly more than the judicial exception for the same reasons to those discussed above with respect to determining that the claim does not integrate the abstract idea into a practical application. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements amount to generic computing components that are merely used
as a tool to perform the recited abstract idea and/or do no more than
generally link the use of the recited abstract idea to a particular
technological environment or field of use. Further, looking at the additional
elements as an ordered combination adds nothing that is not already
present when considering the additional elements individually.
Claim 20 are rejected as disclosing substantially similar limitations as Claim 1.
Claim 8 recites “training the language model”.
Claim 20 recites “A computer program product…a non-transitory computer-readable medium and comprising machine-executable instructions”.
These do not integrate the abstract idea into a practical application or amount to significantly more by analogous reasoning as above.
Claims 2-7, 9-10 do not recite any additional elements beyond those recited in the claims from which they depend, and as a result, Claims 2-7, 9-10 do not include any additional elements that either integrate under Step 2A Prong II or amount to significantly more under Step 2B.
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, 6-7, 10-13, 16, 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Wang(US 20220171936 A1) in view of Hu(Graph Enhanced Contrastive Learning for Radiology Findings Summarization) in further view of Huang(CN 117009490 A).
Claims 1, 11, 20
As to Claim 1, Wang teaches:
A method for generating a report, comprising
In [0079], "Examples of the NLP tasks associated with analysis of the document may include, but are not limited to, an automatic text summarization, a sentiment analysis task, a topic extraction task, a named-entity recognition task, a parts-of-speech tagging task, a semantic relationship extraction task, a stemming task, a text mining task, a machine translation task, and an automated question answering task.".
acquiring, in a processor-based machine learning system, object data associated with a user’s evaluation of an object;
See [0056] for the hardware implementation of neural networks. In [0034], "Typically, analysis of a natural language text in a document may include construction of a parse tree for representation of each sentence in the document... In certain types of documents, such as review documents (e.g., but not limited to, documents associated with product reviews and movie reviews), the document may include multiple sentences that may express opposing opinions.".
generating, in the processor-based machine learning system first text of the object by at least one language model according to the object data;
See [0056] for the hardware implementation of neural networks. In [0004], “The operations may further include generating a document vector for a natural language processing (NLP) task, based on the updated set of features associated with each of the plurality of nodes. The NLP task may correspond to a task associated with an analysis of a natural language text in the document based on a neural network model. The operations may further include displaying an output of the NLP task for the document, based on the generated document vector”. For clarity of the record, we understand the neural network that generates a document vector as outlined above to act as the claimed language model, separate from the GNN.
and generating, in the processor-based machine learning system, the report according to the first text and a graph neural network,
See [0056[ for the hardware implementation of neural networks. We have the graph's role in producing the output in [0058], “The graph neural network (GNN) 206A may comprise suitable logic, circuitry, interfaces, and/or code that may configured to classify or analyze input graph data (for example, the hierarchal graph) to generate an output result for a particular real-time application. For example, a trained GNN model 206A may recognize different nodes (such as, a token node, a sentence node, or a paragraph node) in the input graph data, and edges between each node in the input graph data. The edges may correspond to different connections or relationship between each node in the input graph data (e.g. hierarchal graph). Based on the recognized nodes and edges, the trained GNN model 206A may classify different nodes within the input graph data, into different labels or classes. In an example, the trained GNN model 206A related to an application of sentiment analysis, may use classification of the different nodes to determine key words (i.e. important words), key sentences (i.e. important sentences), and key paragraphs (i.e. important paragraphs) in the document”.
wherein the processor-based machine learning system in generating the report is configured to extract feature vectors from the first text to obtain text tokens, to extract additional features from nodes of the graph neural network as graph tokens, and to process the text tokens and the graph tokens, utilizing at least a portion of the at least one language model, to generate enhanced content;
Regarding the encoding of information as feature vectors, in [0075], “The electronic device 102 may be configured to encode the set of features to generate a feature vector using GNN model 206A. After the encoding, information may be passed between the particular node and the neighboring nodes connected through the edges. Based on the information passed to the neighboring nodes, a final vector may be generated for each node. Such final vector may include information associated with the set of features for the particular node as well as the neighboring nodes, thereby providing reliable and accurate information associated with the particular node. As a result, the GNN model 206A may analyze the document represented as the hierarchal graph”. See [0081-0082] regarding the mechanics of how the resulting analysis of the GNN is used to arrive at said output. This analysis that factors into the document vector generated by another neural network is used to produce output for the NLP task, in [0083], “At block 512, an output of a natural language processing (NLP) task may be displayed. In an embodiment, the processor 204 may be configured to display the output of the NLP task based on the generated document vector. In an embodiment, the NLP task may correspond to a task to analyze the natural language text in the document based on a neural network model. In an example, the displayed output may include an indication of at least one of: one or more important words, one or more important sentences, or one or more important paragraphs in the document (e.g., the first document 110A”. Here, we are considering the neural network that performs natural language processing to be the same language model that generates a document vector as outlined above.
wherein the processor-based machine learning system in generating the report is further configured to combine the enhanced content with at least one template;
We understand the preset output formats to correspond to templates in Wang, in [0083], "At block 512, an output of a natural language processing (NLP) task may be displayed....In an example, the displayed output may include an indication of at least one of: one or more important words, one or more important sentences, or one or more important paragraphs in the document (e.g., the first document 110A)…a representation of the constructed hierarchal graph or a part of the constructed hierarchal graph… ". There is implicit support for rendering other output formats as well, such as plain text, as output can correspond to NLP tasks. In [0079], "Examples of the NLP tasks associated with analysis of the document may include, but are not limited to, an automatic text summarization, a sentiment analysis task…”. These output formats are combined with the output of analysis in [0058], “The graph neural network (GNN) 206A may comprise suitable logic, circuitry, interfaces, and/or code that may configured to classify or analyze input graph data (for example, the hierarchal graph) to generate an output result for a particular real-time application. For example, a trained GNN model 206A may recognize different nodes (such as, a token node, a sentence node, or a paragraph node) in the input graph data, and edges between each node in the input graph data. The edges may correspond to different connections or relationship between each node in the input graph data (e.g. hierarchal graph). Based on the recognized nodes and edges, the trained GNN model 206A may classify different nodes within the input graph data, into different labels or classes. In an example, the trained GNN model 206A related to an application of sentiment analysis, may use classification of the different nodes to determine key words (i.e. important words), key sentences (i.e. important sentences), and key paragraphs (i.e. important paragraphs) in the document”.
Wang does not expressly disclose the remaining limitations.
However, Hu teaches:
wherein the graph neural network is associated with a plurality of objects.
On pg. 4678-4679 of Hu in Section 2.1, "Therefore, we construct a dependency tree to extract the semantic relations between entities and other words, with the direction from their head words to themselves. We also employ the WordPiece to split these words as subwords and connect all the source subwords to the corresponding target words with the original direction. The constructed subword graph is then used to extract relation information, with edges represented by A". In this case, the graph neural network is the graph encoder GNN that is applied to the relation graph. The plurality of objects that the graph neural network is associated with is the extra knowledge combined with the original findings, in the Abstract of Hu, “To address the limitation, we propose a unified framework for exploiting both extra knowledge and the original findings in an integrated way so that
the critical information (i.e., key words and their relations) can be extracted in an appropriate way to facilitate impression generation. In detail, for each input findings, it is encoded by a text encoder, and a graph is constructed
through its entities and dependency tree”. The report is the generated impression section, in the Abstract of Hu, “The impression section of a radiology report summarizes the most prominent observation from the findings section and is the most important section for radiologists to communicate to physicians”.
Wang combined with Huang discloses a system meant to analyze natural language in documents, particularly with the use of neural networks including a graph neural network. Hu discloses a system meant to employ contrastive learning and graph neural networks to summarize medical findings. Each reference discloses applications of graph neural networks for the purpose of text analysis. Extending the contrastive approach as recorded in Hu to the system of Wang combined with Huang is applicable as they are both pertained to performing natural language processing operations aided by a graph neural network.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the contrastive methodology as taught in Hu and apply that to the system of Wang combined with Huang. Motivation to do so comes from the fact that the claim is plainly directed to the predictable result of combining known items in the prior art, with the expected benefit of adopting the contrastive learning being found in pg. 4678 of Hu, "Finally, contrastive learning is introduced to emphasize key words in findings to map positive samples (constructed by masking non-key words) closer and push apart negative ones (constructed by masking key words), as shown in Figure 1. In such a way, key words and their relations are leveraged in an integrated way through the above two modules (i.e., contrastive learning and the graph encoder) to promote AIG. Experimental results on two prevailing datasets (i.e., OpenI and MIMIC-CXR) show that our proposed approach achieves state-of-the-art results compared to existing studies".
Wang combined with Hu does not expressly disclose the remaining limitations.
However, Huang teaches:
and wherein the processor-based machine learning system further comprises a reinforcement learning model configured to adjust one or more characteristics of the at least one template via a reinforcement learning agent in accordance with a computed reward value.
In [0094], "It should be understood that the current step (4) creates a reinforcement learning loop where, in each training episode, the large language model takes several prompts from the training dataset and generates text; its output is then passed to the reward model, which provides a score to assess its consistency with human preferences; after updating, the large language model is subsequently updated to create outputs that score higher in the reward model".
Wang combined with Hu discloses a system meant to analyze natural language in documents, particularly with the use of neural networks. Huang discloses a system meant to fine tune neural networks. Each reference discloses means for leveraging machine learning technology in the field of natural language processing. Extending the fine tuning methodology as recorded in Huang to the system of Wang combined with Hu is applicable as both references are fundamentally pertained to the usage of machine learning algorithms in technology.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the fine tuning as taught in Huang and apply that to the system as taught in Wang combined with Hu. Motivation to do so comes from the fact that the claim is plainly directed to the predictable result of combining known items in the prior art, with the expected benefit found in [0007] of Huang, "The purpose of the present invention is to provide a method and device for training a generative large language model based on knowledge base feedback to address the contradiction between the high cost of manual labeling in the existing technology and the key defect of large language models fabricating facts, which is in conflict with the need for accurate answers in field applications". Note that this is not inherently contradictory to the usage of user feedback, in [0056] of Huang, such benefits are apparent even when only attempting to "partially replace the human feedback in the RLHF, and through the knowledge base and knowledge graph reasoning engine in a specific field, a process of optimizing the training of a large language model in a specific field is obtained"
Claims 11 and 20 are rejected as disclosing substantially similar limitations as Claim 1.
Claim 11 additionally recites “An electronic device, comprising: at least one processor; and a memory coupled to the at least one processor and having instructions stored therein…”.
Claim 20 additionally recites “A computer program product tangibly stored on a non-transitory computer-readable medium and comprising machine-executable instructions that, when executed by a machine, cause the machine to perform actions comprising…”.
These limitations are found in [0041] of Wang, “Examples of the electronic device 102 may include, but are not limited to, a natural language processing (NLP)-capable device, a mobile device, a desktop computer, a laptop, a computer work-station, a computing device, a mainframe machine, a server, such as a cloud server, and a group of servers. In one or more embodiments, the electronic device 102 may include a user-end terminal device and a server communicatively coupled to the user-end terminal device. The electronic device 102 may be implemented using hardware including a processor, a microprocessor (e.g., to perform or control performance of one or more operations), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC). In some other instances, the electronic device 102 may be implemented using a combination of hardware and software”.
Claims 2, 12
As to Claim 2, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
The method according to claim 1, wherein generating the first text of the object comprises: preprocessing the object data to determine features of the object data;
We take encoded positional information to be a feature in [0032], "The system may be further configured to encode first positional information, second positional information, and third positional information. The system may determine a token embedding associated with each of the set of token nodes based on at least one of: the set of first features associated with each of the set of token nodes, the encoded first positional information, the encoded second positional information, and the encoded third positional information. The applying the GNN model on the constructed hierarchal graph may be further based on the determined token embeddings associated with each of the set of token nodes. The first positional information may be associated with relative positions of each of a set of tokens associated with each of a set of words in each of a set of sentences in the document".
encoding the features; and generating the first text of the object according to the encoded features.
In [0033], “The system may be further configured to generate a document vector for an NLP task, based on the updated set of features associated with each of the plurality of nodes. The NLP task may correspond to a task associated with an analysis of a natural language text in the document based on a neural network model (shown in FIG. 2). The generation of the document vector is described further, for example, in FIG. 5. An exemplary operation for a use of the document vector for the analysis of the document for the NLP task is described, for example, in FIG. 14. The system may be further configured to display an output of the NLP task for the document, based on the generated document vector”. Support for understanding the document vector, which is later the basis of a generated NLP task, as a first text is found in the underlined section.
Claim 12 is rejected as disclosing substantially similar limitations as Claim 2.
Claims 3, 13
As to Claim 3, Wang combined with Hu and Huang teaches all the limitations of Claim 2 as discussed above.
Wang teaches:
The method according to claim 2, wherein determining the features of the object data comprises: filtering the object data;
IN [0059], "In some embodiments, the GNN model 206A may correspond to multiple classification layers for classification of different nodes in the input graph data, where each successive layer may use an output of a previous layer as input. Each classification layer may be associated with a plurality of edges, each of which may be further associated with plurality of weights. During training, the GNN model 206A may be configured to filter or remove the edges or the nodes based on the input graph data and further provide an output result (i.e. a graph representation) of the GNN model 206A.".
normalizing the filtered object data;
While it is the weights being normalized in [0136] we understand the correspondence between weights and edges/nodes to signify that we are still performing the normalization operation on filtered data, In [0136], " At 1306, each of the set of weights may be normalized to obtain a set of normalized weights. In an embodiment, the processor 204 may be configured to normalize each of the set of weights to obtain the set of normalized weights. In an embodiment, the normalization of each of the set of weights may be performed to convert each of the set of weights to a normalized value between “0” and “1”. Each of the set of normalized weights may be indicative of an attention coefficient (i.e., “α”) associated with the language attention model. An attention coefficient (e.g., α.sub.ij) associated with the first edge between the first node (node “i”) and the second node (node “j”) may be indicative of an importance of the first edge.".
and extracting the features of the object data according to the normalized filtered object data.
As the normalization and filtering all takes place under the umbrella of applying the GNN model to the constructed hierarchical graph, we look at its relationship with features in [0078], "At block 508, the set of features associated with each of the plurality of nodes of the constructed hierarchal graph may be updated. The processor 204 may be configured to update the set of features associated with each of the plurality of nodes based on the application of the GNN model on the constructed hierarchal graph. The updating of the set of features associated with each of the plurality of nodes is described further, for example, in FIG. 13".
Claim 13 is rejected as disclosing substantially similar limitations as Claim 3.
Claims 6, 16
As to Claim 6, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
the report
In [0079], "Examples of the NLP tasks associated with analysis of the document may include, but are not limited to, an automatic text summarization, a sentiment analysis task, a topic extraction task, a named-entity recognition task, a parts-of-speech tagging task, a semantic relationship extraction task, a stemming task, a text mining task, a machine translation task, and an automated question answering task.".
Wang combined with Hu does not expressly disclose the remaining limitations.
However, Huang teaches:
The method according to claim 1, further comprising: acquiring the user’s feedback about the report;
In [0085], "(3.6) Send the question-answer pair to the user to manually determine the answer score and sort the answer scores; construct the best question-answer pair of question and answer based on the expected answer or predicted answer corresponding to the maximum score after sorting, and use the best question-answer pair to update the knowledge graph of the domain through human feedback".
and modifying ... according to the feedback.
In [0094], "It should be understood that the current step (4) creates a reinforcement learning loop where, in each training episode, the large language model takes several prompts from the training dataset and generates text; its output is then passed to the reward model, which provides a score to assess its consistency with human preferences; after updating, the large language model is subsequently updated to create outputs that score higher in the reward model".
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the fine tuning as taught in Huang and apply that to the system as taught in Wang combined with Hu. Motivation to do so comes from the same rationale as outlined above with respect to Claim 1.
Claim 16 is rejected as reciting substantially similar limitations as Claim 6.
Claim 7
As to Claim 7, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
The method according to claim 1, wherein the evaluation comprises a neutral evaluation, a positive evaluation, or a negative evaluation of the object.
As a specific example of negative sentiment detected in [0154], "Further, the processor 204 may display the result (as 1508) of the sentiment analysis task (e.g., “Sentiment: Negative (73.1%)”) as an annotation associated with the document node 302". Support for sentiment analysis across a longer context is generally found in [0034], "In certain types of documents, such as review documents (e.g., but not limited to, documents associated with product reviews and movie reviews), the document may include multiple sentences that may express opposing opinions. Further, in some cases, a sentence on its own may not express a strong sentiment, however, a paragraph-level context may be indicative of the sentiment of the sentence...The disclosed system captures a global structure of the document in the constructed hierarchal graph and thereby solves the aforementioned problems of the conventional systems". It is implicit that if we are performing sentiment analysis, a review would fall within the rating of neutral, positive, or negative.
Claims 10, 19
As to Claim 10, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
The method according to claim 1, wherein generating the first text of the object comprises: preprocessing the object data to determine features of the object data;
We take encoded positional information to be a feature in [0032], "The system may be further configured to encode first positional information, second positional information, and third positional information. The system may determine a token embedding associated with each of the set of token nodes based on at least one of: the set of first features associated with each of the set of token nodes, the encoded first positional information, the encoded second positional information, and the encoded third positional information. The applying the GNN model on the constructed hierarchal graph may be further based on the determined token embeddings associated with each of the set of token nodes. The first positional information may be associated with relative positions of each of a set of tokens associated with each of a set of words in each of a set of sentences in the document".
encoding the features by a trained language model to obtain enhanced features;
In [0031], "The system may be further configured to update a set of features associated with each of the plurality of nodes based on the application of the GNN model on the constructed hierarchal graph". We understand the GNN's application to be such a trained language model in this context, and the updated features to be enhanced.
and generating the first text according to the enhanced features.
In [0033], "The system may be further configured to generate a document vector for an NLP task, based on the updated set of features associated with each of the plurality of nodes. The NLP task may correspond to a task associated with an analysis of a natural language text in the document based on a neural network model (shown in FIG. 2). The generation of the document vector is described further, for example, in FIG. 5. An exemplary operation for a use of the document vector for the analysis of the document for the NLP task is described, for example, in FIG. 14. The system may be further configured to display an output of the NLP task for the document, based on the generated document vector".
Claim 19 is rejected as disclosing substantially similar limitations as recited by Claim 10.
Claims 4-5, 14-15 are rejected under 35 USC 103 as being unpatentable over Wang(US 20220171936 A1) in view of Hu(Graph Enhanced Contrastive Learning for Radiology Findings Summarization) in further view of Huang(CN 117009490 A) in further view of de Souza(US 20070203935 A1).
Claims 4, 14
As to Claim 4, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
and the plurality of templates being preset;
We have different preset output formats in [0083], "At block 512, an output of a natural language processing (NLP) task may be displayed....In an example, the displayed output may include an indication of at least one of: one or more important words, one or more important sentences, or one or more important paragraphs in the document (e.g., the first document 110A)…a representation of the constructed hierarchal graph or a part of the constructed hierarchal graph… ". There is implicit support for rendering other output formats as well, such as plain text, as output can correspond to NLP tasks. In [0079], "Examples of the NLP tasks associated with analysis of the document may include, but are not limited to, an automatic text summarization, a sentiment analysis task…”.
generating added content by the language model according to the first text and the graph neural network; and generating the report according to the first template and the added content.
Again, we have taken the output formatting to be the template, and the added content to be the actual generated output. We have the graph's role in producing the output in [0058], “The graph neural network (GNN) 206A may comprise suitable logic, circuitry, interfaces, and/or code that may configured to classify or analyze input graph data (for example, the hierarchal graph) to generate an output result for a particular real-time application. For example, a trained GNN model 206A may recognize different nodes (such as, a token node, a sentence node, or a paragraph node) in the input graph data, and edges between each node in the input graph data. The edges may correspond to different connections or relationship between each node in the input graph data (e.g. hierarchal graph). Based on the recognized nodes and edges, the trained GNN model 206A may classify different nodes within the input graph data, into different labels or classes. In an example, the trained GNN model 206A related to an application of sentiment analysis, may use classification of the different nodes to determine key words (i.e. important words), key sentences (i.e. important sentences), and key paragraphs (i.e. important paragraphs) in the document”.
Wang combined with Hu and Huang does not expressly disclose the remaining limitations.
However, de Souza teaches:
The method according to claim 1, wherein generating the report comprises: selecting a first template from a plurality of templates according to the first text, the first template being associated with the evaluation,
In [0041], “In an embodiment of the present invention, report templates are selected from a set of report templates until a suitable subset of report templates is identified. The subset of report templates is created over several iterations. At each round, Selection Module 214, selects the report template which best fits the customers needs as defined by a set of questions. Alternatively, at each round the report template which best fits the customer's needs is defined by a selected metric or measure. Another embodiment of the present invention uses Optimization Module 216 and determines the suitability of all combinations of report templates and returns the optimal subset of report templates. The report templates may be ranked based on a set of financial objectives that the customer wants to optimize or the customer's needs as defined by a set questions. The report templates may be ranked based upon a combination of both needs and financial objectives”. In this case, we view the first text to be the questions and criteria provided to assess which template is most appropriate. The evaluation, or the data that is being summarized, analogizes to the customer’s data that will populate the template.
Wang combined with Hu and Huang discloses a system meant to analyze natural language in documents, particularly with the use of neural networks. De Souza discloses a system meant to automate template selection for report generation. Each reference discloses means for analyzing text for the purpose of performing language processing tasks. Extending the template optimization methodology as recorded in de Souza to the system of Wang combined with Hu and Huang is applicable as both references are fundamentally pertained to natural language analysis.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the report template optimization as taught in de Souza and apply that to the system as taught in Wang combined with Hu and Huang. Motivation to do so comes from the fact that the claim is plainly directed to the predictable result of combining known items in the prior art, with the expected benefit of being able to optimize the selection of various output formats according to the information that needs to be presented.
Claim 14 is rejected as disclosing substantially similar limitations as Claim 4.
Claims 5, 15
As to Claim 5, Wang combined with Hu, Huang and de Souza teaches all the limitations of Claim 4 as discussed above.
Wang teaches:
the first template
We have different preset outputs formats in [0083], "At block 512, an output of a natural language processing (NLP) task may be displayed....In an example, the displayed output may include an indication of at least one of: one or more important words, one or more important sentences, or one or more important paragraphs in the document (e.g., the first document 110A). In another example, the displayed output may include a representation of the constructed hierarchal graph or a part of the constructed hierarchal graph, and an indication of important nodes in the represented hierarchal graph or in the part of the hierarchal graph based on the determined set of weights". There is implicit support for rendering other output formats as well, such as plain text, as output can correspond to NLP tasks. In [0079], "Examples of the NLP tasks associated with analysis of the document may include, but are not limited to, an automatic text summarization, a sentiment analysis task, a topic extraction task, a named-entity recognition task, a parts-of-speech tagging task, a semantic relationship extraction task, a stemming task, a text mining task, a machine translation task, and an automated question answering task."
Wang combined with Hu does not expressly disclose the remaining limitations.
However, Huang teaches:
The method according to claim 4, further comprising: acquiring the user’s feedback about the report; generating a score of the report according to the report and the feedback;
In [0085], "(3.6) Send the question-answer pair to the user to manually determine the answer score and sort the answer scores; construct the best question-answer pair of question and answer based on the expected answer or predicted answer corresponding to the maximum score after sorting, and use the best question-answer pair to update the knowledge graph of the domain through human feedback".
and modifying ... according to the score of the report.
In [0094], "It should be understood that the current step (4) creates a reinforcement learning loop where, in each training episode, the large language model takes several prompts from the training dataset and generates text; its output is then passed to the reward model, which provides a score to assess its consistency with human preferences; after updating, the large language model is subsequently updated to create outputs that score higher in the reward model".
It would have been obvious to one of ordinary skill in the art to combine the fine tuning methodology of Huang and apply it to the system of Wang combined with Hu. Motivation to do some comes from the same rationale as outlined above with respect to Claim 1.
Claim 15 is rejected as reciting substantially similar limitations as Claim 5.
Claims 8-9, 17-18 are rejected under 35 U.S.C. 103 as being unpatentable over Wang(US 20220171936 A1) in view of Hu(Graph Enhanced Contrastive Learning for Radiology Findings Summarization) in further view of Huang(CN 117009490 A) in further view of Li(CLMLF: A Contrastive Learning and Multi-Layer Fusion Method for Multimodal Sentiment Detection).
Claims 8,17
As to Claim 8, Wang combined with Hu and Huang teaches all the limitations of Claim 1 as discussed above.
Wang teaches:
The method according to claim 1, further comprising training the language model, wherein training the language model comprises: determining a loss based on the first text…; and minimizing the loss to train the language model;
In [0150], "At block 1408, the neural network model may be re-trained for the NLP task, based on the document vector, and the generated prediction result. In an embodiment, the processor 204 may be configured to re-train the neural network model for the NLP task based on the document vector, and the generated prediction result. In a training of the neural network model, one or more parameters of each node of the neural network model may be updated based on whether an output of the final layer (i.e., the output layer) for a given input (from a training dataset and/or the document vector) matches a correct result based on a loss function for the neural network model. The above process may be repeated for same or a different input till a minima of loss function may be achieved and a training error may be minimized. Several methods for training are known in art, for example, gradient descent, stochastic gradient descent, batch gradient descent, gradient boost, meta-heuristics, and the like. Control may pass to end".
Wang does not expressly disclose the remaining limitations.
However, Hu teaches:
a first sample…wherein the first sample is generated according to the first text,
On Page 4679, Col 2 of Hu, we are given Algorithm 1, means for generating both positive and negative examples for the purpose of training.
It is given that these samples correspond to the input; we are iterating over tokens of the input sequence as defined on page 4678, Section 2. Here we are understanding the first text to be the input to the model.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the contrastive approach of Hu and apply that to the system of Wang. Motivation to do so comes from the same rationale as outlined above with respect to Claim 5.
Wang combined with Hu does not expressly disclose the remaining limitations.
However, Huang teaches:
and a second sample… and the second sample is acquired from a sample library;
We construe the data sources used to construct the domain-specific knowledge graph to be acquired from libraries, in [0095], "Based on high-quality open domain data or domain-specific databases, a domain-specific knowledge graph is constructed through knowledge extraction and knowledge fusion methods. A reward model used to generate answer scores for a large language model is constructed through knowledge graph pre-training". Note that it is the underlined material pertinent to open domain data and domain-specific databases that is being used to anticipate the claimed second sample.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the model tuning methodology of Huang and apply that to the system of Wang combined with Hu. The motivation to do so comes from the same rationale as outlined above with respect to Claim 1.
Wang combined with Hu and Huang does not expressly disclose the remaining limitations.
However, Li teaches:
and wherein the first sample is emotionally associated with the first text, and the second sample is not emotionally associated with the first text.
In examining a relevant dataset for training sentiment classifiers, we look at Table 4 on pg. 12 of Li. Our data is segmented into Positive, Neutral, and Negative labels.
Wang combined with Hu and Huang discloses a system meant to analyze natural language in documents, particularly with the use of neural networks including a graph neural network and contrastive learning. Li discloses a system meant to leverage the benefits of contrastive learning as applied to multimodal sentiment analysis. Each reference discloses means for leveraging contrastive learning as applied to the performance of natural language processing. Extending the dataset construction as recorded in Li to the system of Wang combined with Hu and Huang is applicable as they are both pertained to optimizing the performance of natural language processing tasks.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the dataset construction as taught in Li and apply that to the system of Wang combined with Hu and Huang. Motivation to do so comes from the fact that the claim is plainly directed to the predictable result of combining known items in the prior art, with the expected benefit that adopting such dataset construction would enable users to leverage the benefits of contrastive learning with respect to the task of sentiment analysis; positive and negative examples in this context can logically analogize to emotional associations.
Claim 17 is rejected as being substantially similar to Claim 8.
Claims 9, 18
As to Claim 9, Wang combined with Hu and Huang and Li teaches all the limitations of Claim 8 as discussed above.
Wang does not expressly disclose the remaining limitations.
However, Hu teaches:
The method according to claim 8, wherein determining the loss comprises: determining an encoded object feature according to the object data;
On Page 4679, Section 2.2, it is explained how the pre-trained Transformer based text encoder implicitly generates encoded feature vectors that are extracted from the relation graph.
determining an encoded first sample feature according to the first sample; determining an encoded second sample feature according to the second sample
On Pages 4679-4680, Section 2.3, it is demonstrated how we derive encoded intermediate state representations from the positive and negative examples, as well as the original data point's feature vector.
and determining the loss based on the encoded object feature, the encoded first sample feature, and the encoded second sample feature.
The bottom of pg. 4680, Col 1 describes a loss function by which contrastive learning can take place, taking into account the encoded samples and original encoded data point's feature vector.
It would have been obvious to one having ordinary skill in the art at the effective filling date of the invention to apply the contrastive approach of Hu and apply that to the system of Wang. Motivation to do so comes from the same rationale as outlined above with respect to Claim 5.
Claim 18 is rejected as disclosing substantially similar limitations as Claim 9.
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 THEODORE L XIE whose telephone number is (571)272-7102. The examiner can normally be reached M-F 9-5.
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, Rutao Wu can be reached at 571-272-6045. 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.
/THEODORE XIE/Examiner, Art Unit 3623
/WILLIAM S BROCKINGTON III/Primary Examiner, Art Unit 3623