DETAILED ACTION
The action is in response to the Amendment filed on 12/22/2025. Claim 26 has been added. Claims 1-5, 7-9, 11-15, 17-19 and 21-26 are pending and have been considered below with claims 1 and 11 being independent. Claims 1, and 11 are amended.
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 .
Priority
Acknowledgment is made of applicant’s claim for priority based on provisional application 63/005,058 filed on April 3, 2020.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
The rejection of claims 1-5, 7-9, 11-15, 17-19 and 21-25 rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement, has been withdrawn as necessitated by the amendment.
The rejection of claims 1-5, 7-9, 11-15, 17-19 and 21-25 rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the enablement requirement, has been withdrawn as necessitated by the amendment.
Claim 26 is rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Claim 26 recites “wherein the training data set is not provided as input to the convolutional neural network model to generate a corresponding output.” , which is not supported by the disclosure. Parent claim 1 indicates that the convolutional neural network “is retrained using a training data set that includes…”. The specification indicates that “machine learning algorithm 430 generates a function that maps an input to an output based on example input- output pairs. For example, machine learning algorithm 430 infers a function from labeled training data consisting of a set of training examples. Each example is a pair consisting of an input object and a desired output value…”, “the above-described techniques can be implemented using supervised learning and/or machine learning algorithms. Supervised learning is the machine learning task of learning a function that maps an input to an output based on example input-output pairs. It infers a function from labeled training data consisting of a set of training examples. Each example is a pair consisting of an input object and a desired output value…” [30, 53]. In other words, the neural network is retrained to improve the output or accuracy of the system. The retraining by definition expects a different output, but an output nonetheless.
Claim 26 is rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the enablement requirement. The claims contain subject matter which was not described in the specification in such a way as to enable one skilled in the art to which it pertains, or with which it is most nearly connected, to make and/or use the invention. the inventors, at the time the application was filed, had possession of the claimed invention. Specifically, the element wherein the training data set is not provided as input to the convolutional neural network model to generate a corresponding output. Parent claim 1 indicates that the convolutional neural network “is retrained using a training data set that includes…”.The specification indicates that “machine learning algorithm 430 generates a function that maps an input to an output based on example input- output pairs. For example, machine learning algorithm 430 infers a function from labeled training data consisting of a set of training examples. Each example is a pair consisting of an input object and a desired output value…”, “the above-described techniques can be implemented using supervised learning and/or machine learning algorithms. Supervised learning is the machine learning task of learning a function that maps an input to an output based on example input-output pairs. It infers a function from labeled training data consisting of a set of training examples. Each example is a pair consisting of an input object and a desired output value…” [30, 53]. In other words, the neural network is retrained to improve the output or accuracy of the system. The retraining by definition expects a different output, but an output nonetheless. Therefore, the disclosure fails to enable one of ordinary skill in the art how to make and use the claimed features.
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 1-5, 7, 8, 11-15, 17, 18 and 21-26 are rejected under 35 U.S.C. 103 as being unpatentable over Weiss et al. (US 2016/0350655 A1, hereinafter Weiss) in view of Lobo (US 2015/0026200 A1), Broudou et al. (US 2021/0081566 A1, hereinafter Broudou), Pajak (US 2021/0286947 A1), Song et al. (US 2018/0225281 A1, hereinafter Song) and Silberman et al. (US 2020/0111022 A1, hereinafter Silberman).
Regarding claim 1, Weiss teaches a computerized method for tagging data strings in electronic documents using a convolutional neural network model, the method comprising: (Weiss, [0057]; “According to some embodiments of the present invention, a neural network based system for word semantic derivation, or semantic role assignment (may also be referred to as 'Semantic Tagging', 'Tagging' or 'Labeling') of dialog utterances (may also be referred to as 'Word Sets' or 'Sentences'), may comprise: (1) An artificial recurrent neural network architecture.” Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells.”).
Weiss does not explicitly teach receiving, by a server computing device, an electronic document in a fixed-layout encoded file format, the electronic document comprising at least a plurality of data strings; converting, by the server computing device, the electronic document from the fixed-layout encoded file format to a structured key-value pair file format, including: storing electronic document data, including a document ID and one or more of the plurality of data strings from the electronic document, in a temporary table…converting the merged electronic document data in the temporary table into structured key-value pairs, generating a new electronic document in a structured key-value
pair file format based upon the structured key-value pairs, and, generating a new electronic document in a structured key-value pair file format based upon the structured key-value pairs, and streaming the new electronic document to a cloud-based database for storage using an application programming interface, and is received from a user computing device operated by a user; generating, by the server computing device, for display the results data on a user device-- (Weiss, [0030]; “In some embodiments, input/output or I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) may be coupled to the system either directly or through intervening I/O controllers.”). However, Lobo, in the area of natural language processing for data extraction, teaches these limitations.
receiving, by a server computing device, an electronic document in a fixed-layout encoded file format, the electronic document comprising at least a plurality of data strings; (Lobo, Abstract; “The method includes the steps of accessing a file in an electronic format from a memory module.” Lobo, [0040]; “System 100 is capable of parsing documents in a wide variety of file formats, including platform neutral file formats such as the portable document format (PDF),” wherein “accessing a file in an electronic format…such as the portable document format (PDF)” encompasses receiving…an electronic document in a fixed-layout encoded file format in accordance with the example embodiment provided at paragraph [00027] of the specification of the claimed invention, “Data extraction module 320 is configured to receive at least one electronic document 310 and extract one or more data strings from the electronic
document 310. In some embodiments, electronic document 310 is a portable document format (PDF).” Lobo, [0041]; “Data extraction module 320 is configured to receive at least one electronic document 310 and extract one or more data strings from the electronic document 310. In some embodiments, electronic document 310 is a portable document format (PDF),” thereby indicating that the electronic document compris[es] at least a plurality of data strings. Lobo, [0056]; “The computer system may include a processor coupled to a memory module and to a mass storage device via a bus or other communication medium; a display or other output device interfacing with the processor,” wherein a “computer system” comprising “a processor 202,
memory 204, storage 206, and communication circuitry 208,” according to the description at paragraph [00025] of the specification of the claimed invention, encompasses a server computing device.). Lobo also teaches a user accessing and interacting with a system using a user interface (38)-- and is received from a user computing device operated by a user, and generating, by the server computing device, for display the results data on a user device.
converting, by the server computing device, the electronic document from the fixed-layout encoded file format to a structured key-value pair file format, including: (Lobo, Abstract; “The method includes the steps of accessing a file in an electronic format from a memory module; extracting data from the file corresponding to a plurality of keys contained within a mapping structure stored in the memory module; organizing the extracted data into values, wherein each value maps to one of the plurality of keys to form a hash map,” wherein “extracting data…to a plurality of keys” and “organizing the extracted data into values” encompasses converting…the electronic document from the fixed-layout encoded file format to a structured key-value pair file format.)
storing electronic document data, including a document ID and one or more of the plurality of data strings from the electronic document, in a temporary table, (Lobo, Fig. 6;
In the figure above, the fields “nb_deal_id” and the values under the “Search Strings” column encompass a document ID and one or more of the plurality of data strings, respectively. Lobo; [0047]; “The mapping structure found within mapping files 110 is the key to locating the relevant data from within the documents. FIG. 6 is an example of a mapping file that may be used by system 100 to extract relevant data from deal documents,” wherein a “mapping file” encompasses a temporary table.) … converting the merged electronic document data in the temporary table into structured key-value pairs (Lobo, Abstract; “The method includes the steps of accessing a file in an electronic format from a memory module; extracting data from the file corresponding to a plurality of keys contained within a mapping structure stored in the memory module; organizing the extracted data into values, wherein each value maps to one of the plurality of keys to form a hash map,” wherein “extracting data…to a plurality of keys” and “organizing the extracted data into values” encompasses converting the merged electronic document data in the temporary table, or “hash map,” into structured key-value pairs.)
generating a new electronic document in a structured key-value pair file format based upon the structured key-value pairs, and (Lobo, Abstract; “The output device allows the user to view a customizable document whose content is derived from the values and keys stored in the database,” wherein “a customizable document whose content is derived from the values and keys” encompasses a new electronic document in a structured key-value format.)
streaming the new electronic document to a cloud-based database for storage using an application programming interface; (Lobo, [0037]; “The invention provides systems and
methods for culling relevant data from documents stored in various electronic formats and for converting the data into a form that can be indexed, manipulated, and stored in a database.”).
Lobo is analogous to the claimed invention as both are from the same field of endeavor, that is, natural language processing for extracting data from electronic documents. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the extraction of key-value pairs and the associated hash map data structure of Lobo with the tokenization and tagging methods of Weiss. The motivation to do so is to package the document data in a manner that is amenable to organized storage and search queries (Lobo, [0063-64]; “The data extracted by PDF parser 210 is written to an EXCEL spreadsheet file by creating a workbook which further creates a sheet which is again divided into rows and columns… EXCEL parser class 212 is illustrated in FIG. 22. EXCEL parser 212 calls the parse function, which extracts the mapping file details into a structure for searching the related information from the source file.”). Also, the access by a user would make it easier to interact and manipulate the data remotely.
Weiss does not explicitly teach merging the electronic document data in the temporary table with compliance review data associated with the electronic document as received from a remote computing device. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0098]; “The final risk value produced by the system may optionally assist the user in identifying the highest risk sections of the portion of the document and may offer suggestions with respect to amending the document to reduce the risk or bringing the document into compliance with, for example, a particular Act or Regulation,” thereby encompassing compliance review data associated with the electronic document. Broudou, [0100]; “A risk matrix may also be generated by the system which may plot each triggered rule to graphically illustrate the level of risk. The risk matrix may compile the results from a number of assessed documents from a particular user or for a particular company. This may allow a graphical output which may highlight the areas of a company which are at risk
of breaching particular laws, regulations or policies,” wherein a “risk matrix” comprising “results from a number of assessed documents” that is used “to graphically illustrate the level of risk” encompasses merging the electronic document in the temporary table with compliance review data.).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the compliance review and associated risk assessment of Broudou with the tokenization and tagging methods of Weiss. The motivation to do so is to automatically modify financial documents to be in line with changing regulatory frameworks thereby avoiding the costs and human error associated with tedious manual review (Broudou, [0046]; “In a preferred embodiment, the system may be adapted for use with the Australian Securities & Investments Commission (ASIC) Regulatory Guide 234 as of 12 Nov. 2012, Advertising financial products and services (including credit): Good practice guidance. It will be appreciated that all future versions of the Regulatory Guide may be compatible with at least one embodiment of the present invention. It will be appreciated that while legislature or rules may change, for example, due to an amendment to a guide, policy or Act, the device, process or system of the present invention may be adapted or dynamically adapt to the amendments of the legislative Act and generate at least one new classifier or rule based on the amendments.”).
Weiss further teaches tokenizing, by the server computing device, each of the plurality of data strings…into a plurality of tokens; (Weiss, Fig. 2B; the figure discloses steps including “receiving a string of one or more characters” and “parsing the data into words and tokenizing
the words.”). Weiss does not explicitly teach in the new electronic document. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0241]; “These images will be subjected to optical character recognition (OCR) process to identify any textual contents in an OCR Text Step in Step 1033. The textual contents are then added back (sandwiched) into the portable document format file in the location of the original image to form a new document with additional data for analysis,” thereby creating the new electronic document. Broudou, [0244]; “The text is converted into tokens by parsing the text through a natural language processing engine,” thereby indicating that the tokenization occurs after the new document is created).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Weiss teaches a method of tokenizing string data extracted from a document but does not explicitly specify that this document has undergone pre-processing or formatting changes. Broudou teaches this limitation. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the electronic document of Weiss using an optical character recognition process to create a new version of the document. The motivation to do so is to allow for necessary modifications and edits to the new document without changing its original structure (Broudou, [0241]; “In an Extract Text Step, the text is then extracted from the resulted PDF file such that the locations of the text from the images relative to text from the PDF are preserved in Step 1035.”).
Weiss does not explicitly teach performing, by the server computing device, preprocessing and feature engineering on the plurality of tokens before any tokens are provided to a convolutional neural network model, wherein performing the preprocessing and feature
engineering identifies a subset of the tokens from the plurality of tokens. However, Pajak, in the area of preprocessing text data for natural language processing tasks, teaches this limitation (Pajak, Abstract; “An apparatus comprises processing circuitry configured to pre-process text data for inputting to a trained model, the pre-processing comprising: receiving a set of text data including numerical information, the set of text data comprising a plurality of tokens, wherein a first subset of the plurality of tokens comprises tokens that do not comprise numerical information, and a second subset of the plurality of tokens comprises tokens that each comprise respective numerical information,” wherein to split the tokens into two subset[s] encompasses identif[ying] a subset of the tokens from the plurality of tokens. Pajak, Abstract; “[t]ransforming each of the plurality of tokens into a respective encoding vector, each of the plurality of tokens in the second subset having a common encoding vector; assigning a respective numerical vector to each of the plurality of tokens, wherein each token in the second subset is assigned a respective numerical vector in dependence on the numerical information in said token; and combining the encoding vectors and numerical vectors to obtain a vector representation of the text data,” wherein “transforming each of the plurality of tokens into a respective encoding vector…and combining the encoding vectors and numerical vectors to obtain a vector representation of the text data” amounts to the creation of a feature embedding and thus is equivalent to feature engineering. Pajak, [0066]; “The trained model may comprise a neural network, for example a convolutional neural network (CNN), a recurrent neural network (RNN), a densely connected network, a transformer language model, or an architecture combining any of the above,” wherein “to pre-process text data for inputting to a trained model” or “convolutional neural network” is equivalent to performing…preprocessing…before any tokens are provided to a convolutional neural network model. Pajak, [0041]; “The text processing apparatus 50 comprises a computing apparatus 52, which in this case is a personal computer (PC) or workstation,” wherein “a personal computer (PC) or workstation” encompasses a server computing device.).
Pajak is analogous to the claimed invention as both are from the same field of endeavor, that is, methods of preprocessing text data for natural language processing tasks. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the
claimed invention, to combine the tokenization and tagging methods of Weiss with the preprocessing and feature embedding steps of Pajak. The motivation to do so is to properly format the text data so that it can be readily input into and processed by a neural network (Pajak, [0003]; “In order to perform natural language processing, text may first be pre-processed to obtain a representation of the text, for example a vector representation. The representation of the text may then be input to a deep learning model.” Pajak, [0008]; “The embedding vectors are used as input to a deep learning model, for example a neural network. The similarity information from the embedding vectors may allow the neural network to generalize over synonyms and closely related terms.”).
Weiss does not explicitly teach and wherein only the identified subset of tokens is provided to the convolutional neural network. However, Song, in the area of semantic token tagging, teaches this limitation (Song, Abstract; “Semantic token tagging can be applied to the specified domain using a subset of the sequence variations that are successfully verified as training data.” Song, [0009]; “The computing system can eliminate a set of sequence variations, in response to failing to verify the set of sequence variations against the external domain, resulting in a subset of sequence variations. The computing system can then train a token tagger by using the subset of sequence variations as training data,” wherein a “subset of sequence variations” used for
“semantic token tagging” encompasses an identified subset of tokens. That this “subset” is a result of a filtering step, it follows that only the identified subset is provided to the “token tagger.” Song, [0010]; “A machine learning architecture can be used to train a token tagger (a machine learning model) and then apply the tagger to perform semantic token tagging. The machine learning architecture can include one or more of Maximum Entropy (MaxEnt), Conditional Random Fields (CRF), Long Short Term Memory (LSTM), Gated Recurrent Unit
(GRU), and Convolutional Neural Network (CNN),” thereby specifying that the “token tagger” is a convolutional neural network.).
Song is analogous to the claimed invention as both are from the same field of endeavor, that is, methods of semantic token tagging. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the tokenization and tagging methods of Weiss to incorporate the sequence filtering preprocessing step of Song. The motivation to do so is to ensure that only high-quality tokens are used to train the tagging model (Song, [0020]; “In order to retain only the high-quality sequence variations, the system can filter the transformed variations using a language model that is built on the POS tags of text from a generalized domain, such as Wikipedia documents. By using word embeddings (that are trained on a generalized domain) as features, the system can train a neural network based semantic token tagger on such filtered sequence variations.”).
Weiss further teaches determining, by the server computing device, a first tag corresponding to a first token of the subset of tokens for a first data string of the plurality of data strings (Weiss, [0011]; “According to some embodiments, character strings representing linguistic user inputs may be fed to the word tokenization and spelling correction model/machine, whereas outputs of the word tokenization and spelling correction model/machine corrected word-
sets/sentences-may be used as inputs for the word semantics derivation model/machine for it to generate outputs in the form of tokenized, corrected and semantically tagged sentences,” wherein “tagged sentences” encompasses ) using the convolutional neural network model; (Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells.” Weiss, [0062]; “According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word semantics derivation model of a system for Deep Learning may include: (1) receiving a set of one or more words (e.g. 'words input', 'word set', 'sentence')… wherein the word semantics derivation model is adapted for: (i) weakly supervising the model learning by providing a substantially small amount of 'right' semantic taggings as learning examples to the model; (ii) assigning markup language semantic tags to at least some, and/or a subset, of the words,” wherein “assigning markup language semantic tags to…a subset, of the words” encompasses applying tag[s] to subset[s] of tokens.)
receiving, by the server computing device, first user response data corresponding to the first tag, wherein the first user response data corresponds to an accuracy of the first tag; (Weiss, Fig. 3B; the figure discloses steps including “assigning markup language semantic tags” and assessing a “level of semantic tagging success.” Thus, the words are evaluated based on an accuracy of the first tag. Weiss, [0068]; “According to some embodiments, the suggested taggings of the candidate words inputs may be altered by the system based on the manual 'feedback' provided,” wherein “manual feedback” encompasses user response data indicating to an accuracy of the first tag. determining, by the server computing device using the convolutional neural network model, a second tag for the first token based on the first user response data indicating the accuracy of the first tag; wherein the convolutional neural network model uses at least the first user response data and the first token as input to determine the second tag as output and the convolutional neural network is retrained (Weiss, Fig. 3B; the figure discloses steps including “assigning markup language semantic tags” and assessing a “level of semantic tagging success.” If the level of success has not met a certain threshold than the prior steps are repeated. In such a case, a second tag is generated.); (Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells.”)
Weiss does not explicitly teach wherein the convolutional neural network model is retrained using a training data set that includes at least the first token, and the second tag that is determined by the convolutional neural network based on the first user response data. However, Silberman, in the area of multi-party token-based decision making, teaches this limitation ( Silberman, [0020]; “In some cases, the token is output along with (and an association is maintained with) at least one token-context value (e.g., one of the features in the set input into the first model, or a value associated therewith, like a zip code, unique ID, or the like, with the desired level of anonymity), which may be shared between the two parties for the models to synchronize. In some cases, the set of input features of the second model includes both data accessible to the second entity and one or more token/token-context value pairs,” wherein a “token-context value” that describes a feature of a token necessarily encompasses a tag. Silberman, [0107]; “the first token, the first value, the second token, and the second value are configured to be part of a training set by which the second machine learning model is trained or re-trained,” wherein “the first value” and “second value” correspond to the first tag and the second tag. Silberman, [0066]; “The machine learning techniques that can be used in this system include the following…Convolutional Neural Network (CNN),” thereby
specifying an embodiment wherein “the machine learning model” is a convolutional neural network.).
Silberman is analogous to the claimed invention as both are from the same field of endeavor, that is, applying tokenization to regulatory compliance scenarios (Silberman, [0016]). Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the tokenization and tagging methods of Weiss with the retraining step of Silberman. The motivation to do so is to improve the accuracy of the convolutional neural network by training it with new data. Silberman highlights the importance of this retraining as it relates to multi-party risk assessment in contexts including cybersecurity and, notably, regulatory compliance (Silberman, [0016-18]; “Two or more entities often wish to share information to facilitate or otherwise improve the quality of joint or independent decision- making. To do so, they often face obstacles of aggregating the information each of the entities own or otherwise hold, which constrains the space of suitable computing architectures. These constraints may arise from a variety of concerns, including regulatory compliance or business issues…To mitigate this issue and related issues in other domains, in some embodiments, a model may be built and trained using data owned or collected by a centralized cybersecurity provider, or by sharing information by (re-) training a model using data owned by individual entities (repeated across many of them, as needed).”). storing, by the server computing device, results data into a database, (Lobo, [0003]; “More particularly, the present invention is directed to a system and method for extracting relevant data from an electronic file and populating a database with the relevant data to allow a user or another program to further analyze and manipulate the data.”) wherein the results data comprises at least the second tag and the first token; and (Weiss, Fig. 3B; “Outputting a tokenized corrected and semantically tagged sentence, based on 'real' user
inputs,” wherein “a tokenized corrected and semantically tagged sentence” encompasses the second tag and the first token per the retagging process explained above.)
Regarding claim 2, the combination of Weiss, Lobo and Broudou teaches the method of claim 1, wherein (and thus the rejection of claim 1 is incorporated).
Weiss further teaches the server computing device is configured to determine a replacement token corresponding to the first token (Weiss, Fig. 2B; The figure illustrates a process with steps including “parsing the data into words and tokenizing the words” and assessing “level of spelling correction success.” If the level of success has not met a certain threshold than the prior steps are repeated. Therefore, a replacement token is generated upon each iteration that does not meet the accuracy threshold.).
Regarding claim 3, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 2, wherein (and thus the rejection of claim 2 is incorporated).
Weiss further teaches the server computing device is configured to generate for
display the replacement token (Weiss, Fig. 2B; the final step in the process illustrated by the figure is “relaying the word tokenization and spelling correction language model's output(s) as input(s) for a word semantics derivation model.” Weiss, Fig. 3B; this model takes as input the replacement token generated from the previous model depicted in Fig. 2B. The final step in the processing of this model is “Outputting a tokenized corrected and semantically tagged sentence,
based on 'real' user inputs.” This output includes the replacement token received from the previous model. Weiss, [0030]; “In some embodiments, input/output or I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) may be coupled to the system either directly or through intervening I/O controllers.” Therefore, the final output of the model, including the replacement token, is generate[d] for display.).
Regarding claim 4, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 3, wherein (and thus the rejection of claim 3 is incorporated).
Weiss further teaches the server computing device is configured to receive second user response data corresponding to the replacement token (Weiss, Fig. 3B; This model takes as input the replacement token generated from the previous model depicted in Fig. 2B. Weiss, [0068]; “According to some embodiments, the suggested taggings of the candidate words inputs may be altered by the system based on the manual 'feedback' provided,” wherein “manual feedback” encompasses user response data correspond[ing] to the replacement token.).
Regarding claim 5, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 4, wherein (and thus the rejection of claim 4 is incorporated).
Weiss further teaches the server computing device is configured to determine a third
tag corresponding to the replacement token (Weiss, Fig. 3B; After taking the replacement token as input, the model depicted in the figure “assign[s] markup language semantic tags,” this step encompassing determin[ing] a third tag corresponding to the replacement token. ) using the convolutional neural network model (Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells.”).
Regarding claim 7, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 1, wherein (and thus the rejection of claim 1 is incorporated).
Weiss does not explicitly teach the server computing device is configured to determine the first tag and the second tag based on regulatory rules. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0046]; “In a preferred embodiment, the system may be adapted for use with the Australian Securities & Investments Commission (ASIC) Regulatory Guide 234 as of 12 Nov. 2012, Advertising financial products and services (including credit): Good practice guidance.” Broudou, [0049]; “In at least one embodiment, between one to seven predetermined rules may be used to determine the potential risk of at least a portion of a document. These rules are preferably professional rules which may be suitable for professional industries such as legal, financial, medical, engineering or any other profession that must adhere to at least one governmental policy or company policy.”).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to employ the tagging method and associated tabular matrix display of Weiss to assess the compliance of text in electronic documents in view of one or more predetermined regulatory rules, as taught by Broudou. The motivation to do so is to identify potentially risky language and to present to a given company various edits or solutions that comply with relevant regulatory standards (Broudou, [0043]; “Assigning a risk value to a distinct marker may assist a user to identify distinct markers which may have a relatively high or unacceptable risk for the user or a company. Reducing the overall risk may increase compliance
with industry regulations or best practices.” Broudou, [0100]; “Each data point on the graph may provide additional details relevant to the risk and may offer suggestions on how to reduce a high risk area of the company.”).
Regarding claim 8, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 1, wherein (and thus the rejection of claim 1 is incorporated). Weiss further teaches the server computing device is configured to tokenize each of the plurality of data strings using natural language processing (Weiss, [0007]; “a word tokenization and spelling correction model/machine may generate corrected word sets outputs based on respective character strings inputs;”).
Regarding claim 21, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 1, further comprising (and thus the rejection of claim 1 is incorporated).
Weiss further teaches teach determining, by the server computing device using the convolutional neural network model, a first predicted tag for each token of the plurality of tokens, (Weiss, [0011]; “According to some embodiments, character strings representing linguistic user inputs may be fed to the word tokenization and spelling correction model/machine, whereas outputs of the word tokenization and spelling correction model/machine corrected word-sets/sentences-may be used as inputs for the word semantics derivation model/machine for it to generate outputs in the form of tokenized, corrected and semantically tagged sentences.” Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least
partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells,” thereby indicating the use of a convolutional neural network.).
Weiss does not explicitly teach wherein the first predicted tag describes whether the corresponding token is compliant or non-compliant in view of one or more rules. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0088]; “Using the above inputs are entered into the system, the rules relevant to the portion of the document being analyzed are retrieved. The rules may be
regulatory/compliance rules, branding rules, profanity/spam and comprise at least one associated classifier,” thereby disclosing an assessment of complian[ce] or non-complian[ce] in view of one or more rules. Broudou, [0244]; “The text is converted into tokens by parsing the text through a natural language processing engine,” indicating that these rules are applied to text data that has been tokenized.).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to employ the tagging method of Weiss to assess the compliance of text in electronic documents, as taught by Broudou. The motivation to do so is to identify and modify language within a document that could potentially result in adverse action to a user or company (Broudou, [0043]; “Assigning a risk value to a distinct marker may assist a user to identify distinct markers which may have a relatively high or unacceptable risk for the user or a company. Reducing the overall risk may increase compliance with industry regulations or best practices.”).
receiving, by the server computing device, first user response data that includes a first recommended tag for each first predicted tag, wherein each first recommended tag indicates an accuracy of the corresponding first predicted tag; (Weiss, Fig. 3B; the figure discloses steps including “assigning markup language semantic tags” and assessing a “level of semantic tagging success.” Thus, the words are evaluated based on an accuracy of the first tag. Weiss, [0068, 57]; “According to some embodiments, the suggested taggings of the candidate words inputs may be altered by the system based on the manual 'feedback' provided” by a system curator, wherein “manual feedback” encompasses user response data thus indicating that the “altered” tag is a first recommended tag indicat[ing] an accuracy of the corresponding first predicted tag.)
determine, using the convolutional neural network model, a second recommended tag for each token of the plurality of tokens based on the corresponding first predicted tag and the first recommended tag (Weiss, Fig. 3B; the figure discloses steps including “assigning markup language semantic tags” and assessing a “level of semantic tagging success.” If the level of success has not met a certain threshold than the prior steps are repeated. In such a case, a second tag is generated based on the corresponding [original] first predicted tag and the first recommended tag. Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network at least partially utilizing Long Short Term Memory (LSTM) gates/blocks/cells,” thereby indicating the use of a convolutional neural network.).
Weiss does not explicitly teach wherein each second recommended tag describes whether the corresponding token is misleading, exaggerated, or promissory in view of the one or more rules; and. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0097]; “A risk value may be dependent
on the rule triggered with respect to the degree of confidence rating,” wherein “a risk value” corresponds to a second recommendation tag…based on the corresponding first predicted tag, or “the [compliance] rule” verdict, and the first recommended tag indicat[ing] an accuracy, or “the degree of confidence rating.” Broudou, [0098-0099]; “The final risk value produced by the system may optionally assist the user in identifying the highest risk sections of the portion of the document… For example, if a user has used the term "baby wraps", which is a complex financial product, the system may issue an alert to the user that industry concepts or jargon such as "baby wraps" may not be understood by customers unless they are within a particular industry,” where “baby wraps” in this context is misleading…in view of the one or more rules.).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to employ the tagging method of Weiss to assess the compliance of text in electronic documents, as taught by Broudou. The motivation to do so is to identify and modify language within a document that could potentially result in adverse action to a user or company (Broudou, [0043]; “Assigning a risk value to a distinct marker may assist a user to identify distinct markers which may have a relatively high or unacceptable risk for the user or a company. Reducing the overall risk may increase compliance with industry regulations or best practices.”).
Weiss does not explicitly teach storing, by the server computing device, an updated training dataset in a database. However, Broudou, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0052-54]; “In the learning stage 200 the system may be configured to determine, receive, be in communication with or otherwise
compile a knowledge base, based on document samples and/or samples from a predetermined act, regulation, guideline or other training document...The association between the distinct markers and the classifiers may be stored in the knowledge base for use in the execution stage 100… Preferably, the knowledge base may be adapted to learn from new documents assessed by the system and update the rules according to user feedback, input or approval” wherein a “knowledge base” encompasses a database. That the “knowledge base may be adapted” in order to “learn from new documents” indicates that it encompasses an updated training dataset.).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the tokenization and tagging method of Weiss to incorporate the self-updating knowledge base of Broudou. The motivation to do so is to teach a model or system new regulatory rules influencing how tokens are to be tagged (Broudou, [0054]; “Preferably, the knowledge base may be adapted to learn from new documents assessed by the system and update the rules according to user feedback, input or approval (210). The feedback may then be aggregated (220) and training samples (230) may be used in combination with the machine learning of the system (240). The samples may teach the system at least one of a new classifier, a new rule or may update existing rules and classifiers to provide an improved degree of certainty.”).
wherein the updated training dataset includes the plurality of tokens, their corresponding first recommended tags, and their corresponding second predicted tags (Weiss, Fig. 3B; the figure discloses steps including “assigning markup language semantic tags” and assessing a
“level of semantic tagging success.” If the level of success has not met a certain threshold than the prior steps are repeated. In such a case, second predicted tags are generated based on the corresponding [original] first recommended tags. The figure further specifies that the model “receiv[es] a set of one or more words” thereby indicating that the token data comprises an updated training dataset.).
Regarding claim 22, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 21, further comprising (and thus the rejection of claim 21 is incorporated).
Weiss does not explicitly teach displaying, by the server computing device, a table on a user device, wherein the table includes the plurality of tokens, their corresponding first predicted tags, their corresponding first recommended tags, and their corresponding second predicted tags; and. However, Broudou in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0100]; “A risk matrix may also be generated by the system which may plot each triggered rule to graphically illustrate the level of risk. The risk matrix may compile the results from a number of assessed documents from a particular user or for a particular company. This may allow a graphical output which may highlight the areas of a company which are at risk of breaching particular laws, regulations or policies. A risk matrix may plot the degree of risk vs the degree of confidence, ” wherein “a graphical output” in the form of “a risk matrix” encompasses displaying, by the server computing device, a table on a user device, and to “plot each triggered rule” and “level of risk” or “risk value” against “the degree of confidence” is equivalent to includ[ing] the plurality of tokens, their corresponding first predicted tags, their corresponding first recommended tags, and their corresponding second predicted tags).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to employ the tagging method and associated tabular matrix display of Weiss to assess the compliance of text in electronic documents, as taught by Broudou. The motivation to do so is to identify potentially risky language and to present to a given company various edits or solutions that comply with relevant regulatory standards (Broudou, [0043]; “Assigning a risk value to a distinct marker may assist a user to identify distinct markers which may have a relatively high or unacceptable risk for the user or a company. Reducing the overall risk may increase compliance with industry regulations or best practices.” Broudou, [0100]; “Each data point on the graph may provide additional details relevant to the risk and may offer suggestions on how to reduce a high risk area of the company.”).
Weiss further teaches training, by the server computing device, the convolutional neural network model using the updated training dataset (Weiss, [0043]; “According to some embodiments, the unsupervised training mechanism may be further adapted for adjusting the neural network.” Weiss, [0049]; “According to some embodiments, the neural-network may, for example, be a Recurrent and/or a Convolutional Neural Network.” Weiss, [0052]; “According to some embodiments of the present invention, as part of an auto-encoding process, one or more known artificial 'noise' elements may be introduced into the linguistic input learning data, wherein the linguistic input learning data may be unsupervised training data,” thereby training…the convolutional neural network model using the updated training dataset. Weiss, Fig. 3B; In the case when the model has not “reached a threshold level of semantic tagging
success” the “set of one or more words” including the second predicted tags encompasses an updated training dataset.).
Claims 11-15, 17, 18, 23 and 24 are system claims corresponding to the steps of claims 1-5, 7, 8, 21 and 22 and are therefore rejected for the same reasons. The additional hardware limitation, a server computing device communicatively coupled to a database and a user device, the server computing device configured to, is also taught by Lobo (Lobo, [0056]; “The computer system may include a processor coupled to a memory module and to a mass storage device via a bus or other communication medium; a display or other output device interfacing with the processor; and a keyboard, mouse, touchpad, or other input device that receives input from a user and interfaces with the processor,” wherein a “computer system” corresponds to a server computing device, “a mass storage device” corresponds to a database and “a keyboard, mouse, or other input device” encompasses a user device.).
Regarding claim 25, the combination of Weiss, Lobo, Broudou and Pajak teaches [t]he computerized method of claim 1 (and thus the rejection of claim 1 is incorporated).
Weiss does not explicitly teach wherein performing the feature engineering comprises identifying, for a particular token, a set of features including an indication of whether the particular token has previously been classified as misleading, promissory, or exaggerated. However, Goel, in the area of reviewing financial documents using natural language processing, teaches this limitation (Broudou, [0036]; “Classifier: a predetermined category which may be associated with at least one distinct marker based on the key terms, phrases or other predetermined text or symbols of the distinct marker,” wherein “a predetermined category” encompasses a categorical variable and, thus, a feature. Broudou, [0043]; “The risk value of a
portion of a document may be determined by comparing the analyzed portion of the document with classifiers stored in a knowledge base, each classifier may be associated with at least one associated rule, such that a portion of a document is assigned at least one classifier. The rules may be applied to the at least one classifier and assign a risk value thereto based on whether a predetermined rule or number or rules have been triggered or breached,” wherein feature
engineering encompasses applying rules “to the at least one classifier,” or feature, “and assigning a risk value thereto.” Broudou, [0069]; “Further, the system may be configured to provide a blacklist or exclusion list of symbols or terms for a document based if a predetermined classifier is associated with at least one distinct marker or if a predetermined rule is triggered,” wherein a “blacklist” term provides an indication of whether the particular token, or “distinct marker,” has previously been classified as misleading. Broudou, [0071]; “For example, if the term ‘free’ is used and at least one fee or charge is found to be associated with the term ‘free’ a high risk value may be assigned to the use of the term ‘free’ as this may provide incorrect information or misleading information,” wherein the term “free” was previously classified as misleading and placed on the “blacklist.” Its status as a “blacklist” term results in the term being assigned “a high risk value” in its given context.).
Broudou is analogous to the claimed invention as both are from the same field of endeavor, that is, assessing the compliance of financial documents using natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the combined token preprocessing and feature engineering of Weiss and Pajak to incorporate the predetermined blacklist terms of Broudou. The motivation to do so is to utilize a predefined list of problematic words in assessing the risk of a compliance or regulatory breach in financial or legal documents (Broudou, [0069]; “For
example, if a term or symbol from the blacklist is detected by the system, a risk value may be applied to the blacklist term which indicates a high or unacceptable risk value.” Broudou, [0089]; “Other predetermined thresholds may be the presence of absence of trigger or blacklist terms or phrases, text length, the number of ambiguous terms, the use of personal advice or how colloquial the language of the document may be.”).
Regarding claim 26, and in light of the 35 USC 112a rejection above, Weiss, discloses “According to some embodiments, the suggested taggings of the candidate words inputs may be altered by the system based on the manual 'feedback' provided,”; the figure discloses steps including “assigning markup language semantic tags” and assessing a “level of semantic tagging success.” If the level of success has not met a certain threshold than the prior steps are repeated. In such a case, a second tag is generated [0068, Fig. 3B ]-- wherein the training data set is not provided as input to the convolutional neural network model to generate a corresponding output.
Claims 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Weiss, Lobo, Broudou, Pajak, Song, Silberman and Xue et al. (US 11,429,834 B1, hereinafter Xue).
Regarding claim 9, the combination of Weiss, Lobo, Broudou and Pajak teaches the method of claim 8, wherein (and thus the rejection of claim 8 is incorporated).
Weiss does not explicitly teach the server computing device is configured to tokenize each of the plurality of data strings using lemmatization. However, Xue, in the area of building automated customer support agents using natural language processing models, teaches this limitation (Xue, Col. 4, lines 47-50; “In some examples, support texts 102 are first converted to all lowercase before performing any other pre-processing operation, such as stemming, lemmatization, tokenization, term replacement, or others.”).
Xue is analogous to the claimed invention as both are from the same field of endeavor, that is, natural language processing. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to combine the tokenization and tagging method of Weiss with the lemmatization of Xue. The motivation to do so would be to cut down on processing time by grouping together terms that share the same root (Xue, Cols. 3-4, lines 62-8; “For example, pre-processing 104 may include stemming support texts 102, which is
the process of reducing inflected (or sometimes derived) words to their word stem, base, or root form… Pre-processing 104 may also include lemmatization, which is a more complex approach to determining a stem of a word that involves first determining the part of speech of a word, and applying different normalization rules for each part of speech. In lemmatization, the part of speech is first detected prior to attempting to find the root since for some languages, the stemming rules change depending on a word's part of speech.”).
Claim 19 is a system claim corresponding to claim 9, and is therefore rejected for the same reasons as claim 9.
Response to Arguments
Applicant’s arguments and amendments, filed 12/22/2025, regarding the rejections
from the previous office action made under 35 U.S.C. 112(a) have been fully considered but are not persuasive.
Applicant’s argues that “described in Weiss is equivalent to the claimed "user response data." The Applicant respectfully disagrees. Weiss explains that manual feedback is part of internal training and corpus refinement, not feedback from a user indicating the accuracy of a model-assigned tag. See Weiss, paragraph [0068]. This is confirmed in the following paragraph, paragraph [0069] of Weiss, that further describes how the seeding of the manual feedback is used during training. The manual feedback in Weiss is not used by the model to assign a tag during inference or evaluation, and it is not accuracy feedback for a particular token by a user as required by the claims.”(p.13-14). The examiner disagrees, because Weiss teaches “According to some embodiments, the suggested taggings of the candidate words inputs may be altered by the system based on the manual 'feedback' provided” by a system curator, wherein “manual feedback” [0068, 57]-- user response data. The manual feedback is provided by a user or curator who provides feedback as to the accuracy or correctness of the output including token streams.
Applicant indicates that “Weiss does not generate two different tags for the same token, let alone in the claimed manner in which the second tag is determined based on the accuracy of the first tag.”(page 13). The examiner disagrees because Weiss teaches generating multiple taggings of each of the word being tagged(68). Each of the taggings can be provided with manual feedback and can be altered based on the manual feedback. The word(s) are repetitively fed through the model until a correct tagging is achieved (119).
Additionally, Applicant indicates that “Moreover, amended claim 1 requires that the user response data be provided as input to a convolutional neural network model to generate the second tag for the same token. Weiss does not disclose anything of this nature. As explained above, the manual feedback in Weiss is for training and is never supplied as input to the model and used by the model to generate a tag for a particular token. In other words, training data used to train a model is very different than input data provided to a model to generate a corresponding output.”(page 14). The examiner disagrees because Weiss teaches that each of the taggings can be provided with manual feedback indicating the similarity or accuracy of the taggings in relation to seed words and, and that the taggings for the same words can be altered based on the manual feedback. The word(s) are repetitively fed through the model until a correct tagging is achieved (fig.3B, 119, parag.68) through the model until a correct tagging is achieved (119).
For at least the reasons outlined above, the claims are taught by the prior art of record.
Conclusion
THIS ACTION IS MADE FINAL. 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 CESAR PAULA whose telephone number is (571)272-4128. The examiner can normally be reached Monday - Friday, 6.30am- 4:30 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, David Wiley can be reached at (571)272-3923. 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.
/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
Prosecution Insights