DETAILED ACTION
This communication is in response to the Application filed on 12/28/2023. Claims 1, 2, 4-11, and 13-20 are pending and have been examined. Hence, this Action has been made FINAL.
Any previous objection/rejection not mentioned in this Office Action has been withdrawn by the examiner.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 13, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
Applicant's arguments filed 12/08/2025 have been fully considered but they are not persuasive.
With respect to the 35 U.S.C. 101 rejections, regarding Step 2A Prong one the applicant asserts that these claims do not recite merely "reading" or "sorting" by a mailroom worker (human mind), rather they recite an ordered set of technical operations that operate on machine-interpretable representations of text and metadata, and that are performed by computer hardware running specifically configured machine-learning models. Specifically, each of these above indicated steps presents a concrete, technical operation that (a) requires a non-human computational process (e.g., transformer embedding generation, token batching, hyperparameter-based fine-tuning, softmax probability distributions), and (b) cannot be performed in the human mind in the manner claimed (for example, a human cannot execute transformer attention layers, generate multi-dimensional embeddings, or perform hyperparameter grid search and automated retraining loops on streams of labeled data). These limitations define a structured computational process for transforming unstructured electronic mail data into accurate, categorized outputs through a coordinated multi-model machine-learning architecture. This architecture converts noisy, heterogeneous email data into structured, actionable intelligence in real time.
Examiner respectfully disagrees, each limitation in the claim is a task that can be performed by the human mind combined with “apply it” language saying that it is done by a processor or MLM. The mailroom example was meant to illustrate how the broadest reasonable interpretation of the claims could be performed by the human mind and the additional components are merely used to apply the method via a computer. The examples given to claim this cannot be done by the human mind (transformer embedding generation, token batching, hyperparameter-based fine-tuning, softmax probability distributions, executing transformer attention layers, generating multi-dimensional embeddings, and the hyperparameter grid search) are not explicitly stated within the independent claim language. The current claim language does not include enough specificity on any processes that would make them unreasonable for the human mind to perform.
With respect to Step 2A Prong One, the applicant further asserts that unlike Electric Power Group (collecting and analyzing data), the present claims implement a domain-specific machine-learning pipeline that performs automated intent detection, reason-code classification, and entity standardization to improve the technological process of enterprise email understanding. This represents a concrete improvement to existing electronic communication workflows, akin to McRO, Inc. v. Bandai Namco Games Am., 837 F.3d 1299 (Fed. Cir. 2016), where rule-based automation was non-abstract. For example, "tokenization" as recited is the present claim is not the simple act of "labeling" that the Examiner has analogized in the office action. The claims require converting text and metadata into tokens, constructing vector embeddings, conducting multilabel transformer inference and model fusion operations that require arithmetic on high-dimensional vectors, GPU/TPU matrix multiplications, and trained attention weights. These are machine-level processes that are not "mental steps" for the reasons discussed in McRO, Inc. v. Bandai Namco Games Am., 837 F.3d 1299 (Fed. Cir. 2016) and Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016).
Examiner respectfully disagrees, McRo is not directly relevant to the instant application as the subject matter. McRo is related to improving 3D animation which is a process performed by computers. This application is in regards to categorizing text based on the intent and then applying the method via a computer. As currently claimed, the specification of processors, emails, and MLMs are merely used to apply the method via a computer rather than being essential to it. Furthermore, the processes listed as machine-level process are not explicitly present in the claim language and thus cannot be relied upon to as a step the human mind could not perform.
With respect to Step 2A Prong Two, the applicant asserts that the Examiner conflates generic articulation of components with recited, integrated, and concrete functional improvements. The claims do more than list generic components (processor, UI, storage). They recite a specific architecture and cooperative interaction among multiple transformer based machine learning model (a DistilRoBERTa intent model, a spaCy+RoBERTa reason-code classifier, and a spaCy-based RoBERTa NER system), orchestrated in an inference pipeline that (a) reuses embeddings across stages, (b) asynchronously batches and tokenizes heterogeneous inputs (metadata + latest message content), (c) applies entity- masking, lemmatization, and stopword filtering to improve tokenization quality, and (d) includes an automated retraining feedback loop driven by UI-provided labels and performance metrics. The integration is not "merely" presenting data on a UI; the UI is an integral component of a closed-loop retraining system that materially enhances model accuracy and system scalability.
Examiner respectfully disagrees, pre-trained models such as DistilRoBERTa, spaCy, and RoBERTa are considered general purpose due to their wide spread use and availability. Tasks A-C listed above represent tasks such models would are designed to perform and do so regularly do. Task D represents specific retraining of the model which could serve to differentiate it from the base models, however, the independent claim language does not explicitly state any form of specialized training for the models.
With respect to Step 2A Prong Two, the applicant further asserts these operations constitute a concrete technical implementation, not mere data presentation. The user interface is an active element of the closed-loop retraining subsystem that materially enhances model accuracy, adaptivity, and scalability. As per the USPTO's 2019 Revised Guidance and Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016) and DDR Holdings v. Hotels.com, 773 F.3d 1245 (Fed. Cir. 2014)), claims that improve computer performance or provide a technical solution to a technical problem are integrated into a practical application. The present invention demonstrates multiple such improvements: … Collectively, these additional elements apply the alleged abstract concept of "email categorization" in a specific, non-generic technological manner that enhances computer operation itself. This satisfies the "practical-application" standard of Enfish and DDR Holdings and accords with MPEP § 2106.05(a)-(c), which recognize improvements to computer functionality as evidence of integration into a practical application.
Examiner respectfully disagrees, the user interface is the component associated with the post-solution activity of data presentment. This is due to the current independent claim only mentioning this interface in regards to providing output. The remaining additional components (processors and MLMs) are considered “apply it” components as they are merely being used to apply the method via a computer rather than by the human mind. In regards to Enfish and DDR Holdings, as stated by the applicant, these are meant to show claims that improve computer performance or provide a technical solution to a technical problem. The issue here is, as currently claimed, identifying intent and categorizing text based on it is not inherently a technical problem. The recitation of the additional components (processor, MLM, interface) is merely being used to apply a mental process to a computer.
With respect to Step 2B, the applicant further asserts that when considered as an ordered combination, these elements define a non-conventional ML pipeline specifically tailored to large-scale enterprise electronic-mail understanding and classification. This is not the mere automation of a mail-sorting practice on a generic computer, but a domain-specific, technically controlled ML workflow that improves inference efficiency, accuracy, adaptivity, and scalability in environments characterized by unstructured, irregular, and high-volume communication data. This mirrors BASCOM Global Internet Servs., Inc. v. AT&T Mobility LLC, 827 F.3d 1341 (Fed. Cir. 2016), where a non-generic arrangement of known components was found inventive. Here, the integration of asynchronous tokenization, contextual embedding reuse, and transformer-based classification across multiple ML stages, combined with entity normalization and automated retraining through UI-driven feedback, within a real-time enterprise email-analysis workflow, achieves a technical improvement in inference accuracy, computational efficiency, and adaptive scalability of the computer system itself. Accordingly, the claimed invention represents significantly more than the alleged abstract idea. The claims recite an inventive, non-conventional, and technically integrated machine- learning architecture that materially improves computer functionality in handling unstructured language data.
Examiner respectfully disagrees, the ordered combination in the applicants’ arguments is not present within the claim language. The three MLMs, as currently claimed, do not show a multi-stage machine learning subsystem, but rather, a series of steps is listed in which some of those steps are done “based on an MLM”. Processes such as pre-processing, token optimization, asynchronous batching, retraining, hyperparameter tuning, and hardware level orchestration are also not present within the current claim. As stated in the non-final office action the processor is only used in the context of applying the method via a computer and as described in paragraph 64 of the specification it can take many general-purpose forms. The first and second MLM’s are described using exemplary language such as “may be a DistilRoBERTa Base model” (Paragraph 81), “In an embodiment” (Paragraph 99). Such language does not distinguish the component from a general-purpose MLM as it is not given definitive structure. Furthermore, the third MLM is described as being a pre-trained RoBERTa in paragraph 107 making it clear that it is generic/general-purpose. In order to show that the components are more than general-purpose it must be explicitly shown in both the claim language and the specification.
With respect to the 35 U.S.C. 103 rejections, the applicant asserts that the examiner has relied on columns 15 line 6-9 and column 3 lines 49-51. The examiner further asserts that "the system utilizes an intent analyzer for finding the intent of the emails, furthermore it is stated that operations of this systems can be done by an AI or ML." In this regard, Applicant respectfully submits that Kadatoka, at column 15, lines 6-9, merely provides a generic boilerplate statement that the operations disclosed may be conducted or enhanced by artificial intelligence or machine learning, without specifying any details of how or what machine learning techniques, architectures, or models may be utilized. This single, high-level reference to AI/ML enhancement does not constitute an enabling disclosure of a "machine learning model" as claimed in the present invention. The passage is purely speculative and non-limiting in nature, serving only as an optional statement of possible implementation improvements, rather than a teaching of any defined model or algorithm.
Examiner respectfully disagrees, the current claim language does not specify the implementation of the MLM in any meaningful way. The claim states “analyzing, by the one or more hardware processors, the intent of the one or more electronic mails based on a first machine learning model.” Analyzing an intent based on a MLM is not distinct from an invention that discloses analyzing an intent and states that it can be done using machine learning techniques. The current claim language does, for each of the three MLMs, merely states that some action is done based on the MLM. This level of broadness and generality can be interpreted in many ways which is why the prior art can be relied upon to teach such limitations. Overall, saying “doing X based on Y” is not distinct from “doing X, X can be done using Y”.
The applicant further asserts that column 3, lines 49-51 of Kadatoka disclose that the system 100 for classifying electronic messages "may comprise a Mail Room Intent Analyzer 102." As described in the accompanying detailed description (see, e.g., column 4, lines 40-67, FIG. 1), the Mail Room Intent Analyzer 102 functions merely as a communication and coordination node in the system, interfacing with the user device, the Message Validation Module (MVM 106), Stripping Module (108), Case Generation Engine (110), Parsing Module (112), Analysis Module (114), and Database (116). Its role is to orchestrate and trigger downstream processing steps, not to perform semantic interpretation or machine-learning-based inference. Contrary to the Examiner's assertion, Kadatoka does not disclose or suggest that the Mail Room Intent Analyzer 102 "finds" the semantic intent of an email, nor does it implement or invoke any artificial intelligence or machine learning model for such purpose. Instead, the "intent analyzer" operates as a routing controller that passes messages and metadata between deterministic rule-based modules. The text provides no description of any model training, feature extraction, tokenization, embedding generation, or classification algorithms comparable to the claimed first machine learning model. The mail room intent analyzer is responsible for classifying electronic messages and does so with communications it receives from the analysis module as can be seen here “In various embodiments, the configuration data 122 and/or user data 120 may include an association table 606 between the metadata elements (310, 312, 314) and the case type data 118. In various embodiments, the analysis module 114 may populate a possible case type set 608 comprising a plurality of case types based on the association table and the received metadata elements (310, 312, 314) (step 607)” (Col. 7, Line 65 to Col. 8, Line 3).
Examiner respectfully disagrees, as stated the mail room intent analyzer is categorizing emails into case types. This is representative of an intent as the case types represent intentions behind a received email. While the mail room intent analyzer is sending and receiving information from different components it is overall responsible for assigning this case type and thus it is analyzing the intent of the email. (In various embodiments, the configuration data 122 and/or user data 120 may include an association table 606 between the metadata elements (310, 312, 314) and the case type data 118. In various embodiments, the analysis module 114 may populate a possible case type set 608 comprising a plurality of case types based on the association table and the received metadata elements (310, 312, 314) (step 607).) (Col. 7, Line 64 to Col. 8, Line 3). The components above that are housing and selecting case types are both orchestrated by the mail room intent analyzer as can be seen in Fig. 1.
The applicant further asserts that the examiner has relied on Col. 6 lines 35-40 and FIG. 5 of Kadatoka. Col. 6 lines 35-40 and FIG. 5 merely discloses that the system may tokenize the body meat into a plurality of tokenized sentences 502' via a tokenizer sub-process (step 502). For example, the parsing module 112 may apply an unsupervised learning algorithm, which may include boundary detection, to extract individual (i.e., tokenized) sentences for further processing." It is submitted that the cited disclosure concerns rule-based sentence segmentation using boundary detection, punctuation or capitalization heuristics to isolate sentences. The reference to "unsupervised learning" in Kadatoka is statistical segmentation, not a neural-network tokenization step that generates contextual embeddings. The present invention, by contrast, applies transformer-based tokenization within a first transformer-based machine learning model to optimize text data by encoding contextual semantics across tokens. Kadatoka does not teach or suggest optimizing text data through ML token embeddings or feature weighting. Its "tokenization" is merely preprocessing for rule-based parsing, not machine-learning analysis.
Examiner respectfully disagrees, the claim language does not explicitly state the type of tokenization process. Under the broadest reasonable interpretation of the claims Kadatoka is teaching converting text to tokens using a tokenization process. Further specification would need to be amended to the claims in order to show a difference in methodology. (The system may tokenize the body meat into a plurality of tokenized sentences 502′ via a tokenizer subprocess (step 502). For example, the parsing module 112 may apply an unsupervised learning algorithm, which may include boundary detection, to extract individual (i.e. tokenized) sentences for further processing.) (Column 6, Lines 35 40).
The applicant further asserts that on page 26 of the Office Action, the Examiner further contends that Kadatoka discloses "converting, by the one or more hardware processors, the one or more text data associated with at least one of: the one or more metadata and the one or more first textual contents, into one or more tokens, based on the tokenization process, wherein the first machine learning model is configured to detect one or more semantic relationships on the one or more text data based on the one or more tokens." In support, the Examiner relies on Kadatoka, Column 7, lines 25-27 and Column 8, lines 20-26. Kadatoka in col7 lines 25-27 discloses generating a set of valid events 522. In this regard, the system may extract certain sets of events from the message body. Col 8 lines 20-26 discloses this is achieved by mapping words into a meaningful space where the distance between words is related to semantic similarity ... to generate respective word vectors (step 704). Applicant respectfully submits that these disclosures describe only a static, post- tokenization mapping of extracted event terms into pre-trained GloVe embedding space, where pairwise distances reflect co-occurrence statistics from an external corpus. This mapping is a fixed mathematical projection, not a learned or adaptive inference process. Kadatoka therefore employs non-trainable, static embeddings, rather than a machine learning model that dynamically detects and updates semantic relationships among tokens.
Examiner respectfully disagrees, “based on one or more tokens” does not explicitly state that the text is in a tokenized form when the semantic relationships are found. The flowchart in Fig. 5 of Kadatoka clearly shows the text being tokenized and then events being found from them. Regardless of the form they’re in when those relationships are found, the method is and order of operations are the same. Furthermore, as stated previously, Kadatoka specifies that these operation can be performed using machine learning techniques.
The applicant further asserts that on pages 27-28 of the Office Action, the Examiner contends that Kadatoka discloses "analyzing, by the one or more hardware processors, the intent of the one or more electronic mails by categorizing the one or more electronic mails into one or more distinct intent categories based on a multilabel classification process." In this regard, the Examiner relies on Kadatoka, Column 8, lines 6-12, which states that "the helper tags comprise a natural language text string which may be descriptive of the case type. ... For example, a first case type may be 'Account Changes' and associated with helper tags such as 'update my account, ‘change account, ‘revise account information,' and the like," and Column 8, lines 32-37, which notes that "Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712." In Kadatoka, the "case types" and "helper tags" are static label templates that operate within a rule-based case generation engine. The system merely computes vector similarities between pre-defined tag embeddings and extracted text features to assign a single case type. This is a one-to-one keyword-based matching process, not a probabilistic multilabel classification performed by a trained machine learning model.
Examiner respectfully disagrees, “one or more distinct intent categories” under broadest reasonable interpretation is not any different that case types or helper tags taught by Kadatoka. They are distinct as they each have their own meaning and intent associated with them, they are categories that emails are classified into, and they have multiple labels that can be applied (case types). As currently claimed, the limitation is not distinct from the method taught by Kadatoka. Furthermore, as stated previously, Kadatoka recites that these methods can be performed via machine learning techniques. (The helper tags comprise a natural language text string which may be descriptive of the case type. … For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like.) (Column 8, Lines 6-12)
The applicant further asserts that on page 27, the Examiner indicates that "Kadatoka in view of Liberty does not explicitly teach: wherein the one or more distinct intent categories comprise at least one of- procure-to-pay (P2P), dispute, follow up, account statement request, invoice request, payment confirmation, query and miscellaneous". … The Examiner has relied on paragraphs 13 and 35 of Tong, paragraphs 13 and 35 merely describe exemplary enterprise communication workflows in which natural-language inputs (such as a customer's email response) may correspond to transactional actions-e.g., disputing an invoice, confirming payment, or requesting a statement. Tong's "NLU" is invoked at a conceptual level to determine which business rule applies within a preconfigured process (a "smartflow"), but there is no teaching of multilabel ML classification, DistilRoBERTa-like transformer architectures, or semantic token embeddings. Tong's examples are procedural triggers within an enterprise automation system, not data-driven model classifications of multiple simultaneous intents.
Examiner respectfully disagrees, as discussed above, Kadatoka already teaches the method presented by the claims. Tong is brought in merely to show that these specific intent categories are not novel and the method of Kadatoka could. Tong shows that emails can be classified into one or more of the categories listed in the limitation and there is motivation to combine due to the closely related subject matter. (interactions with smartflows executed by an integration platform can involve natural language understanding (NLU) to interpret the messages and decide what actions are requested. For example, a smartflow may send notification of payment due to a customer. The customer may ignore the message, or take any of a number of other actions in response, such as dispute an item on invoice, request a copy of the invoice, notify that they will pay in a certain number of days, etc. In many embodiments of the invention, the notification can be an email (e.g., from a server) and the response can be an email (e.g., by a customer's device).) (Tong et al. Paragraph 35).
Aspects such as the DistilRoBERTa-like transformer architecture are not explicitly stated in the claim language and the multilabel classification and semantic token embedding are taught by Kadatoka. (The helper tags comprise a natural language text string which may be descriptive of the case type. … For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like.) (Kadatoka et al. Column 8, Lines 6-12). (The system may tokenize the body meat into a plurality of tokenized sentences 502′ via a tokenizer subprocess (step 502). For example, the parsing module 112 may apply an unsupervised learning algorithm, which may include boundary detection, to extract individual (i.e. tokenized) sentences for further processing.) (Kadatoka et al. Column 6, Lines 35 40).
The applicant further asserts that on page 18 of the Office Action, the Examiner asserts that Kadatoka, discloses "classifying, by the one or more hardware processors, the one or more electronic mails into one or more reason codes based on a second machine learning model. In this regard, the Examiner has relied on Kadatoka, Col. 8, lines 1-14 for its "analysis module 114" and its corresponding logic. Kadatoka in col 8 lines 1-14 discloses that " the analysis module 114 may populate a possible case type set 608 comprising a plurality of case types based on the association table and the received metadata elements (310, 312, 314) (step 607). In this regard, the system may generate a possible case type set based on the metadata. ... the case type may be a text string. For example, a first case type may be 'Account Changes' and associated with helper tags, such as 'update my account, ‘change account, ‘account changes, ‘revise account information,' and/or the like. The helper tags may be stored as configuration data 122 in database 116. Applicant respectfully submits that the cited disclosure does not teach or suggest the claimed classification into reason codes based on a second machine learning model. The Kadatoka passage clearly describes a rule-based population of static case-type sets from an association table and pre-configured helper-tag strings. The "helper tags" are literal text patterns ("update my account, “change account," etc.) stored as configuration data, which are matched deterministically against message metadata. This is a string-matching and lookup mechanism, not a trained model performing probabilistic inference or contextual reasoning.
Examiner respectfully disagrees, the claim language is not clear as to what a reason code is. Under broadest reasonable interpretation, assigning the helper tags to the electronic message (even if the helper tags are text patterns) to assist in the classification of the message is the same as classifying the message into reason codes. (the analysis module 114 may populate a possible case type set 608 comprising a plurality of case types based on the association table and the received metadata elements (310, 312, 314) (step 607). In this regard, the system may generate a possible case type set based on the metadata. … the case type may be a text string. For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like. The helper tags may be stored as configuration data 122 in database 116.) (Column 8, Lines 1-14).
The helper tags represent reasons that a message would fall under a certain case type. As discussed previously, saying “based on an MLM” is not distinct from saying “these method can be performed by an MLM”.
The applicant further asserts that on page 18 of the Office Action, the Examiner contends that Kadatoka discloses "extracting, by the one or more hardware processors, information associated with one or more named entities from textual information associated with one or more unstructured and unlabeled electronic mails based on a third machine learning model" In support, the Examiner relies on Kadatoka, Column 7, lines 41-46, and Figure 3, which describes that "the system may extract named entities 510 and events 522 from the message body. For example, the parsing module 112 may apply pattern recognition to detect entities such as names, dates, account numbers, or transaction identifiers" Applicant respectfully submits that this disclosure does not teach or suggest the claimed extraction of named entities based on a third machine learning model. The cited Kadatoka passage clearly describes rule-based pattern recognition that is, applying deterministic regular expressions or keyword heuristics to identify fixed categories of information (names, dates, account numbers, etc.). This approach does not involve any learning, feature embedding, or adaptive inference.
Examiner respectfully disagrees, as discussed previously, the current claim language provides no specificity as to how the MLMs are used, it merely states “based on an MLM” at the end of a method step. Due to this non-specific language the implementation of an MLM can be interpreted very broadly. Thus, when Kadatoka describes that their method can be performed using machine learning techniques it is sufficient evidence that this aspect of the claim is being taught.
The applicant further asserts that on pages 18-19 of the office action, the Examiner contends that Kadatoka discloses "standardizing, by the one or more hardware processors, the extracted information associated with the one or more named entities by refining and harmonizing the information associated with the one or more named entities". In this regard, the Examiner has relied on Col. 8 lines 20-26 of Kadatoka. It is submitted that Kadatoka discloses that "...the model is an unsupervised learning algorithm for obtaining vector representations for words. This is achieved by mapping words into a meaningful space where the distance between words is related to semantic similarity ... to any of the possible case type set 608, the set of valid events 522, and/or the named entities 510 to generate respective word vectors (step 704) ..."34 Applicant respectfully submits that this passage describes a vector embedding procedure, not a standardization process. Kadatoka's system applies a GloVe-style unsupervised model to produce numeric word vectors representing co-occurrence statistics within the corpus. Such embedding is a mathematical encoding step performed to compute similarity scores for case-type classification. It does not perform the refining or harmonizing of entity information as claimed. The generated word vectors in Kadatoka remain separate, unrefined numerical representations and are not used to correct, unify, or reconcile inconsistent entity instances.
Examiner respectfully disagrees, “refining and harmonizing” does not provide a clear description as to how the standardizing is occurring. By converting the text to vector representations, the text is being “refined” and by mapping them based on semantic similarity the vectors are being “harmonized”. In this sense, Kadatoka teaches the standardization of the extracted information. (Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712.) (Col. 8, Lines 32-36)
The applicant further asserts that on page 19 of the Office action, the Examiner contends that the Kadatoka discloses "grouping, by the one or more hardware processors, the one or more electronic mails into one or more categories based on at least one of. the intent of the one or more electronic mails, the one or more reason codes and the standardized information associated with the one or more named entities". In this regard, the Examiner has relied on Col. 8 lines 32-37 and Fig. 7 of Kadatoka. It is submitted that Kadatoka in the said portion discloses "Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712." Applicant respectfully submits that Kadatoka does not disclose or suggest the claimed grouping operation. The described "matching" is a vector similarity computation used to associate an email message with one case type based on pre-defined helper tags. This is a classification step that produces a single case-type label, not a grouping process based on multiple layers of analytical criteria. The "helper tags" serve as keyword descriptors for pre-set categories ("account changes,""update my account," etc.) but are not used to organize multiple emails or to form dynamic, data-driven categories across datasets. In contrast, the present invention expressly recites that the electronic mails are grouped into categories based simultaneously on multiple analytical dimensions including (i) the detected intent, (ii) the reason codes and (iii) the standardized named entity information. This grouping is performed dynamically to form multi-criteria clusters that reflect semantic, functional, and contextual relationships among emails. Furthermore, Kadatoka does not utilize the standardized or harmonized entity information in any subsequent grouping or classification process; entities 510 are static inputs to event extraction, not organizing parameters.
Examiner respectfully disagrees, the electronic mails are grouped as they are put into categories as a case type (intent). Regardless of the exact method of grouping, it is doing a grouping operation based on intent, reason codes, and standardized information. As discussed previously the standardized information is represented by the named entities as these are standardized based on a vectorization process. This gropuing can be seen in Fig. 7 of Kadatoka where the case types (intent), helper tags (reason codes), and standardized information (final entities) are combined to select the most likely case type for the email. (Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712.) (Col. 8, Lines 32-36)
The applicant further asserts that with regards to Liberty, it is submitted that Liberty is directed to a rule-driven email prioritization and organization system, fundamentally and functionally distinct from the claimed multi-model machine-learning architecture of the present invention. Specifically, Liberty discloses filtering, ranking, and routing of emails based on predetermined criteria such as sender identity, message thread participation, static keywords, and user-defined priority settings. Liberty executes deterministic logic or user- authored rules to sort or highlight messages; however, it does not perform tokenization, context embedding, or deep-learning-based semantic intent detection. Nor does Liberty disclose any multi-level transformer-based models, any multilabel classification process, or any pipeline that extracts, standardizes, and harmonizes named entities.
Examiner respectfully disagrees, Liberty is not meant to teach every aspect of the claimed method, but rather to teach preprocessing the email with the context of previously received/stored emails. Liberty does do email processing based on previously received emails and there is motivation to combine based on both references performing processing on received emails. The remaining aspects are taught by Kadatoka and other references detailed below.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 2, 4-11, and 13-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Claims 1, 10, and 19 recite A machine-learning based (ML-based) computing method for analyzing an intent of one or more electronic mails. the ML-based computing method comprising:
Receiving, by [one or more hardware processors], one or more metadata associated with the one or more electronic mails from one or more databases,
wherein the one or more metadata associated with the one or more electronic mails comprise at least one of one or more primary electronic mail keys. one or more raw electronic mail links, one or more conversation identities, one or more electronic mail receipt dates. and one or more sender identities:
extracting, by the one or more hardware processors, one or more first textual contents from one or more last received electronic mails within one or more electronic mail chains stored in one or more electronic mail files by preprocessing the one or more last received electronic mails:
analyzing, by the one or more hardware processors, the intent of the one or more electronic mails based on a [first machine learning model], comprises:
obtaining, by the one or more hardware processors, at least one of: the one or more metadata and the one or more first textual contents extracted from the one or more last received electronic mails;
analyzing, by the one or more hardware processors. at least one of: the one or more metadata and the one or more first textual contents to optimize one or more text data associated with at least one of: the one or more metadata and the one or more first textual contents, based on a tokenization process:
converting, by the one or more hardware processors, the one or more text data associated with at least one of: the one or more metadata and the one or more first textual contents, into one or more tokens, based on the tokenization process,
wherein the first machine learning model is configured to detect one or more semantic relationships on the one or more text data based on the one or more tokens: and
analyzing, by the one or more hardware processors, the intent of the one or more electronic mails by categorizing the one or more electronic mails into one or more distinct intent categories based on a multilabel classification process,
wherein the one or more distinct intent categories comprise at least one of: procure-to-pay (P2P), dispute, follow up, account statement request, invoice request, payment confirmation, and query and miscellaneous.
Classifying, by the one or more hardware processors. the one or more electronic mails into one or more reason codes based on a [second machine learning model]:
extracting, by the one or more hardware processors, information associated with one or more named entities from textual information associated with one or more unstructured and unlabeled electronic mails based on a [third machine learning model]:
wherein the first machine-learning model, the second machine-learning model, and the third machine-learning model comprises one or more transformer-based machine learning models:
standardizing, by the one or more hardware processors. the extracted information associated with the one or more named entities by refining and harmonizing the information associated with the one or more named entities:
grouping, by the one or more hardware processors, the one or more electronic mails into one or more categories based on at least one of: the intent of the one or more electronic mails, the one or more reason codes, and the standardized information associated with the one or more named entities: and
providing, by the one or more hardware processors, an output of one or more categorized electronic mails to one or more users on a [user interface] associated with one or more electronic devices.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. The metadata associated with an electronic mail is equivalent to the information that would be labeled on the front of physical mail. For example, someone reading the name of the senders on letters would be the human mind gathering metadata from the mail. Then some can extract information from a letter using their mind and reading comprehension skills. In this example, preprocessing the data could be discarding the pieces of mail they assume aren't useful. A human can analyze the intent of a piece of mail using the prior knowledge they possess. A human could tokenize the letter by assigning certain labels or numbers to predetermined words or phrases in order to help organize it. The tokens can then be used to find semantic relationships by putting words/phrases with the same label/number together. A human could then form a multilabel classification process by grouping letters with certain label/number combinations together when organizing mail. A human would be capable of assigning labels to letters based on categories such as account statement or invoice requests. A human classifying mail into reason codes could be someone working in a mailroom and sorting letters based on where in the building they need to go. A human mind is capable of finding named entities in a piece of mail. Can standardize the named entities by using prior knowledge and context from the letter to group together letters for the same person or from the same sender. Using the mail room example again the worker can organize the letters based on where they need to go and what type of letter it is, thus classifying it. Finally, the letters can be provided to someone or multiple people once they are organized. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. In particular, claims 1, 10 and 19 recite processors, a first, second, and third machine learning model, and a user interface. The processor is described in Paragraph 64 of the specification and is given a generic description of the component. The first MLM is described in Paragraph 81 of the specification with a general-purpose example of DistilRoBERTa Base Model given. The second MLM is described in paragraph 99 of the specification as a combination of general-purpose models spaCy and RoBERTa. The third MLM is described on Paragraph 107 with a general-purpose example of a RoBERTa Base transformer model given. The user interface is described in Paragraph 150 of the specification with various generic examples provided. Furthermore, the user interface is considered pre-solution and/or post-solution activity as it is merely gathering or presenting data for all the steps that occur in the method. Claim 10 specifically lists the additional component of a memory which is described in Paragraph 65 of the specification with a generic description of the component. Claim 19 specifically lists the additional component of a non-transitory computer readable storage medium which is described in Paragraph 65 of the specification with a generic description of the component. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 2 and 11 recite wherein preprocessing the one or more last received electronic mails comprises: identifying. by the one or more hardware processors, the one or more last received electronic mails by analyzing one or more timestamps associated with the one or more electronic mails based on an electronic mail identification process;
detecting, by the one or more hardware processors, the one or more first textual contents within the one or more electronic mail files by applying a regular expression text extraction process.
wherein the regular expression text extraction process is configured to be applied on one or more electronic mail formatting settings for detecting at least one of: one or more plain textual contents and one or more hypertext markup language (HTML) embedded contents:
converting, by the one or more hardware processors, the one or more hypertext markup language (HTML) embedded contents to the one or more plain textual contents by parsing the one or more hypertext markup language (HTML) embedded contents based on an hypertext markup language (HTML) parsing process:
and extracting, by the one or more hardware processors. one or more sender information from the one or more electronic mail files to identify and remove one or more second textual contents associated with the one or more electronic mail senders.
wherein the one or more second textual contents associated with the one or more electronic mail senders comprise at least one of: greetings, salutations. and signatures.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. A human can interpret the timeframe letters were received by reading the date labeled on them. Humans are capable of detecting context in the plain text of letters, furthermore, humans can read and interpret HTML as it is syntax designed by humans. A human who understands HTML could interpret what the code says and write what it means in written language. A human could also remove greetings, salutations, or signatures from a letter by either ignoring it or literally removing it from the letter using white-out or scissors. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 4 and 13 recite Wherein the first machine learning model is fine-tuned for categorizing the one or more electronic mails into one or more distinct intent categories:
the first machine learning model is optimized to categorize the one or more electronic mails into one or more distinct intent categories by selecting one or more hyperparameters comprising at least one of: learning rates, batch sizes, and regularization strength; and
the first machine learning model is trained to learn from the one or more electronic mails being labelled under the one or more distinct intent categories.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. A human can categorize mail based on what they believe the intent of the letter to be. The decision to use certain types of parameters when training an MLM is a design decision which the human mind is capable of making. This also applies to the design decision to use mail that has already been labeled into distinct categories for training the MLM. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 5 and 14 recite further comprising generating, by the first machine learning model, one or more confidence scores for each intent analyzed for the one or more electronic mails by providing probability distribution over the one or more tokens based on a softmax function,
wherein the one or more tokens with optimum probability are selected as the analyzed intent of the one or more electronic mails.
The limitations in these claims, as drafted, is a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind and mathematical concepts being performed by generic computer components. Selecting the highest probability assigned using a SoftMax function is merely a mathematical equation being performed by a computer. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. If a claim limitation, under its broadest reasonable interpretation, covers mathematical relationships or equations being performed by generic computer components, then it falls within the “Mathematical Concepts” grouping of abstract ideas. Accordingly, the claim recites an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 6 and 15 recite wherein classifying, by the second machine leaning model, the one or more electronic mails into the one or more reason codes comprises: obtaining, by the one or more hardware processors, the one or more first textual contents;
converting, by the one or more hardware processors, the one or more first textual contents into the one or more tokens, based on the tokenization process:
applying, by the one or more hardware processors, entity masking process on the one or more first textual contents to normalize one or more words essential in the one or more first textual contents:
applying, by the one or more hardware processors, at least one of: a lemmatization process and a stopword process to refine the one or more first textual contents;
extracting, by the one or more hardware processors, one or more first linguistic information associated with the one or more tokens corresponding to the one or more first textual contents:
generating, by the one or more hardware processors, one or more contextually relevant embeddings capturing intricate semantics related to the one or more electronic mails by applying the extracted one or more first linguistic information on the first machine learning model:
assigning, by the one or more hardware processors, one or more scores to the one or more reason codes based on at least one of: the one or more contextually relevant embeddings and the one or more first linguistic information associated with the one or more electronic mails: and
classifying, by the one or more hardware processors, the one or more reason codes for the one or more electronic mails based on one or more optimum scores assigned to the one or more reason codes.
wherein the one or more reason codes comprise at least one of: one or more product reason codes, one or more service reason codes, one or more customer reason codes, one or more account reason codes, one or more transaction reason codes, one or more location reason codes, one or more payment reason codes.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. A human can obtain the body of text from a letter and convert it into tokens by assigning labels/numbers to corresponding words or phrases. A human could also identify named entities in a piece of mail and either group them with matching entities or remove them from the body of the text. A human is also capable of simplifying the language of a letter by replacing words and removing repetitive words that are deemed unnecessary either by ignoring them or literally removing them from the mail. A human can extract tokens from the mail by looking at the previously assigned labels/numbering to and knowing what they mean either with prior knowledge or using a reference. A human can assign intrinsic semantics to these tokens using their prior knowledge to determine how the words/phrases relate and are significant to the piece of mail. Reason codes could then be assigned by doing further labeling/numbering of the mail based on the labels/numbers added for the individual terms. An optimum score in this instance could be a level of importance of the piece of mail based on how it was labeled or a level of certainty of its category based on the specific labels that were within the letter. A human is capable of providing a label to mail that is related to service, customers, accounts, or payments. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 7, 16, and 20 recite wherein the third machine learning model is an integration of at least one of: the first machine learning model and the second machine learning model. and
wherein extracting, by the third machine learning model, the information associated with the one or more named entities comprises. obtaining, by the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails:
scanning, be the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails for extracting the one or more named entities:
applying, by the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails on the third machine learning model: and
extracting, by the one or more hardware processors, one or more patterns indicating the one or more named entities by leveraging one or more second linguistic information associated with the one or more unstructured and unlabeled electronic mails,
wherein the information associated with the one or more named entities comprises at least one of: names of one or more electronic mail senders, one or more organizations. one or more dates, one or more times. one or more invoice numbers, and one or more amounts.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. The third MLM's intended use and the decision to integrate from another MLM are design decisions that could be made by the human mind. A human can obtain the information in mail by reading it. A human can read the mail and look for named entities within it. A human can find patterns relating to the second textual information by, for example, looking for the name of the sender within the body of the letter. A human could find this information for senders, organization, dates, times, invoice numbers, or account numbers. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 8 and 17 recite further comprising re-training, by the one or more hardware processors, one or more machine learning models comprising at least one of:
the first machine leaning model, the second machine teaming model, and the third machine learning model, over a plurality of time intervals based on one or more training data,
wherein re-training the one or more machine learning models over the plurality of time intervals comprises: receiving, by the one or more hardware processors, the one or more training data associated with the one or more categorized electronic mails, from an output subsystem;
adding, by the one or more hardware processors, the one or more training data with one or more original training datasets to generate one or more updated training datasets.
re-training, by the one or more hardware processors, the one or more machine learning models, by adjusting at least one of: the one or more hyperparameters and one or more training configurations of at least one of: an electronic mail intent analyzing subsystem, an electronic mail reason code classification subsystem, and a named entity recognition subsystem:
and executing, by the one or more hardware processors, the re-trained one or more machine learning models in at least one of: the electronic mail intent analyzing subsystem. the electronic mail reason code classification subsystem, and the named entity recognition subsystem, to output the one or more categorized electronic mails
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. The design decision to re-train models over a plurality of intervals could be done by the human mind. As could the decision to use categorized letters as additional training data. The decision to adjust specific hyperparameters is another design decision on how to train the MLM and the subsystems associated with that MLM. Finally, utilizing the MLM to execute its purpose after re-training it is another design decision as to the training of an MLM which a human is capable of making. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
Claims 9 and 18 recite wherein the one or more metadata associated with the one or more electronic mails, are extracted from the one or more databases based on one or more techniques comprising at least one of: data normalization, data anonymization, data aggregation, data analysis, and data storage:
and wherein the one or more databases comprises at least one of: one or more relational databases, one or more object-oriented databases, one or more data warehouses, and one or more cloud-based databases.
The limitations in these claims, as drafted, are a process that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. A human can extract information from a piece of mail using a technique such as data normalizations by removing duplicates or renaming things that are meant for the same person/place but were spelled incorrectly. Furthermore, a human is capable of obtaining mail from a database such as a bin in a mailroom. The design decision as to specific types of databases to potentially pull information from could be made by a human. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
This judicial exception is not integrated into a practical application. The claims do not recite any additional components that were not present in the independent claims. Accordingly, the claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The claims are not patent eligible.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 9, 10, 18, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view of US Patent Publication US 8856249 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), and US Patent Publication US 11914630 B2 (Wu et al.).
Regarding Claims 1, 10, and 19, Kadatoka et al. teaches A machine-learning based (ML-based) computing method for analyzing an intent of one or more electronic mails. the ML-based computing method comprising:
(The present disclosure generally relates to systems and methods for classifying electronic messages.) (Column 1, Lines 5-6)
(the operations may be machine operations or any of the operations may be conducted or enhanced by artificial intelligence (AI) or machine learning.) (Column 15, Lines 6-9)
Receiving, by one or more hardware processors, one or more metadata associated with the one or more electronic mails from one or more databases,
(The stripping module may comprise an 1×ml parser configured to separate the metadata and the message body. The metadata parser 306 may be configured to extract elements of a structured message header such as a sender 310, a receiver 312, a copy to (CC) field 314, and a subject field 316.) (Column 5, Lines 8-13)
Metadata associated with the emails is used, as can be seen in Fig. 3.
wherein the one or more metadata associated with the one or more electronic mails comprise at least one of one or more primary electronic mail keys. one or more raw electronic mail links, one or more conversation identities, one or more electronic mail receipt dates. and one or more sender identities:
(The stripping module may comprise an 1×ml parser configured to separate the metadata and the message body. The metadata parser 306 may be configured to extract elements of a structured message header such as a sender 310, a receiver 312, a copy to (CC) field 314, and a subject field 316.) (Column 5, Lines 8-13)
Metadata includes sender receiver and CC info which is conversation identities and/or sender identities. Once again, this can be visualized in Fig. 3
analyzing, by the one or more hardware processors, the intent of the one or more electronic mails based on a first machine learning model, comprises:
(the operations may be machine operations or any of the operations may be conducted or enhanced by artificial intelligence (AI) or machine learning.) (Column 15, Lines 6-9)
(In various embodiments, and with reference to FIG. 1, a system 100 for classifying electronic messages may comprise a mail room intent analyzer 102) (Column 3, Lines 49-51)
The system utilizes an intent analyzer for finding the intent of the emails, furthermore, it is stated that the operations of this system can be done by AI or ML.
obtaining, by the one or more hardware processors, at least one of: the one or more metadata and the one or more first textual contents extracted from the one or more last received electronic mails;
(the operations may be machine operations or any of the operations may be conducted or enhanced by artificial intelligence (AI) or machine learning.) (Column 15, Lines 6-9)
(The stripping module may comprise an 1×ml parser configured to separate the metadata and the message body. The metadata parser 306 may be configured to extract elements of a structured message header such as a sender 310, a receiver 312, a copy to (CC) field 314, and a subject field 316.) (Column 5, Lines 8-13)
In Kadatoka et al, as can be seen in Fig. 3, the metadata is extracted from the emails and any process in this system can be conducted via machine learning.
analyzing, by the one or more hardware processors. at least one of: the one or more metadata and the one or more first textual contents to optimize one or more text data associated with at least one of: the one or more metadata and the one or more first textual contents, based on a tokenization process:
(The system may tokenize the body meat into a plurality of tokenized sentences 502′ via a tokenizer subprocess (step 502). For example, the parsing module 112 may apply an unsupervised learning algorithm, which may include boundary detection, to extract individual (i.e. tokenized) sentences for further processing.) (Column 6, Lines 35 40).
In Kadatoka et al. the body meat represents the first textual contents and it is tokenized. This can be visualized in Fig. 5.
converting, by the one or more hardware processors, the one or more text data associated with at least one of: the one or more metadata and the one or more first textual contents, into one or more tokens, based on the tokenization process,
(The system may tokenize the body meat into a plurality of tokenized sentences 502′ via a tokenizer subprocess (step 502). For example, the parsing module 112 may apply an unsupervised learning algorithm, which may include boundary detection, to extract individual (i.e. tokenized) sentences for further processing.) (Column 6, Lines 35 40).
In Kadatoka et al. the body meat represents the first textual contents and it is tokenized. This can be visualized in Fig. 5.
wherein the first machine learning model is configured to detect one or more semantic relationships on the one or more text data based on the one or more tokens: and
(generate a set of valid events 522. In this regard, the system may extract certain sets of events from the message body.) (Column 7, Lines 25-27)
(This is achieved by mapping words into a meaningful space where the distance between words is related to semantic similarity) to any of the possible case type set 608, the set of valid events 522, and/or the named entities 510 to generate respective word vectors (step 704).) (Column 8, Lines 20-26)
In Kadatoka et al, In Fig. 5 it can be seen that tagging the tokenized sentences eventually leads to the Final Valid Events component 522. Then in Fig. 7 it shown by the Glove algorithm that semantic relationships are identified on component 522
analyzing, by the one or more hardware processors, the intent of the one or more electronic mails by categorizing the one or more electronic mails into one or more distinct intent categories based on a multilabel classification process,
(Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712.) (Column 8, Lines 32-37)
(The helper tags comprise a natural language text string which may be descriptive of the case type. … For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like.) (Column 8, Lines 6-12)
In Kadatoka et al, the case types represent possible intents for an email. In Fig. 7 it can be seen how the word vectors from the tokenized sentences representing entities and events are matched with helper tags. This matching represents several distinct intent categories as a result of a multilabel classification process.
Classifying, by the one or more hardware processors. the one or more electronic mails into one or more reason codes based on a second machine learning model:
(the analysis module 114 may populate a possible case type set 608 comprising a plurality of case types based on the association table and the received metadata elements (310, 312, 314) (step 607). In this regard, the system may generate a possible case type set based on the metadata. … the case type may be a text string. For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like. The helper tags may be stored as configuration data 122 in database 116.) (Column 8, Lines 1-14)
In this case the reason codes are represented by helper tags which can take the form of strings fulfilling the same purpose of categorization.
extracting, by the one or more hardware processors, information associated with one or more named entities from textual information associated with one or more unstructured and unlabeled electronic mails based on a third machine learning model:
(The system may extract a set of named entities 510 from the body meat 320 via an entity extractor subprocess (step 508). Step 508 includes searching the body meat 320 for entities such as, for example, a merchant ID, a merchant name, a user name, a place name, and/or the like.) (Column 7, Lines 41-46)
It can be seen in Fig. 3 that the named entities are extracted from the body text of the email.
standardizing, by the one or more hardware processors. the extracted information associated with the one or more named entities by refining and harmonizing the information associated with the one or more named entities:
(The model is an unsupervised learning algorithm for obtaining vector representations for words. This is achieved by mapping words into a meaningful space where the distance between words is related to semantic similarity) to any of the possible case type set 608, the set of valid events 522, and/or the named entities 510 to generate respective word vectors (step 704).) (Column 8, Lines 20-26).
The named entities are standardized into a vector representation along with other information gathered.
grouping, by the one or more hardware processors, the one or more electronic mails into one or more categories based on at least one of: the intent of the one or more electronic mails, the one or more reason codes, and the standardized information associated with the one or more named entities: and
(Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712.) (Col. 8, Lines 32-36)
Fig. 7 shows the combination of the components into final case types which represent the emails being grouped into categories. The intent of the emails is represented by the possible case types in element 608. The reason codes are represented by the helper tags in element 608. The standardized information with the named entities is represented by both the final entities and the final events in elements 510 and 522. The components are grouped in the analysis module 704 where a combined word vector is paired with helper tags. The pairings are then ranked to determine the correct classification for the email.
providing, by the one or more hardware processors, an output of one or more categorized electronic mails to one or more users on a user interface associated with one or more electronic devices.
(a display device coupled to the processor and memory for displaying information derived from digital data processed by the processor.) (Column 14, Lines 53-55).
Information of this system is displayed to the user. This can be visualized in Fig. 10.
Kadatoka et al. does not explicitly teach: extracting, by the one or more hardware processors, one or more first textual contents from one or more last received electronic mails within one or more electronic mail chains stored in one or more electronic mail files by preprocessing the one or more last received electronic mails: wherein the one or more distinct intent categories comprise at least one of: procure-to-pay (P2P), dispute, follow up, account statement request, invoice request, payment confirmation, and query and miscellaneous. wherein the first machine-learning model, the second machine-learning model, and the third machine-learning model comprises one or more transformer-based machine learning models;
However, Liberty et al. teaches extracting, by the one or more hardware processors, one or more first textual contents from one or more last received electronic mails within one or more electronic mail chains stored in one or more electronic mail files by preprocessing the one or more last received electronic mails:
(For example, in one embodiment, the email server 122 determines the time that the email message was sent and compares this time to the time associated with the previous email that may have caused this email. If the difference in time period is above a given threshold, then this email message is probably not related to the previous email message. Other parameters that the email server 122 can review include the vendor that sent the email messages (e.g., both email messages are from eBay), a confidence value that this email message is caused by the previous email message, matching order numbers, matching email information, etc) (Column 7, Lines 55-66).
Liberty et al. performs forms of preprocessing on the current email and previously received ones to determine if there’s a cause-and-effect relationship.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. to include processing of previous emails as taught by Liberty et al. This would have been an obvious improvement add an additional metric to group emails according to, the metric being relation through previous emails (Liberty et al. Column 2, Lines 49-57).
Kadatoka et al. in view of Liberty et al. does not explicitly teach: wherein the one or more distinct intent categories comprise at least one of: procure-to-pay (P2P), dispute, follow up, account statement request, invoice request, payment confirmation, and query and miscellaneous. wherein the first machine-learning model, the second machine-learning model, and the third machine-learning model comprises one or more transformer-based machine learning models;
However, Tong et al. teaches a wherein the one or more distinct intent categories comprise at least one of: procure-to-pay (P2P), dispute, follow up, account statement request, invoice request, payment confirmation, and query and miscellaneous.
(interactions with smartflows executed by an integration platform can involve natural language understanding (NLU) to interpret the messages and decide what actions are requested. For example, a smartflow may send notification of payment due to a customer. The customer may ignore the message, or take any of a number of other actions in response, such as dispute an item on invoice, request a copy of the invoice, notify that they will pay in a certain number of days, etc. In many embodiments of the invention, the notification can be an email (e.g., from a server) and the response can be an email (e.g., by a customer's device).) (Paragraph 35).
Tong et al. teaches a system of identifying intent in emails which includes intent categories such as disputes, invoice requests, payment confirmation, or misc.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al. to include distinct categories having to do with payment and account information as taught by Tong et al. This would have been an obvious improvement as businesses commonly need to organize large amounts of information related to transactions and accounts (Tong et al. Paragraph 13).
Kadatoka et al. in view of Liberty et al. and Tong et al. does not explicitly teach: wherein the first machine-learning model, the second machine-learning model, and the third machine-learning model comprises one or more transformer-based machine learning models;
However, Wu et al. teaches a wherein the first machine-learning model, the second machine-learning model, and the third machine-learning model comprises one or more transformer-based machine learning models;
(Application of the unsupervised machine learning algorithm to the unlabeled text data generates labeled text data where the text data is labelled (e.g., annotated) with the probabilistic labels. The text data and the probabilistic labels are then provided to another machine learning algorithm (e.g., a transformer-based machine learning algorithm such as BERT (Bidirectional Encoder Representations from Transformers)) for training of the machine learning algorithm. … The trained classifiers may then be implemented by the transformer-based machine learning algorithm (or another transformer-based machine learning algorithm) in a classification process on text data from any source (such as chat messages, email messages, text messages, social media applications, etc.). Output for the text data may also be in a matrix showing probabilities for each category.) (Col. 4, Line 45 to Col. 5, Line 2).
Wu et al. teaches a method for classifying text data such as emails in which transformer-based machine learning models are used to the label the text data.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al. and Tong et al. to implement transformer-based machine learning models as taught by Wu et al. This would have been an obvious improvement Kadatoka et al. already says machine learning techniques can perform the method and machine learning models can be trained to provide fast and high-quality categorization data (Wu et al. Col. 3, Lines 53-67).
Regarding Claims 9 and 18, Kadatoka et al. in view of Liberty et al, Tong et al, and Wu et al. teaches the system of claims 1 and 10.
Furthermore, Kadatoka et al. teaches wherein the one or more metadata associated with the one or more electronic mails, are extracted from the one or more databases based on one or more techniques comprising at least one of: data normalization, data anonymization, data aggregation, data analysis, and data storage:
(The mail room intent analyzer 102 may pull the electronic message generated at 202 from the shared mailbox by, for example, an IMAP pull request (step 206).) (Column 4, Lines 47-49).
(The stripping module 108 may receive the plain text from the mail room intent analyzer 102 (step 302), and may extract a message metadata and a message body from the plain text (step 304). In various embodiments, the stripping module 108 may comprise a metadata parser 306 and a body parser 308. … The body parser 308 may comprise an extraction pattern matching algorithm configured to identify sentence fragments of the text based on content and relative position within the text.) (Column 5, Lines 1-24).
(In various embodiments, the system 100 may store the extracted metadata elements (310, 312, 314, 316) and message body elements (318, 320, 322, 324, 326) in database 116.) (Column 5, Lines 49-52).
The system pulls initial messages from a shared mailbox which could be described as a cloud database, it then extracts the metadata and stores it in the database used for the system. The above quotes show how data storage techniques are in use. Furthermore, Fig.2 shows database information going into an analysis bot which shows the concept of database analysis.
and wherein the one or more databases comprises at least one of: one or more relational databases, one or more object-oriented databases, one or more data warehouses, and one or more cloud-based databases.
(Any databases discussed herein may include relational, hierarchical, graphical, blockchain, object-oriented structure, and/or any other database configurations.) (Column 17, Lines 28-30).
Both relations and object-orientated databases are listed here.
Claims 2 and 11 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view of US Patent Publication US 8856249 B2 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), US Patent Publication US 11914630 B2 (Wu et al.), and further in view of “Stably Extracting Text Contents from Email Messages with Python” (Sun).
Regarding Claims 2 and 11, Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. teaches the system of claims 1 and 10. wherein preprocessing the one or more last received electronic mails comprises: identifying. by the one or more hardware processors, the one or more last received electronic mails by analyzing one or more timestamps associated with the one or more electronic mails based on an electronic mail identification process;
(For example, in one embodiment, the email server 122 determines the time that the email message was sent and compares this time to the time associated with the previous email that may have caused this email. If the difference in time period is above a given threshold, then this email message is probably not related to the previous email message.) (Column 7, Lines 55-61).
Liberty et al. uses timestamps to associate an email with a previously received email
detecting, by the one or more hardware processors, the one or more first textual contents within the one or more electronic mail files by applying a regular expression text extraction process.
(The body parser 308 may comprise an extraction pattern matching algorithm configured to identify sentence fragments of the text based on content and relative position within the text. For example, a body text may read: Hello Jane, I would like to update my bank account number from 123 to 456. Thanks, Tom. DISCLAIMER—emails may contain viruses, only open emails from a trusted source.) (Column 5, Lines 21-31).
Kadatoka et al. parses the body of an email (first textual contents) and extracts expression.
and extracting, by the one or more hardware processors. one or more sender information from the one or more electronic mail files to identify and remove one or more second textual contents associated with the one or more electronic mail senders.
(The body parser 308 may extract any text between the greeting 318 and the signature 322, for example, “I would like to update my bank account number from 123 to 456” as the body meat 320.) (Column 5, Lines 38-41)
In Kadatoka et al. the body parser removes greetings and signatures, which represent second textual information, and leaves the body meat, which represents the first textual information.
wherein the one or more second textual contents associated with the one or more electronic mail senders comprise at least one of: greetings, salutations. and signatures.
(The body parser 308 may extract any text between the greeting 318 and the signature 322, for example, “I would like to update my bank account number from 123 to 456” as the body meat 320.) (Column 5, Lines 38-41)
In Kadatoka et al. greetings and signatures are listed as parts of the email that are not in the first textual contents.
Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. does not explicitly teach: wherein the regular expression text extraction process is configured to be applied on one or more electronic mail formatting settings for detecting at least one of: one or more plain textual contents and one or more hypertext markup language (HTML) embedded contents: converting, by the one or more hardware processors, the one or more hypertext markup language (HTML) embedded contents to the one or more plain textual contents by parsing the one or more hypertext markup language (HTML) embedded contents based on an hypertext markup language (HTML) parsing process:
However, Sun teaches a wherein the regular expression text extraction process is configured to be applied on one or more electronic mail formatting settings for detecting at least one of: one or more plain textual contents and one or more hypertext markup language (HTML) embedded contents:
(Accurately extracting text content: Complying with HTML rules as strictly as possible to strip HTML presentation information as much as possible and leave text content as much as possible) (Section 3, Paragraphs 1-2).
Suns method of is able to detect HTML and plain text in emails in order to do the conversion.
converting, by the one or more hardware processors, the one or more hypertext markup language (HTML) embedded contents to the one or more plain textual contents by parsing the one or more hypertext markup language (HTML) embedded contents based on an hypertext markup language (HTML) parsing process:
(Accurately extracting text content: Complying with HTML rules as strictly as possible to strip HTML presentation information as much as possible and leave text content as much as possible) (Section 3, Paragraphs 1-2).
Sun shows a method of extracting plain text from and HTML format.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. to convert HTML to plain text as taught by Sun. This would have been an obvious improvement as HTML is common format for email presentation (Sun, Section 1, Paragraph 6).
Claims 4, 8, 13 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view of US Patent Publication US 8856249 B2 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), US Patent Publication US 11914630 B2 (Wu et al.), and further in view of US Patent Publication US 11481734 B2 (Venkatasubramanian et al.).
Regarding Claims 4 and 13, Kadatoka et al. in view of Liberty et al, Tong et al, and Wu et al. teaches the system of claims 1 and 10.
Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. does not explicitly teach: Wherein the first machine learning model is fine-tuned for categorizing the one or more electronic mails into one or more distinct intent categories: the first machine learning model is optimized to categorize the one or more electronic mails into one or more distinct intent categories by selecting one or more hyperparameters comprising at least one of: learning rates, batch sizes, and regularization strength; and the first machine learning model is trained to learn from the one or more electronic mails being labelled under the one or more distinct intent categories.
However, Venkatasubramanian et al. teaches a Wherein the first machine learning model is fine-tuned for categorizing the one or more electronic mails into one or more distinct intent categories:
(The user interface allows the label to be selected and changed. The system may then transmit the changed label and corresponding email as feedback to the machine learning classifier to retrain the machine learning classifier.) (Column 11, Lines 26-30)
Re-training the MLM is a form of fine-tuning it.
the first machine learning model is optimized to categorize the one or more electronic mails into one or more distinct intent categories by selecting one or more hyperparameters comprising at least one of: learning rates, batch sizes, and regularization strength; and
(The following parameter values were selected for hyperparameter tuning for regularization with 5-fold cross-validation after a grid search over a range of values: eta=0.2, gamma=6, max_depth=3, min_child_weight=1, subsample=0.5, colsample_bytree=0.5, nrounds=92) (Column 6, Lines 29-38).
An eta can be considered a learning rate as it is specifying the timeframe for the training. Also, the nrounds can be considered a batch size as it is an alternative method of limiting the amount of training that occurs.
the first machine learning model is trained to learn from the one or more electronic mails being labelled under the one or more distinct intent categories.
(The user interface allows the label to be selected and changed. The system may then transmit the changed label and corresponding email as feedback to the machine learning classifier to retrain the machine learning classifier.) (Column 11, Lines 26-30)
Labelled emails are used as training data for the retraining. It can be seen in Fig. 4 element 480 that emails either receive a label of litigious or non-litigious which can be considered distinct intent categories in regards to this invention.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. to include a method of retraining the models as taught by Venkatasubramanian et al. This would have been an obvious improvement to improve the model based on user feedback (Venkatasubramanian et al. Column 10, Lines 25-37).
Regarding Claims 8 and 17, Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. teaches the system of claims 1 and 10.
Furthermore, Venkatasubramanian et al. teaches further comprising re-training, by the one or more hardware processors, one or more machine learning models comprising at least one of:
(The ensemble model 130 may also continuously learn from user-feedback that helps to validate results, which is then fed back into the system for retraining.) (Column 9, Lines 34-36)
In Venkatasubramanian et al. re-training occurs on the model.
the first machine leaning model, the second machine teaming model, and the third machine learning model, over a plurality of time intervals based on one or more training data,
(The ensemble model 130 may also continuously learn from user-feedback that helps to validate results, which is then fed back into the system for retraining.) (Column 9, Lines 34-36)
In Venkatasubramanian et al, continuous learning based on feeding results back into the system occurs. This is equivalent to retraining over a plurality of intervals.
wherein re-training the one or more machine learning models over the plurality of time intervals comprises: receiving, by the one or more hardware processors, the one or more training data associated with the one or more categorized electronic mails, from an output subsystem;
(The user interface allows the label to be selected and changed. The system may then transmit the changed label and corresponding email as feedback to the machine learning classifier to retrain the machine learning classifier.) (Column 11, Lines 26-30)
In Venkatasubramanian et al. after classifying emails, they are output back into the machine learning classifier for further training.
adding, by the one or more hardware processors, the one or more training data with one or more original training datasets to generate one or more updated training datasets.
(A continuously learning process is implemented to retrain the ensemble model 130 with the new feedback data that changes previous labels. The ensemble model 130 receives the label changes and other feedback data as input to be combined and retrained with the existing training dataset of classified data (block 155). This feedback data 155 is used to re-train the ensemble model 130 with the previous and newly labeled correspondence bodies. The retrained ensemble model 130 will replace the existing model if the retrained model outperforms the existing model.) (Column 11, Lines 46-55).
In Venkatasubramanian et al, new results are combined with previous training data when re-training.
re-training, by the one or more hardware processors, the one or more machine learning models, by adjusting at least one of: the one or more hyperparameters and one or more training configurations of at least one of: an electronic mail intent analyzing subsystem, an electronic mail reason code classification subsystem, and a named entity recognition subsystem:
(The following parameter values were selected for hyperparameter tuning for regularization with 5-fold cross-validation after a grid search over a range of values: eta=0.2, gamma=6, max_depth=3, min_child_weight=1, subsample=0.5, colsample_bytree=0.5, nrounds=92) (Column 6, Lines 29-38).
In Venkatasubramanian et al. Hyperparameter tuning occurs in the retraining process. Furthermore, in the primary reference of Kadatoka et al. these subsystems exist and are performed by machine learning models, see Fig. 7 of Kadatoka et al. for evidence of a named entity recognition subsystem, intent analyzing subsystem, and reason code classification subsystem.
and executing, by the one or more hardware processors, the re-trained one or more machine learning models in at least one of: the electronic mail intent analyzing subsystem. the electronic mail reason code classification subsystem, and the named entity recognition subsystem, to output the one or more categorized electronic mails
(The retrained ensemble model 130 will replace the existing model if the retrained model outperforms the existing model. This may be based on executing a number of comparison tests to determine the model's accuracy in predictions. Using this feedback mechanism, the risk detection system 100 will learn to classify correspondences more accurately over a period of time.) (Column 11, Lines 53-60).
In Venkatasubramanian et al, by replacing the existing model and doing this process continuously the re-trained model is executed on the system. In the primary reference Kadatoka et al. the above-mentioned subsystems exist. See Fig. 7 for the Analysis module and ranking module which is equivalent to the electronic mail intent analyzing subsystem. See Figs. 6-7 for an equivalent to the reason code classification subsystem in the form of helper tags. See Figs. 5 and 7 for the Entity Extractor module which is equivalent to the named entity recognition subsystem.
Claims 5 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view of US Patent Publication US 8856249 B2 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), US Patent Publication US 11914630 B2 (Wu et al.), and further in view of US Patent Application Publication US 20240203404 A1 (Shabat et al.).
Regarding Claims 5 and 14, Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. teaches the system of claims 1 and 10. further comprising generating, by the first machine learning model, one or more confidence scores for each intent analyzed for the one or more electronic mails by providing (probability distribution) (Taught by Rice et al.) over the one or more tokens based on a (softmax function) (Taught by Rice et al.),
(Step 718 may include calculating a Term Frequency-Inverse Document Frequency (TFIDF) score for each element of the match set 714, calculating a proximity score for each element of the match set 717, or calculating an accuracy score for each element of the match set 714. In various embodiments, the rank of each element may be determined based on a proximity score cutoff, an accuracy score cutoff, the TFIDF score, or a combination thereof. Step 718 may include calculating a classification score for each element of the ranked match set 720.) (Column 8, Lines 38-47)
In Kadatoka et al. it can be seen in Fig. 7 that a form of confidence score is calculated on the analyzed intent at the end of the system.
wherein the one or more tokens with (optimum probability) (Taught by Rice et al.) are selected as the analyzed intent of the one or more electronic mails.
(The analysis module 114 may apply a document matching algorithm to the word vectors (706, 708, 710, 712) to generate a match set 714 between the word vectors (step 716).) (Column 8, Lines 29-32)
(Where the classification score is greater than the classification score cutoff, the system may retrieve the associated case type for the highest ranked element of the ranked match set 720, and the case generation engine 110 may generate the new case event in case management system 810 having the associated case type (step 814).) (Column 9, Lines 13-19).
In Kadatoka et al. a match set refers to the possible case types being combined with the vectorized entities and events; thus, the match set represents the possible analyzed intents of the system. Then a calculation is performed and the highest ranked set is the analyzed intent.
Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. does not explicitly teach: providing probability distribution over the one or more tokens based on a softmax function.
wherein the one or more tokens with optimum probability are selected
However, Shabat et al. teaches the use of a probability distribution, a softmax function, and optimum probabilities.
(Put another way, the softmax module 350 can process the aggregated encoding to provide a probability distribution over a plurality of outcomes. For instance, the probability distribution may be over a plurality of possible embeddings (e.g. indicative of intents). In this way, the output of the softmax module 350 can be indicative of one or more intent 360 corresponding to the spoken utterance.) (Paragraph 57)
In Shabat et al. it can be seen in the above quote and Fig.3 that a softmax function is used for a probability distribution with the goal of finding intent among a plurality of possible intents.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. to calculate probabilities of possible intent classifications using a softmax function as taught by Kumar et al. This would have been an obvious replacement of steps as the softmax function and associated probability distribution is a method of choosing among a plurality of possibilities (Shabat et al. Paragraph 57).
Claims 6 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view US Patent Publication US 8856249 B2 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), US Patent Publication US 11914630 B2 (Wu et al.), and further in view of US Patent Application Publication US 20220207483 A1 (Agarwal et al.).
Regarding Claims 6 and 15, Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. teaches the system of claims 1 and 10. wherein classifying, by the second machine leaning model, the one or more electronic mails into the one or more reason codes comprises: obtaining, by the one or more hardware processors, the one or more first textual contents;
(The stripping module 108 may receive the plain text from the mail room intent analyzer 102 (step 302), and may extract a message metadata and a message body from the plain text (step 304).) (Column 5, Lines 2-6)
In Kadatoka et al. the message body represents the first textual contents.
converting, by the one or more hardware processors, the one or more first textual contents into the one or more tokens, based on the tokenization process:
(the system may tokenize the body meat into a plurality of tokenized sentences via a tokenizer subprocess.) (Column 1, Lines 55-56)
In Kadatoka et al. the body text, which represents the first textual contents, is tokenized.
extracting, by the one or more hardware processors, one or more first linguistic information associated with the one or more tokens corresponding to the one or more first textual contents:
(The tagging algorithm may generate the tagged sentence set 512′ based on a custom defined electronic message corpus (e.g., all electronic messages received at the shared mailbox) and a standard language corpus such as, for example, the Brown University Standard Corpus of Present-Day American English. The tagged sentence set 512′ comprises tokenized sentences wherein each word is tagged and associated with a part-of-speech attribute.) (Column 6, Lines 47-55)
In Kadatoka et al. linguistic information in the form of part-of-speech attributes is extracted from the tokens.
generating, by the one or more hardware processors, one or more contextually relevant embeddings capturing intricate semantics related to the one or more electronic mails by applying the extracted one or more first linguistic information on the first machine learning model:
(Process 500 includes running an event of interest extraction subprocess to extract an event set from the tagged sentence set 512′ (step 516).) (Column 7, Lines 3-9)
(This is achieved by mapping words into a meaningful space where the distance between words is related to semantic similarity) to any of the possible case type set 608, the set of valid events 522, and/or the named entities 510 to generate respective word vectors (step 704).) (Column 8, Lines 20-26)
In Kadatoka et al. the events and named entities have a specific meaning related to the email and are mapped together based on a semantic relationship found by the system.
assigning, by the one or more hardware processors, one or more scores to the one or more reason codes based on at least one of: the one or more contextually relevant embeddings and the one or more first linguistic information associated with the one or more electronic mails: and
(each case type may be associated with a set of helper tags. The helper tags comprise a natural language text string which may be descriptive of the case type. … For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like.) (Column 8, Lines 5-12).
(Step 716 may include combining the events word vector 706, the entities word vector 708, and the case types word vector 710 into a combined word vector and matching between the combined word vector and the helper tags word vector 712. The analysis module may apply a ranking algorithm to the match set 714 to generate a ranked match set 720 (step 718).) (Column 8, Lines 32-38).
In Kadatoka et al. the reason codes are represented by the helper tags. Fig. 7 shows how the helper tags are assigned to combinations of the event and entity word vectors. Those pairings are then scored in step 718 of Fig. 7.
classifying, by the one or more hardware processors, the one or more reason codes for the one or more electronic mails based on one or more optimum scores assigned to the one or more reason codes.
(Where the classification score is greater than the classification score cutoff, the system may retrieve the associated case type for the highest ranked element of the ranked match set 720,) (Column 9, Lines 13-16).
In Kadatoka et al. the decision to retrieve a case type based on the highest scoring match set is equivalent to assigning an intent based on the score given to a reason code.
wherein the one or more reason codes comprise at least one of: one or more product reason codes, one or more service reason codes, one or more customer reason codes, one or more account reason codes, one or more transaction reason codes, one or more location reason codes, one or more payment reason codes.
(The helper tags comprise a natural language text string which may be descriptive of the case type. In various embodiments, the case type may be a text string. For example, a first case type may be “Account Changes” and associated with helper tags, such as “update my account,” “change account,” “account changes,” “revise account information,” and/or the like.) (Column 8, Lines 6-14).
In Kadatoka et al. the example helper tags here can be described as account reason codes as they are all various intents related to a user account.
Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. does not explicitly teach: applying, by the one or more hardware processors, entity masking process on the one or more first textual contents to normalize one or more words essential in the one or more first textual contents: applying, by the one or more hardware processors, at least one of: a lemmatization process and a stopword process to refine the one or more first textual contents;
However, Agarwal et al. teaches applying, by the one or more hardware processors, entity masking process on the one or more first textual contents to normalize one or more words essential in the one or more first textual contents:
(For instance, an email with full content “the reporter went to the reporter meeting and met another reporter” will have T(“reporter”)=3. Tokens may be obtained, for example, by indexing each email and removing stop words (i.e., common English words such as “the” that are not semantically meaningful) and purely numerical tokens.) (Paragraph 112).
In this case the entity being masked is “reporter” as it is being singled out as an essential word to the email’s contents.
applying, by the one or more hardware processors, at least one of: a lemmatization process and a stopword process to refine the one or more first textual contents;
(For instance, an email with full content “the reporter went to the reporter meeting and met another reporter” will have T(“reporter”)=3. Tokens may be obtained, for example, by indexing each email and removing stopwords (i.e., common English words such as “the” that are not semantically meaningful) and purely numerical tokens.) (Paragraph 112)
The email contents are refined using a stopword process.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification method as taught by Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. to normalize the text using a stopword method as taught by Agarwal et al. This would have been an obvious improvement help the tokenization process by removing common words that can be assumed to be insignificant (Agarwal et al. Paragraph 112).
Claims 7, 16, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over US Patent Publication US 11200267 B2 (Kadatoka et al.) in view US Patent Publication US 8856249 B2 (Liberty et al.), US Patent Application Publication US 20240062016 A1 (Tong et al.), US Patent Publication US 11914630 B2 (Wu et al.), and further in view of US Patent Application Publication US 20240062011 A1 (Kanuga et al.).
Regarding Claims 7, 16, and 20, Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. teaches the system of claims 1, 10, and 19.
wherein extracting, by the third machine learning model, the information associated with the one or more named entities comprises. obtaining, by the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails:
(The body parser 308 may extract any text between the greeting 318 and the signature 322, for example, “I would like to update my bank account number from 123 to 456” as the body meat 320.) (Column 5, Lines 38-41)
(The system may extract a set of named entities 510 from the body meat 320 via an entity extractor subprocess (step 508). Step 508 includes searching the body meat 320 for entities such as, for example, a merchant ID, a merchant name, a user name, a place name, and/or the like.) (Column 7, Lines 41-46).
In Kadatoka et al. the body meat represents the unstructured text of the email and named entities are extracted from this portion.
scanning, be the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails for extracting the one or more named entities:
(The system may extract a set of named entities 510 from the body meat 320 via an entity extractor subprocess (step 508). Step 508 includes searching the body meat 320 for entities such as, for example, a merchant ID, a merchant name, a user name, a place name, and/or the like.) (Column 7, Lines 41-46).
In Kadatoka et al. searching for named entities is equivalent to scanning.
applying, by the one or more hardware processors, the textual information associated with the one or more unstructured and unlabeled electronic mails on the third machine learning model: and
In Kadatoka et al. it can be seen in Fig. 3 that the unstructured emails are fed into an analysis bot which represents a third MLM. The third MLM will be further addressed below.
extracting, by the one or more hardware processors, one or more patterns indicating the one or more named entities by leveraging one or more second linguistic information associated with the one or more unstructured and unlabeled electronic mails,
(Where the subject 316 does not contain a case management system record ID, the system may pass the extracted metadata elements and body data elements to the parsing module 112 and/or analysis module 114 for further processing (step 408).) (Column 6, Lines 3-7).
(Step 508 includes searching the body meat 320 for entities such as, for example, a merchant ID, a merchant name, a user name, a place name, and/or the like.) (Column 7, Lines 43-46)
In Kadatoka et al. the body meat represents the unstructured text and the merchant ID, name, etc. represent the second linguistic information. Thus, patterns from the second linguistic information are found in the unstructured text when extracting named entities.
wherein the information associated with the one or more named entities comprises at least one of: names of one or more electronic mail senders, one or more organizations. one or more dates, one or more times. one or more invoice numbers, and one or more amounts.
(Step 508 includes searching the body meat 320 for entities such as, for example, a merchant ID, a merchant name, a user name, a place name, and/or the like.) (Column 7, Lines 43-46)
In Kadatoka et al. mail senders and organizations are listed as examples of entities in the form of merchant ID, merchant name, or a user name.
Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. does not explicitly teach: wherein the third machine learning model is an integration of at least one of: the first machine learning model and the second machine learning model. and
However, Kanuga et al. teaches a wherein the third machine learning model is an integration of at least one of: the first machine learning model and the second machine learning model. and
(The training stage 705 builds and trains one or more machine learning models 730a-730n (‘n’ represents any natural number) to be used by the other stages (which may be referred to herein individually as a model 730 or collectively as the models 730). For example, the models 730 can include a first model for translating a natural language text to a logical form, a second model for recognizing and extracting named entities from a natural language text, and third model for translating the logical form to a particular system query language.) (Paragraph 136)
As the third model is generated to handle the output of the first model it can be described as an integration of the two.
It would have been obvious to a person having ordinary skill in the art before the time of the effective filing date of the claimed invention of the instant application to modify the email classification system as taught by Kadatoka et al. in view of Liberty et al., Tong et al., and Wu et al. to integrate aspects of earlier models in the system into later models as taught by Kanuga et al. This would have been an obvious improvement to avoid training downstream models from scratch (Kanuga et al. Paragraph 124).
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 NICHOLAS DANIEL LOWEN whose telephone number is (571)272-5828. The examiner can normally be reached Mon-Fri 8:00am - 4:00pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Paras D Shah can be reached at (571) 270-1650. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/NICHOLAS D LOWEN/Examiner, Art Unit 2653
/Paras D Shah/Supervisory Patent Examiner, Art Unit 2653
02/20/2026