DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claims 1-20 are pending and have been examined. Claims 1, 10, and 19 are independent.
Apparent priority: April 1, 2025.
Specification
The disclosure is objected to because of the following informalities:
a. In paragraph [0005], “FIG. 3 illustrates a machine learning engine for training and execution related to in accordance with some examples” should apparently read “FIG. 3 illustrates a machine learning engine for training and execution related to predicting customer complaints in electronic communications in accordance with some examples.”
b. In paragraph [0021], “training and execution related to determining an attack surface for a malicious attack on an application” should apparently read “training and execution related to predicting customer complaints in electronic communications.”
c. In paragraph [0028], “C9.5” should apparently read “C4.5.”
d. In paragraphs [0046], [0055], [0064], and [0066], “includes, a customer complaint” should apparently read “includes a customer complaint.”
e. In paragraphs [0069]-[0072], “to implement of any of Examples 1-23” should apparently read “to implement any of Examples 1-23.”
Appropriate correction is required.
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 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1 (Statutory Category). Claims 1-9 are directed to a method (a process), claims 10-18 are directed to a system (a machine), and claims 19 and 20 are directed to at least one non-transitory machine readable medium (a manufacture). Claims 1-20 are therefore directed to a statutory category of invention.
Step 2A, Prong One (Recitation of a Judicial Exception). Claim 1 recites receiving a plurality of electronic mail messages; preprocessing the messages to create structured text data of each message; tokenizing the structured text data of each message to generate a set of tokens corresponding to each word; determining a term frequency within each message of respective tokens; determining an inverse document frequency of each token; generating a uniqueness score for each token based on the term frequency and the inverse document frequency; and training a machine learning model to predict whether an electronic mail message includes a customer complaint using the uniqueness score for each token as an input. The determining and generating limitations are mathematical calculations: the specification computes the term frequency as the number of times a word appears in a document divided by the total number of words in it, the inverse document frequency as a log of the number of documents in the corpus divided by the number of documents including the word, and the score as their product (paragraph [0009], Equation 1). The preprocessing, tokenizing, and predicting limitations, as drafted, cover activities a person performs with pen and paper. Restated as a person would carry it out: a reviewer collects a stack of customer emails, strikes out the parts of each that are not the customer's own words, lists the words of each email, counts how often each word appears in each email and in how many of the emails it appears, works out a score for each word from those two counts, and uses the scored words to judge whether a new email contains a complaint. Claim 1 therefore recites mathematical concepts and a mental process.
Step 2A, Prong Two (Integration into a Practical Application). The judicial exception is not integrated into a practical application. Beyond the calculations and mental steps, claim 1 recites receiving the electronic mail messages and training a machine learning model on the scores. Receiving the messages is data gathering. Training is recited by its result alone, with no architecture, training objective, or update rule, and the specification describes the model as any of a list of generic supervised learning algorithms (paragraph [0028]; see also paragraph [0021], listing Random Forest, AdaBoost, XGBoost, LightGBM, CatBoost, Gradient Boosting, and Histogram-Based Gradient Boosting as example model types). The improvement the specification describes, fewer false positives and false negatives in complaint predictions (paragraph [0013]), is an improvement in the result of the abstract analysis, not in the functioning of a computer or any other technology. The additional elements do not impose any meaningful limits on practicing the abstract idea, and claim 1 is directed to an abstract idea.
Step 2B (Inventive Concept). Claim 1 does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to integration, receiving the messages is insignificant extra-solution activity and the training of a machine learning model is recited generically. The specification itself describes TF-IDF as “a statistical technique used in natural language processing” (paragraph [0008]) and the model types as interchangeable, conventional choices (paragraphs [0021] and [0028]). Mere instructions to apply an exception using generic machine learning cannot provide an inventive concept. Claim 1 is not patent eligible.
Independent claims 10 and 19. Claims 10 and 19 recite the operations of claim 1 in system and machine readable medium form, respectively. Beyond those operations, claim 10 adds one or more hardware processors and at least one memory storing instructions, and claim 19 adds a non-transitory machine readable medium and processing circuitry. The specification describes these as generic computing components, a machine 500 that may be “any machine capable of executing instructions” (paragraph [0038]) with a hardware processor 502 and memories 504 and 506 (paragraph [0040]), and a machine readable medium that may include any medium capable of storing instructions (paragraph [0043]). Implementing the abstract idea on generic computer components neither integrates it into a practical application nor amounts to significantly more. Claims 10 and 19 are not patent eligible for the reasons given for claim 1.
Dependent claims. The dependent claims do not cure the deficiency. Claims 2, 11, and 20 and claims 3 and 12 recite what is removed from the text of each message in preprocessing, which a person does by striking the material out. Claims 4 and 13 recite grouping equivalent words under common tokens and normalizing acronym and company-name variations, which a person does by treating the variants as one word when listing and counting. Claims 5 and 14 recite multiplying the term frequency by the inverse document frequency, a further mathematical calculation. Claims 6 and 15 recite that the model is a gradient boosting model, itself a mathematical technique for combining weak learners, and claims 7 and 16 recite tuning its hyperparameters by grid search, a further mathematical procedure of trying combinations of values and keeping the best. Claims 8 and 17 and claims 9 and 18 recite generating a prediction that a later email includes a complaint or has a risk level, which is the evaluation a person makes on reading the email. None of these adds an element that integrates the exception into a practical application or amounts to significantly more. Claims 1-20 are not patent eligible.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 5, 8-10, 14, and 17-19 are rejected under 35 U.S.C. 103 as being unpatentable over Mann (U.S. 11,790,411), hereinafter Mann, in view of Yadav (U.S. 11,087,179), hereinafter Yadav.
The recited inverse document frequency “in the structured text data of each token” is read as the inverse document frequency of each token across the structured text data of the received plurality of messages, consistent with the specification, which computes the inverse document frequency against a set of stored emails (paragraph [0015]) as a log of the number of documents in the corpus divided by the number of documents including the word (paragraph [0009]). The recited uniqueness score is read as the value computed for a token from its term frequency and inverse document frequency, which the specification states gives an indication of uniqueness of the word within the document and corpus (paragraph [0009]).
Regarding claim 1, Mann discloses:
1. A method for predicting customer complaints in electronic communications, comprising:
“In general, this disclosure describes techniques for performing complaint identification and classification using a machine learning model for customer communications associated with a business or organization.” (Mann, col. 1, ll. 33-36.)
receiving a plurality of electronic mail messages;
“Communication analysis system 22 may receive a plurality of messages (402) from message routing system 20.” (Mann, col. 17, ll. 42-43.)
“Additionally, or alternatively, message center 12 may allow customers to submit messages (e.g., service complaints) via text channels such as email, text messaging, and social media messaging.” (Mann, col. 3, ll. 40-44.)
preprocessing the plurality of electronic mail messages to create structured text data of each message of the plurality of electronic mail messages;
“communication analysis system 200 may parse the data indicative of the incoming message to identify a sequence of items including any one or combination of words, phrases, characters (e.g., punctuation), and numerical values.” (Mann, col. 8, ll. 54-58.)
The sequence of items Mann parses from the data of each message, its words, phrases, characters, and numerical values in order, is the structured text data of that message.
tokenizing the structured text data of each message to generate a set of tokens corresponding to each word in the structured text data of each message;
“Token unit 212 may determine, for each word of the sequence of words, a numerical token corresponding to the respective word.” (Mann, col. 13, ll. 15-17.)
“communication analysis system 22 may identify a set of words in the string of characters and look up a respective token corresponding to each word of the set of words in a token dictionary stored in a memory.” (Mann, col. 7, ll. 1-4.)
training a machine learning model to predict whether an electronic mail message includes a customer complaint using the uniqueness score for each token as an input.
“Training unit 230 may train machine learning model 210 using training data 209 (504).” (Mann, col. 18, ll. 53-54.)
“At least some of the plurality of messages may represent customer complaints submitted to message center 209.” (Mann, col. 10, ll. 47-49.)
“The plurality of messages may include a group of high-risk messages and a group of low-risk messages, where each message of the plurality of messages is either known to be high-risk or known to be low-risk.” (Mann, col. 10, ll. 49-52.)
“A machine learning algorithm or function (e.g., a word embedding algorithm) is trained to create machine learning model 210 configured to accept an input sequence of tokens associated with a message and produce, using complaint classification unit 220, any one or combination of an output classification of whether the incoming message is high-risk or low-risk” (Mann, col. 11, ll. 52-58.)
“Risk unit 224 may be configured to identify a probability that an incoming message is a high-risk message based on the set of token vectors that are identified using the message data.” (Mann, col. 13, ll. 58-61.)
“Elevating a high-risk message may include flagging the high-risk message as including a rather urgent, grave, or serious complaint or criticism of one or more services or aspects of the business or organization that provides message center 12.” (Mann, col. 15, ll. 63-67.)
Mann’s training of machine learning model 210, on messages each known to be high-risk or low-risk, to output whether an incoming message is high-risk reads on training a machine learning model to predict whether an electronic mail message includes a customer complaint.
Mann does not teach the term-frequency, inverse-document-frequency, and uniqueness-score aspects of the claim.
Yadav teaches:
determining a term frequency within each message of respective tokens of the set of tokens; (“Term frequency (tf) measures how frequently a term occurs in a document. This can either be as a count or as a proportion of the words in a document. When the proportion approach is used, the count for a term is divided by the total count of words in the document.” (Yadav, col. 6, ll. 59-64).)
Yadav’s terms are the words of each document (Yadav, col. 6, ll. 47-52), the units Mann tokenizes, so Yadav’s term frequency taken over the tokens of each message is the recited term frequency within each message of respective tokens.
determining an inverse document frequency in the structured text data of each token in the set of tokens; (“Inverse document frequency (idf) measures the discriminating power of a term.” (Yadav, col. 6, ll. 64-65); “can be logarithmically scaled by taking a log of the total number of documents (in the universe being considered) divided by the number of documents with term” (Yadav, col. 7, ll. 1-4).)
generating a uniqueness score for each token in the set of tokens based on the term frequency and the inverse document frequency; and (“The tf-idf weight is a product of two terms: term frequency (tf) and inverse document or domain frequency (idf).” (Yadav, col. 6, ll. 55-56); “The importance increases proportionally to the number of times a word appears in the document but is offset by the frequency of the word in the corpus.” (Yadav, col. 6, ll. 49-52).)
training a machine learning model to predict whether an electronic mail message includes a customer complaint using the uniqueness score for each token as an input. (“The raw document data is converted to document features for input to the multi-label document classifier 161. An example of document features is frequency features based on term frequency-inverse document frequency (TF-IDF).” (Yadav, col. 5, ll. 59-63); “For each combination of hyper parameter values, the multi-label document classifier 161 is trained using labels from label classes databases 159 and document features in the document features labeled database 331.” (Yadav, col. 8, ll. 13-17).)
Mann and Yadav pertain to classification of text documents by trained machine learning models using features computed from the words of the documents. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify Mann to determine, for each token of each message, the term frequency, the inverse document frequency across the received messages, and their product as taught by Yadav, and to add that tf-idf weight of each token to the inputs on which Mann’s machine learning model is trained, alongside Mann’s token vectors. One of ordinary skill would have been motivated to make this modification in order to improve the accuracy with which Mann’s trained model classifies each message as high-risk or low-risk, because the tf-idf weight measures how important each word is to its message within the collection (Yadav, col. 6, ll. 47-52) and gives the classifier the discriminating power of each word: Yadav explains that words appearing in virtually every document “have little discriminating power for classification,” so that idf down scales them and up scales the rare words that set one document apart from the others (Yadav, col. 6, l. 64 to col. 7, l. 1).
Regarding claim 5, Mann does not teach the recited multiplication. Yadav teaches:
5. The method of claim 1, wherein generating the uniqueness score comprises multiplying the term frequency and the inverse document frequency to generate a term frequency inverse document frequency (TF-IDF) score for each token.
“The tf-idf weight is a product of two terms: term frequency (tf) and inverse document or domain frequency (idf).” (Yadav, col. 6, ll. 55-56.)
The rationale for the combination is similar to the one provided for claim 1.
Regarding claim 8, Mann discloses:
8. The method of claim 1, further comprising generating, using the machine learning model, a prediction identifying that an electronic mail message in a second plurality of electronic mail messages includes a customer complaint.
“Communication analysis system 22 may classify, using a machine learning model (e.g., machine learning model 210 of FIG. 2), a risk level of each message of the plurality of messages (404).” (Mann, col. 17, ll. 45-48.)
“Risk unit 224 may be configured to identify a probability that an incoming message is a high-risk message based on the set of token vectors that are identified using the message data.” (Mann, col. 13, ll. 58-61.)
“Elevating a high-risk message may include flagging the high-risk message as including a rather urgent, grave, or serious complaint or criticism of one or more services or aspects of the business or organization that provides message center 12.” (Mann, col. 15, ll. 63-67.)
“A machine learning algorithm or function (e.g., a word embedding algorithm) is trained to create machine learning model 210 configured to accept an input sequence of tokens associated with a message and produce, using complaint classification unit 220, any one or combination of an output classification of whether the incoming message is high-risk or low-risk” (Mann, col. 11, ll. 52-58.)
The messages Mann receives and classifies with the trained model (col. 17, ll. 42-48) are a second plurality of electronic mail messages, and Mann’s output classification by trained model 210 that such a message is high-risk reads on a prediction identifying that an electronic mail message in the second plurality includes a customer complaint.
Regarding claim 9, Mann discloses:
9. The method of claim 1, further comprising generating, using the machine learning model, a prediction identifying a risk level of an electronic mail message in a second plurality of electronic mail messages.
“Communication analysis system 22 may receive a plurality of messages (402) from message routing system 20.” (Mann, col. 17, ll. 42-43.)
“Communication analysis system 22 may classify, using a machine learning model (e.g., machine learning model 210 of FIG. 2), a risk level of each message of the plurality of messages (404).” (Mann, col. 17, ll. 45-48.)
Claim 10 is a system claim with limitations corresponding to the limitations of Claim 1 and is rejected under similar rationale. Additionally:
Regarding claim 10, Mann discloses:
10. A system for predicting customer complaints in electronic communications, comprising: one or more hardware processors; and at least one memory storing instructions that, when executed by the one or more hardware processors, cause the system to perform operations comprising: ...
“Processors 202 may include, for example, microprocessors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate array (FPGAs), or equivalent discrete or integrated logic circuitry” (Mann, col. 9, ll. 41-45.)
“Communication analysis system 200 also includes a machine learning model 210, a complaint classification unit 220, and a training unit 230, which may be implemented as program instructions and/or data stored in storage units 206 and executable by processors 202.” (Mann, col. 9, ll. 21-25.)
Processors 202 are the one or more hardware processors, and storage units 206 storing the program instructions executable by processors 202 are the at least one memory storing instructions of the claim.
Claim 14 is a system claim with limitations corresponding to the limitations of Claim 5 and is rejected under a similar rationale.
Claim 17 is a system claim with limitations corresponding to the limitations of Claim 8 and is rejected under a similar rationale.
Claim 18 is a system claim with limitations corresponding to the limitations of Claim 9 and is rejected under a similar rationale.
Claim 19 is a machine readable medium claim with limitations corresponding to the limitations of Claim 1 and is rejected under similar rationale. Additionally:
Regarding claim 19, Mann discloses:
19. At least one non-transitory machine readable medium including instructions for predicting customer complaints in electronic communications, which when executed by processing circuitry, cause the processing circuitry to perform operations comprising: ...
“computer-readable media generally may correspond to (1) tangible computer-readable storage media which is non-transitory” (Mann, col. 19, ll. 18-20.)
“Instructions may be executed by one or more processors, such as one or more DSPs, general purpose microprocessors, ASICs, FPGAs, or other equivalent integrated or discrete logic circuitry” (Mann, col. 19, ll. 53-56.)
The tangible, non-transitory computer-readable storage media storing the instructions are the at least one non-transitory machine readable medium, and the processors and logic circuitry executing them are the processing circuitry of the claim.
Claims 2, 11, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Mann (U.S. 11,790,411) in view of Yadav (U.S. 11,087,179), further in view of Bašaragin et al. (“Improving Customer Service with Automatic Topic Detection in User Emails,” arXiv:2502.19115, Feb. 2025), hereinafter Bašaragin, and further in view of Consumer Financial Protection Bureau, Office of Consumer Response, “Narrative Scrubbing Standard” (Mar. 2015), hereinafter CFPB.
Regarding claim 2, Mann and Yadav do not teach the recited removals. Bašaragin teaches:
2. The method of claim 1, wherein preprocessing the plurality of electronic mail messages comprises, removing: non-business content, spam content, echo content, signature content, automated response content, reply chains, names, phone numbers, social security numbers, references to money, times, email addresses, URLs, dates, and numbers.
“After preprocessing, we filtered the dataset to eliminate irrelevant or noisy emails, such as spam or automated emails (e.g., out-of-office notices), which could negatively impact topic modeling.” (Bašaragin, sec. 3.1, p. 4.)
“Removal of closing phrases, signatures, and disclaimers: Closing remarks (e.g., Best regards) appear across a wide range of emails and typically signify the end of an email and the start of signatures and disclaimers.” (Bašaragin, sec. 3.1, p. 4.)
“with an additional step of detecting and deleting parts of emails that are a reply or forwarded.” (Bašaragin, sec. 5, p. 10.)
“the emails had been anonymized by replacing personal information (e.g., names, locations, organizations) with placeholders indicating their category (e.g., PER, LOC, ORG).” (Bašaragin, sec. 3.1, p. 3.)
“Anonymization placeholder removal: We remove anonymization placeholders from the emails since they appear in all emails.” (Bašaragin, sec. 3.1, p. 4.)
“Punctuation, numbers, and special characters removal: We removed punctuation, individual numbers, and special characters (e.g. symbols and emojis) since they are not informative in search of latent topics in the text.” (Bašaragin, sec. 3.1, p. 3.)
The parts of an email that are a reply or forwarded are the echo content and the reply chain of the email, and the disclaimers removed with the signature are non-business content of the email.
Mann, Yadav, and Bašaragin pertain to the analysis of customer electronic mail text by models built from the words of the messages; Bašaragin prepares customer-service emails for training a model. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the preprocessing of the combination of Mann and Yadav to remove spam and automated emails, signatures and disclaimers, reply and forwarded parts, names, and numbers as taught by Bašaragin. One of ordinary skill would have been motivated to make this modification in order to improve the model trained on the messages by training it only on the informative text of each message, as Bašaragin removes numbers because they are “not informative,” removes closing phrases, signatures, and disclaimers because they appear across a wide range of emails, and filters out spam and automated emails because they “could negatively impact” the model built from the emails (Bašaragin, sec. 3.1, pp. 3-4).
Mann, Yadav, and Bašaragin do not teach removing phone numbers, social security numbers, references to money, times, email addresses, URLs, and dates. CFPB teaches:
phone numbers, social security numbers, references to money, times, email addresses, URLs, dates,
“Location and contact information: Addresses, phone and fax numbers, emails, IP addresses, sensitive websites and URLs,” (CFPB, p. 4.)
“Date and time information: Dates (except for the year), times, and references to specific holidays.” (CFPB, p. 5.)
“Sensitive numbers: Social security, account, credit or debit card, confirmation, case” (CFPB, p. 5.)
“Dollar amounts: We round dollar amounts to prevent others from using exact dollar amounts to identify you.” (CFPB, p. 5.)
“We remove information from your narrative that could be used to identify you.” (CFPB, p. 4.)
Removing the exact dollar amount from each reference to money in a complaint narrative (CFPB) and removing individual numbers (Bašaragin) reads on removing references to money.
Mann, Yadav, Bašaragin, and CFPB pertain to the processing of customer communications, including complaints about financial products, before the text is analyzed; CFPB prepares consumer complaint narratives about financial products for publication. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the preprocessing of the combination of Mann, Yadav, and Bašaragin to remove phone numbers, social security numbers, exact dollar amounts, times, email addresses, URLs, and dates as taught by CFPB. One of ordinary skill would have been motivated to make this modification in order to protect the privacy of the customers whose messages are stored and analyzed, by removing “information from your narrative that could be used to identify you” (CFPB, p. 4), the same benefit for which Bašaragin anonymizes its emails “[t]o comply with data protection regulations and safeguard user privacy” (Bašaragin, sec. 3.1, p. 3).
Claim 11 is a system claim with limitations corresponding to the limitations of Claim 2 and is rejected under a similar rationale.
Claim 20 is a machine readable medium claim with limitations corresponding to the limitations of Claim 2 and is rejected under a similar rationale.
Claims 6, 7, 15, and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Mann (U.S. 11,790,411) in view of Yadav (U.S. 11,087,179), and further in view of Nag (U.S. 2021/0256436), hereinafter Nag.
Regarding claim 6, Mann and Yadav do not teach a gradient boosting model. Nag teaches:
6. The method of claim 1, wherein the machine learning model comprises a gradient boosting model.
“a gradient boosting classifier XGBoost 120, which uses the Gradient Boosting framework;” (Nag, para. [0023].)
Mann, Yadav, and Nag pertain to classification of electronic correspondence by machine learning models trained on features computed from its words. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to use a gradient boosting classifier as taught by Nag as the machine learning model of the combination of Mann and Yadav. One of ordinary skill would have been motivated to make this modification because tree-based ensemble methods such as XGBoost evaluate each feature as a potential splitting feature, “which makes them robust to unimportant/irrelevant variables” (Nag, para. [0053]), an advantage where every token of every message contributes a tf-idf feature.
Regarding claim 7, Mann and Yadav do not teach the recited tuning. Nag teaches:
7. The method of claim 6, further comprising tuning hyperparameters of the gradient boosting model using grid search optimization to improve prediction accuracy.
“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” (Nag, para. [0054].)
“The ‘AUC’ under the ‘ROC curve’ was used as the evaluation metric.” (Nag, para. [0062].)
Selecting the hyperparameter values of the gradient boosting classifier by a grid search evaluated by the area under the ROC curve is tuning the hyperparameters using grid search optimization to improve prediction accuracy.
The rationale for the combination is similar to the one provided for claim 6.
Claim 15 is a system claim with limitations corresponding to the limitations of Claim 6 and is rejected under a similar rationale.
Claim 16 is a system claim with limitations corresponding to the limitations of Claim 7 and is rejected under a similar rationale.
Claims 4 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Mann (U.S. 11,790,411) in view of Yadav (U.S. 11,087,179), further in view of Nag (U.S. 2021/0256436), and further in view of Bonev et al. (“Statistical sentiment analysis performance in Opinum,” arXiv:1303.0446, 2013), hereinafter Bonev.
Regarding claim 4, Mann and Yadav do not teach the recited grouping and normalizing. Nag teaches:
4. The method of claim 1, wherein tokenizing the structured text data comprises: grouping semantically equivalent words under common tokens; and
“If the email includes at least one word with greater than 4 letters, then each word in the email is stemmed to their root word.” (Nag, para. [0100].)
“All remaining words in the text were stemmed to their original root words. Then the sum of term frequencies-inverse document frequencies (TF-IDF scores) of only potentially stemmed litigious words” (Nag, para. [0137].)
Stemming each word to its root word groups the inflected forms of a word, which are semantically equivalent, under the common root token.
Mann, Yadav, and Nag pertain to classification of electronic correspondence by machine learning models trained on features computed from its words. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the tokenizing of the combination of Mann and Yadav to stem each word of each message to its root word as taught by Nag, before the tf-idf values are computed. One of ordinary skill would have been motivated to make this modification in order that each word is counted as one term in whatever inflected form it takes, so that its term frequency and inverse document frequency reflect all of its occurrences rather than being split across its forms, as Nag stems the words of each email to their root words before computing their TF-IDF scores (Nag, para. [0137]).
Mann, Yadav, and Nag do not teach normalizing variations of acronyms and company names through a masking process. Bonev teaches:
normalizing variations of acronyms and company names through a masking process.
“We also replace named entities by wildcards: person_entity, organization_entity and company_entity.” (Bonev, sec. 3, p. 5.)
“For a generalization purpose we make the texts independent of concrete entities. We do make a difference between names of places, people and organizations/companies.” (Bonev, sec. 3.2, pp. 6-7.)
Bonev’s substitution of the single wildcard company_entity for every name of a company, whatever form the name takes in the text, reads on normalizing variations of acronyms and company names through a masking process.
Mann, Yadav, Nag, and Bonev pertain to classification of customer text by models trained on features computed from its words; Bonev classifies customer opinions in the financial field, “including banks, savings accounts, loans, mortgages, investments, credit cards” (Bonev, sec. 4, p. 10). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the tokenizing of the combination of Mann, Yadav, and Nag to replace every name of a company in each message with a common wildcard as taught by Bonev, before the tf-idf values are computed. One of ordinary skill would have been motivated to make this modification in order to reduce the complexity of the training data and to keep the trained model from associating a complaint with a particular company rather than with what the customer says, as Bonev replaces named entities with wildcards to “reduce the complexity of the training corpus” (Bonev, abstract, p. 1) and so that negative constructions are not “associated to concrete entities” (Bonev, sec. 3, p. 5), making the texts “independent of concrete entities” (Bonev, sec. 3.2, p. 6).
Claim 13 is a system claim with limitations corresponding to the limitations of Claim 4 and is rejected under a similar rationale.
Claims 3 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Mann (U.S. 11,790,411) in view of Yadav (U.S. 11,087,179), further in view of Nag (U.S. 2021/0256436), further in view of Chandramouli et al. (U.S. 10,692,016), hereinafter Chandramouli, and further in view of Yu (U.S. 7,644,127).
Regarding claim 3, Mann and Yadav do not teach the recited removals and case conversion. Nag teaches:
3. The method of claim 1, wherein preprocessing the plurality of electronic mail messages comprises, removing: punctuations, acronyms, non-ASCII characters, HTML tags and entities, and excessive white spaces; and
“each email may be cleaned by removal of all non-Latin alphabet characters, html tags, punctuations, numbers and stop words.” (Nag, para. [0100].)
Non-Latin alphabet characters are non-ASCII characters.
converting upper case letters to lower case.
“The correspondence text from each record is cleaned by removal of stop words, punctuations, numbers and HTML tags. Remaining words are stemmed to their root word with all lowercase characters.” (Nag, para. [0035].)
Mann, Yadav, and Nag pertain to classification of electronic correspondence by machine learning models trained on features computed from its words. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the preprocessing of the combination of Mann and Yadav to remove non-Latin alphabet characters, HTML tags, and punctuation from each message and to convert its remaining words to lowercase as taught by Nag. One of ordinary skill would have been motivated to make this modification in order that the tokens, and the term frequency and inverse document frequency computed for them, reflect only the words of each message, with a word counted as the same token whether or not it is capitalized, as Nag cleans each email of these characters and tags before the text is tokenized (Nag, para. [0100]) and keeps the remaining words in lowercase (Nag, para. [0035]).
Mann, Yadav, and Nag do not teach removing acronyms. Chandramouli teaches:
acronyms,
“Examples of stopwords identified by the system 100 are first names (e.g., “Barbara,” “David,” “Alison”), entity-specific acronyms or strings (e.g., ID numbers), or generic, frequently-used words, such as ‘the,’ ‘and,’ ‘a’ and the like. The stopword extraction engine 111 identifies and removes (208) the tokens that are stopwords in the sets of tokens.” (Chandramouli, col. 8, ll. 44-49.)
Mann, Yadav, Nag, and Chandramouli pertain to the processing of customer communications before the text is analyzed; Chandramouli tokenizes customer interaction text and removes stopwords from the tokens (Chandramouli, col. 8, ll. 35-49). It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the preprocessing of the combination of Mann, Yadav, and Nag to remove entity-specific acronyms from each message as taught by Chandramouli. One of ordinary skill would have been motivated to make this modification in order to improve the quality of the message text from which the tokens and their weights are computed, as Chandramouli removes such stopwords because they are “noise” that “add little value to analytics” and are removed “to improve the quality of the unstructured computer text” (Chandramouli, col. 8, ll. 38-42).
Mann, Yadav, Nag, and Chandramouli do not teach removing HTML entities and excessive white spaces. Yu teaches:
HTML tags and entities, and excessive white spaces;
“Other translations are possible during rule pre-processing of a message such as deleting ‘whitespace’ characters (e.g., spaces, tabs, line returns, etc.)” (Yu, col. 5, ll. 7-9.)
“Another process identifies and removes characters or character codes that do not generate visible symbols in a message displayed to a user.” (Yu, col. 2, ll. 18-20.)
“This table has mappings for all the possible ascii art and non-unique ways of representing letters including html entities that can be represented as characters.” (Yu, col. 9, ll. 4-6.)
Yu’s removal of character codes that do not generate visible symbols, which Yu’s character codes include in the form of html entities, reads on removing HTML entities, and Yu’s deletion of whitespace characters reads on removing excessive white spaces.
Mann, Yadav, Nag, Chandramouli, and Yu pertain to preparing the text of electronic mail messages for analysis by a classifier. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to further modify the preprocessing of the combination of Mann, Yadav, Nag, and Chandramouli to remove character codes that do not generate visible symbols and to delete whitespace characters as taught by Yu. One of ordinary skill would have been motivated to make this modification in order that each message is analyzed more efficiently, as Yu translates message text “into a standardized format that is easier to analyze efficiently” (Yu, col. 2, ll. 13-15).
Claim 12 is a system claim with limitations corresponding to the limitations of Claim 3 and is rejected under a similar rationale.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure:
a. Attaran Rezaei et al. (U.S. 2007/0198506): a preprocessing function that removes irrelevant content, noise, and duplicate data from collected content before tokenization, and word frequencies computed after removing common words and word noise and divided by frequency across the corpus.
b. Gelfand et al. (U.S. 10,318,617): extraction of content from an email or email thread for use by recommendation engines, removing replied content, signatures, and markup.
c. Muthuswamy Sivaraman et al. (U.S. 10,984,388): identification of complaints in customer emails, with variations of terms grouped into dictionaries to prepare a training dataset for supervised models.
d. Shi et al. (U.S. 10,956,835): training a gradient boosting tree predictive model with hyperparameters selected by an automatic tuning process that minimizes a validation error estimate.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to YUVAL H. LEVENTAL whose telephone number is (571) 270-3130. The examiner can normally be reached Monday-Friday, 8:00 AM - 5:00 PM.
If attempts to reach the examiner by telephone are unsuccessful, the examiner's supervisor, PIERRE-LOUIS DESIR, can be reached at (571) 272-7799. 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.
/YUVAL HAIM LEVENTAL/
Examiner, Art Unit 2659
/PIERRE LOUIS DESIR/Supervisory Patent Examiner, Art Unit 2659
Prosecution Insights