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 .
Claim Objections
Claim 8 is objected to because of the following informalities: claim 8 recites in line 1, “for the comprising”. Examiner suggests correcting this phrase to “further comprising”. Appropriate correction is required.
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.
Claim(s) 1-2, 5-9, 11-12, and 15-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Studiawan et al. (NPL: Sentiment Analysis in a Forensic Timelines With Deep Learning, published March 2020, hereinafter “Studiawan”) in view of Udupa et al. (US Pub. No. 2023/0039338, published Feb. 2023, hereinafter “Udupa”).
Regarding claim 1, Studiawan teaches a method comprising:
receiving log data (Studiawan, Fig. 2 and Section III Paragraph 1 – “In the training step, we first preprocess the log files to extract messages containing negative or positive sentiments.” – teaches receiving log data);
processing the log data with a text generation model to generate synthesized text (Studiawan, Fig. 2, Section III Subsection D Last Paragraph – “To improve the sentence representation to handle more complex log messages, we add the sentence representation from the embedding layer vs and the sentence representation from the content attention layer vsr , resulting in vf . Finally, vf is the final representation of the aspect, considering the context and sentiment information of the log messages from the previous context attention layer” – teaches processing log data with a text generation model (processes log data with embedding and attention layers of model as in Fig. 2) to generate synthesized text (generates sentence representation vector vf, and in Specification of claimed invention at [0030] – states that synthesized text may be stored as vector data));
processing the synthesized text with a sentiment prediction model to generate a sentiment prediction (Studiawan, Fig. 2 and in Section III Paragraph 1 – “We then use a softmax layer to determine whether a message sentiment is positive or negative. In the last step of the training phase, we save the sentiment model.” – teaches processing the synthesized text (sentiment model processes synthesized text from attention layers) with a sentiment prediction model (as in Fig. 2, softmax layer & sentiment model) to generate a sentiment prediction (determines whether a message sentiment is positive or negative, thus generating a sentiment prediction))
presenting the sentiment prediction (Studiawan, Fig. 2 and in Section IV Subsection E Paragraph 1 – “Finally, the negative messages are displayed in a timeline to assist the forensic investigation.” – teaches presenting the sentiment prediction (negative sentiments are displayed in a timeline))
Studiawan fails to explicitly teach wherein the sentiment prediction model is trained with a training label received responsive to a similarity score of a training vector meeting a similarity threshold.
However, analogous to the field of the claimed invention, Udupa teaches:
processing the synthesized text with a sentiment prediction model to generate a sentiment prediction (Udupa, [0034] – “For example, a target outcome can include users or organizations (e.g., entities) that, based on survey responses to a digital survey, satisfy a certain parameter, objective, goal, sentiment, estimate, constraint, count, or number of instances. Example target outcomes include a range of customer satisfaction scores”, [0040], [0055] – “As shown in FIG. 2, the emerging user segment system 104 identifies a target outcome 210…’” – teaches processing synthesized text with a sentiment prediction model to generate a sentiment prediction (emerging user segment system 104 with user-segment-machine-learning model that identifies target outcome that includes a range of customer satisfaction scores, and as in [0040], customer satisfaction scores are digital metrics indicating a sentiment)), wherein the sentiment prediction model is trained with a training label received responsive to a similarity score of a training vector meeting a similarity threshold (Udupa, [0080-0081], [0089], and in [0090] – “As shown in FIG. 4A, in particular embodiments, the training attributes 402 comprise training labels that correspond to the attributes of users associated with the ground truth user segment 410.” – teaches wherein the sentiment prediction model (emerging-user-segment-machine-learning model 404 of emerging user segment system 104 as in Udupa at [0089]) is trained with a training label (training attributes comprise training labels) received responsive to a similarity score of a training vector (determines similarity using cosine similarity, vectorization, etc., thus teaching training vectors for determining similarities) meeting a similarity threshold (as in Udupa at [0081], teaches determining a similarity score for attributes of non-respondents and comparing the similarity score to a similarity score threshold to determine that attributes of non-respondents correspond to a subset of respondent attributes. Thus determining attributes, which comprise labels, for non-respondents by determining their similarity to respondent attributes using a similarity score compared to a similarity threshold, and thus training the sentiment prediction model with training labels received responsive to a similarity score meeting a similarity threshold)); and
presenting the sentiment prediction (Udupa, [0039-0040], [0054] – “Based on the emerging user segment, the emerging user segment system 104 can provide a segment visualization for display within a graphical user interface. In accordance with one or more embodiments, FIG. 2 illustrates the emerging user segment system 104 generating an emerging user segment 215 for providing a segment visualization.” – teaches presenting the sentiment prediction (provides segment visualization for display, and as in Udupa at [0039-0040] the segment can include customer satisfaction scores which are digital metrics indicating sentiment of one or more users))).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the sentiment prediction model and training method of Udupa to the log processing, text generation, and sentiment prediction model of Studiawan. Doing so would provide models that dynamically predict users that have same or similar sentiments without their interaction information (Udupa, [0005]).
Claims 11 and 20 incorporate substantively all the limitations of claim 1 in a system and a method, and are rejected on similar grounds as above. Udupa teaches the processors and application of claim 11 at [0133]. Claim 20 recites the limitation regarding “transmitting log data,” which Studiawan teaches at Fig. 2 and in Section IV Subsection A Paragraph 1 – “Therefore, the operating system records several events, which are associated with these incidents. We extract these log messages and analyze their sentiments with the proposed method”.
Regarding claim 2, the combination of Studiawan and Udupa teaches the method of claim 1, wherein training the sentiment prediction model comprises:
training the text generation model to process training log data to generate training text, wherein the text generation model is updated using the training text (Studiawan, Fig. 2, Section III Subsection D Last Paragraph – “To improve the sentence representation to handle more complex log messages, we add the sentence representation from the embedding layer vs and the sentence representation from the content attention layer vsr , resulting in vf . Finally, vf is the final representation of the aspect, considering the context and sentiment information of the log messages from the previous context attention layer”, and in Section III Subsection E – teaches training the text generation model (deep learning model stacked with embedding, content, and context attention layer, where content attention layer generates training text in form of vector, model is trained to minimize cross-entropy loss using generated vector from attention layer, thus the model is updated using the training text));
processing the training text with the sentiment prediction model to generate a training prediction (Studiawan, Fig. 2 and in Section III Paragraph 1 – “We then use a softmax layer to determine whether a message sentiment is positive or negative. In the last step of the training phase, we save the sentiment model.” – teaches processing the training text (sentiment model processes vector from attention layer) with the sentiment prediction model (as in Fig. 2, softmax layer & sentiment model) to generate a training prediction (determines whether a message sentiment is positive or negative, thus generating a sentiment prediction during training))
comparing the training prediction to the training label corresponding to the training log data (Studiawan, Fig. 2 and Section III Subsection E – “In the model training step, we minimize the cross-entropy loss H, which is calculated by: Eq. (11) where g is the ground-truth distribution and s is the estimated distribution from the softmax function.” – teaches comparing the training prediction to the training label corresponding to the training log data (compares estimated distribution s from softmax to ground-truth distribution g in cross-entropy loss)); and
updating the sentiment prediction model responsive to comparing the training prediction to the training label (Studiawan, Fig. 2 and Section III Subsection E – “In the model training step, we minimize the cross-entropy loss H, which is calculated by: Eq. (11) where g is the ground-truth distribution and s is the estimated distribution from the softmax function.” – teaches updating the sentiment prediction model responsive to comparing the training prediction to the training label (trains model using cross-entropy loss, thus updating the sentiment model responsive to comparing the estimated distribution to the ground-truth distribution)).
Claim 12 is similar to claim 2, hence similarly rejected.
Regarding claim 5, the combination of Studiawan and Udupa teaches the method of claim 1, wherein in the sentiment prediction model comprises one or more of a natural language processing (NLP) model and a transformer model (Udupa, [0080] – “To illustrate, when performing the act 312, the emerging user segment system 104 uses one or more natural language processing algorithms to determine a semantic similarity between attributes of non-respondents and the subset of respondent attributes…” – teaches wherein the sentiment prediction model (emerging user segment system 104) comprises one or more a natural language processing model (system uses one or more natural language processing algorithms)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the natural language processing model of Udupa to further modify the sentiment prediction model of Studiawan and Udupa. Doing so would provide algorithms for predicting sentiment of survey responses of respondent and non-respondent users (Udupa, [0034] and [0080]).
Claim 15 is similar to claim 5, hence similarly rejected.
Regarding claim 6, the combination of Studiawan and Udupa teaches the method of claim 1, wherein the sentiment prediction comprises a sentiment classification (Studiawan, Fig. 2 and Section III Paragraph 1 – “We then use a softmax layer to determine whether a message sentiment is positive or negative.” – teaches wherein the sentiment prediction comprises a sentiment classification (classifies message sentiment as positive or negative using softmax layer)).
Claim 16 is similar to claim 6, hence similarly rejected.
Regarding claim 7, the combination of Studiawan and Udupa teaches the method of claim 1, wherein the text generation model comprises one or more of an image to text model, a transformer model, a generative adversarial model, and a generative diffusion model (Studiawan, Fig. 2 and Section III Subsection D Last Paragraph – “To improve the sentence representation to handle more complex log messages, we add the sentence representation from the embedding layer vs and the sentence representation from the content attention layer vsr , resulting in vf . Finally, vf is the final representation of the aspect, considering the context and sentiment information of the log messages from the previous context attention layer” – teaches wherein the text generation model comprises one or more of a transformer model (sentiment model uses embedding and attention layers to generate sentence representations from preprocessed logs, thus the sentiment model comprises a transformer consistent with the Specification of the claimed invention at [0051])).
Claim 17 is similar to claim 7, hence similarly rejected.
Regarding claim 8, the combination of Studiawan and Udupa teaches the method of claim 1, for the comprising:
obtaining log statistics from system logs to form the log data (Studiawan, Section III Subsection A Paragraph 2 – “Unlike the existing approaches, which generally use regular expressions to parse log files, we employ the nerlogparser tool [28]. It can automatically split each entity in a log entry using a pretrained deep learning model, namely, bidirectional long short-term memory. The output of the nerlogparser is a JSON file containing the entity names and values for all records in a log file or a dictionary data structure.” – teaches obtaining log statistics (obtains values for all records) from system logs to form the log data).
Claim 18 is similar to claim 8, hence similarly rejected.
Regarding claim 9, the combination of Studiawan and Udupa teaches the method of claim 1, wherein the log data comprises system logs with log events comprising successful start of a software service, successful termination of software service, and intermediate operation logs (Studiawan, Fig. 7 – teaches wherein the log data comprises system logs with log events comprising successful start of a software service (as in Fig. 7, shows “session opened for user…”), successful termination of a software service (as in Fig. 7, shows “session closes for user…”), and intermediate operation logs (as in Fig. 7, shows logs such as failed password, authentication failure as intermediate operation logs)).
Claim 19 is similar to claim 9, hence similarly rejected.
Claim(s) 3-4 and 13-14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Studiawan and Udupa as applied to claims 1, 11, and 20 above, and further in view of Jacobs et al. (NPL: Active Learning for Reducing Labeling Effort in Text Classification Tasks, published Jan. 2022, hereinafter “Jacobs”).
Regarding claim 3, the combination of Studiawan and Udupa teaches the method of claim 1, wherein receiving the training label includes requesting feedback by:
processing training log data with vector generation model to generate a training vector (Udupa, [0080] – “For example, the emerging user segment system 104 compares attributes of non-respondents and the subset of respondent attributes using one or more of an edit distance algorithm (also referred to as the Levenshtein distance algorithm), cosine similarity, vectorization, bag of words, term frequency and inverse document frequency, text normalization, naïve Bayes algorithm, word embedding, long short-term memory, etc.” – teaches processing training log data with vector generation model to generate a training vector (emerging user segment system processes attributes of non-respondents using one or more of vectorization, cosine similarity, and word embedding, thus processing training data with vector generation model to generate a training vector));
processing the training vector with a vector similarity model to calculate the similarity score (Udupa, [0080] – “For example, the emerging user segment system 104 compares attributes of non-respondents and the subset of respondent attributes using one or more of an edit distance algorithm (also referred to as the Levenshtein distance algorithm), cosine similarity, vectorization, bag of words, term frequency and inverse document frequency, text normalization, naïve Bayes algorithm, word embedding, long short-term memory, etc.” and in [0081] – “For example, the emerging user segment system 104 can determine similarity scores for the attributes of non-respondents.” – teaches processing the training vector with a vector similarity model to calculate the similarity score (uses one or more of cosine similarity, vectorization, and word embedding to process vectors with a vector similarity model to calculate a similarity score));
determining the similarity score meets the similarity threshold indicating the training vector does not match a previous vector in a database (Udupa, [0081] – “Otherwise, if a similarity score of an attribute of a non-respondent fails to satisfy a threshold similarity score, the attribute does not correspond to the subset of the respondent attributes 208.” – teaches determining the similarity score meets the similarity indicating the training vector does not match a previous vector in a database (determines that similarity score fails to meet a threshold and thus indicating that the attribute, which is a word embedding and/or vectorized as in Udupa at [0080], does not match a previous vector in the respondent attributes));
The combination of Studiawan and Udupa fails to explicitly teach requesting feedback corresponding to the training log data responsive to determining the similarity score meets the similarity threshold; and generating the training label from the feedback to identify a sentiment identifier corresponding to the training log data.
However, analogous to the field of the claimed invention, Jacobs teaches:
requesting feedback corresponding to the training log data responsive to determining the similarity score meets the similarity threshold (Jacobs, Section 3.4 – “The assumption made is that semantically similar data conveys the same type of information to the model. The examples are selected based on their cosine similarity to other examples. RP is looped through and examples are only added to QP if their cosine similarity to all other points that are already in QP is lower than the chosen threshold l.” and in Section 1.2 A.2 Algorithm 4 – teaches requesting feedback corresponding to the training log data responsive to determining the similarity score meets the similarity threshold (as in Algorithm 4, if unlabeled data similarity score is lower than chosen threshold l, it is added to unlabeled dataset U and requests feedback from an Oracle)); and
generating the training label from the feedback to identify a sentiment identifier corresponding to the training log data (Jacobs, Section 1.2 A.2 Algorithm 4 – teaches generating the training label from the feedback to identify a sentiment identifier corresponding to the training log data (as in Algorithm 4, lets Oracle generate training label for identified unlabeled dataset U, thus generating the training label from the feedback to identify a sentiment identifier)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the feedback requests responsive to similarity scores meeting a similarity threshold and generation of training labels from feedback requests of Jacobs to the training method, similarity thresholds, log data, and sentiment prediction models of Studiawan and Udupa. Doing so would reduce labeling effort by only using the data which Natural Language Processing models deem most informative (Jacobs, Introduction).
Claim 13 is similar to claim 3, hence similarly rejected.
Regarding claim 4, the combination of Studiawan and Udupa teaches the method of claim 1.
The combination of Studiawan and Udupa fails to explicitly teach wherein feedback, from which the training label is generated, comprises a rating.
However, analogous to the field of the claimed invention, Jacobs teaches: wherein
feedback, from which the training label is generated, comprises a rating (Jacobs, Section 3.5 Paragraph 1 – “Only these full sentences were used in the experiments, and the sentiment labels were mapped to five categories in the following way: –0≤label < 0.2: very negative– 0.2 ≤ label < 0.4: negative– 0.4 ≤ label ≤ 0.6: neutral– 0.6 < label ≤ 0.8: positive– 0.8 < label ≤ 1: very positive” – teaches wherein feedback, from which the training label is generated, comprises a rating (sentiment labels mapped to five categories wherein the labels are ratings, and as in Section 1.2 A.2 Algorithm 4, the labels may be received from feedback from an Oracle)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the feedback and ratings of Jacobs to the training method, training labels, log data, and sentiment prediction model of Studiawan and Udupa. Doing so would utilize feedback from a human labeler to label informative data points (Jacobs, Introduction) and provide methods to map sentiment labels to categories based on ratings (Jacobs, Section 3.5).
Claim 14 is similar to claim 4, hence similarly rejected.
Claim(s) 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Studiawan and Udupa as applied to claims 1, 11, and 20 above, and further in view of Zhong et al. (from IDS: US Pub. No. 2023/0206255, filed Dec. 2021, hereinafter “Zhong”).
Regarding claim 10, the combination of Studiawan and Udupa teaches the method of claim 1, wherein the log data further comprises log statistics comprising geographical spread of usage of a service (Udupa, [0035] – “To illustrate, respondent attributes can include qualities, such as age, gender, location, type of computing device, type of operating system, subscription status with respect to an online service or computer application, interaction event (e.g. an event from an interaction history), purchase event (e.g., an event from a purchase history), preference, or interest.” – teaches wherein the log data further comprises log statistics comprising geographical spread of usage of a service (attributes can include location, subscription status, interaction events, preference, or interest, and thus teaches wherein the data comprises geographical spread of usage of a service by location and subscription status/interaction event))
The combination of Studiawan and Udupa fails to explicitly teach wherein the log data further comprises log statistics comprising duration of service use and frequency of use.
However, analogous to the field of the claimed invention, Zhong teaches:
wherein the log data further comprises log statistics comprising duration of service use and frequency of use (Zhong, [0007] – “The non-textual metadata may include at least one of a length of time the customer has been associated with the business, a quantity of the one or more interactions, or a subscription level associated with the customer.” – teaches wherein the log data further comprises log statistics comprising duration of service use (length of time customer associated with business/service), frequency of use (quantity of one or more interactions)).
Therefore, it would have been obvious to a person of ordinary skill in the art, before the effective filing date of the claimed invention, to incorporate the log data comprising duration of service use and frequency of use of Zhong to the log data of Studiawan and Udupa. Doing so would improve sentiment prediction using textual and non-textual data representative of one or more interactions between and customer and a business (Zhong, [0006]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Margatina et al. (NPL: Active Learning by Acquiring Contrastive Examples, published Nov. 2021) teaches methods for active learning by selecting data points that are similar in model feature space but output maximally different predictive likelihoods, and labeling the identified data points. Teaches performing the active learning method using sentiment analysis datasets.
Gupta et al. (NPL: Data Augmentation for Low Resource Sentiment Analysis Using Generative Adversarial Networks, published 2019) teaches methods for sentiment analysis using generative adversarial networks for data augmentation for improving sentiment classifier generation. Teaches synthesizing text with a text generation model comprising a generative adversarial network and classifying the synthesized text using a discriminator of the GAN.
Syeda-Mahmood et al. (US Patent No. 12,086,565, filed Feb. 2022) teaches systems and methods for implementing a text encoder to learn a sense and similarity preserving embedding. Teaches encoding natural language text to generate an embedding to cause the downstream computing system to perform a computer natural language processing operation on the embedding.
Shukla et al. (US Patent No. 11,620,320, published April 2023) teaches systems and methods for generating a summary of a document and performing sentiment analysis. Teaches wherein the document comprises a survey associated with one or more IT assets, and performing sentiment analysis on the document by using a “context retainer” built using self-attention weights for generating a representation of the data that may be used in an AI model to classify sentiments of aspect terms.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LOUIS C NYE whose telephone number is 571-272-0636. The examiner can normally be reached Monday - Friday 9:00AM - 5:00PM.
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, MATT ELL can be reached at 571-270-3264. 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.
/LOUIS CHRISTOPHER NYE/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights