DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of Claims
This action is in reply to the claims and arguments/remarks filed on 7 July 2026.
Claims 3, 10 and 17 were canceled.
Claims 1, 8, and 15 have been amended.
Claims 1-2, 4-9, 11-16, and 18-20 are currently pending and have been examined.
Response to Arguments
Applicant's arguments, see Page(s) 7-8, filed 7 July 2026, with respect to the 35 USC § 103 rejection(s) of claim(s) 1-2, 4-9, 11-16, and 18-20 have been fully considered but they are not persuasive. Applicant argues that Vu in view of Xu does not teach transfer learning. The Examiner respectfully disagrees.
Transfer learning is a technique where a model trained on one task is reused for a related task, especially when the new task has limited data. Vu Paragraph [0351] recites:
The benefits of the disclosed techniques were demonstrated by conducting various experiments. To implement the experiments, the disclosed techniques were used to fine-tune the mDistilBERT pretrained multi-lingual model using two datasets, including without Spanish and with translated Spanish. The “without Spanish” dataset included labeled English data only and the “with translated Spanish” dataset included labeled English data and the corresponding Spanish translation.
Paragraph [0351] shows how an initial generic sentiment analysis model such as mDistilBERT is fine-tuned to perform a specific task such as sentiment analysis of a specific language. Paragraph [0038] of Vu further recites:
The performance of an ML model is only as good as its training. To properly train a model that can accurately and reliably predict sentiment in different languages, a large amount of training data is needed to ensure that the model is accurate and reliable in predicting sentiments from utterances and/or documents. The training data also has to be diverse covering various situations and different types of expressions. The availability of such training data is presently very limited for some languages. There are several reasons for this. A large volume of diverse training documents covering diverse situations is not easily available in low-resource languages. Additionally, each training document has to be annotated (e.g., labeled). These annotations are typically done manually. This is a very tedious and time-consuming task.
Paragraph [0038] shows that the problem being solved is where the new task (predicting sentiment in a specific language) has limited data.
Therefore, Vu teaches transfer learning by showing that a model is trained on one task (i.e., general sentiment analysis) and is reused for a related task (i.e., sentiment analysis of a specific language) when the new task as limited data (i.e., there is not enough training data of that specific language). This general sentiment analysis model can be reused for to generate models for different languages.
Vu does not teach that the large language model is a generative model or that the LLM is adapted into an embedding model with improved embedding performance. However, the Examiner is relying on new prior art Muennighoff to teach this limitation, as explained in the below 103 rejection of claim 1.
With regards to claims 8 and 15 the applicant argues these claims are allowable due to their similarities to claim 1. As stated in the arguments above, the Examiner is maintaining the rejections for claims 1. Therefore, claims 8 and 15 remain rejected. With regards to claims 2, 4-7, 9, 11-14, 16, and 18-20 the applicant argues these claims are allowable due to their dependence on claims 1, 8, and 15. As stated in the arguments above, the Examiner is maintaining the rejections for claims 1, 8, and 15. Therefore, claims 2, 4-7, 9, 11-14, 16, and 18-20 remain rejected.
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 1-2, 4-9, 11-16, and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Patent Application Publication No. 2024/0135116 A1 to Duy Vu et al. (hereinafter Vu) in view of Non-Patent Literature HSimCSE: Improving Contrastive Learning of Unsupervised Sentence Representation with Adversarial Hard Positive and Dual Hard Negatives to Bo Xu et al. (hereinafter Xu), in further view of Muennighoff (see attached NPL).
Regarding claim 1, A method of training a neural network model, the method comprising: (Vu teaches training a multi-lingual multi-task model wherein at least a portion of the machine learning model is a neural network. Vu at ¶¶ [0140] - [0143] and [0153] - [0156].)
receiving, via a data interface, a first plurality of data samples; (Vu teaches accessing, obtaining, or generating a plurality of datasets that differ from each other in one or more differentiation attributes. (i.e., a plurality of data samples). Vu at ¶¶ [0159] - [0160].)
generating a plurality of batches using the first plurality of data samples, wherein a first batch includes data samples associated with a single first task, and wherein a second batch includes data samples associated with a single second task; (Vu teaches generating training batches used for ABSA and SLSA training tasks (i.e., a first batch for a first task and a second batch for a second task.) Vu at ¶¶ [0160] - [0167].)
and performing a first training process to the neural network model using the plurality of batches, wherein the first training process fine-tunes a pre-trained
wherein the performing the first training process includes:
generating a first loss objective function for the first batch based on the first task; generating a second loss objective function for the second batch based on the second task; (Vu teaches, using the training batches, minimizing a cross-entropy loss function and updating model parameters based on the loss function. Vu at ¶¶ [0375] - [0380]. The training batches are associated with specific tasks, and because training is performed for each task, wherein a loss is minimized based upon the training batches used for that task, then this constitutes generating a first loss objective function for the first batch based on the first task. Further, the training batches apply to multiple tasks (ABSA and SLSA) therefore a new loss is used for each task (i.e., a second task, a second loss function, and a second batch based on the second task). Vu at ¶¶ [0160] - [0167] and [0375] - [0380].)
computing a first loss based on the first loss objective function; computing a second loss based on the second loss objective function; (Vu teaches minimizing a loss based on each task at hand as a process of training for the task. Vu at ¶¶ [0375] - [0380].)
and updating parameters of the neural network model based on the first loss and the second loss via backpropagation; (Vu teaches updating parameters of a machine learning model based on the loss functions used for training. Vu at ¶¶ [0375] - [0380]. Further, Vu teaches using backpropagation to train a machine learning model (i.e., the model trained by updating parameters). Vu at ¶¶ [0140] - [0141].)
Vu, however, does not alone teach:
and performing a first training process to the neural network model using the plurality of batches, wherein the first training process fine-tunes a pre-trained generative large language model of the neural network model via multi-task transfer learning using the plurality of batches to adapt the pre-trained generative large language model into an embedding model with improved embedding performance,
wherein the neural network model trained by the first training process is used to perform a text retrieval task based on text embedding.
However, In a similar field of endeavor (e.g., training neural networks for performing specific tasks including text retrieval/question answering wherein embeddings are used), Xu teaches and wherein the neural network model trained by the first training process is used to perform a text retrieval task based on text embedding. (Xu teaches sentence (i.e., text) embeddings being used for question answering (i.e., text retrieval) wherein training is performed to achieve such a task. Xu at sections I. Introduction, II. Overview, and III. Method, subsections A - D.)
It would have been prima facie obvious to one of ordinary skill in the art before the effective filing date to combine the teachings of Vu with the teachings of Xu to provide the limitations of claim 1. Doing so would have improved the performance of the model of Vu by increasing performance over traditional models using hard positive and hard negative sampling to enhanced performance of the models as recognized by Xu at section IV. Experiment, B. Performance Comparison, and C. Ablation Study.
Vu in view of Xu does not teach:
and performing a first training process to the neural network model using the plurality of batches, wherein the first training process fine-tunes a pre-trained generative large language model of the neural network model via multi-task transfer learning using the plurality of batches to adapt the pre-trained generative large language model into an embedding model with improved embedding performance.
However, Muennighoff teaches:
adapting a generative large language model into an embedding model with improved embedding performance. (Page 9, section 5 “SGPT-BE uses position-weighted mean pooling and fine-tuning of only bias tensors. At scale, it produces new state-of-the-art sentence embeddings. The model can be used for semantic search or other embedding tasks.”; Abstract, Page 1 “At 5.8 billion parameters SGPT improves on the previously best sentence embeddings by a margin of 7% and outperforms a concurrent method with 175 billion parameters as measured on the BEIR search benchmark.” of Muennighoff)
This step of Muennighoff is applicable to the method of Vu in view of Xu as they both share characteristics and capabilities, namely, they are directed to using large language models to fine tune models for more efficient performance. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to have modified the method of Vu in view of Xu by substituting the encoder-based LLM with a pre-trained generative LLM that is adapted into an embedding model as taught by Muennighoff. One of ordinary skill in the art before the effective filling date of the claimed invention would have been motivated to modify Vu in view of Xu in order to prevent the need to train and maintain separate encoder and decoder models and prevent additional costs to the environment (see Introduction, Page 1 of Muennighoff).
Regarding claim 2, Vu in view of Xu in view of Muennighoff (hereinafter Vu-Xu-Muennighoff) teaches all the limitations of claim 1. Further, Vu does not teach, but Xu teaches the method of claim 1, wherein prior to the first training process, the neural network model is trained using a second training process using a second plurality of data samples. (Xu teaches the sentence encoder is a pre-trained model (i.e., trained using a second training process with different data prior to the training Xu). Xu at Section II. Overview, A. Problem Formulation.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 1.
Regarding claim 4, Vu-Xu-Muennighoff teaches all the limitations of claim 1 as laid out above. Further, Vu does not teach, but Xu teaches the method of claim 1, wherein the text retrieval task is different from the first task and the second task. (Xu teaches that the sentence encoder, and specifically learning sentence representations themselves, is used in downstream applications (i.e., tasks that are not part of the training process) that includes question answering (i.e., text retrieval). Xu at Section I. Introduction. Therefore, the text-retrieval task is a different task than the training tasks as it is a downstream task, not a training task.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 1.
Regarding claim 5, Vu-Xu-Muennighoff teaches all the limitations of claim 1 as laid out above. Further, Vu does not teach, but Xu teaches the method of claim 1, wherein the first loss objective function includes a first contrastive loss customized to the first task, and wherein the second loss objective function includes a second contrastive loss customized to the second task. (Xu teaches minimizing the normalized temperature-scaled cross-entropy loss NT-Xent which is recognized as the "Core of contrastive learning". Xu at section III. Method, A. Backbone. Further, Xu teaches using additionally using a contrastive quadruplet loss function to train the sentence encoder (i.e., a second training task separate from the first that minimized cross-entropy loss). Xu at section III. Method, sections A, B, and D.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 1.
Regarding claim 6, Vu-Xu-Muennighoff teaches all the limitations of claim 1 as laid out above. Further, Vu does not teach, but Xu teaches the method of claim 1, wherein the performing the first training process includes: generating a plurality of hard negatives for the first task; (Xu teaches retrieving the closest samples to anchor samples as hard-negative samples (i.e., generating a plurality of hard negatives.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
selecting a predetermined number of hard negatives from the plurality of hard negatives for the first task; (Xu teaches selecting the top-K closest samples as the hard-negatives, then removing the top-1 sample to avoid the false-negative problem. Xu at section III. Method, C. Dual Negative Sample Selection Module.)
and updating the first batch using the selected predetermined number of hard negatives. (Xu teaches using the hard-negatives to alter the Faiss library to solve the false-negative problem by removing the top-1 closest negative sample and obtaining a mixed global negative sample from the remaining top-K hard negative samples (i.e., updating the batch based on the predetermined number.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 1.
Regarding claim 7, Vu-Xu-Muennighoff teaches all the limitations of claim 6 as laid out above. Further, Vu does not teach, but Xu teaches the method of claim 6, wherein a pre-trained second neural network model is used to generate the plurality of hard negatives for the first task. (Xu teaches using RoBERTa to encode sentences then retrieving the top-K closest negative samples from the resulting encodings (i.e., RoBERTa is used as a secondary model specifically for the generation of the hard-negative samples.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 1.
Claim(s) 8 is/are directed to a system. Claim(s) 8 recite limitations parallel in nature as those addressed above for claim(s) 1, which are directed towards a method. Claim(s) 8 is/are therefore rejected for the same reasons as set above for claim(s) 1. Claim 8 further recites: a memory that stores a neural network model and a plurality of processor-executable instructions; (Vu teaches the system implemented on computer readable memory connected to processors. Vu at ¶ [0158].); a communication interface that receives a first plurality of data samples; (Vu teaches the system implemented with communication mechanisms. Vu at ¶ [0158]. (Vu teaches accessing, obtaining, or generating a plurality of datasets that differ from each other in one or more differentiation attributes. (i.e., a plurality of data samples). Vu at ¶¶ [0159] - [0160].); and one or more hardware processors that read and execute the plurality of processor-executable instructions from the memory to perform operations comprising: (Vu teaches multiple processors that read and execute instructions containing the system. Vu at ¶ [0158].)
Regarding claim 9, Vu-Xu-Muennighoff teaches all the limitations of claim 8 as laid out above. Further, Vu does not teach, but Xu teaches the system of claim 8, wherein prior to the first training process, the neural network model is trained using a second training process using a second plurality of data samples. (Xu teaches the sentence encoder is a pre-trained model (i.e., trained using a second training process with different data prior to the training Xu). Xu at Section II. Overview, A. Problem Formulation.) The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 7.
Regarding claim 11, Vu-Xu-Muennighoff teaches all the limitations of claim 8 as laid out above. Further, Vu does not teach, but Xu teaches the system of claim 8, wherein the text retrieval task is different from the first task and the second task. (Xu teaches that the sentence encoder, and specifically learning sentence representations themselves, is used in downstream applications (i.e., tasks that are not part of the training process) that includes question answering (i.e., text retrieval). Xu at Section I. Introduction. Therefore, the text-retrieval task is a different task than the training tasks as it is a downstream task, not a training task.). The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 7.
Regarding claim 12, Vu-Xu-Muennighoff teaches all the limitations of claim 8 as laid out above. Further, Vu does not teach, but Xu teaches the system of claim 8, wherein the first loss objective function includes a first contrastive loss customized to the first task, and wherein the second loss objective function includes a second contrastive loss customized to the second task. (Xu teaches minimizing the normalized temperature-scaled cross-entropy loss NT-Xent which is recognized as the "Core of contrastive learning". Xu at section III. Method, A. Backbone. Further, Xu teaches using additionally using a contrastive quadruplet loss function to train the sentence encoder (i.e., a second training task separate from the first that minimized cross-entropy loss). Xu at section III. Method, sections A, B, and D.). The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 7.
Regarding claim 13, Vu-Xu-Muennighoff teaches all the limitations of claim 8 as laid out above. Further, Vu does not teach, but Xu teaches the system of claim 8, wherein the performing the first training process includes: generating a plurality of hard negatives for the first task; (Xu teaches retrieving the closest samples to anchor samples as hard-negative samples (i.e., generating a plurality of hard negatives.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
selecting a predetermined number of hard negatives from the plurality of hard negatives for the first task; (Xu teaches selecting the top-K closest samples as the hard-negatives, then removing the top-1 sample to avoid the false-negative problem. Xu at section III. Method, C. Dual Negative Sample Selection Module.)
and updating the first batch using the selected predetermined number of hard negatives. (Xu teaches using the hard-negatives to alter the Faiss library to solve the false-negative problem by removing the top-1 closest negative sample and obtaining a mixed global negative sample from the remaining top-K hard negative samples (i.e., updating the batch based on the predetermined number.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 7.
Regarding claim 14, Vu-Xu-Muennighoff teaches all the limitations of claim 13 as laid out above. Further, Vu does not teach, but Xu teaches the system of claim 13, wherein a pre-trained second neural network model is used to generate the plurality of hard negatives for the first task. (Xu teaches using RoBERTa to encode sentences then retrieving the top-K closest negative samples from the resulting encodings (i.e., RoBERTa is used as a secondary model specifically for the generation of the hard-negative samples.) Xu at section III. Method, C. Dual Negative Sample Selection Module.) The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 7.
Claim(s) 15 is/are directed to a system. Claim(s) 15 recite limitations parallel in nature as those addressed above for claim(s) 1, which are directed towards a method. Claim(s) 15 is/are therefore rejected for the same reasons as set above for claim(s) 1. Claim 15 further recites: a non-transitory machine-readable medium comprising a plurality of machine-executable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform operations comprising: (Vu teaches the system implemented on computer readable memory connected to processors. Vu at ¶ [0158].)
Regarding claim 16, Vu-Xu-Muennighoff teaches all the limitations of claim 15 as laid out above. Further, Vu does not teach, but Xu teaches the non-transitory machine-readable medium of claim 15, wherein prior to the first training process, the neural network model is trained using a second training process using a second plurality of data samples. (Xu teaches the sentence encoder is a pre-trained model (i.e., trained using a second training process with different data prior to the training Xu). Xu at Section II. Overview, A. Problem Formulation.) The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 15.
Regarding claim 18, Vu-Xu-Muennighoff teaches all the limitations of claim 15 as laid out above. Further, Vu does not teach, but Xu teaches the non-transitory machine-readable medium of claim 15, wherein the text retrieval task is different from the first task and the second task. (Xu teaches that the sentence encoder, and specifically learning sentence representations themselves, is used in downstream applications (i.e., tasks that are not part of the training process) that includes question answering (i.e., text retrieval). Xu at Section I. Introduction. Therefore, the text-retrieval task is a different task than the training tasks as it is a downstream task, not a training task.). The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 15.
Regarding claim 19, Vu-Xu-Muennighoff teaches all the limitations of claim 15 as laid out above. Further, Vu does not teach, but Xu teaches the non-transitory machine-readable medium of claim 15, wherein the first loss objective function includes a first contrastive loss customized to the first task, and wherein the second loss objective function includes a second contrastive loss customized to the second task. (Xu teaches minimizing the normalized temperature-scaled cross-entropy loss NT-Xent which is recognized as the "Core of contrastive learning". Xu at section III. Method, A. Backbone. Further, Xu teaches using additionally using a contrastive quadruplet loss function to train the sentence encoder (i.e., a second training task separate from the first that minimized cross-entropy loss). Xu at section III. Method, sections A, B, and D.). The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 15.
Regarding claim 20, Vu-Xu-Muennighoff teaches all the limitations of claim 15 as laid out above. Further, Vu does not teach, but Xu teaches the non-transitory machine-readable medium of claim 15, wherein the performing the first training process includes:
generating a plurality of hard negatives for the first task; (Xu teaches retrieving the closest samples to anchor samples as hard-negative samples (i.e., generating a plurality of hard negatives.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
selecting a predetermined number of hard negatives from the plurality of hard negatives for the first task; (Xu teaches selecting the top-K closest samples as the hard-negatives, then removing the top-1 sample to avoid the false-negative problem. Xu at section III. Method, C. Dual Negative Sample Selection Module.)
and updating the first batch using the selected predetermined number of hard negatives. (Xu teaches using the hard-negatives to alter the Faiss library to solve the false-negative problem by removing the top-1 closest negative sample and obtaining a mixed global negative sample from the remaining top-K hard negative samples (i.e., updating the batch based on the predetermined number.) Xu at section III. Method, C. Dual Negative Sample Selection Module.)
The motivation for making this modification to the teachings of Vu is the same as that set forth above, in the rejection of claim 15.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DANIELLE ELIZABETH ZEVITZ whose telephone number is (703)756-1070. The examiner can normally be reached Mo-Th 10am-6pm.
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, Andrew Flanders can be reached at (571) 272-7516. 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.
/DANIELLE ELIZABETH ZEVITZ/Examiner, Art Unit 2655
/ANDREW C FLANDERS/Supervisory Patent Examiner, Art Unit 2655
Prosecution Insights