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 Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
The terms “less computationally efficient decoding method” and “more computationally efficient decoding method” in claims 1 and 18 are a relative term which renders the claim indefinite. The terms “less computationally efficient decoding method” and “more computationally efficient decoding method” are not defined by the claim, the specification does not provide a standard for ascertaining the requisite degree, and one of ordinary skill in the art would not be reasonably apprised of the scope of the invention. Under the broadest reasonable interpretation, the terms “less computationally efficient decoding method” and “more computationally efficient decoding method” may be directed to any process, program, or algorithm that performs a task efficiently by factors such as amount of time, memory usage, or energy consumption. The Specification does not provide a clear definition as to what constitutes as “less computationally efficient decoding method”. In addition, the Specification does not provide a standard for ascertaining the requisite degree for the limitation "wherein the less computationally efficient decoding method is less computationally efficient than a more computationally efficient decoding method”. This may encompass a process that requires more computing resources when compared to a different process, a program that stores a large amount of data for processing, and/or an algorithm that consists of more steps to process. It is not clear what constitutes as the standard when comparing two decoding methods to determine that one is less computationally efficient than the other.
Examiner interprets “less computationally efficient decoding method” as any system, model, or framework that has access to large amounts of data, long computing time, and/or requires a lot of computing capacity. Examiner interprets “more computationally efficient decoding method” as any system, model, or framework that has access to a limited amount of data, low computing time requirement, and/or low computing capacity requirement. The two methods can be compared based on either data usage, compute time, or computing resource usage to determine whether one method is less computationally efficient.
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, 4-11, 14-16 and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Guo “Self-Distillation Mixup Training for Non-autoregressive Neural Machine Translation” in view of Fernandes “Quality-Aware Decoding for Neural Machine Translation”. Fernandes is provided as a reference in the IDS dated 09/20/2024.
Regarding claim 1, Guo teaches:
“A method implemented by one or more processors, the method comprising” (abstract, [pg. 6, Section 5.1, par. 1-2], An enhanced training strategy is proposed to improve the training of student model on distilled data generated by AT teacher models. It is implied that the experiments of training the models are conducted on a computer that consists of one or more processors.)
“for each of a plurality of source inputs: processing the source input, using a trained generative model, to generate corresponding generative model output” ([pg. 2-3, Section 2.3, par. 1-2; pg. 4, Section 4.1, par. 1-3, pg. 5, Figure 3], In sequence-level knowledge distillation, a pretrained large teacher model is used to train a small sequence generation model. A translation model generates k-candidates based on the input sentence. From Figure 3, the autoregressive (AT) teacher model is trained on raw data prior to generating the distilled data for training of the non-autoregressive (NAT) student model.)
“processing, using a less computationally efficient decoding method, the corresponding generative model output to: ” ([pg. 2-3, Section 2.3, par. 2; pg. 4, Section 4.1, par. 1-3, pg. 5, Figure 3], The translation model generates k-candidates based on the input and an AT teacher model may re-score the candidates. The reference does not explicitly disclose a sampling method during candidate generate to generate the corresponding candidate generative predictions.)
“select a corresponding single prediction from the corresponding candidate generative predictions” ([pg. 4-5, Section 4.1, par. 3-4; pg. 5, Figure 3], The framework filter incorrect train set out to generate the filtered teacher-distilled data. This is done by calculating the TER score and keeping the sentence pair lower than a threshold.)
“wherein the less computationally efficient decoding method is less computationally efficient than a more computationally efficient decoding method” ([pg. 1, Section 1, par. 1; pg. 2, Section 2, par. 1-2; pg. 5, Figure 3], AT teacher model decode tokens one-by-one and the inference is slow. NAT student model generates all outputs in parallel and it speeds up the inference.)
“storing, as a corresponding training instance, the source segment along with the corresponding single prediction” ([pg. 4, Section 3.2, par. 3; pg. 4-5, Section 4.1, par. 2-4; pg. 5, Figure 3], The sentence pair consisting of the input sentence and the generated candidate is selected to generate the filtered teacher-distilled data. It is implied that the filtered distilled data is stored in memory for finetuning of the NAT model.)
“fine-tuning, using the corresponding instances of training data, the trained generative model or an additional trained generative model, to generate a fine-tuned student generative model” ([pg. 4-5, Section 4.1, par. 2-4; pg. 5, Figure 3], The NAT student model is finetuned using the filtered distilled data.)
“subsequent to generating the fine-tuned student generative model: utilizing, in an inference time environment, the fine-tuned student generative model and the more computationally efficient decoding method in generating generative predictions” ([pg. 6, Section 5.1, par. 1-2; pg. 7-8, Section 5.3, par. 1-7; pg. 8, Table 13], The dataset consists of test data. A NAT-CRF model performs to the inference to generate the results. NAT-CRF performs better on the filtered teacher-distilled data than on a dataset without filtering.)
“wherein the less computationally efficient decoding method is not utilized in generating the generative predictions in the inference time environment” ([pg. 4-5, Section 4.1, par. 3-4; pg. 5, Figure 3; pg. 7, Table 7], The AT teacher model is used to rerank the candidates generated by the translation model. The filtered teacher-distilled data is used to finetune the student NAT model, which is used to perform the predictions on the test dataset.)
Guo does not explicitly disclose an implementation of “processing, using a less computationally efficient decoding method, the corresponding generative model output to: generate multiple corresponding candidate generative predictions”. However, Fernandes discloses in the same field of endeavor:
“processing, using a less computationally efficient decoding method, the corresponding generative model output to: generate multiple corresponding candidate generative predictions” ([pg. 2, Section 2.1, par. 1-4], A NMT model defines a probability distribution over a set of hypotheses conditioned on a source sentence. Stochastic candidate method like p-nucleus sampling is used to draw samples from the probability distribution.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “processing, using a less computationally efficient decoding method, the corresponding generative model output to: generate multiple corresponding candidate generative predictions” from Fernandes into the teaching of Guo. Doing so improves the performance of neural machine translation model by implementing quality-aware decoding techniques that consists of candidate generation and ranking methods (Fernandes, abstract).
Regarding claim 2, Guo in view of Fernandes teaches:
“sampling from a sequence of probability distributions, reflected by the corresponding generative model output, to generate the corresponding candidate generative predictions” ([Fernandes, pg. 2, Section 2.1, par. 1-4], A NMT model defines a probability distribution over a set of hypotheses conditioned on a source sentence. Stochastic candidate method like p-nucleus sampling is used to draw samples from the probability distribution.)
“applying each of the corresponding candidate generative predictions to an objective utility function to generate a corresponding utility metric for each of the corresponding candidate generative predictions” ([Guo, pg. 5, Section 4.2, par. 1-3], Bayesian Decision Theory is used to analyze the distilled data based on the Bayes risk. The Bayes risk is minimize for each input. In addition, Fernandes (pg. 3, Section 2.2.2, par. 1) teaches the candidates are ranked using minimum bayes risk decoding that uses a reference-based metric.)
“selecting the corresponding single prediction based on the corresponding utility metric for the corresponding single prediction” ([Guo, pg. 5, Section 4.2, par. 1-3], The selection of training data is based on minimizing the conditional risk and it is based on Equation 6.)
Regarding claim 4, Guo teaches:
“wherein the objective utility function includes a reference-based objective utility function” ([pg. 5, Section 4.2, par. 1-3], Bayesian Decision Theory is used to analyze the distilled data based on the Bayes risk. The Bayes risk is minimize for each input based on the function provided in Equation 5.)
Regarding claim 5, Guo in view of Fernandes teaches:
“applying the given prediction in multiple passes using the reference-based objective utility function to generate multiple individual reference-based utility metrics, each of the multiple passes being used to generate a corresponding one of the individual reference-based utility metrics based on a pairing of the given prediction with a corresponding other of the corresponding candidate generative predictions” ([Fernandes, pg. 3, Section 2.2.2, par. 1], MBR decoding uses reference-based metrics to rank candidates. The translation with the highest expected utility can be computed by comparing each hypothesis to all the other hypotheses in the set. Thus. MBR decoding consists of multiple passes of a single hypothesis to generate a utility metric for every pairing of another hypothesis in the set.)
“wherein the corresponding utility metric, for the given prediction, is based on the individual reference-based utility metrics generated for the given prediction” ([Fernandes, pg. 3, Section 2.2.2, par. 1], MBR decoding seeks for the maximum utility metric computed for each pairing of hypotheses.)
Regarding claim 6, Guo in view of Fernandes teaches:
“wherein the reference-based objective utility function is a Minimum Bayes’ Risk (MBR) scoring function” ([Guo, pg. 5, Section 4.2, par. 1-3], Bayesian Decision Theory is used to analyze the distilled data based on the Bayes risk. The Bayes risk is minimize for each input as shown in Equation 5 and 6. Fernandes (pg. 3, Section 2.2.2, par. 1) teaches the candidates are ranked using minimum bayes risk decoding that uses a reference-based metric)
Regarding claim 7, Guo in view of Fernandes teaches:
“wherein the objective utility function includes a reference-free objective utility function” ([Fernandes, pg. 2, Section 2.2, par. 1-3], The framework evaluates 2 different ranking strategies. N-best reranking uses a reference-free metric. In one embodiment, a single feature is used and the candidate that maximizes the score is picked as the final translation.)
Regarding claim 8, Guo in view of Fernandes teaches:
“applying the given prediction and the source input using the reference-free objective utility function to generate a reference-free utility metric” ([Fernandes, pg. 2, Section 2.2, par. 1-3; pg. 4, Section 3.2, par. 1-6], The framework uses N-best reranking to rank candidate translations for a source sentence. In one embodiment, COMET-QE is used as a reference free metric for the N-best reranking.)
“wherein the corresponding utility metric, for the given prediction, is based on the reference-free utility metric generated for the given prediction” ([Fernandes, pg. 2, Section 2.2, par. 1-3; pg. 4, Section 3.2, par. 1-6], COMET-QE is used as the utility metric to score and re-rank candidates and selecting the sentence that maximizes the predicted quality score.)
Regarding claim 9, Guo in view of Fernandes teaches:
“wherein the reference-free objective utility function is a quality estimation (QE) scoring function” ([Fernandes, pg. 2, Section 2.2, par. 1-3; pg. 4, Section 3.2, par. 1-6], COMET-QE is used as the utility metric to score and re-rank candidates and selecting the sentence that maximizes the predicted quality score. N-best reranking may use Equation 2 and 3 to determine the best candidates.)
Regarding claim 10, Guo teaches:
“wherein the source inputs are natural language source inputs and the corresponding candidate generative predictions are corresponding natural language generative predictions” ([pg. 6, Section 5.1, par. 1-2], The dataset of WMT14 consists of natural language inputs and the AT teacher model generates predictions in natural language.)
Regarding claim 11, Guo teaches:
“wherein the natural language source inputs are in a first spoken language, the corresponding natural language generative predictions are translation predictions in a second spoken language, and the fine-tuned student generative model is a neural machine translation (NMT) model” ([pg. 6, Section 5.1, par. 1-2], The dataset of WMT14 En-De consists of natural language inputs for translating text from English to German. The student model is a NAT model, which is a specific type of neural machine translation.)
Regarding claim 14, Guo teaches:
“wherein the more computationally efficient decoding method, utilized in the inference time environment in generating the generative predictions is a beam search method, a greedy search method, or a sampling method” ([pg. 2, Section 2.2, par. 1-3; pg. 6, Section 5.1, par. 1-2; pg. 7, Table 7], The study used different NAT model that includes Vanilla NAT (V-NAT) to implement the proposed framework. V-NAT models are well known to decode using greedy search method.)
Regarding claim 15, Guo teaches:
“receiving new source input” ([pg. 6, Section 5, par. 1-2], The study used five standard datasets to train and test the framework. The student NAT model is tested on different testing data after it has been trained using the training dataset.)
“processing new source input, using the fine-tuned student generative model, to generate new generative model output” ([pg. 2, Section 2.2, par. 5; pg. 6, Section 5, par. 1-2], The study used five standard datasets to train and test the framework. In one embodiment, the student NAT model is an iterative refinement NAT model, which processes the outputs of the previous step as input during each refinement step.)
“processing the new generative model output using the more computationally efficient decoding method to determine a new generative prediction” ([pg. 2, Section 2.2, par. 5; pg. 6, Section 5, par. 1-2], The iterative refinement NAT model is more computationally efficient compared to AT teacher models because it generates all outputs in parallel.)
“providing the new generative prediction responsive to receiving the new source input” ([pg. 7-8, Section 5.3, par. 1-5], The study evaluates the predictions generated by the NAT models based on the test dataset.)
Regarding claim 16, Guo teaches:
“wherein the new source input is generated based on user interface input at a client device and wherein providing the new generative prediction comprises causing the new generative prediction to be rendered at the client device” ([pg. 6, Section 5, par. 1-2; pg. 7-8, Section 5.3, par. 1-5], It is implied that the experiments are conducted on a computer with a user interface to input data and output results. The datasets are collected by a user and the datasets are provided to the model. The results of the models’ predictions and evaluation of the results are displayed for the user.)
Regarding claim 18:
Claim 18 recites a system that performs the same process as described in Claim 1. Therefore claim 18 is rejected under the same reasons mentioned for claim 1. The additional elements of claim 18 are addressed below by Guo:
“A system, comprising: memory storing instructions; one or more processors operable to execute the instructions to” ([abstract, pg. 6, Section 5, par. 1-3], It is implied that the framework is implemented on a computer system consisting of a memory and one or more processors to perform the proposed method.)
Regarding claim 19, Guo teaches:
“wherein in causing the fine-tuned student generative model to be utilized, in the inference time environment, along with the more computationally efficient decoding method, one or more of the processors are to transmit the fine-tuned student generative model to one or more devices” ([abstract, pg. 6, Section 5, par. 1-3], It is implied that the framework is implemented on a computer system consisting of a memory and one or more processors to perform the proposed method. Under the broadest reasonable interpretation, one or more devices can constitute input devices, output devices, storage devices, and processing devices of a computer system. When the student NAT model is executed, the processor may transmit the code to execute the model from a hard drive to RAM.)
Regarding claim 20, Guo teaches:
“wherein in causing the fine-tuned student generative model to be utilized, in the inference time environment, along with the more computationally efficient decoding method, one or more of the processors are further to transmit, to the one or more devices, instructions that specify to utilize the more computationally efficient decoding method in the inference time environment and/or that specify to not utilize the less computationally efficient decoding method” ([abstract, pg. 6, Section 5, par. 1-3], It is implied that the framework is implemented on a computer system consisting of a memory and one or more processors to perform the proposed method. Under the broadest reasonable interpretation, one or more devices can constitute input devices, output devices, storage devices, and processing devices of a computer system. When the student NAT model is executed, the processor may transmit the code to execute the model from a hard drive to RAM. The processor also loads instructions that is specific to running the student NAT model and these instructions are different from the code to execute the AT teacher model.)
Claims 3 and 12-13 are rejected under 35 U.S.C. 103 as being unpatentable over Guo, “Self-Distillation Mixup Training for Non-autoregressive Neural Machine Translation” in view of Fernandes, “Quality-Aware Decoding for Neural Machine Translation” and Hewitt, “Truncation Sampling as Language Model Desmoothing”. Fernandes and Hewitt are provided as a reference in the IDS dated 09/20/2024.
Regarding claim 3, Guo in view of Fernandes does not explicitly disclose an implementation of “wherein the sampling is epsilon sampling”. However, Hewitt discloses in the same field of endeavor:
“wherein the sampling is epsilon sampling” ([abstract; pg. 5, Section 4.4, par. 1-2], The framework discloses η-sampling, which truncates words below an entropy-dependent probability threshold.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein the sampling is epsilon sampling” from Hewitt into the teaching of Guo in view of Fernandes. Doing so improves the performance of sampling algorithms in language model by implementing a new sampling technique that truncate samples below a probability threshold (Hewitt, abstract).
Regarding claim 12, Guo in view of Fernandes teaches:
“wherein the trained generative model is a ” ([Guo, pg. 4, Section 4.1, par. 1-2; pg. 6, Section 5.1, par. 1-2; pg. 5, Figure 3], In one embodiment, a language model that is trained on the original target-side corpus is used to re-score the self-distilled data. The fine-tuning is performed on the NAT student model. A Transformer-based model is used as the teacher model to create AT-distilled data.)
Guo in view of Fernandes does not explicitly disclose an implementation of “wherein the trained generative model is a large language model (LLM)”. However, Hewitt discloses in the same field of endeavor:
“wherein the trained generative model is a large language model (LLM) …” ([pg. 5-6, Section 5, par. 2; pg. 6, Section 5.1, par. 1-2], The framework uses GPT-2 models to perform the different sampling strategies, such as top-p. GPT-2 is built on a transformer model architecture and the language model taught in Guo and Fernandes may be modified to use a LLM.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein the trained generative model is a large language model (LLM)” from Hewitt into the teaching of Guo in view of Fernandes. Doing so improves the performance of sampling algorithms in natural language tasks by implementing a new sampling technique that truncate samples below a probability threshold using a LLM (Hewitt, abstract).
Regarding claim 13, Guo in view of Fernandes teaches:
“wherein the fine-tuned student generative model is a ” ([Guo, pg. 4, Section 4.1, par. 1-2; pg. 6, Section 5.1, par. 1-2; pg. 5, Figure 3], The fine-tuning is performed on the NAT student model, which is a specific type of Transformer model.)
Guo in view of Fernandes does not explicitly disclose an implementation of “wherein the fine-tuned student generative model is a large language model (LLM)”. However, Hewitt discloses in the same field of endeavor:
“wherein the fine-tuned ” ([pg. 6, Section 5.1, par. 1-2; pg. 9, Section 6, par. 1], The framework uses GPT-2 models to perform decoding. GPT-2 is built on a transformer model architecture and the language model taught in Guo and Fernandes may be modified to use a LLM.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein the fine-tuned student generative model is a large language model (LLM)” from Hewitt into the teaching of Guo in view of Fernandes. Doing so improves the performance of sampling algorithms in natural language tasks by implementing a new sampling technique that truncate samples below a probability threshold using a LLM (Hewitt, abstract).
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Guo, “Self-Distillation Mixup Training for Non-autoregressive Neural Machine Translation” in view of Fernandes, “Quality-Aware Decoding for Neural Machine Translation” and Wuebker (US10878201B1).
Regarding claim 17, Guo in view of Fernandes does not explicitly disclose an implementation of “wherein processing the new source input, processing the new generative model output, and providing the new generative prediction are performed by one or more server devices, wherein the new source input is received in a request transmitted to the one or more server devices, and wherein providing the new generative prediction comprises transmitting the new generative prediction”. However, Wuebker discloses in the same field of endeavor:
“wherein processing the new source input, processing the new generative model output, and providing the new generative prediction are performed by one or more server devices, wherein the new source input is received in a request transmitted to the one or more server devices, and wherein providing the new generative prediction comprises transmitting the new generative prediction” ([abstract; col. 3, lines 28-53], The neural machine translation system consists of a server and client devices. The server receives the source sentence from the client device and performs encoding and decoding operations to generate a translation suggestion. The results are provided to the client over the network.)
It would be obvious to one of the ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teaching of “wherein processing the new source input, processing the new generative model output, and providing the new generative prediction are performed by one or more server devices, wherein the new source input is received in a request transmitted to the one or more server devices, and wherein providing the new generative prediction comprises transmitting the new generative prediction” from Wuebker into the teaching of Guo in view of Fernandes. Doing so improves the performance of neural machine translation tasks by implementing the system onto a server and connecting to a network to collect input data from a plurality of client devices (Wuebker, abstract).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to GARY MAC whose telephone number is (703)756-1517. The examiner can normally be reached Monday - Friday 8:00 AM - 5:00 PM.
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, Abdullah Kawsar can be reached at (571) 270-3169. 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.
/GARY MAC/Examiner, Art Unit 2127
/ABDULLAH AL KAWSAR/Supervisory Patent Examiner, Art Unit 2127
