CTNF 18/458,540 CTNF 95989 Notice of Pre-AIA or AIA Status 07-03-aia AIA 15-10-aia 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 § 103 07-20-aia AIA 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. 07-21-aia AIA Claims 1-3, 6, 8-14, 16-18, and 20 are r ejected under 35 U.S.C. 103 as being unpatentable over L ewis et al (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, herein Lewis) in view of Luong et al (US 20210089724 A1, herein Luong). R egarding claim 1, Lewis teaches a context generation component that selects, using a first machine learning model, an information source from a group of information sources based on relevance of the information source to an input prompt, resulting in a selected information source (fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z” (i.e., using a first machine learning model to select information sources based on top-K relevance to an input prompt)) ; and a response formulation component that transforms, using a second machine learning model that is not the first machine learning model, the input prompt into a human-readable response, the human-readable response being constructed by applying parameters of the second machine learning model to information in the selected information source (fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z”. Section 2.3 recites “The generator component p θ (yi|x, z, y1:i−1) could be modelled using any encoder-decoder. We use BART-large, a pre-trained seq2seq transformer with 400M parameters. To combine the input x with the retrieved content z when generating from BART, we simply concatenate them” (i.e., transforming the input prompt into a human readable response from a second machine learning model different from the first machine learning model)) . However, while one of ordinary skill in the art would recognize that the methods taught by Lewis would be implemented by a computer system, Lewis does not explicitly teach a system, comprising: a memory that stores executable components; and a processor that executes the executable components stored in the memory. Luong teaches a system, comprising: a memory that stores executable components; and a processor that executes the executable components stored in the memory (para. [0063] recites “The user computing device 102 includes one or more processors 112 and a memory 114. The one or more processors 112 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, a FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. The memory 114 can include one or more non-transitory computer-readable storage mediums, such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. The memory 114 can store data 116 and instructions 118 which are executed by the processor 112 to cause the user computing device 102 to perform operations” (i.e., a system comprising a processor and memory)) . Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine these teachings by implementing the retrieval-augmented generation model from Lewis using the computing system from Luong. Lewis and Luong are both directed to methods of utilizing generative pre-trained transformer machine learning models to automatically generate text based on a corpus of training data. One of ordinary skill in the art would recognize that the known computing system from Luong could be utilized to implement the known retrieval-augmented generation model from Lewis. Regarding claim 2, the combination of Lewis and Luong teaches the system of claim 1, wherein the first machine learning model is a question answering model (Lewis section 2.2 recites “This retriever was trained to retrieve documents which contain answers to TriviaQA questions and Natural Questions” (i.e., the first machine learning model is a question answering model)) . Regarding claim 3, the combination of Lewis and Luong teaches the system of claim 1, wherein the second machine learning model is a generative pre-trained transformer model (Lewis section 2.3 recites “The generator component p θ (yi|x, z, y1:i−1) could be modelled using any encoder-decoder. We use BART-large, a pre-trained seq2seq transformer with 400M parameters. To combine the input x with the retrieved content z when generating from BART, we simply concatenate them” (i.e., the second machine learning model is a BART model, or a generative pre-trained transformer (GPT) model)) . Regarding claim 6, the combination of Lewis and Luong teaches the system of claim 1, wherein: the context generation component selects, using a group of first machine learning models comprising the first machine learning model, information sources comprising the selected information source from subgroups of the group of information sources, resulting in a group of selected information sources comprising the selected information source, and the subgroups of the group of information sources are associated with respective ones of the group of first machine learning models (Lewis fig. 1 and section 2.1 recite “In the RAG-Token model we can draw a different latent document for each target token and marginalize accordingly. This allows the generator to choose content from several documents when producing an answer. Concretely, the top K documents are retrieved using the retriever, and then the generator produces a distribution for the next output token for each document, before marginalizing, and repeating the process with the following output token” (i.e., subgroups of information sources are selected and analyzed by the first machine learning model)) . Regarding claim 8, the combination of Lewis and Luong teaches the system of claim 6, wherein the selected information source is a first selected information source, wherein the information in the selected information source is first information, and wherein the response formulation component constructs the human-readable response using the first information and second information in a second selected information source of the group of selected information sources (Lewis fig. 1 and section 2.1 recite “In the RAG-Token model we can draw a different latent document for each target token and marginalize accordingly. This allows the generator to choose content from several documents when producing an answer. Concretely, the top K documents are retrieved using the retriever, and then the generator produces a distribution for the next output token for each document, before marginalizing, and repeating the process with the following output token” (i.e., the human readable response is constructed using at least first and second information from multiple selected information sources)) . Regarding claim 9, the combination of Lewis and Luong teaches the system of claim 1, wherein respective ones of the group of information sources are of a source type selected from a group of source types comprising a text document, an image, a video, and an audio recording (the abstract of Lewis recites “We explore a general-purpose fine-tuning recipe for retrieval-augmented generation (RAG)—models which combine pre-trained parametric and non-parametric memory for language generation”. Lewis section I para. 2 recites “The retriever (Dense Passage Retriever, henceforth DPR) provides latent documents conditioned on the input, and the seq2seq model (BART) then conditions on these latent documents together with the input to generate the output” (i.e., the information sources comprise text documents)) . Claim 10 is a method claim and its limitation is included in claim 1. The only difference is that claim 10 requires a method (Luong para. [0006] recites “One example aspect of the present disclosure is directed to a computer-implemented method to train a machine-learned language encoder model”) . Therefore, claim 10 is rejected for the same reasons as claim 1. Claim 11 is a method claim and its limitation is included in claim 2. Claim 11 is rejected for the same reasons as claim 2. Claim 12 is a method claim and its limitation is included in claim 3. Claim 12 is rejected for the same reasons as claim 3. Claim 14 is a method claim and its limitation is included in claim 6. Claim 14 is rejected for the same reasons as claim 6. Regarding claim 16, Lewis teaches in response to obtaining a question input and using a first machine learning model, selecting an information source from a group of information sources based on a context score assigned to the information source by the first machine learning model, resulting in a context information source, wherein the context score is representative of an amount of context information pertaining to the question input contained in the context information source (fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z”. (i.e., using a first machine learning model to select information sources that are determined to be relevant to an input prompt based on a top-K scores as described in the description of figure 1 and the context of previously analyzed tokens)) ; and using a second machine learning model that is not the first machine learning model, forming a human-readable response output to the question input by applying parameters of the second machine learning model to the question input and the context information (fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z”. Section 2.3 recites “The generator component p θ (yi|x, z, y1:i−1) could be modelled using any encoder-decoder. We use BART-large, a pre-trained seq2seq transformer with 400M parameters. To combine the input x with the retrieved content z when generating from BART, we simply concatenate them” (i.e., transforming the input prompt into a human readable response from a second machine learning model different from the first machine learning model)) . However, while one of ordinary skill in the art would recognize that the methods taught by Lewis would be implemented by a computer system, Lewis does not explicitly teach a non- transitory machine-readable medium comprising computer executable instructions that, when executed by a processor, facilitate performance of operations. Luong teaches a non-transitory machine-readable medium comprising computer executable instructions that, when executed by a processor, facilitate performance of operations (para. [0063] recites “The user computing device 102 includes one or more processors 112 and a memory 114. The one or more processors 112 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, a FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. The memory 114 can include one or more non-transitory computer-readable storage mediums, such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. The memory 114 can store data 116 and instructions 118 which are executed by the processor 112 to cause the user computing device 102 to perform operations” (i.e., a non-transitory computer readable medium)) . See claim 1 for motivation to combine. Claim 17 is a non-transitory machine-readable medium claim and its limitation is included in claim 2. Claim 17 is rejected for the same reasons as claim 2. Claim 18 is a non-transitory machine-readable medium claim and its limitation is included in claim 3. Claim 18 is rejected for the same reasons as claim 3. Regarding claim 20 the combination of Lewis and Luong teaches the non-transitory machine-readable medium of claim 16, wherein the context information source is a first context information source, wherein the group of information sources is a first group of information sources, wherein the context score is a first context score, wherein the amount of context information is a first amount (Lewis fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z” (i.e., using a first machine learning model to select information sources that are determined to be relevant to an input prompt based on a top-K scores as described in the description of figure 1 and the context of previously analyzed tokens)) , and wherein the operations further comprise: in further response to the question input and using a third machine learning model that is not the first machine learning model or the second machine learning model, selecting a second context information source from a second group of information sources based on a second context score assigned to the second context information source by the third machine learning model, wherein the second context score is representative of a second amount of the context information pertaining to the question input contained in the second context information source (Lewis fig. 1 and section 2 para. 1 recite “We explore RAG models, which use the input sequence x to retrieve text documents z and use them as additional context when generating the target sequence y. As shown in Figure 1, our models leverage two components: (i) a retriever pn (z|x) with parameters n that returns (top-K truncated) distributions over text passages given a query x and (ii) a generator pθ(yi|x, z, y1:i-1) parametrized by θ that generates a current token based on a context of the previous i-1 tokens y1:i-1, the original input x, and a retrieved passage z”. Lewis section 2.3 recites “The generator component p θ (yi|x, z, y1:i−1) could be modelled using any encoder-decoder. We use BART-large, a pre-trained seq2seq transformer with 400M parameters. To combine the input x with the retrieved content z when generating from BART, we simply concatenate them” (i.e., transforming the input prompt into a human readable response from a second machine learning model different from the first machine learning model. Examiner notes that as Lewis states that the steps associated with generating a response using generator component could be modelled using any encoder-decoder and further describes applying the model to multiple kinds of information sources (see at least section 1 para. 4), that one of ordinary skill in the art would recognize that a different “third” generator model could be utilized to analyze a second group of information sources)) . 07-21-aia AIA Claim s 4-5 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis et al (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, herein Lewis) in view of Luong et al (US 20210089724 A1, herein Luong), in further view of Gao et al (“Enabling Large Language Models to Generate Text with Citations”, herein Gao) . Regarding claim 4, the combination of Lewis and Luong teaches the system of claim 1, wherein the human-readable response is a first human-readable response (Lewis section 2.1 recites “the top K documents are retrieved using the retriever, and then the generator produces a distribution for the next output token for each document, before marginalizing, and repeating the process with the following output token” (i.e., a human readable response can be generated based on the input prompt to identify the selected information sources)) . However, the combination of Lewis and Luong does not explicitly teach wherein the response formulation component further generates a second human-readable response including an identification of the selected information sources. Gao teaches wherein the response formulation component further generates a second human-readable response including an identification of the selected information sources (Gao fig. 1 and section 2 para. 1 recite “Our task is formalized as follows: Given a query q and a corpus of text passages D, the system is required to return an output S, which consists of statements s1, ..., sn, and each statement si cites a list of passages Ci = {ci1, ci2, . . .}, where cij ϵ D” (i.e., the response formulation includes a second human readable response including an identification of the selected information sources)) . Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine these teachings by utilizing the source citation method from Gao to supplement the retrieval-augmented generation model from Lewis (as modified by Luong). Lewis and Gao are both directed to retrieval-augmented generation methods, but the text generator from Lewis does not teach the capability to let a user know where the generated text was sourced from. One of ordinary skill in the art would be motivated to extend the generator from Lewis with Gao’s ability to cite the retrieved sources that were relied upon by the generator to more clearly show a user how the generated text was created by the model. Regarding claim 5, the combination of Lewis, Luong, and Gao teaches the system of claim 4, wherein the input prompt is a first input prompt, and wherein the response formulation component generates the second human-readable response in response to a second input prompt received by the response formulation component subsequent to the first input prompt and based on the selected information source received from the context generation component in response to the first input prompt (Lewis section 2.1 recites “the top K documents are retrieved using the retriever, and then the generator produces a distribution for the next output token for each document, before marginalizing, and repeating the process with the following output token”. Gao section 4.1 recites “In our experiments, we explore simple, off-the-shelf retrievers: We use dense retrievers like GTR and DPR”. Gao section 4.3 recites “POSTCITE: We also provide a post-hoc cite option to models, where for each statement, we find the best matching passage among the top-100 retrieved passages using GTR and cite it” (i.e., a second human readable response can be generated based on the first input prompt to identify the selected information sources)) . Claim 13 is a method claim and its limitation is included in claim 4. Claim 13 is rejected for the same reasons as claim 4. Regarding claim 19, the combination of Lewis and Luong teaches the non-transitory machine-readable medium of claim 16, wherein the question input is a first question input, wherein the human-readable response output is first human-readable response output (Lewis section 2.1 recites “the top K documents are retrieved using the retriever, and then the generator produces a distribution for the next output token for each document, before marginalizing, and repeating the process with the following output token” (i.e., a human readable response can be generated based on the input prompt to identify the selected information sources)). However, the combination of Lewis and Luong does not explicitly teach wherein in response to a second question input that follows the first question input and using the second machine learning model, forming a second human-readable response output, the second human-readable response output comprising a location of the context information source within a data storage system. Gao teaches wherein in response to a second question input that follows the first question input and using the second machine learning model, forming a second human-readable response output, the second human-readable response output comprising a location of the context information source within a data storage system (fig. 1 and section 2 para. 1 recite “Our task is formalized as follows: Given a query q and a corpus of text passages D, the system is required to return an output S, which consists of statements s1, ..., sn, and each statement si cites a list of passages Ci = {ci1, ci2, . . .}, where cij ϵ D”. Section 4.1 recites “In our experiments, we explore simple, off-the-shelf retrievers: We use dense retrievers like GTR and DPR”. Gao section 4.3 recites “POSTCITE: We also provide a post-hoc cite option to models, where for each statement, we find the best matching passage among the top-100 retrieved passages using GTR and cite it” (i.e., the response formulation includes a second human readable response including an identification of the selected information sources)) . See claim 4 for motivation to combine . 07-21-aia AIA Claim s 7 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Lewis et al (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, herein Lewis) in view of Luong et al (US 20210089724 A1, herein Luong), in further view of Glass et al (“Re2G: Retrieve, Rerank, Generate”, herein Glass) . Regarding claim 7, the combination of Lewis and Luong teaches the system of claim 6. However, the combination of Lewis and Luong does not teach wherein a first one of the group of first machine learning models is of a first model type, and wherein a second one of the group of first machine learning models is of a second model type that is not the first model type. Glass teaches wherein a first one of the group of first machine learning models is of a first model type, and wherein a second one of the group of first machine learning models is of a second model type that is not the first model type (Glass fig. 3 and section I para. 4-5 recite “We build on this line of research, pioneered by REALM and RAG, and propose a new approach that we call Re2G (Retrieve, Rerank, Generate), which combines both neural initial retrieval and reranking into a BART-based sequence-to-sequence generation. There are two particular aspects on which our approach is different from the previous works. Firstly, our reranking approach permits merging retrieval results from sources with incomparable scores, e.g. enabling an ensemble of BM25 and neural initial retrieval” (i.e., a second retrieval machine learning model from a group can be a different model type from a first retrieval model in the group)) Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine these teachings by modifying the retriever model from Lewis (as modified by Luong) to utilize the ensemble retriever from Glass. Glass and Lewis are both directed to retrieval-augmented generation methods for natural language generation and question answering. One of ordinary skill in the art would be motivated to apply the ensemble retriever from Glass to expand the kinds of sources available to the retriever from Lewis, as Glass states in at least section I, “our reranking approach permits merging retrieval results from sources with incomparable scores, e.g. enabling an ensemble of BM25 and neural initial retrieval”. Claim 15 is a method claim and its limitation is included in claim 7. Claim 15 is rejected for the same reasons as claim 7 . Conclusion 07-96 AIA The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. “Dense Passage Retrieval for Open-Domain Question Answering” (Karpukhin et al) teaches a dense passage retriever model for open domain question answering. “REALM: Retrieval-Augmented Language Model Pre-Training” (Guu et al) teaches a method for augmenting language model pre-training with a latent knowledge retriever, which allows the model to retrieve and attend over documents from a large corpus. “Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks” (Reimers et al) teaches a method for modification of the pretrained BERT network that uses Siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. Any inquiry concerning this communication or earlier communications from the examiner should be directed to LEAH M FEITL whose telephone number is (571) 272-8350. The examiner can normally be reached on M-F 0900-1700 EST. 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, Viker Lamardo can be reached on (571) 270-5871. 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. /L.M.F./ Examiner, Art Unit 2147 /VIKER A LAMARDO/Supervisory Patent Examiner, Art Unit 2147 Application/Control Number: 18/458,540 Page 2 Art Unit: 2147 Application/Control Number: 18/458,540 Page 3 Art Unit: 2147 Application/Control Number: 18/458,540 Page 4 Art Unit: 2147 Application/Control Number: 18/458,540 Page 5 Art Unit: 2147 Application/Control Number: 18/458,540 Page 6 Art Unit: 2147 Application/Control Number: 18/458,540 Page 7 Art Unit: 2147 Application/Control Number: 18/458,540 Page 8 Art Unit: 2147 Application/Control Number: 18/458,540 Page 9 Art Unit: 2147 Application/Control Number: 18/458,540 Page 10 Art Unit: 2147 Application/Control Number: 18/458,540 Page 11 Art Unit: 2147 Application/Control Number: 18/458,540 Page 12 Art Unit: 2147 Application/Control Number: 18/458,540 Page 13 Art Unit: 2147 Application/Control Number: 18/458,540 Page 14 Art Unit: 2147 Application/Control Number: 18/458,540 Page 15 Art Unit: 2147 Application/Control Number: 18/458,540 Page 16 Art Unit: 2147 Application/Control Number: 18/458,540 Page 17 Art Unit: 2147
Prosecution Insights