DETAILED ACTION
This communication is in response to amendments and arguments filed on January 26th, 2026. Claims 6, 13 and 20 have been cancelled. Claims 1-4, 8-11, and 15-18 are pending and have been examined.
Any objection/rejections not mentioned in this communication have been withdrawn by the Examiner.
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 .
Response to Amendments and Arguments
Regarding the rejections made under 35 U.S.C. 103, Applicant argues, “Cheng fails to disclose or suggest that the input-specific subnetworks includes a generalist model, where the each of the specific subnetworks and the generalist model share the same model parameters. That is, Cheng fails to disclose or suggest ‘the knowledge selector determines a sequence-to-expert assignment in a mixture-of-experts (MoE) architecture, the MoE architecture comprising a plurality of expert semi-parametric language models and at least one generalist semi-parametric language model, wherein each semi-parametric language model shares the same model parameters,’ as recited in claim 1,” (page 10 of Applicant’s Remarks).
Examiner respectfully disagrees. Cheng teaches an architecture wherein both shared and specialized transformer blocks are used to process questions and passages; “the transformer-based text encoder described herein includes at least one ‘shared’ transformer block, and at least one ‘specialized’ transformer block. The shared transformer block is trained both for questions and passages, while the specialized transformer block includes different input-specific subnetworks used for different encoding tasks,” and, “Use of shared and specialized transformer blocks in this manner provides the technical benefit of reducing consumption of computer resources, as at least some network parameters are shared between the encoding operations applied to questions and passages.”
Cheng’s system makes use of specialized and non-specialized models to process queries and return results. Accordingly, the rejections under 35 U.S.C. 103 are maintained in view of Cheng. Further details are provided below.
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, 8-9, 11, 15-16 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over "Augmenting Transformers with KNN-Based Composite Memory for Dialog" by Fan et al. (hereinafter, "Fan") in view of U.S. Patent Application Publication 2024/0126993 to Cheng et al. (hereinafter, "Cheng") and in view of "An Efficient Memory-Augmented Transformer for Knowledge-Intensive NLP Tasks" by Wu et al. (hereinafter, "Wu").
Regarding claims 1, 8 and 15, Fan teaches a method, an apparatus and a computer-readable medium comprising: receiving an input comprising natural language texts (Fan section 3 KNN-Based Information Fetching Modules, "Broadly, the KIF module assumes an encoder model M can access inputs
X
=
{
x
1
,
x
2
,
.
.
.
,
x
n
}
. For example, X can be a collection of sentences, and xi represents an individual sentence.");
retrieving one or more helpful knowledge pieces from the selected knowledge category […], wherein retrieving the one or more helpful knowledge pieces from the selected knowledge category comprises:
converting the one or more helpful knowledge pieces into natural language sentences as values (Fan section 3 KNN-Based Information Fetching Modules, "However, in many tasks, additional information is present, represented as
E
=
{
e
1
,
e
2
,
.
.
.
,
e
m
}
. We encode each element of X and E into a vector representation using the encoder. To identify the closest information in E that is relevant to xi, our general approach will be to use KNN by comparing the representation of xi with the representation of each element in the set E.");
augmenting the input using the one or more helpful knowledge pieces (Fan section 3 KNN-Based Information Fetching Modules, "Then, the KIF module output for an element xi is the set of all re-encoded representations of the KNN-retrieved knowledge:
K
I
F
x
i
=
M
'
e
e
∈
K
N
N
i
These elements are weighted by their normalized nearest neighbor scores and then summed. This is subsequently concatenated to
M
'
(
x
i
)
to form the final encoder output:");
feeding the augmented input into a parametric backbone text-to-text model that is a text-to-text Transfer Transformer (T5) or Bidirectional and Auto-Regressive Transformer (BART) (Fan section 3 KNN-Based Information Fetching Modules, "This concatenation represents the output of the encoder M′ and can be used for various purposes, such as providing the encoder output to a decoder in a sequence to sequence model," and Wu section 2 Efficient OQDA Systems, "One simple approach to accelerate ODQA is Closed-Book QA (CBQA) – a sequence-to-sequence model (Sutskever et al., 2014; Kalchbrenner et al., 2014) such as T5 (Raffel et al., 2020) or BART (Lewis et al., 2020a) is fine-tuned on ODQA data, by training it to produce the answer given the question."); and
generating an output answer based on the text-to-text model (Fan section 4.1 KIF for Generative Dialog, "Standard generative dialog models use a Transformer neural network as the encoder M and want to produce an output that is an appropriate response to the conversation.").
Fan does not explicitly teach the use of sub-models or “selecting, via a knowledge selector, one of a plurality of knowledge categories from an external memory based on a context of the input, each knowledge category associated with a respective semi-parametric language model, wherein the knowledge selector determine a sequence-to-expert assignment in a mixture-of-experts (MoE) architecture, the MoE architecture comprising a plurality of expert semi-parametric language models and at least one generalist semi-parametric language model, wherein each semi-parametric language model shares the same model parameters,” and thus, Cheng is introduced.
Cheng teaches selecting, via a knowledge selector, one of a plurality of knowledge categories from an external memory based on a context of the input, each knowledge category associated with a respective semi-parametric language model, wherein the knowledge selector determine a sequence-to-expert assignment in a mixture-of-experts (MoE) architecture, the MoE architecture comprising a plurality of expert semi-parametric language models and at least one generalist semi-parametric language model (Cheng paragraph [0037], "In general, a specialized transformer block includes two or more input-specific subnetworks, and a routing function to select an input-specific subnetwork of the two or more input-specific subnetworks for each of the computer-readable text representations. In FIG. 4B, the specialized transformer block includes a routing function 408, a question-specific subnetwork 410Q, and a passage specific subnetwork 410P. In some embodiments, the specialized transformer block can be described as using a mixture-of-experts (MOE) approach, in which multiple expert FFN sub-layers are used in parallel, and a routing function is used to choose among these expert FFN sub-layers," and paragraph [0026], "The plurality of transformer blocks of the transformer-based text encoder include at least a shared transformer block 208 and a specialized transformer block 210. The shared transformer block is trained for both the computer-readable question representations and the computer-readable passage representations. The specialized transformer block includes two or more different input-specific subnetworks—e.g., one question-specific subnetwork and one passage-specific subnetwork."), wherein each semi-parametric language model shares the same model parameters (Cheng paragraph [0019], "Use of shared and specialized transformer blocks in this manner provides the technical benefit of reducing consumption of computer resources, as at least some network parameters are shared between the encoding operations applied to questions and passages. This is contrary to bi-encoder architectures, in which the question encoder and passage encoder are separately parameterized.").
Fan and Cheng are considered analogous because they are each concerned with natural language question answering. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified the knowledge retrieval of Fan with the mixture of experts as taught by Cheng for the purpose of improving language model response efficiency or answer validity. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
The combination of Fan and Cheng does not teach “wherein retrieving the one or more helpful knowledge pieces from the selected knowledge category comprises:… encoding the natural language sentences into dense vectors as keys using a sentence encoder, the keys corresponding to utterances with background context as corresponding values,” and thus Wu is introduced.
Wu teaches encoding the natural language sentences into dense vectors as keys using a sentence encoder, the keys corresponding to utterances with background context as corresponding values (Wu section 3.1 Key-Value Memory, "The key-value memory M = (K, V) contains representations of keys K and values V, with each key ki mapping to one value vi. Since we use PAQ (Lewis et al., 2021b) as our knowledge source, each key represents a question, and its value represents the corresponding answer.").
Fan, Cheng and Wu are considered analogous because they are each concerned with question answering. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Fan and Cheng with the teachings of Wu for the purpose of improving language model response efficiency. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Regarding claims 2, 9 and 16, Fan teaches a method, apparatus and computer-readable medium wherein both the input and the output answer are in natural language forms after prompting (section 1 Introduction, "In this work, we focus on human–machine dialog and how to efficiently retrieve external knowledge that is relevant to the dialog."). The teachings of Fan are directed to human-machine dialog wherein a user is engaged in conversation, characterizing the utterances therein as natural language, and therefore the teachings read on the claim limitation.
Regarding claims 4, 11 and 18, Fan teaches a method, apparatus and computer-readable medium further comprising adapting the knowledge selector based on an input instance (section 3 KNN-Based Information Fetching Modules, "To identify the closest information in E that is relevant to xi, our general approach will be to use KNN by comparing the representation of xi with the representation of each element in the set E. KNN is a fully differentiable operation (Plötz and Roth, 2018), so can be incorporated in a straightforward way into neural models. The most relevant information in E will then be available in the model."). The teachings of Fan illustrate that the knowledge selections are dependent on the given input at any execution, and therefore the teachings read on the claim limitation.
Claims 3, 10 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Fan, Cheng and Wu as applied to claims 1, 8 and 15 above, and further in view of U.S. Patent Application Publication 2023/0092736 to Jiang et al. (hereinafter, "Jiang").
Regarding claims 3, 10 and 17, The combination of Fan, Cheng and Wu does not teach “the plurality of knowledge categories comprises an entity knowledge category, a dictionary knowledge category, a commonsense knowledge category, an event knowledge category, a script knowledge category, and a causality knowledge category,” and thus, Jiang is introduced.
Jiang teaches the plurality of knowledge categories comprises an entity knowledge category, a dictionary knowledge category, a commonsense knowledge category, an event knowledge category, a script knowledge category, and a causality knowledge category (paragraph [0050], "A specific number of the understanding module layers included in the multilayer understanding model in the present embodiment may be predetermined before a training operation, and may be, for example, 3, 5, or another positive integer greater than or equal to 2. A number of the knowledge bases may be greater than or equal to 2, and the knowledge bases may be divided according to fields, such as a basic language field, a communication field, a radio field, a finance field, a sports field, an entertainment field, or the like. Certainly, the knowledge bases may be divided according to other types, which are not repeated herein.").
Fan, Cheng, Wu and Jiang are considered analogous because they are each concerned with question answering. Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to have modified Fan, Cheng and Wu with the teachings of Jiang for the purpose of improving language model response accuracy. Given that all the claimed elements were known in the prior art, one skilled in the art could have combined the elements by known methods with no change in their respective functions, and the combination would have yielded nothing more than predictable results.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
U.S. Patent Application Publication 2013/0218564 to Cox et al.
U.S. Patent Application Publication 2019/0034811 to Cuddihy et al.
U.S. Patent Application Publication 2019/0205726 to Khabiri et al.
U.S. Patent Application Publication 2021/0374168 to Srinivasan et al.
U.S. Patent Application Publication 2021/0144107 to Liang et al.
U.S. Patent Application Publication 2022/0300570 to Ostadzadeh et al.
U.S. Patent Application Publication 2023/0229960 to Zhu et al.
U.S. Patent Application Publication 2023/0394333 to Jiao et al.
U.S. Patent Application Publication 2024/0073159 to Bhatia et al.
U.S. Patent 11,989,214 to Osmon et al.
International Publication WO 2021/248125 to Riquelme Ruiz et al.
China Invention Application 112084314 to Zhu et al.
“WiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representations” by Pilehvar and Camacho-Collados.
“Improving Question Answering with External Knowledge” by Pan et al.
“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” by Lewis et al.
“Retrieval Augmented Language Model Pre-Training” by Guu et al.
“Dense Passage Retrieval for Open-Domain Question Answering” by Karpukhin et al.
“Improving Sequential Recommendation with Knowledge-Enhanced Memory Networks” by Huang et al.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SEAN T SMITH whose telephone number is (571)272-6643. The examiner can normally be reached Monday - Friday 8:00am - 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, PIERRE-LOUIS DESIR can be reached at (571) 272-7799. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/SEAN THOMAS SMITH/Examiner, Art Unit 2659
/PIERRE LOUIS DESIR/Supervisory Patent Examiner, Art Unit 2659
Prosecution Insights