DETAILED ACTION
This communication is in response to the Application No. 18/655,184 filed on May 03, 2024
in which Claims 1 – 20 are presented for examination.
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 .
Double Patenting
The nonstatutory double patenting rejection is based on a judicially created doctrine grounded in public policy (a policy reflected in the statute) so as to prevent the unjustified or improper timewise extension of the “right to exclude” granted by a patent and to prevent possible harassment by multiple assignees. A nonstatutory double patenting rejection is appropriate where the conflicting claims are not identical, but at least one examined application claim is not patentably distinct from the reference claim(s) because the examined application claim is either anticipated by, or would have been obvious over, the reference claim(s). See, e.g., In re Berg, 140 F.3d 1428, 46 USPQ2d 1226 (Fed. Cir. 1998); In re Goodman, 11 F.3d 1046, 29 USPQ2d 2010 (Fed. Cir. 1993); In re Longi, 759 F.2d 887, 225 USPQ 645 (Fed. Cir. 1985); In re Van Ornum, 686 F.2d 937, 214 USPQ 761 (CCPA 1982); In re Vogel, 422 F.2d 438, 164 USPQ 619 (CCPA 1970); In re Thorington, 418 F.2d 528, 163 USPQ 644 (CCPA 1969).
A timely filed terminal disclaimer in compliance with 37 CFR 1.321(c) or 1.321(d) may be used to overcome an actual or provisional rejection based on nonstatutory double patenting provided the reference application or patent either is shown to be commonly owned with the examined application, or claims an invention made as a result of activities undertaken within the scope of a joint research agreement. See MPEP § 717.02 for applications subject to examination under the first inventor to file provisions of the AIA as explained in MPEP § 2159. See MPEP § 2146 et seq. for applications not subject to examination under the first inventor to file provisions of the AIA . A terminal disclaimer must be signed in compliance with 37 CFR 1.321(b).
The filing of a terminal disclaimer by itself is not a complete reply to a nonstatutory double patenting (NSDP) rejection. A complete reply requires that the terminal disclaimer be accompanied by a reply requesting reconsideration of the prior Office action. Even where the NSDP rejection is provisional the reply must be complete. See MPEP § 804, subsection I.B.1. For a reply to a non-final Office action, see 37 CFR 1.111(a). For a reply to final Office action, see 37 CFR 1.113(c). A request for reconsideration while not provided for in 37 CFR 1.113(c) may be filed after final for consideration. See MPEP §§ 706.07(e) and 714.13.
The USPTO Internet website contains terminal disclaimer forms which may be used. Please visit www.uspto.gov/patent/patents-forms. The actual filing date of the application in which the form is filed determines what form (e.g., PTO/SB/25, PTO/SB/26, PTO/AIA /25, or PTO/AIA /26) should be used. A web-based eTerminal Disclaimer may be filled out completely online using web-screens. An eTerminal Disclaimer that meets all requirements is auto-processed and approved immediately upon submission. For more information about eTerminal Disclaimers, refer to www.uspto.gov/patents/apply/applying-online/eterminal-disclaimer.
Claims 1-2, 4-6, 8-11, and 18 are provisionally rejected on the ground of nonstatutory double patenting as being unpatentable over Claims 1-3, 7-10, and 17-18 of copending Application No. 18/655,187 in view of Akyurek et al. (a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”), and Claim 3 is provisionally rejected on the ground of nonstatutory double patenting as being unpatentable over Claim 8 of copending Application No. 18/655,187 in view of Akyurek et al. (a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) and in further view of Burns et al. (a non-patent literature reference titled “Discovering Latent Knowledge in Language Models Without Supervision”). Although the claims at issue are not identical, they are not patentably distinct from each other because the subject matter claimed in the instant application is disclosed in copending Application No. 18/655,187 since the instant application claims common subject matter.
This is a provisional nonstatutory double patenting rejection.
Note: The bolded portions below highlight the differences between the instant application and
the copending application, which illustrates the obvious and anticipatory relationship of the claim
limitations at issue:
Instant App. (18/655,184)
Copending App. (18/655,187)
Claim 1: A computer-implemented method for verifying responses to requests, the method comprising:
processing a first request via a trained machine learning model to generate a first response;
and performing one or more operations to verify the first response based on first data used to train the trained machine learning model.
Claim 7: The computer-implemented method of claim 1, further comprising:
processing a first request via the trained machine learning model to generate a first response;
and performing one or more operations to verify the first response based on the first data.
Copending Application 18/655,187 discloses the same limitations as the instant application except
that the instant application specifies performing one or more operations to verify the first response based on first data used to train the trained machine learning model, rather than
performing one or more operations to verify the first response based on first data. However, Akyurek et al. (a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) teaches performing one or more operations based on first data used to train the trained machine learning model. (Akyurek, Page 1 – Abstract, “identifying which training examples taught an LM to generate a particular factual assertion”, thus performing an operation based on first data used to train the trained machine learning model is disclosed, because Akyurek teaches identifying the training examples that taught the language model to generate a particular factual assertion).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the process of performing one or more operations to verify the first response, as disclosed by Copending Application 18/655,187, so that the first data comprises data used to train the trained machine learning model, as disclosed by Akyurek. A POSTIA would have been motivated to apply this modification so that performing the one or more operations to verify the first response is based on first data used to train the trained machine learning model, thereby enabling the first response to be traced back to the training examples that taught the language model to generate the corresponding factual assertion (Akyurek, Page 1 – Abstract, “But when an LM generates an assertion, it is often difficult to determine where it learned this information and whether it is true. In this paper, we propose the problem of fact tracing: identifying which training examples taught an LM to generate a particular factual assertion)
Claim 2: The computer-implemented method of claim 1, wherein performing the one or more operations to verify the first response comprises:
generating a first embedding based on the first response;
generating a second embedding based on the first data;
and computing a similarity between the first embedding and the second embedding.
Claim 8: The computer-implemented method of claim 7, wherein performing the one or more operations to verify the first response comprises:
generating a first embedding based on the first response;
generating a second embedding based on the first data;
and computing a similarity between the first embedding and the second embedding.
Claim 3: The computer-implemented method of claim 1, wherein performing the one or more operations to verify the first response comprises:
generating a first embedding based on a negation of the first request and the first response;
generating a second embedding based on the first data;
and computing a similarity between the first embedding and the second embedding.
Claim 8: The computer-implemented method of claim 7, wherein performing the one or more operations to verify the first response comprises:
generating a first embedding based on the first response;
generating a second embedding based on the first data;
and computing a similarity between the first embedding and the second embedding.
Copending Application 18/655,187 discloses the same limitations as the instant application except
that the instant application specifies generating a first embedding based on a negation of the first request and the first response, rather than generating a first embedding based on the first response. However, Burns et al. (a non-patent literature reference titled “Discovering Latent Knowledge in Language Models Without Supervision”) teaches generating a first embedding based on a negation of the first request and the first response. (Burns, Page 1 – Abstract, “finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values,” & Page 1 – Section 1, “Contrast-Consistent Search (CCS), a method that learns a linear projection of the hidden states that is consistent across negations,” & Page 3 – Section 2.2, “CCS works by (1) answering each question qi as both “Yes” (x + i ) and “No” (x − i ), (2) computing the representations φ(x + i ) and φ(x − i ) of each answer, (3) mapping the answer representations to probabilities p + i and p − i of being true, then (4) optimizing that mapping so that the probabilities are both consistent and confident,” & “for each question qi , we answer qi both as “Yes”, resulting in the new natural language statement x + i , and as “No”, resulting in the natural language statement x − i,” thus generating a first embedding based on a negation of the first request and the first response is disclosed because Burns teaches forming opposite natural language statements based on a question and its possible answers, where the statements are negations of one another and have opposite truth values, and computing representations of those statement).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the process of generating a first embedding based on the first request and the first response, as disclosed by Copending Application 18/655,187, so that the first embedding can be generated based on a negation of the first request, as disclosed by Burns. A POSITA would have been motivated to apply this modification so that a first embedding can generated based on a negation of the first request and the first response and compared with the second embedding based on the first data, thereby allowing the verification process to leverage the logical consistency property that a statement and its negation have opposite truth values and to maintain accurate verification even when the language model generates an incorrect response (Burns, Page 1 – Abstract, “It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question answering datasets, it outperforms zero-shot accuracy by 4% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers”)
Claim 4: The computer-implemented method of claim 1, wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response.
Claim 9: The computer-implemented method of claim 7, wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data and at least one of the first request or the first response.
Claim 5: The computer-implemented method of claim 1, wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data and the first response.
Claim 9: The computer-implemented method of claim 7, wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data and at least one of the first request or the first response.
Claim 6: The computer-implemented method of claim 1, wherein performing the one or more operations to verify the first response comprises:
processing the first response via another trained machine learning model to generate a second response that indicates whether the first response is verified.
Claim 10: The computer-implemented method of claim 7, wherein performing the one or more operations to verify the first response comprises:
processing the first response via another trained machine learning model to generate a second response that indicates whether the first response is verified.
Claim 8: The computer-implemented method of claim 1, further comprising:
appending context information to one or more portions of the first data to generate second data;
and performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model.
Claim 1: A computer-implemented method for training a machine learning model, the method comprising:
appending context information to at least one portion of first data to generate second data;
and performing one or more operations to train the machine learning model based on the second data to generate a trained machine learning model.
Claim 9: The computer-implemented method of claim 1, wherein the context information that is appended to each portion of the one or more portions comprises a token indicating one or more contexts associated with the portion.
Claim 2: The computer-implemented method of claim 1, wherein the context information appended to each portion included in the at least one portion comprises a token indicating one or more contexts associated with the portion.
Claim 10: The computer-implemented method of claim 1, wherein the context information that is appended to each portion of the one or more portions comprises a hierarchy of one or more contexts associated with the portion.
Claim 3: The computer-implemented method of claim 1, wherein the context information appended to each portion included in the at least one portion indicates a hierarchy of one or more contexts associated with the portion.
Claim 11: One or more non-transitory computer readable media including instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of:
processing a first request via a trained machine learning model to generate a first response;
and performing one or more operations to verify the first response based on first data used to train the trained machine learning model.
Claim 17: The one or more non-transitory computer readable media of claim 11, wherein the instructions, when executed by the one or more processors, further cause the one or more processors to perform the steps of:
processing a request via the trained machine learning model to generate a response;
and performing one or more operations to verify the response based on the first data.
Copending Application 18/655,187 discloses the same limitations as the instant application except
that the instant application specifies performing one or more operations to verify the first response based on first data used to train the trained machine learning model, rather than
performing one or more operations to verify the first response based on first data. However, Akyurek et al. (a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) teaches performing one or more operations based on first data used to train the trained machine learning model. (Akyurek, Page 1 – Abstract, “identifying which training examples taught an LM to generate a particular factual assertion”, thus performing an operation based on first data used to train the trained machine learning model is disclosed, because Akyurek teaches identifying the training examples that taught the language model to generate a particular factual assertion).
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the process of performing one or more operations to verify the first response, as disclosed by Copending Application 18/655,187, so that the first data comprises data used to train the trained machine learning model, as disclosed by Akyurek. A POSTIA would have been motivated to apply this modification so that performing the one or more operations to verify the first response is based on first data used to train the trained machine learning model, thereby enabling the first response to be traced back to the training examples that taught the language model to generate the corresponding factual assertion (Akyurek, Page 1 – Abstract, “But when an LM generates an assertion, it is often difficult to determine where it learned this information and whether it is true. In this paper, we propose the problem of fact tracing: identifying which training examples taught an LM to generate a particular factual assertion)
Claim 18: The one or more non-transitory computer readable media of claim 11, wherein the trained machine learning model comprises a language model.
Claim 18: The one or more non-transitory computer readable media of claim 11, wherein the trained machine learning model comprises a language model.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claim 1-20 are rejected under 35 U.S.C. 101 because these claimed inventions are directed to an abstract idea without significantly more.
Regarding Claim 1:
Step 1: Claim 1 is a method type claim. Therefore, Claims 1-10 fall within one of the four statutory
categories (i.e., process, machine, manufacture, or composition of matter).
2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance
of the limitation in the mind but for the recitation of generic computer components, then it falls within
the “Mental Processes” grouping of abstract ideas. If a claim limitation, under its broadest reasonable
interpretation, covers performance of the limitation by mathematical calculation but for the recitation
of generic computer components, then it falls within the “Mathematical Concepts” grouping of abstract
ideas.
processing a first request […] to generate a first response (mental process – processing a first request to generate a first response may be performed mentally or using pen and paper by a user observing/analyzing the first request, identifying the information sought by the first request, recalling or looking up information relevant to the first request, and accordingly using judgment/evaluation to formulate an answer as the first response based on said analysis)
performing one or more operations to verify the first response based on first data used to train the trained machine learning model (mental process – performing one or more operations to verify the first response based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the first response, reviewing the first data, comparing the content of the first response to the content of the first data, and accordingly using judgment/evaluation to determine whether the first response is supported by the first data based on said analysis)
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
Step 2B: The claim does not include additional elements considered individually and in combination that
are sufficient to amount to significantly more than the judicial exception.
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
For the reasons above, Claim 1 is rejected as being directed to an abstract idea without significantly more. This rejection applies equally to dependent claims 1 - 10. The additional limitations of the dependent claims are addressed below.
Regarding Claim 2:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 2 depends on.
generating a first embedding based on the first response (mental process – generating a first embedding based on the first response may be performed mentally or using pen and paper by a user observing/analyzing the words of the first response, identifying the terms appearing in the first response, and accordingly using judgment/evaluation to record a numerical representation of the first response (e.g., a list of values indicating which terms of a vocabulary appear in the first response and how often) based on said analysis)
generating a second embedding based on the first data (mental process - generating a second embedding based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the words of the first data, identifying the terms appearing in the first data, and accordingly using judgment/evaluation to record a numerical representation of the first data (e.g., a list of values indicating which terms of the vocabulary appear in the first data and how often) based on said analysis)
computing a similarity between the first embedding and the second embedding (mental process - computing a similarity between the first embedding and the second embedding may be performed mentally or using pen and paper by a user observing/analyzing the numerical values of the first embedding and the second embedding, comparing the corresponding values of the two embeddings, and accordingly using judgment/evaluation to determine a degree of similarity between the first response and the first data based on said comparison)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 3:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 3 depends on.
generating a first embedding based on a negation of the first request and the first response (mental process – generating a first embedding based on a negation of the first request and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first request, rewriting the first request in negated form (e.g., inserting "not" into the question), combining the negated request with the first response, identifying the terms appearing in the combined text, and accordingly using judgment/evaluation to record a numerical representation of the combined text based on said analysis)
generating a second embedding based on the first data (mental process - generating a second embedding based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the words of the first data, identifying the terms appearing in the first data, and accordingly using judgment/evaluation to record a numerical representation of the first data (e.g., a list of values indicating which terms of the vocabulary appear in the first data and how often) based on said analysis)
computing a similarity between the first embedding and the second embedding (mental process - computing a similarity between the first embedding and the second embedding may be performed mentally or using pen and paper by a user observing/analyzing the numerical values of the first embedding and the second embedding, comparing the corresponding values of the two embeddings, and accordingly using judgment/evaluation to determine a degree of similarity between the first response and the first data based on said comparison)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 4:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 4 depends on.
wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response (mental process – computing an entailment score based on the first data, the first request, and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first data, the first request, and the first response, comparing the content of the first request and the first response to the content of the first data, determining whether the first data supports (entails), contradicts, or is neutral with respect to the first request and the first response, and accordingly using judgment/evaluation to assign a score indicating the extent of that relationship based on said analysis)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 5:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 5 depends on.
wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data and the first response (mental process – computing an entailment score based on the first data and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first data and the first response, comparing the content of the first response to the content of the first data, determining whether the first data supports (entails), contradicts, or is neutral with respect to the first response, and accordingly using judgment/evaluation to assign a score indicating the extent of that relationship based on said analysis)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 6:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 6 depends on.
processing the first response […] to generate a second response that indicates whether the first response is verified (mental process – processing the first response to generate a second response that indicates whether the first response is verified may be performed mentally or using pen and paper by a user observing/analyzing the first response, evaluating whether the content of the first response is accurate, and accordingly using judgment/evaluation to record a second response indicating whether the first response is verified based on said analysis)
Step 2A Prong 2 & Step 2B:
[…] via another trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the
abstract idea into practical application because it does not impose any meaningful limits on practicing
the abstract idea, as discussed above in the rejection of claim 1. The claim does not include additional
elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 7:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 7 depends on.
Step 2A Prong 2 & Step 2B:
displaying the first response and an indication of whether the first response is verified (Adding insignificant extra-solution activity to the judicial exception – see MPEP 2106.05(g). Further, MPEP 2106.05(d)(II) indicates that "Presenting offers and gathering statistics" is a well-understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim). Thereby, a conclusion that the claimed limitation is well-understood, routine, conventional activity is supported under Berkheimer)
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 1. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 8:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 8 depends on.
appending context information to one or more portions of the first data to generate second data (mental process – appending context information to one or more portions of the first data to generate second data may be performed mentally or using pen and paper by a user observing/analyzing the portions of the first data (e.g., sentences of a book), identifying the source or context from which each portion originates (e.g., the book and the collection or genre to which the book belongs), and accordingly using judgment/evaluation to write the corresponding context information at the end of each portion, thereby generating the second data based on said analysis)
Step 2A Prong 2 & Step 2B:
performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: recited at a high-level of generality (i.e., generic training of a generic machine learning model, with no detail as to the training technique, the model architecture, or how the appended context information is used during training) such that it amounts to no more than mere instructions to apply the exception using a generic machine learning model as a tool)
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the
abstract idea into practical application because it does not impose any meaningful limits on practicing
the abstract idea, as discussed above in the rejection of claim 1. The claim does not include additional
elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 9:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 9 depends on.
Step 2A Prong 2 & Step 2B:
wherein the context information that is appended to each portion of the one or more portions comprises a token indicating one or more contexts associated with the portion (Field of Use – limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception does not amount to significantly more than the exception itself, and cannot integrate a judicial exception into a practical application; in this case specifying that the context information that is appended to each portion of the one or more portions comprises a token indicating one or more contexts associated with the portion does not integrate the exception into a practical application nor amount to significantly more – See MPEP 2106.05(h))
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 1. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 10:
Step 2A Prong 1: See the rejection of Claim 1 above, which Claim 10 depends on.
Step 2A Prong 2 & Step 2B:
wherein the context information that is appended to each portion of the one or more portions comprises a hierarchy of one or more contexts associated with the portion (Field of Use – limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception does not amount to significantly more than the exception itself, and cannot integrate a judicial exception into a practical application; in this case specifying that the context information that is appended to each portion of the one or more portions comprises a hierarchy of one or more contexts associated with the portion does not integrate the exception into a practical application nor amount to significantly more – See MPEP 2106.05(h))
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 1. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 11:
Step 1: Claim 11 is a non-transitory computer readable media type claim. Therefore, Claims 11-19 fall within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance
of the limitation in the mind but for the recitation of generic computer components, then it falls within
the “Mental Processes” grouping of abstract ideas. If a claim limitation, under its broadest reasonable
interpretation, covers performance of the limitation by mathematical calculation but for the recitation
of generic computer components, then it falls within the “Mathematical Concepts” grouping of abstract
ideas.
processing a first request […] to generate a first response (mental process – processing a first request to generate a first response may be performed mentally or using pen and paper by a user observing/analyzing the first request, identifying the information sought by the first request, recalling or looking up information relevant to the first request, and accordingly using judgment/evaluation to formulate an answer as the first response based on said analysis)
performing one or more operations to verify the first response based on first data used to train the trained machine learning model (mental process – performing one or more operations to verify the first response based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the first response, reviewing the first data, comparing the content of the first response to the content of the first data, and accordingly using judgment/evaluation to determine whether the first response is supported by the first data based on said analysis)
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
One or more processors (recited at a high-level of generality (i.e., a generic processor, computer-readable storage medium, a communication interface, a user interface and memory) such that it amounts to no more than mere instructions to apply the exception using generic computer components)
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
Step 2B: The claim does not include additional elements considered individually and in combination that
are sufficient to amount to significantly more than the judicial exception.
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
For the reasons above, Claim 11 is rejected as being directed to an abstract idea without significantly more. This rejection applies equally to dependent claims 11 - 19. The additional limitations of the dependent claims are addressed below.
Regarding Claim 12:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 12 depends on.
generating a first embedding based on the first response (mental process – generating a first embedding based on the first response may be performed mentally or using pen and paper by a user observing/analyzing the words of the first response, identifying the terms appearing in the first response, and accordingly using judgment/evaluation to record a numerical representation of the first response (e.g., a list of values indicating which terms of a vocabulary appear in the first response and how often) based on said analysis)
generating a second embedding based on the first data (mental process - generating a second embedding based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the words of the first data, identifying the terms appearing in the first data, and accordingly using judgment/evaluation to record a numerical representation of the first data (e.g., a list of values indicating which terms of the vocabulary appear in the first data and how often) based on said analysis)
computing a similarity between the first embedding and the second embedding (mental process - computing a similarity between the first embedding and the second embedding may be performed mentally or using pen and paper by a user observing/analyzing the numerical values of the first embedding and the second embedding, comparing the corresponding values of the two embeddings, and accordingly using judgment/evaluation to determine a degree of similarity between the first response and the first data based on said comparison)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 13:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 13 depends on.
generating a first embedding based on a negation of the first request and the first response (mental process – generating a first embedding based on a negation of the first request and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first request, rewriting the first request in negated form (e.g., inserting "not" into the question), combining the negated request with the first response, identifying the terms appearing in the combined text, and accordingly using judgment/evaluation to record a numerical representation of the combined text based on said analysis)
generating a second embedding based on the first data (mental process - generating a second embedding based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the words of the first data, identifying the terms appearing in the first data, and accordingly using judgment/evaluation to record a numerical representation of the first data (e.g., a list of values indicating which terms of the vocabulary appear in the first data and how often) based on said analysis)
computing a similarity between the first embedding and the second embedding (mental process - computing a similarity between the first embedding and the second embedding may be performed mentally or using pen and paper by a user observing/analyzing the numerical values of the first embedding and the second embedding, comparing the corresponding values of the two embeddings, and accordingly using judgment/evaluation to determine a degree of similarity between the first response and the first data based on said comparison)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 14:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 14 depends on.
wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response (mental process – computing an entailment score based on the first data, the first request, and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first data, the first request, and the first response, comparing the content of the first request and the first response to the content of the first data, determining whether the first data supports (entails), contradicts, or is neutral with respect to the first request and the first response, and accordingly using judgment/evaluation to assign a score indicating the extent of that relationship based on said analysis)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 15:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 15 depends on.
wherein performing the one or more operations to verify the first response comprises computing an entailment score based on the first data and the first response (mental process – computing an entailment score based on the first data and the first response may be performed mentally or using pen and paper by a user observing/analyzing the first data and the first response, comparing the content of the first response to the content of the first data, determining whether the first data supports (entails), contradicts, or is neutral with respect to the first response, and accordingly using judgment/evaluation to assign a score indicating the extent of that relationship based on said analysis)
Step 2A Prong 2 & Step 2B:
Accordingly, under Step 2A Prong 2 and Step 2B, there are no additional elements that integrate the abstract idea into practical application. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 16:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 16 depends on.
processing the first response […] to generate a second response that indicates whether the first response is verified (mental process – processing the first response to generate a second response that indicates whether the first response is verified may be performed mentally or using pen and paper by a user observing/analyzing the first response, evaluating whether the content of the first response is accurate, and accordingly using judgment/evaluation to record a second response indicating whether the first response is verified based on said analysis)
Step 2A Prong 2 & Step 2B:
[…] via another trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the
abstract idea into practical application because it does not impose any meaningful limits on practicing
the abstract idea, as discussed above in the rejection of claim 11. The claim does not include additional
elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 17:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 17 depends on.
Step 2A Prong 2 & Step 2B:
wherein the trained machine learning model comprises an artificial neural network (Field of Use – limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception does not amount to significantly more than the exception itself, and cannot integrate a judicial exception into a practical application; in this case specifying that the trained machine learning model comprises an artificial neural network does not integrate the exception into a practical application nor amount to significantly more – See MPEP 2106.05(h))
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 11. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 18:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 18 depends on.
Step 2A Prong 2 & Step 2B:
wherein the trained machine learning model comprises a language model (Field of Use – limitations that amount to merely indicating a field of use or technological environment in which to apply a judicial exception does not amount to significantly more than the exception itself, and cannot integrate a judicial exception into a practical application; in this case specifying that the trained machine learning model comprises a language model does not integrate the exception into a practical application nor amount to significantly more – See MPEP 2106.05(h))
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 11. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 19:
Step 2A Prong 1: See the rejection of Claim 11 above, which Claim 19 depends on.
Step 2A Prong 2 & Step 2B:
displaying the first response and an indication of whether the first response is verified, and an indication of the first data (Adding insignificant extra-solution activity to the judicial exception – see MPEP 2106.05(g). Further, MPEP 2106.05(d)(II) indicates that "Presenting offers and gathering statistics" is a well-understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim). Thereby, a conclusion that the claimed limitation is well-understood, routine, conventional activity is supported under Berkheimer)
Accordingly, under Step 2A Prong 2 and Step 2B, this additional element does not integrate the abstract idea into practical application because it does not impose any meaningful limits on practicing the abstract idea, as discussed above in the rejection of claim 11. The claim does not include additional elements considered individually and in combination that are sufficient to amount to significantly more than the judicial exception.
Regarding Claim 20:
Step 1: Claim 20 is a system type claim. Therefore, Claim 20 falls within one of the four statutory categories (i.e., process, machine, manufacture, or composition of matter).
2A Prong 1: If a claim limitation, under its broadest reasonable interpretation, covers performance
of the limitation in the mind but for the recitation of generic computer components, then it falls within
the “Mental Processes” grouping of abstract ideas. If a claim limitation, under its broadest reasonable
interpretation, covers performance of the limitation by mathematical calculation but for the recitation
of generic computer components, then it falls within the “Mathematical Concepts” grouping of abstract
ideas.
processing a first request […] to generate a first response (mental process – processing a first request to generate a first response may be performed mentally or using pen and paper by a user observing/analyzing the first request, identifying the information sought by the first request, recalling or looking up information relevant to the first request, and accordingly using judgment/evaluation to formulate an answer as the first response based on said analysis)
performing one or more operations to verify the first response based on first data used to train the trained machine learning model (mental process – performing one or more operations to verify the first response based on the first data may be performed mentally or using pen and paper by a user observing/analyzing the first response, reviewing the first data, comparing the content of the first response to the content of the first data, and accordingly using judgment/evaluation to determine whether the first response is supported by the first data based on said analysis)
Step 2A Prong 2: This judicial exception is not integrated into a practical application.
One or more memories […]; one or more processors […] (recited at a high-level of generality (i.e., a generic processor, computer-readable storage medium, a communication interface, a user interface and memory) such that it amounts to no more than mere instructions to apply the exception using generic computer components)
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
Step 2B: The claim does not include additional elements considered individually and in combination that
are sufficient to amount to significantly more than the judicial exception.
[…] via a trained machine learning model […] (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: high level recitation of using a trained machine learning model without significantly more)
For the reasons above, Claim 20 is rejected as being directed to an abstract idea without significantly more.
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, 6-7, 11-12, and 16-20 are rejected under 35 U.S.C. 103 as being unpatentable over Akyurek et al. (hereinafter Akyurek, a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) in view of Kryscinski et al. (hereinafter Kryscinski) (US 20210124876).
Regarding Claim 1, Akyurek teaches:
processing a first request via a trained machine learning model to generate a first response (Akyurek, Page 1 – Section 1, Figure 1,
PNG
media_image1.png
559
639
media_image1.png
Greyscale
, thus processing a first request via a trained machine learning model to generate a first response is disclosed because Akyurek teaches providing a query to a language model (LM) and generating a corresponding prediction in response to the query, as illustrated in Figure 1 by the query “The United States was founded in ___” and the prediction “1776”)
[…] based on first data used to train the trained machine learning model (Akyurek, Page 1 – Abstract, “identifying which training examples taught an LM to generate a particular factual assertion”, thus based on first data used to train the trained machine learning model is disclosed, because Akyurek teaches identifying the training examples that taught the language model to generate a particular factual assertion. Akyurek’s training examples correspond to the first data used to train the trained machine learning model, and the generated factual assertion corresponds to the first response)
Akyurek does not explicitly teach performing one or more operations to verify the first response […],
However, Kryscinski teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Akyurek’s teaching of processing a first request via a trained machine learning model to generate a first response and identifying the training examples that taught the language model to generate the first response, with Kryscinski’s teaching of performing one or more operations to verify generated text by determining whether the generated text is factually consistent with source data. Therefore, a POSITA would have been motivated to incorporate Kryscinski’s factual consistency verification into Akyurek’s training data attribution process so that one or more operations are performed to verify the first response based on first data used to train the trained machine learning model, thereby enabling the first response to be verified based on the identified training data and allowing factual inconsistencies in the generated response to be identified (Kryscinski, Par. [0020], “One such problem is that of verifying factual consistency between source documents and generated summaries: a factually consistent summary should contain only statements that are entailed by the source document. However, studies have shown that a substantial number of summaries generated by abstractive models contain factual inconsistencies. Such high levels of factual inconsistency render automatically generated summaries virtually useless in practice”)
Regarding Claim 2, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Akyurek further teaches:
generating a first embedding based on the first response (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a first embedding based on the first response is disclosed, because Akyurek teaches generating a single vector representation from the intermediate layer outputs of a Transformer language model and using the representation in determining the association with a model prediction. Akyurek’s model prediction corresponds to the first response, and the single vector representation associated with the model prediction corresponds to the first embedding based on the first response)
generating a second embedding based on the first data (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a second embedding based on the first data is disclosed, because Akyurek teaches generating a single vector representation for a training example and using the representation in determining the association between the training example and a model prediction. Akyurek’s training example corresponds to the first data, and the single vector representation of the training example corresponds to the second embedding based on the first data)
computing a similarity between the first embedding and the second embedding (Akyurek, Page 4 – Section 2.2, “Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus computing a similarity between the first embedding and the second embedding is disclosed, because Akyurek teaches determining the association between a training example and a model prediction using a cosine product. Therefore, the cosine product between the vector representation of the model prediction and the vector representation of the training example corresponds to computing a similarity between the first embedding and the second embedding)
Regarding Claim 6, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Kryscinski further teaches:
processing […the first response…] via another trained machine learning model to generate a second response that indicates whether […the first response…] is verified (Kryscinski, Par. [0063], “After training, the model (e.g., implementing factual consistency module 150) can be applied or used to check for factual consistency of text summarizations generated by one or more summarization models for respective source documents,” & Par. [0065], “the factual consistency neural network model (e.g., factual consistency module 150) is provided with or receives (e.g., as input 160) one or more source documents and text summarizations for the same. The text summarization may be generated by a summarization model from a respective source document,” & Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document,” thus processing the first response via another trained machine learning model to generate a second response that indicates whether the first response is verified is disclosed because Kryscinski teaches providing a text summarization generated by one summarization model as input to a separately trained factual consistency neural network model, which determines whether the text summarization is factually consistent and classifies it as consistent/correct or inconsistent/incorrect. Kryscinski’s generated text summarization corresponds to the first response, the factual consistency neural network model corresponds to another trained machine learning model, and the consistent or inconsistent classification corresponds to the second response indicating whether the first response is verified)
Regarding Claim 7, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Kryscinski further teaches:
displaying […the first response…] and an indication of whether […the first response…] is verified (Kryscinski, Par. [0065], “In some embodiments, the text summarization is in the form of a claim sentence. An example of such source document (e.g., article) and claim sentence is illustrated in table 600 of FIG. 6,” & Par. [0066], “the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either ‘CONSISTENT’ (or correct) or ‘INCONSISTENT’ (or incorrect) with the source document. Referring to the example of FIG. 6, the factual consistency module 150 classifies the claim sentence as INCONSISTENT or incorrect,” & Par. [0068], “Referring to the example shown in FIG. 6, bold text in the claim sentence indicates the span of the claim sentence (i.e., ‘Angela Moore’) that is identified as INCONSISTENT or incorrect,” thus displaying the first response and an indication of whether the first response is verified is disclosed because Kryscinski teaches illustrating a generated claim sentence in Figure 6 and visually identifying, using bold text, a portion of the claim sentence that is classified as inconsistent or incorrect. Kryscinski’s claim sentence corresponds to the first response, and the visual identification and inconsistent /incorrect classification correspond to an indication of whether the first response is verified)
Regarding Claim 11, Akyurek teaches:
processing a first request via a trained machine learning model to generate a first response (Akyurek, Page 1 – Section 1, Figure 1,
PNG
media_image1.png
559
639
media_image1.png
Greyscale
, thus processing a first request via a trained machine learning model to generate a first response is disclosed because Akyurek teaches providing a query to a language model (LM) and generating a corresponding prediction in response to the query, as illustrated in Figure 1 by the query “The United States was founded in ___” and the prediction “1776”)
[…] based on first data used to train the trained machine learning model (Akyurek, Page 1 – Abstract, “identifying which training examples taught an LM to generate a particular factual assertion”, thus based on first data used to train the trained machine learning model is disclosed, because Akyurek teaches identifying the training examples that taught the language model to generate a particular factual assertion. Akyurek’s training examples correspond to the first data used to train the trained machine learning model, and the generated factual assertion corresponds to the first response)
Akyurek does not explicitly teach performing one or more operations to verify the first response […],
However, Kryscinski teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Akyurek’s teaching of processing a first request via a trained machine learning model to generate a first response and identifying the training examples that taught the language model to generate the first response, with Kryscinski’s teaching of performing one or more operations to verify generated text by determining whether the generated text is factually consistent with source data. Therefore, a POSITA would have been motivated to incorporate Kryscinski’s factual consistency verification into Akyurek’s training data attribution process so that one or more operations are performed to verify the first response based on first data used to train the trained machine learning model, thereby enabling the first response to be verified based on the identified training data and allowing factual inconsistencies in the generated response to be identified (Kryscinski, Par. [0020], “One such problem is that of verifying factual consistency between source documents and generated summaries: a factually consistent summary should contain only statements that are entailed by the source document. However, studies have shown that a substantial number of summaries generated by abstractive models contain factual inconsistencies. Such high levels of factual inconsistency render automatically generated summaries virtually useless in practice”)
Regarding Claim 12, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Akyurek further teaches:
generating a first embedding based on the first response (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a first embedding based on the first response is disclosed, because Akyurek teaches generating a single vector representation from the intermediate layer outputs of a Transformer language model and using the representation in determining the association with a model prediction. Akyurek’s model prediction corresponds to the first response, and the single vector representation associated with the model prediction corresponds to the first embedding based on the first response)
generating a second embedding based on the first data (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a second embedding based on the first data is disclosed, because Akyurek teaches generating a single vector representation for a training example and using the representation in determining the association between the training example and a model prediction. Akyurek’s training example corresponds to the first data, and the single vector representation of the training example corresponds to the second embedding based on the first data)
computing a similarity between the first embedding and the second embedding (Akyurek, Page 4 – Section 2.2, “Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus computing a similarity between the first embedding and the second embedding is disclosed, because Akyurek teaches determining the association between a training example and a model prediction using a cosine product. Therefore, the cosine product between the vector representation of the model prediction and the vector representation of the training example corresponds to computing a similarity between the first embedding and the second embedding)
Regarding Claim 16, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
processing […the first response…] via another trained machine learning model to generate a second response that indicates whether […the first response…] is verified (Kryscinski, Par. [0063], “After training, the model (e.g., implementing factual consistency module 150) can be applied or used to check for factual consistency of text summarizations generated by one or more summarization models for respective source documents,” & Par. [0065], “the factual consistency neural network model (e.g., factual consistency module 150) is provided with or receives (e.g., as input 160) one or more source documents and text summarizations for the same. The text summarization may be generated by a summarization model from a respective source document,” & Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document,” thus processing the first response via another trained machine learning model to generate a second response that indicates whether the first response is verified is disclosed because Kryscinski teaches providing a text summarization generated by one summarization model as input to a separately trained factual consistency neural network model, which determines whether the text summarization is factually consistent and classifies it as consistent/correct or inconsistent/incorrect. Kryscinski’s generated text summarization corresponds to the first response, the factual consistency neural network model corresponds to another trained machine learning model, and the consistent or inconsistent classification corresponds to the second response indicating whether the first response is verified)
Regarding Claim 17, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
wherein the trained machine learning model comprises an artificial neural network (Kryscinski, Par. [0061], “the systems and methods for factual consistency checking disclosed herein (e.g., factual consistency module 150 of FIG. 1) can be implemented at least in part by one or more neural network models. In some embodiments, the neural network model can comprise or adopt a language representational model operable to perform one or more natural language understanding (NLU) tasks (including natural language inference),” & Par. [0062], “the neural network language model coming uses a pre-trained transformer-based models such as, for example, a Bidirectional Encoder Representations from Transformers (BERT) model,” thus the trained machine learning model comprising an artificial neural network is disclosed because Kryscinski teaches implementing the machine learning model as a neural network language model, including a pre-trained transformer-based BERT model. Therefore, Kryscinski’s neural network language model corresponds to the artificial neural network)
Regarding Claim 18, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
wherein the trained machine learning model comprises a language model (Kryscinski, Par. [0024], “The unannotated source documents and the labeled, transformed sentences can be provided to a neural network language model for training on checking or verifying factual consistency,” & Par. [0062], “the neural network language model coming uses a pre-trained transformer-based models such as, for example, a Bidirectional Encoder Representations from Transformers (BERT) model,” thus the trained machine learning model comprising a language model is disclosed because Kryscinski teaches training a neural network language model for checking or verifying factual consistency and further teaches implementing the neural network language model using a pre-trained transformer-based BERT model. Therefore, Kryscinski’s trained neural network language model corresponds to the trained machine learning model comprising a language model)
Regarding Claim 19, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
displaying […the first response…] and an indication of whether […the first response…] is verified, and indication of the first data (Kryscinski, Par. [0065], “In some embodiments, the text summarization is in the form of a claim sentence. An example of such source document (e.g., article) and claim sentence is illustrated in table 600 of FIG. 6,” & Par. [0066], “the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either ‘CONSISTENT’ (or correct) or ‘INCONSISTENT’ (or incorrect) with the source document. Referring to the example of FIG. 6, the factual consistency module 150 classifies the claim sentence as INCONSISTENT or incorrect,” & Par. [0068], “Referring to the example shown in FIG. 6, bold text in the claim sentence indicates the span of the claim sentence (i.e., ‘Angela Moore’) that is identified as INCONSISTENT or incorrect,” thus displaying the first response and an indication of whether the first response is verified is disclosed because Kryscinski teaches illustrating a generated claim sentence in Figure 6 and visually identifying, using bold text, a portion of the claim sentence that is classified as inconsistent or incorrect. Kryscinski’s claim sentence corresponds to the first response, and the visual identification and inconsistent /incorrect classification corresponds to an indication of whether the first response is verified)
Regarding Claim 20, Akyurek teaches:
processing a first request via a trained machine learning model to generate a first response (Akyurek, Page 1 – Section 1, Figure 1,
PNG
media_image1.png
559
639
media_image1.png
Greyscale
, thus processing a first request via a trained machine learning model to generate a first response is disclosed because Akyurek teaches providing a query to a language model (LM) and generating a corresponding prediction in response to the query, as illustrated in Figure 1 by the query “The United States was founded in ___” and the prediction “1776”)
[…] based on first data used to train the trained machine learning model (Akyurek, Page 1 – Abstract, “identifying which training examples taught an LM to generate a particular factual assertion”, thus based on first data used to train the trained machine learning model is disclosed, because Akyurek teaches identifying the training examples that taught the language model to generate a particular factual assertion. Akyurek’s training examples correspond to the first data used to train the trained machine learning model, and the generated factual assertion corresponds to the first response)
Akyurek does not explicitly teach performing one or more operations to verify the first response […],
However, Kryscinski teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to combine Akyurek’s teaching of processing a first request via a trained machine learning model to generate a first response and identifying the training examples that taught the language model to generate the first response, with Kryscinski’s teaching of performing one or more operations to verify generated text by determining whether the generated text is factually consistent with source data. Therefore, a POSITA would have been motivated to incorporate Kryscinski’s factual consistency verification into Akyurek’s training data attribution process so that one or more operations are performed to verify the first response based on first data used to train the trained machine learning model, thereby enabling the first response to be verified based on the identified training data and allowing factual inconsistencies in the generated response to be identified (Kryscinski, Par. [0020], “One such problem is that of verifying factual consistency between source documents and generated summaries: a factually consistent summary should contain only statements that are entailed by the source document. However, studies have shown that a substantial number of summaries generated by abstractive models contain factual inconsistencies. Such high levels of factual inconsistency render automatically generated summaries virtually useless in practice”)
Claims 3 and 13 are rejected under 35 U.S.C. 103 as being unpatentable over Akyurek et al. (hereinafter Akyurek, a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) in view of Kryscinski et al. (hereinafter Kryscinski) (US 20210124876), and further in view of Burns et al. (hereinafter Burns, a non-patent literature reference titled “Discovering Latent Knowledge in Language Models Without Supervision”).
Regarding Claim 3, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Akyurek further teaches:
generating a second embedding based on the first data (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a second embedding based on the first data is disclosed, because Akyurek teaches generating a single vector representation for a training example and using the representation in determining the association between the training example and a model prediction. Akyurek’s training example corresponds to the first data, and the single vector representation of the training example corresponds to the second embedding based on the first data)
computing a similarity between the first embedding and the second embedding (Akyurek, Page 4 – Section 2.2, “Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus computing a similarity between the first embedding and the second embedding is disclosed, because Akyurek teaches determining the association between a training example and a model prediction using a cosine product. Therefore, the cosine product between the vector representation of the model prediction and the vector representation of the training example corresponds to computing a similarity between the first embedding and the second embedding)
Akyurek combined with Kryscinski does not explicitly teach generating a first embedding based on a negation of […the first request…] and […the first response…].
However, Burns teaches:
generating a first embedding based on a negation of […the first request…] and […the first response…] (Burns, Page 1 – Abstract, “finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values,” & Page 1 – Section 1, “Contrast-Consistent Search (CCS), a method that learns a linear projection of the hidden states that is consistent across negations,” & Page 3 – Section 2.2, “CCS works by (1) answering each question qi as both “Yes” (x + i ) and “No” (x − i ), (2) computing the representations φ(x + i ) and φ(x − i ) of each answer, (3) mapping the answer representations to probabilities p + i and p − i of being true, then (4) optimizing that mapping so that the probabilities are both consistent and confident,” & “for each question qi , we answer qi both as “Yes”, resulting in the new natural language statement x + i , and as “No”, resulting in the natural language statement x − i,” thus generating a first embedding based on a negation of the first request and the first response is disclosed because Burns teaches forming opposite natural language statements based on a question and its possible answers, where the statements are negations of one another and have opposite truth values, and computing representations of those statements. Burns further teaches that CCS operates on hidden state representations that are consistent across negations. Burns’s question corresponds to the first request, the answer corresponds to the first response, the oppositely stated question answer combination corresponds to the negation of the first request and the first response, and the computed representation of that statement corresponds to the first embedding)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Burns’s teaching of generating representations based on a statement and its negation. Akyurek teaches generating vector representations of training examples and model predictions and determining an association between the representations using a cosine product, while Kryscinski teaches verifying generated text for factual consistency. Burns further teaches generating representations of opposite natural language statements based on a question and its answers and using the representations to determine truth based on negation consistency. Therefore, a POSITA would have been motivated to incorporate Burns’s negation based representation into the factual consistency verification of Akyurek and Kryscinski so that a first embedding is generated based on a negation of the first request and the first response and compared with the second embedding based on the first data, thereby allowing the verification process to leverage the logical consistency property that a statement and its negation have opposite truth values and to maintain accurate verification even when the language model generates an incorrect response (Burns, Page 1 – Abstract, “It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question answering datasets, it outperforms zero-shot accuracy by 4% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers”)
Regarding Claim 13, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Akyurek further teaches:
generating a second embedding based on the first data (Akyurek, Page 4 – Section 2.2, “we extract the intermediate layer outputs of a Transformer language model, and average over decoding time-steps to obtain a single vector representation for any example. In our experiments, we consider representations at different layers of the Transformers, as well as their concatenations. Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus generating a second embedding based on the first data is disclosed, because Akyurek teaches generating a single vector representation for a training example and using the representation in determining the association between the training example and a model prediction. Akyurek’s training example corresponds to the first data, and the single vector representation of the training example corresponds to the second embedding based on the first data)
computing a similarity between the first embedding and the second embedding (Akyurek, Page 4 – Section 2.2, “Similar to the case of gradient-based methods, the association between a training example and a model prediction is defined by a cosine product”, thus computing a similarity between the first embedding and the second embedding is disclosed, because Akyurek teaches determining the association between a training example and a model prediction using a cosine product. Therefore, the cosine product between the vector representation of the model prediction and the vector representation of the training example corresponds to computing a similarity between the first embedding and the second embedding)
Akyurek combined with Kryscinski does not explicitly teach generating a first embedding based on a negation of […the first request…] and […the first response…].
However, Burns teaches:
generating a first embedding based on a negation of […the first request…] and […the first response…] (Burns, Page 1 – Abstract, “finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values,” & Page 1 – Section 1, “Contrast-Consistent Search (CCS), a method that learns a linear projection of the hidden states that is consistent across negations,” & Page 3 – Section 2.2, “CCS works by (1) answering each question qi as both “Yes” (x + i ) and “No” (x − i ), (2) computing the representations φ(x + i ) and φ(x − i ) of each answer, (3) mapping the answer representations to probabilities p + i and p − i of being true, then (4) optimizing that mapping so that the probabilities are both consistent and confident,” & “for each question qi , we answer qi both as “Yes”, resulting in the new natural language statement x + i , and as “No”, resulting in the natural language statement x − i,” thus generating a first embedding based on a negation of the first request and the first response is disclosed because Burns teaches forming opposite natural language statements based on a question and its possible answers, where the statements are negations of one another and have opposite truth values, and computing representations of those statements. Burns further teaches that CCS operates on hidden state representations that are consistent across negations. Burns’s question corresponds to the first request, the answer corresponds to the first response, the oppositely stated question answer combination corresponds to the negation of the first request and the first response, and the computed representation of that statement corresponds to the first embedding)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Burns’s teaching of generating representations based on a statement and its negation. Akyurek teaches generating vector representations of training examples and model predictions and determining an association between the representations using a cosine product, while Kryscinski teaches verifying generated text for factual consistency. Burns further teaches generating representations of opposite natural language statements based on a question and its answers and using the representations to determine truth based on negation consistency. Therefore, a POSITA would have been motivated to incorporate Burns’s negation based representation into the factual consistency verification of Akyurek and Kryscinski so that a first embedding is generated based on a negation of the first request and the first response and compared with the second embedding based on the first data, thereby allowing the verification process to leverage the logical consistency property that a statement and its negation have opposite truth values and to maintain accurate verification even when the language model generates an incorrect response (Burns, Page 1 – Abstract, “It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question answering datasets, it outperforms zero-shot accuracy by 4% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers”)
Claims 4-5 and 14-15 are rejected under 35 U.S.C. 103 as being unpatentable over Akyurek et al. (hereinafter Akyurek, a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) in view of Kryscinski et al. (hereinafter Kryscinski) (US 20210124876), and further in view of Pandey et al. (hereinafter Pandey) (US 20210110277).
Regarding Claim 4, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Kryscinski further teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
Akyurek combined with Kryscinski does not explicitly teach computing an entailment score […].
However, Pandey teaches:
computing an entailment score based on […the first data…], […the first request…], and […the first response…] (Pandey, Par. [0021], “The entailment data organizer may obtain a query from a user. The query may be indicating a data entailment requirement comprising entailment data and associated with entailment operations. The entailment data organizer may implement an artificial intelligence component to identify a premise from the entailment data. The premise may be comprising a first word data set associated with the data entailment requirement,” & Par. [0022], “The hypothesis generator may implement a first cognitive learning operation to determine a plurality of hypothesis corresponding to the premise. Each of the plurality of hypothesis may be comprising a second-word data set and indicating an inference associated with the premise. The hypothesis generator may generate a confidence index for each of the plurality of hypothesis based on a comparison of the hypothesis index with the premise index,” & Par. [0023], “The modeler implement may a second cognitive learning operation to determine an entailment value based on the confidence index for each of the plurality of hypothesis. The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” thus computing an entailment score based on the first data, the first request, and the first response is disclosed because Pandey teaches obtaining a query, identifying a premise comprising data associated with the query, determining a hypothesis corresponding to the premise, comparing the hypothesis with the premise, and determining an entailment value indicating the probability that the hypothesis is positively associated with the premise. Pandey’s query corresponds to the first request, Pandey’s hypothesis corresponds to the first response, and Pandey’s entailment value corresponds to the entailment score)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Pandey’s teaching of computing an entailment score based on data, a request, and a response. Akyurek teaches first data used to train the trained machine learning model and a first response generated by the trained machine learning model, while Kryscinski teaches performing one or more operations to verify the first response. Pandey further teaches determining an entailment value based on a query, a premise, and a corresponding hypothesis. Therefore, a POSITA would have been motivated to incorporate Pandey’s entailment value determination into the verification process of Akyurek and Kryscinski so that performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response, thereby providing a quantitative indication of the relationship between the first response and the data used for verification and enabling the entailment relationship between the relevant text fragments to be determined in an efficient and cost effective manner (Pandey, Par. [0023], “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” & Par. [0025], “the present disclosure aims to provide a textual entailment system that may account for the various factors mentioned above, amongst others, to multi-dimensional relationships between various text fragments from a text document in an efficient, and cost-effective manner. Furthermore, the present disclosure may categorically analyze various parameters that may have an impact on deciding an appropriate entailment relationship amongst various text fragments from a given text document”)
Regarding Claim 5, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above and Kryscinski further teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
Akyurek combined with Kryscinski does not explicitly teach computing an entailment score […].
However, Pandey teaches:
computing an entailment score based on […the first data…] and […the first response…] (Pandey, Par. [0038], “The modeler 150 implement may a second cognitive learning operation to determine an entailment value based on the confidence index for each of the plurality of hypothesis. In an example, the second cognitive learning operation may be the NLP model deploying various word embedding techniques, for example, Word2vec® and Glove® vectors to train a neural network to determine if a hypothesis entails a text. In an example, a long short-term memory (LSTM) model may be used to determine an entailment value,” & “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” thus computing an entailment score based on the first data and the first response is disclosed because Pandey teaches determining an entailment value indicating a probability that a hypothesis is positively associated with a premise. Pandey’s hypothesis corresponds to the first response, Pandey’s premise corresponds to the data against which the first response is evaluated, and Pandey’s entailment value corresponds to the entailment score)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Pandey’s teaching of computing an entailment score based on data, a request, and a response. Akyurek teaches first data used to train the trained machine learning model and a first response generated by the trained machine learning model, while Kryscinski teaches performing one or more operations to verify the first response. Pandey further teaches determining an entailment value based on a query, a premise, and a corresponding hypothesis. Therefore, a POSITA would have been motivated to incorporate Pandey’s entailment value determination into the verification process of Akyurek and Kryscinski so that performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response, thereby providing a quantitative indication of the relationship between the first response and the data used for verification and enabling the entailment relationship between the relevant text fragments to be determined in an efficient and cost effective manner (Pandey, Par. [0023], “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” & Par. [0025], “the present disclosure aims to provide a textual entailment system that may account for the various factors mentioned above, amongst others, to multi-dimensional relationships between various text fragments from a text document in an efficient, and cost-effective manner. Furthermore, the present disclosure may categorically analyze various parameters that may have an impact on deciding an appropriate entailment relationship amongst various text fragments from a given text document”)
Regarding Claim 14, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
Akyurek combined with Kryscinski does not explicitly teach computing an entailment score […].
However, Pandey teaches:
computing an entailment score based on […the first data…], […the first request…], and […the first response…] (Pandey, Par. [0021], “The entailment data organizer may obtain a query from a user. The query may be indicating a data entailment requirement comprising entailment data and associated with entailment operations. The entailment data organizer may implement an artificial intelligence component to identify a premise from the entailment data. The premise may be comprising a first word data set associated with the data entailment requirement,” & Par. [0022], “The hypothesis generator may implement a first cognitive learning operation to determine a plurality of hypothesis corresponding to the premise. Each of the plurality of hypothesis may be comprising a second-word data set and indicating an inference associated with the premise. The hypothesis generator may generate a confidence index for each of the plurality of hypothesis based on a comparison of the hypothesis index with the premise index,” & Par. [0023], “The modeler implement may a second cognitive learning operation to determine an entailment value based on the confidence index for each of the plurality of hypothesis. The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” thus computing an entailment score based on the first data, the first request, and the first response is disclosed because Pandey teaches obtaining a query, identifying a premise comprising data associated with the query, determining a hypothesis corresponding to the premise, comparing the hypothesis with the premise, and determining an entailment value indicating the probability that the hypothesis is positively associated with the premise. Pandey’s query corresponds to the first request, Pandey’s hypothesis corresponds to the first response, and Pandey’s entailment value corresponds to the entailment score)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Pandey’s teaching of computing an entailment score based on data, a request, and a response. Akyurek teaches first data used to train the trained machine learning model and a first response generated by the trained machine learning model, while Kryscinski teaches performing one or more operations to verify the first response. Pandey further teaches determining an entailment value based on a query, a premise, and a corresponding hypothesis. Therefore, a POSITA would have been motivated to incorporate Pandey’s entailment value determination into the verification process of Akyurek and Kryscinski so that performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response, thereby providing a quantitative indication of the relationship between the first response and the data used for verification and enabling the entailment relationship between the relevant text fragments to be determined in an efficient and cost effective manner (Pandey, Par. [0023], “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” & Par. [0025], “the present disclosure aims to provide a textual entailment system that may account for the various factors mentioned above, amongst others, to multi-dimensional relationships between various text fragments from a text document in an efficient, and cost-effective manner. Furthermore, the present disclosure may categorically analyze various parameters that may have an impact on deciding an appropriate entailment relationship amongst various text fragments from a given text document”)
Regarding Claim 15, Akyurek combined with Kryscinski teaches all the limitations of claim 11 as cited above and Kryscinski further teaches:
performing one or more operations to verify the first response […] (Kryscinski, Par. [0066], “factual consistency model determines or classifies whether the text summarization or claim sentence (i.e., “Angela Moore was back home resting and enjoying time with his grandchildren.”) remains factually consistent with the source document. In some embodiments, the model may perform two-way classification—e.g., using a single-layer classifier based on the [CLS] token—to classify the claim sentence as either “CONSISTENT” (or correct) or “INCONSISTENT” (or incorrect) with the source document”, thus performing one or more operations to verify the first response is disclosed, because Kryscinski teaches determining whether a generated text summarization or claim sentence is factually consistent with the source document and classifying the generated text as either consistent/correct or inconsistent/incorrect. Kryscinski’s text summarization or claim sentence corresponds to the first response, and the factual consistency determination and classification correspond to performing one or more operations to verify the first response)
Akyurek combined with Kryscinski does not explicitly teach computing an entailment score […].
However, Pandey teaches:
computing an entailment score based on […the first data…] and […the first response…] (Pandey, Par. [0038], “The modeler 150 implement may a second cognitive learning operation to determine an entailment value based on the confidence index for each of the plurality of hypothesis. In an example, the second cognitive learning operation may be the NLP model deploying various word embedding techniques, for example, Word2vec and Glove vectors to train a neural network to determine if a hypothesis entails a text. In an example, a long short-term memory (LSTM) model may be used to determine an entailment value,” & “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” thus computing an entailment score based on the first data and the first response is disclosed because Pandey teaches determining an entailment value indicating a probability that a hypothesis is positively associated with a premise. Pandey’s hypothesis corresponds to the first response, Pandey’s premise corresponds to the data against which the first response is evaluated, and Pandey’s entailment value corresponds to the entailment score)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Pandey’s teaching of computing an entailment score based on data, a request, and a response. Akyurek teaches first data used to train the trained machine learning model and a first response generated by the trained machine learning model, while Kryscinski teaches performing one or more operations to verify the first response. Pandey further teaches determining an entailment value based on a query, a premise, and a corresponding hypothesis. Therefore, a POSITA would have been motivated to incorporate Pandey’s entailment value determination into the verification process of Akyurek and Kryscinski so that performing the one or more operations to verify the first response comprises computing an entailment score based on the first data, the first request, and the first response, thereby providing a quantitative indication of the relationship between the first response and the data used for verification and enabling the entailment relationship between the relevant text fragments to be determined in an efficient and cost effective manner (Pandey, Par. [0023], “The entailment value may be indicating a probability of a hypothesis from the plurality of hypothesis being positively associated with the premise,” & Par. [0025], “the present disclosure aims to provide a textual entailment system that may account for the various factors mentioned above, amongst others, to multi-dimensional relationships between various text fragments from a text document in an efficient, and cost-effective manner. Furthermore, the present disclosure may categorically analyze various parameters that may have an impact on deciding an appropriate entailment relationship amongst various text fragments from a given text document”)
Claims 8, 9, and 10 are rejected under 35 U.S.C. 103 as being unpatentable over Akyurek et al. (hereinafter Akyurek, a non-patent literature reference titled “Towards Tracing Factual Knowledge in Language Models Back to the Training Data”) in view of Kryscinski et al. (hereinafter Kryscinski) (US 20210124876), and further in view of Mekala et al. (hereinafter Mekala, a non-patent literature reference titled “Coarse2Fine: Fine-grained Text Classification on Coarsely-grained Annotated Data”).
Regarding Claim 8, Akyurek combined with Kryscinski teaches all the limitations of claim 1 as cited above:
Akyurek combined with Kryscinski does not explicitly teach appending context information to one or more portions of […the first data…] to generate second data and performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model.
However, Mekala teaches:
appending context information to one or more portions of […the first data…] to generate second data (Mekala, Page 2 – Introduction, “we inform the language models with label information by adding the label surface names at the beginning of each document,” & Page 4 – Section 3.2.1, “We designate the label surface names as the special token sequences and append them in the beginning to their respective documents with another special token separating the label sequence and document. For example, a sample document ‘Messi plays for FC Barcelona’ belonging to ‘soccer’ is modified to ‘soccer Messi plays for FC Barcelona’,” thus appending context information to one or more portions of the first data to generate second data is disclosed because Mekala teaches adding label information to the beginning of each document and specifically appending label surface names as special token sequences to their respective documents to form modified, label-conditioned documents. Mekala’s label surface names correspond to the context information, the respective documents correspond to the first data, and the modified label-conditioned documents correspond to the second data)
performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model (Mekala, Page 2 – Introduction, “We propose to fine-tune language models in a label conditioned, hierarchy-aware manner. Specifically, we inform the language models with label information by adding the label surface names at the beginning of each document. We further incorporate a regularization objective into the fine-tuning process that captures the constraints derived from the label hierarchy,” & Page 4 – Section 3.2.1, “To continuously train GPT-2 in a label conditioned way, one has to maximize P(D|l) instead of P(D). We designate the label surface names as the special token sequences and append them in the beginning to their respective documents with another special token separating the label sequence and document,” thus performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model is disclosed, because Mekala teaches fine-tuning and continuously training a language model using documents that have been modified by adding label information to the documents. Mekala’s label-conditioned documents correspond to the second data, the pre-trained GPT-2 corresponds to the machine learning model, and the fine-tuned and continuously trained GPT-2 corresponds to the trained machine learning model)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to further combine Akyurek and Kryscinski with Mekala’s teaching of appending context information to first data to generate second data and performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model. Akyurek teaches first data used to train a trained machine learning model, while Kryscinski teaches verifying a response generated by the trained machine learning model. Mekala further teaches appending context information to one or more portions of first data to generate second data and performing one or more operations to train a machine learning model based on the second data to generate the trained machine learning model. Therefore, a POSITA would have been motivated to incorporate Mekala’s context-conditioned training into the system of Akyurek and Kryscinski so that context information is appended to one or more portions of the first data to generate second data and one or more operations are performed to train a machine learning model based on the second data to generate the trained machine learning model, thereby allowing the trained machine learning model to better incorporate the context information associated with the first data and generate higher-quality data for the relevant context (Mekala, Page 3 – Section 3, “Our major contributions lie in how to better incorporate the label names and their hierarchical relations into the language model and therefore generate more high-quality psuedo training data,” & Page 4 – Section 3.2.1, “We designate the label surface names as the special token sequences and append them in the beginning to their respective documents with another special token <labelsep> separating the label sequence and document”)
Regarding Claim 9, Akyurek and Kryscinski combined with Mekala teaches all the limitations of claim 1 as cited above and Mekala further teaches:
wherein the context information that is appended to each portion of the one or more portions comprises a token indicating one or more contexts associated with the portion (Mekala, Page 4 – Section 3.2.1, “We designate the label surface names as the special token sequences and append them in the beginning to their respective documents with another special token <labelsep> separating the label sequence and document. For example, a sample document ‘Messi plays for FC Barcelona’ belonging to ‘soccer’ is modified to ‘soccer <labelsep> Messi plays for FC Barcelona’,” & “Note that, the l here could be the label surface name of a coarse-grained or fine-grained class. One can view our formulation as asking the label token sequence to play the role of prompt and the document D to be the continuation,” thus the context information that is appended to each portion comprises a token indicating one or more contexts associated with the portion is disclosed because Mekala teaches representing label information as special token sequences and appending the label token sequence to the document to indicate the coarse-grained or fine-grained class associated with that document. Mekala’s label surface name corresponds to the context associated with the portion, the special label token sequence corresponds to the token indicating the context, and the respective document corresponds to the portion of the first data)
Regarding Claim 10, Akyurek and Kryscinski combined with Mekala teaches all the limitations of claim 1 as cited above and Mekala further teaches:
wherein the context information that is appended to each portion of the one or more portions comprises a hierarchy of one or more contexts associated with the portion (Mekala, Page 2 – Figure 2,
PNG
media_image2.png
372
1125
media_image2.png
Greyscale
, & Page 4 – Section 3.2.1, “During the continuous training process, we have access to both the gold, coarse-grained labels and weak, fine-grained labels. Examples included in weak supervision give rise to two label-conditioned documents — one by prefixing with the coarse-grained, gold label and the other with the weak, fine-grained one,” & Page 4 – Section 3.2.2, “fine-grained labels are more specific to coarse-grained labels,” and “we propose to add a regularization to the language model with constraints derived from hierarchy,” thus the context information appended to a portion comprises a hierarchy of one or more contexts associated with the portion is disclosed because Mekala teaches associating the same document with both a coarse-grained label and a more specific fine-grained label, where the coarse and fine-grained labels form a hierarchical relationship, and prefixing the respective labels to the document during label-conditioned training. Therefore, Mekala’s coarse-grained and fine-grained labels correspond to the hierarchy of contexts associated with the portion, and the prefixed label information corresponds to the context information appended to the portion)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MAHLIET ADMASU whose telephone number is (571)272-0034. The examiner can normally be reached Mon-Fri, 8am-5pm.
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, Alexey Shmatov can be reached at (571)270-3428. 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.
/M.T.A./Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights