DETAILED ACTION
This communication is in response to the Amendments and Arguments filed on 12/11/2025.
Claim(s) 1-20 are pending and have been examined. Hence, this action has been made FINAL.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments and Amendments
Amendments to the claims by the Applicant have been considered and addressed below.
With respect to the 35 USC § 101, 102, and 103 rejections, the Applicant provides several arguments in which the Examiner will respond accordingly, below.
35 USC § 101 rejection(s)
Arguments in pages 11-12 of Remarks filed on 12/11/2025.
Examiner response to Arguments:
Applicant’s arguments, with respect to the rejection(s) of independent claim(s) 1, 11, and 20 under 35 USC 101 have been fully considered but are not persuasive.
The Applicant argues that:
“The present claims purport to improve computer capabilities. The improvement of the computer capabilities may be seen in the technical solution described in Appellant's specification.
[…]
Thus, one or more embodiments improve the capabilities of a computer, by permitting a computer to return an acceptable result where no direct match exists. Accordingly, one or more embodiments solve a technical problem by means of a technical solution.”
However, the Examiner respectfully disagrees with the arguments and provides more details on the rationale used to examine the claims rejected under 35 U.S.C. § 101 of the Instant Application for clarification, below.
Please see detailed analysis below (Prong Two) for more details on how the Examiner understands the independent claims do not recite additional elements that integrate the judicial exception into a practical application. Hence, not qualifying as patent eligible subject matter under 35 U.S.C. § 101.
Please refer to MPEP 2106.04(1): Eligibility Step 2A: Whether a Claim is Directed to a Judicial Exception: Prong One.
“Prong One asks does the claim recite an abstract idea, law of nature, or natural phenomenon? In Prong One examiners evaluate whether the claim recites a judicial exception, i.e. whether a law of nature, natural phenomenon, or abstract idea is set forth or described in the claim. While the terms "set forth" and "described" are thus both equated with "recite", their different language is intended to indicate that there are two ways in which an exception can be recited in a claim. For instance, the claims in Diehr, 450 U.S. at 178 n. 2, 179 n.5, 191-92, 209 USPQ at 4-5 (1981), clearly stated a mathematical equation in the repetitively calculating step, and the claims in Mayo, 566 U.S. 66, 75-77, 101 USPQ2d 1961, 1967-68 (2012), clearly stated laws of nature in the wherein clause, such that the claims "set forth" an identifiable judicial exception. Alternatively, the claims in Alice Corp., 573 U.S. at 218, 110 USPQ2d at 1982, described the concept of intermediated settlement without ever explicitly using the words "intermediated" or "settlement."”
“An example of a claim that recites a judicial exception is "A machine comprising elements that operate in accordance with F=ma." This claim sets forth the principle that force equals mass times acceleration (F=ma) and therefore recites a law of nature exception. Because F=ma represents a mathematical formula, the claim could alternatively be considered as reciting an abstract idea. Because this claim recites a judicial exception, it requires further analysis in Prong Two in order to answer the Step 2A inquiry. An example of a claim that merely involves, or is based on, an exception is a claim to "A teeter-totter comprising an elongated member pivotably attached to a base member, having seats and handles attached at opposing sides of the elongated member." This claim is based on the concept of a lever pivoting on a fulcrum, which involves the natural principles of mechanical advantage and the law of the lever. However, this claim does not recite these natural principles and therefore is not directed to a judicial exception (Step 2A: NO). Thus, the claim is eligible at Pathway B without further analysis.”
From this analysis, in Step 2A, Prong One, the Examiner has evaluated the independent claims accordingly and determined that the amended independent claims as drafted indeed describe a judicial exception (i.e., an abstract idea), which represent a mental process (which can be performed by a human with pen and paper).
More specifically, similar to what was discussed in the Non-Final Rejection mailed on 09/11/2025:
The limitations of claims 1, 11, and 20, as drafted cover a human (mental process).
More specifically, the independent claim(s) recite(s):
1. (Currently Amended) A method comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table;
executing, by the processor, a large language model on the query to generate a query vector in a multi-dimensional dense vector space, wherein:
the query vector comprises a query data structure storing a semantic meaning of the query, the language model comprises a transformer-based large language model pre-trained on sentence datasets, and the large language model is programmed to map phrases to the multi- dimensional dense vector space for vector similarity comparison;
executing, by the server controller, a semantic matching algorithm on both the query vector and a lookup vector, wherein:
the lookup vector comprises a lookup data structure storing the plurality of entries of [[a]] the lookup table and encoded semantic meanings of the plurality of entries, and the semantic matching algorithm compares the query vector to the lookup vector and returns, as a result of comparing, a found entry in the lookup table, the found entry comprising a closest semantic match to the query vector;
looking up, by the server controller, using the found entry in the lookup table, the target entry in the lookup table; and
returning, by the server controller, the target entry.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Applying another set of predetermined rules to the query list and another predefined list/table
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
11. (Currently Amended) A system comprising:
a computer processor;
a data repository in communication with the computer processor and storing:
a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table,
a query vector comprising query data structure storing a semantic meaning of the query in a multi-dimensional dense vector space,
a lookup table,
a found entry in the lookup table and [[a]] the target entry in the lookup table, and
a lookup vector comprising a lookup data structure storing the plurality of entries of [[a]] the lookup table and encoded semantic meanings of the plurality of entries,
a large language model which, when applied by the processor to the query, generates the query vector, wherein the large language model comprises a transformer-based large language model pre-trained on sentence datasets and wherein the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison;
a semantic matching algorithm which, when applied by the processor to both the query vector and the lookup vector, compares the query vector to the lookup vector and returns, as a result of comparing, the found entry in the lookup table; and
a lookup algorithm which, when applied by the processor to the lookup table using the found entry, looks up the target entry in the lookup table and returns the target entry, the found entry comprising a closest semantic match to the query vector.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Applying another set of predetermined rules to the query list and another predefined list/table;
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
20. (Currently Amended) A method comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table;
executing, by the processor, a large language model to a lookup table to generate a lookup vector, wherein the lookup vector comprises a lookup data structure storing a plurality of semantic meanings of a plurality of entries of the lookup table and encoded semantic meanings of the plurality of entries;
executing, by the processor and after applying the large language model to the lookup table, a large language model [[to a]] on the query to generate a query vector in a multi-dimensional dense vector space, wherein:
the query vector comprises a query data structure storing a semantic meaning of the query,
the language model comprises a transformer-based large language model pre-trained on sentence datasets, and
the large language model is programmed to map phrases to the multi- dimensional dense vector space for vector similarity comparison;
executing, by the server controller, a semantic matching algorithm [[to]]on both the query vector and a lookup vector, and wherein the semantic matching algorithm further performs:
comparing the query vector to the lookup vector and returning a plurality of semantic distances between the query vector and a plurality of entries in the lookup table, comparing the plurality of semantic distances to a threshold value, adding a set of entries, from the plurality of entries, to a list of candidate entries when a corresponding semantic distance in the plurality of semantic distances satisfies the threshold value, and transmitting the list of candidate entries to a remote user device;
receiving, at the server controller, a user selection of one of the candidate entries as being a found entry in the lookup table;
looking up, by the server controller, using the found entry in the lookup table, [[a]]the target entry in the lookup table, the found entry comprising a closest semantic match to the query vector; and
returning, by the server controller, the target entry.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Wherein the another list includes semantic meaning;
Applying another set of predetermined rules to the list and another a query list;
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Comparing the semantic distances based on a threshold (e.g., mathematical concept);
Adding words to a list based on the distances satisfying the threshold;
Receiving a selection (e.g., of a word);
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
Please also refer to MPEP 2106.05(f)(2): Whether the claim invokes computers or other machinery merely as a tool to perform an existing process, and MPEP 2106.06(b): Clear Improvement to a Technology or to Computer Functionality.
Please refer to MPEP 2106.04(2): Eligibility Step 2A: Whether a Claim is Directed to a Judicial Exception: Prong Two.
“Prong Two asks does the claim recite additional elements that integrate the judicial exception into a practical application? In Prong Two, examiners evaluate whether the claim as a whole integrates the exception into a practical application of that exception. If the additional elements in the claim integrate the recited exception into a practical application of the exception, then the claim is not directed to the judicial exception (Step 2A: NO) and thus is eligible at Pathway B. This concludes the eligibility analysis. If, however, the additional elements do not integrate the exception into a practical application, then the claim is directed to the recited judicial exception (Step 2A: YES), and requires further analysis under Step 2B (where it may still be eligible if it amounts to an ‘‘inventive concept’’). For more information on how to evaluate whether a judicial exception is integrated into a practical application, see MPEP § 2106.04(d)(2).”
From this analysis, in Step 2A, Prong Two, the Examiner has evaluated the independent claims accordingly and determined that the amended independent claims as drafted that the claims as a whole do not include additional elements that integrate the exception into a practical application of that exception. (i.e., an abstract idea). As discussed in the Non-Final Rejection mailed on 09/11/2025:
This judicial exception is not integrated into a practical application because for example: claim 1 recites “a server controller”, “a processor”, “atransformer-based large language model” , claim 11 recites ”, “a processor”, “a transformer-based large language model”, and claim 20 recites recites “a server controller”, “transformer-based large language model”, “a computer processor” and “a data repository”. As an example, in [0089-0091] of the as filed specification, it is disclosed that “[0089] The input device(s) (510) may include a touchscreen, keyboard, mouse, microphone, touchpad, electronic pen, or any other type of input device… [0090] Further, the output device(s) (512) may include a display device, a printer, external storage, or any other output device… [0091] Software instructions in the form of computer readable program code to perform embodiments may be stored, in whole or in part, temporarily or permanently, on a non-transitory computer readable medium such as a solid state drive (SSD), compact disk (CD), digital video disk (DVD), storage device, a diskette, a tape, flash memory, physical memory, or any other computer readable storage medium. Specifically, the software instructions may correspond to computer readable program code that, when executed by the computer processor(s) (502), is configured to perform one or more embodiments, which may include transmitting, receiving, presenting, and displaying data and messages described in the other figures of the disclosure.” Therefore, a general-purpose computer or computing device is described and mainly used as an application thereof. Accordingly, these additional elements do not integrate the abstract idea into a practical idea because it does not impose any meaningful limits on practicing the abstract idea.
Please also refer to MPEP 2106.05(f)(2): Whether the claim invokes computers or other machinery merely as a tool to perform an existing process.
Finally, please refer to MPEP 2106.05(A): Relevant Considerations For Evaluating Whether Additional Elements Amount To An Inventive Concept
“Limitations that the courts have found not to be enough to qualify as "significantly more" when recited in a claim with a judicial exception include:
i. Adding the words "apply it" (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, e.g., a limitation indicating that a particular function such as creating and maintaining electronic records is performed by a computer, as discussed in Alice Corp., 573 U.S. at 225-26, 110 USPQ2d at 1984 (see MPEP § 2106.05(f));
ii. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception, e.g., a claim to an abstract idea requiring no more than a generic computer to perform generic computer functions that are well-understood, routine and conventional activities previously known to the industry, as discussed in Alice Corp., 573 U.S. at 225, 110 USPQ2d at 1984 (see MPEP § 2106.05(d));”
From this analysis, in Step 2B, the Examiner has evaluated the independent claims accordingly and determined that the independent claims as drafted have limitations that the courts have found not to be enough to qualify as "significantly more" when recited in a claim with a judicial exception. Similar to what was discussed in the Non-Final Rejection mailed on 09/11/2025:
The claim(s) does/do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to the integration of the abstract idea into a practical application, the additional elements of using a computer is listed as a general computing device as noted. The claim is not patent eligible.
Lastly, the Examiner refers the Applicant to MPEP 2106.05(a):
“It is important to note that in order for a method claim to improve computer functionality, the broadest reasonable interpretation of the claim must be limited to computer implementation. That is, a claim whose entire scope can be performed mentally, cannot be said to improve computer technology. Synopsys, Inc. v. Mentor Graphics Corp., 839 F.3d 1138, 120 USPQ2d 1473 (Fed. Cir. 2016) (a method of translating a logic circuit into a hardware component description of a logic circuit was found to be ineligible because the method did not employ a computer and a skilled artisan could perform all the steps mentally). Similarly, a claimed process covering embodiments that can be performed on a computer, as well as embodiments that can be practiced verbally or with a telephone, cannot improve computer technology. See RecogniCorp, LLC v. Nintendo Co., 855 F.3d 1322, 1328, 122 USPQ2d 1377, 1381 (Fed. Cir. 2017) (process for encoding/decoding facial data using image codes assigned to particular facial features held ineligible because the process did not require a computer).” (Emphasis added)
In summary, the Examiner respectfully disagrees with the arguments above.
For more details, please refer to updated 35 U.S.C. § 103 rejections for claims 1-20, below.
35 USC § 102/103 rejection(s)
Arguments in pages 13-15 of Remarks filed on 12/11/2025.
Examiner’s Response to Arguments:
Applicant’s arguments with respect to claim(s) 1, 11, and 20 under 35 U.S.C. § 102, and/or 103 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made
in view of Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) for claims 1 and 11, and
in view of Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and Lindrup et al. (US 20250046330 A1) for claim 20.
For more details, please refer to updated 35 U.S.C. § 103 rejections for claims 1-20, below.
Claim Objections
Claims 9 and 18 objected to because of the following informalities: the second occurrence of the limitation “ a multi-dimensional dense vector space” should now read: the multi-dimensional dense vector space. Appropriate correction is required.
Claim 20 objected to because of the following informalities: the second occurrence of the limitation “a large language model” should read: the large language model. Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1-10 and 20 rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claims 1 and 20 recite the limitation "the language model" in line 12 of claim 1 and line 17 of claim 20. There is insufficient antecedent basis for this limitation in the claim.
Claims 2-10 depend on claim 1 and inherit indefiniteness. Hence, are also rejected.
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.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. More specifically directed to the abstract idea grouping of: mental process and/or mathematical concept.
The independent claim(s) recite(s):
1. (Currently Amended) A method comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table;
executing, by the processor, a large language model on the query to generate a query vector in a multi-dimensional dense vector space, wherein:
the query vector comprises a query data structure storing a semantic meaning of the query, the language model comprises a transformer-based large language model pre-trained on sentence datasets, and the large language model is programmed to map phrases to the multi- dimensional dense vector space for vector similarity comparison;
executing, by the server controller, a semantic matching algorithm on both the query vector and a lookup vector, wherein:
the lookup vector comprises a lookup data structure storing the plurality of entries of [[a]] the lookup table and encoded semantic meanings of the plurality of entries, and the semantic matching algorithm compares the query vector to the lookup vector and returns, as a result of comparing, a found entry in the lookup table, the found entry comprising a closest semantic match to the query vector;
looking up, by the server controller, using the found entry in the lookup table, the target entry in the lookup table; and
returning, by the server controller, the target entry.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Applying another set of predetermined rules to the query list and another predefined list/table
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
11. (Currently Amended) A system comprising:
a computer processor;
a data repository in communication with the computer processor and storing:
a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table,
a query vector comprising query data structure storing a semantic meaning of the query in a multi-dimensional dense vector space,
a lookup table,
a found entry in the lookup table and [[a]] the target entry in the lookup table, and
a lookup vector comprising a lookup data structure storing the plurality of entries of [[a]] the lookup table and encoded semantic meanings of the plurality of entries,
a large language model which, when applied by the processor to the query, generates the query vector, wherein the large language model comprises a transformer-based large language model pre-trained on sentence datasets and wherein the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison;
a semantic matching algorithm which, when applied by the processor to both the query vector and the lookup vector, compares the query vector to the lookup vector and returns, as a result of comparing, the found entry in the lookup table; and
a lookup algorithm which, when applied by the processor to the lookup table using the found entry, looks up the target entry in the lookup table and returns the target entry, the found entry comprising a closest semantic match to the query vector.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Applying another set of predetermined rules to the query list and another predefined list/table;
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
20. (Currently Amended) A method comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table;
executing, by the processor, a large language model to a lookup table to generate a lookup vector, wherein the lookup vector comprises a lookup data structure storing a plurality of semantic meanings of a plurality of entries of the lookup table and encoded semantic meanings of the plurality of entries;
executing, by the processor and after applying the large language model to the lookup table, a large language model [[to a]] on the query to generate a query vector in a multi-dimensional dense vector space, wherein:
the query vector comprises a query data structure storing a semantic meaning of the query,
the language model comprises a transformer-based large language model pre-trained on sentence datasets, and
the large language model is programmed to map phrases to the multi- dimensional dense vector space for vector similarity comparison;
executing, by the server controller, a semantic matching algorithm [[to]]on both the query vector and a lookup vector, and wherein the semantic matching algorithm further performs:
comparing the query vector to the lookup vector and returning a plurality of semantic distances between the query vector and a plurality of entries in the lookup table, comparing the plurality of semantic distances to a threshold value, adding a set of entries, from the plurality of entries, to a list of candidate entries when a corresponding semantic distance in the plurality of semantic distances satisfies the threshold value, and transmitting the list of candidate entries to a remote user device;
receiving, at the server controller, a user selection of one of the candidate entries as being a found entry in the lookup table;
looking up, by the server controller, using the found entry in the lookup table, [[a]]the target entry in the lookup table, the found entry comprising a closest semantic match to the query vector; and
returning, by the server controller, the target entry.
This reads on a human (e.g., mentally and/or using pen and paper):
Receiving language from another human (e.g., text or speech) having indirect reference to a word (e.g., drink instead of water)
Applying a predetermined set of rules to the query (or received language from the other human) to generate (e.g., writing down) a list comprising semantic meaning of the query;
Wherein the another list includes semantic meaning;
Applying another set of predetermined rules to the list and another a query list;
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Comparing the semantic distances based on a threshold (e.g., mathematical concept);
Adding words to a list based on the distances satisfying the threshold;
Receiving a selection (e.g., of a word);
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
This judicial exception is not integrated into a practical application because for example: claim 1 recites “a server controller”, “a processor”, “atransformer-based large language model” , claim 11 recites ”, “a processor”, “a transformer-based large language model”, and claim 20 recites recites “a server controller”, “transformer-based large language model”, “a computer processor” and “a data repository”. As an example, in [0089-0091] of the as filed specification, it is disclosed that “[0089] The input device(s) (510) may include a touchscreen, keyboard, mouse, microphone, touchpad, electronic pen, or any other type of input device… [0090] Further, the output device(s) (512) may include a display device, a printer, external storage, or any other output device… [0091] Software instructions in the form of computer readable program code to perform embodiments may be stored, in whole or in part, temporarily or permanently, on a non-transitory computer readable medium such as a solid state drive (SSD), compact disk (CD), digital video disk (DVD), storage device, a diskette, a tape, flash memory, physical memory, or any other computer readable storage medium. Specifically, the software instructions may correspond to computer readable program code that, when executed by the computer processor(s) (502), is configured to perform one or more embodiments, which may include transmitting, receiving, presenting, and displaying data and messages described in the other figures of the disclosure.” Therefore, a general-purpose computer or computing device is described and mainly used as an application thereof. Accordingly, these additional elements do not integrate the abstract idea into a practical idea because it does not impose any meaningful limits on practicing the abstract idea.
The claim(s) does/do not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to the integration of the abstract idea into a practical application, the additional elements of using a computer is listed as a general computing device as noted. The claim is not patent eligible.
With respect to claims 2 and 12, the claim(s) recite:
wherein comparing the query vector to the lookup vector comprises identifying the found entry in the lookup vector as having a least semantic distance to the query vector, relative to other entries in the lookup table.
This reads on a human:
Wherein the comparison consists of identifying entries (e.g., words) in the list(s) and having at least a predefined semantic distance (e.g., mathematical concept)
No additional limitations are present.
With respect to claims 3 and 13, the claim(s) recite:
wherein comparing the query vector to the lookup vector comprises:
identifying the found entry in the lookup vector as having a semantic distance to the query vector;
comparing the semantic distance to a threshold value; and
returning the found entry when the semantic distance satisfies the threshold value.
This reads on a human:
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Comparing the semantic distances based on a threshold (e.g., mathematical concept);
Writing the predefined term down (if it satisfies the threshold).
No additional limitations are present.
With respect to claims 4 and 14, the claim(s) recite:
wherein comparing the query vector to the lookup vector comprises:
identifying the found entry in the lookup vector as having a semantic distance to the query vector;
comparing the semantic distance to a first threshold value;
comparing, responsive to the semantic distance failing to satisfy the first threshold value, the semantic distance to a second threshold value;
adding, responsive to the semantic distance satisfying the second threshold value, the found entry to a list of candidate entries comprising additional entries in the lookup vector;
transmitting, to a user device, the list of candidate entries; and
receiving, from the user device, a selection of the found entry from the list of candidate entries.
This reads on a human:
Wherein the another predefined list/table comprises of semantic meaning details and the predetermined set of rules include a comparison of the query list and the another predefined list/table;
Comparing the semantic distances based on a threshold (e.g., mathematical concept);
Comparing the semantic distances based on a second threshold (e.g., mathematical concept);
Adding words to a list based on the distances satisfying the threshold;
Receiving a selection (e.g., of a word);
Finding a predefined term in the predefined list/table; and
Writing the predefined term down.
No additional limitations are present.
With respect to claims 5 and 15, the claim(s) recite:
wherein looking up the target entry comprises:
looking up, using the found entry, a plurality of second entries in the lookup table,
wherein the target entry is among the plurality of second entries;
transmitting, to a user device, the plurality of second entries; and
receiving, from the user device, a selection of the target entry in the lookup table.
This reads on a human:
Wherein finding a predefined term in the predefined list/table includes:
Finding a plurality terms;
The plurality of terms including the predefined term
Writing the predefined term(s) down for display to another human
Receiving a selection of the predefined term from the other human
No additional limitations are present.
With respect to claims 6 and 16, the claim(s) recite:
6. The method of claim 1, further comprising:
applying, prior to applying the semantic matching algorithm, the large language model to the lookup table to generate the lookup vector.
16. The system of claim 11, wherein the large language model, when applied by the processor to the lookup table prior to applying the semantic matching algorithm, generates the lookup vector.
This reads on a human:
Applying predetermined set of rules to generate (e.g., writing down) another list.
No additional limitations are present.
With respect to claims 7 and 17, the claim(s) recite:
7. The method of claim 1, further comprising:
receiving, prior to applying the semantic matching algorithm, a new entry to a new lookup table; and
applying, prior to applying the semantic matching algorithm, the large language model to the new lookup table to generate the lookup vector,
wherein the new lookup table is the lookup table when looking up the target entry.
17. The system of claim 11, wherein the data repository further stores a new lookup table,
the large language model, when applied by the processor to the new lookup table prior to applying the semantic matching algorithm, generates the lookup vector, and
the new lookup table is the lookup table when the lookup algorithm returns the target entry.
This reads on a human:
Receiving/writing down a new list;
Applying a predetermined set of rules to generate (e.g., writing down) another list
Finding a predefined term in the predefined list/table; and
No additional limitations are present.
With respect to claim 8, the claim(s) recite:
8. The method of claim 1, wherein the large language model comprises a transformer-based large language model that is pre-trained on sentence data sets.
This reads on a human:
Further defining the predetermined set of rules (i.e., using sentences as reference).
No additional limitations are present.
With respect to claims 9 and 18, the claim(s) recite:
wherein the large language model is programmed to map phrases to a multi-dimensional dense vector space suitable for a computer to perform vector similarity comparisons.
This reads on a human:
Wherein the predetermined set of rules include multiple lists/vectors on which mathematical concepts like similarity comparisons can be performed
No additional limitations are present.
With respect to claims 10 and 19, the claim(s) recite:
10. The method of claim 1, wherein returning comprises providing the target entry to a data processing algorithm programmed to process the target entry to generate a secondary result.
19. The system of claim 11, wherein the system further comprises:
a data processing algorithm which, when applied by the processor to the target entry, processes the target entry to generate a secondary result.
This reads on a human:
Wherein writing down the predefined term includes using said term to determine a secondary term based on predetermined set of rules.
No additional limitations are present.
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.
1-2, 6-7, 9-12, and 16-19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and further in view of Laprise et al. (US 20240419907 A1).
As to independent claim 1, Kotaru et al. teaches:
1. A method (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above…” and ¶ [0135]: “FIG. 16 is a flowchart 1600 of an illustrative method that may be performed when implementing the present conversational AI system in which a foundation model is adapted for information synthesis of wireless communication specifications…”) comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028]: “FIG. 18 is a block diagram of an illustrative server or computing device that may be used at least in part to implement the present conversational AI system;” and ¶ [0054]: “FIG. 3 shows an illustrative architecture of the present conversational AI system 300. Prior to delving into the specifics, the following provides a concise summary of the operational procedure of the system. When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.”) , (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028 and 0054] citation(s) as in limitation above, more specifically: “FIG. 18 is a block diagram of an illustrative server or computing device that may be used at least in part to implement the present conversational AI system;” and ¶ [0054]: “… When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.” and further ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”);
executing, by the processor, a large language model on the query to generate a query vector in a multi-dimensional dense vector space (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citation(s) as in limitation(s) above and further ¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs. OpenAI embedding model openai-textembedding-ada-002 is used to transform each of the text samples into word embedding vectors. The resulting vector representation for each sample is stored in an index for efficient querying. The resulting domain-specific database of domain-specific word embedding vectors consists of 86 million tokens.”, ¶ [0066-0067]: “[0066] When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically. [0067] To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.”, ¶ [0137]: “Block 1610 includes providing a base foundation large language model (LLM) for identifying relevant information from the domain-specific database in response to queries from the user made through a user interface (UI) to the conversational AI system …” and ¶ [0138]: “Block 1625 includes generating tokenized text for a query received from the user at the UI from which word embedding vectors are generated. ”),
wherein the query vector comprises a query data structure storing a semantic meaning of the query (see ¶ [0138] citation as in limitation above and further: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”);
executing, by the server controller, a semantic matching algorithm on both the query vector and a lookup vector (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, 0064, and 0066-0067, and 0137-0138] citation(s) as in limitation(s) above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138]: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”), wherein:
the lookup vector comprises a lookup data structure storing the plurality of entries of the lookup table and encoded semantic meanings of the plurality of entries (see ¶ [0066-0067 and 0138] citations as in limitation(s) above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”), and the semantic matching algorithm compares the query vector to the lookup vector and returns (see ¶ [0066-0067 and 0138] citations as in limitation(s) above. More specifically: ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138]: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”), as a result of comparing, a found entry in the lookup table, (see ¶ [0067 and 0138] citations as in limitation(s) above and further ¶ [0068 and 0139]: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors. [0139] Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.”)
the found entry comprising a closest semantic match to the query vector (see ¶ [0067 and 0138] citations as in limitation(s) above, more specifically: ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and further ¶ [0068 and 0139]: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt);
looking up, by the server controller, using the found entry in the lookup table, a target entry in the lookup table (see ¶ [0067-0068 and 0138-0139] citations as in limitation(s) above. More specifically: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors.”
Here, the Examiner notes that the most similar samples would read on the found entry in the lookup table while the created prompt using the similar samples would read on the target entry in the lookup table.); and
returning, by the server controller, the target entry (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, 0064, and 0066-0067, and 0137-0139] citation(s) as in limitation(s) above. More specifically: “[0139] Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.” and further ¶ [0070]: “The present conversational AI system incorporates few-shot learning techniques by creating a prompt template with few pre-defined example queries, relevant context obtained from specifications and reference responses. A prompt is then generated, following the template illustrated in Table 2, with the { query } variable assigned to the new user query and the { context } variable assigned to the top-ranked samples in the database according to the similarity metric with the user query. The LangChain library is used to create the prompt according to the template and the OpenAI text-davinci-003 model is used as the foundation LLM.”
Here, the Examiner notes that the created prompt using the similar samples would read on the target entry in the lookup table being returned.).
However, Kotaru et al. does not explicitly teach, but Manjunath et al. does teach:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry (see ¶ [0055]: “In some embodiments, where a direct match is not found in the list of entities based on a search of the latent or semantic space for a similar embedding, a large language model (LLM) may be queried with the text from the query, the closest entities found, and/or additional information, and the LLM may be used to help determine a closest entity. For example, the LLM may be queried for synonyms or slang terms of an item, topping, size, etc. in the query, and the synonyms or slang terms may be converted to an embedding, and a search may be performed in the embedding space to determine if any entities match any of the synonyms or slang terms. In this way, an embedding may be found for one of the alternative synonyms or slang words, and the corresponding entity may be associated with the query. As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”
and ¶ [0079]: “The CPU(s) 1006 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 1000 to perform one or more of the methods and/or processes described herein. The CPU(s) 1006 may each include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) that are capable of handling a multitude of software threads simultaneously. The CPU(s) 1006 may include any type of processor, and may include different types of processors depending on the type of computing device 1000 implemented (e.g., processors with fewer cores for mobile devices and processors with more cores for servers). ”),
wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table (see ¶ [0055] citation as in limitation above, more specifically: “…As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”);
the language model comprises a transformer-based large language model pre-trained on sentence datasets (see ¶ [0061]: “Referring back to the example of FIG. 1, in addition to, or alternatively from, the response component 124 using the templates to generate the responses, in other examples, the response component 124 may use one or more models, such as one or more language models (e.g., LLMs), to generate the responses. For instance, as described herein, the model(s) may include, but is not limited to, a Bidirectional Encoder Representations from Transformers (BERT) model, a Generative Pre-Trained Transformer 2 (GPT-2) model, a XLNet model, and/or any other type of language model. When using such a model(s), the response component 124 may use the similarity data 122, the intent/entity data 112, the text data 108, and/or the templates data 126 to generate the responses. ”),
Kotaru et al. and Manjunath et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Manjunathet al. of receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table; and the language model comprises a transformer-based large language model pre-trained on sentence datasets which provides the benefit of being able to interpret the query from the user, even if the entity is not an exact match to a stored entity ([abstract] of Manjunathet al.).
However, Kotaru et al. in combination with Manjunath et al. do not explicitly teach, but Laprise et al. does teach:
the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison (see ¶ [0035-0037, and 0051]: “[0035] In at least one embodiment, a language model can be used to generate representations of operational design domains (ODDs)—such as intersections—using a tokenized representation, written in a language such as Road Topology Language (RTL). An embedding can be generated for each such ODD, allowing each ODD to be represented by, for example, a point in an n-dimensional latent space. Similar ODDs, such as similar intersections, will have similar embeddings. When these embeddings are considered in a latent space, for example, these similar ODD embeddings will be represented by points that are proximate, or located a small distance from, each other in the latent space. This proximity of points for similar ODDs allows for clustering of similar ODDs, where the clustering can be performed at various levels, allowing for clusters and sub-clusters of varying granularities. Generated embeddings can capture semantic, geometric, topological, and/or other information for the ODDs. Such similarity-based clustering of ODDs does not require a dependence on human input or a predefined and fixed terminology, and can be constructed without supervision. [0036] In at least one embodiment, clustering of similar ODDs based on proximity of corresponding points in a latent space (or other such determinable similarity or similarity criterion) allows for automatic labeling of ODDs based at least in part upon labels applied to similar ODDs, such as those applied to a cluster in which an embedding generated for a new or current ODD is contained. These applied labels can be useful to improve map quality, for example, and can also be used as ground truth training data for fine-tuning, or adjusting parameters of, generative and/or language models. The labels selected to be applied in at least one embodiment may not include all possible labels for a given environment or domain, but may instead include those labels determined to be most useful, informative, or challenging to specific algorithms or models, as may vary between use cases. [0037] The availability of such embeddings, feature vectors, latent points, or other such representations also allows for similarity-based querying. This can include querying of map-related data, where sensor data (or other observations or data representative of a location) can be fed to a language model that can generate an embedding for a new or current ODD, and this embedding can be used to locate information for similar ODDs by locating similar or proximate embeddings in, for example, the latent space. In at least this context, an embedding can be from a single token (or other discrete object or component that may be comprised of text and/or other data), sentence, paragraph, document, graph, or other such format. An embedding can be generated from other sources as well, such as image patches and audio clips, among other such options. In other embodiments, similarity searching can be performed for a feature vector against feature vectors in a vector database (VDB) for an environment, among other such options. A query to be used for the search can correspond to an example rather than a set of specifics, where the example can correspond to, or be based at least in part on, a text (e.g., RTL)-based embedding (or other tokenized description) generated automatically from the sensor data using a language model, for example, without the need for a user to provide a specific query representative of the location, or provide any manual labeling or description of the location for which the query is to be performed. The example tokenized description is generated using a language model trained for a specific domain, and can include semantic, topology, and geometry data that can be leveraged in the similarity search. Such an approach allows for similar representations to be identified without a user having to predefine or specify a category or attribute of interest from a pre-defined and fixed set. The ability to determine annotations or other information automatically, without user input, can provide benefits for various types of operations, such as may be directed to the improvement of the generation of maps (or similar environment representation). [0051] … As a trained LLM will know how to manipulate or fill in a sentence in natural language, so can an LLM learn to fill in a text string in a structured representation language. The LLM can also infer relationships between objects based on its understanding of the language. The LLM can then generate a unified text representation of an environment that can include information that was not present or determinable from the input alone but that allows the environment to be more realistic and to comply with real world rules and/or constraints…”
and ¶ [0091-0093 and 0114]: “[0091] Performing unsupervised clustering of points determined using descriptions generated using RTL or another such language can also help to identify clusters that may not have been previously considered, or that may correspond to similarities that may not be easily explainable using human language. For example, there may be regions with features whose representative vectors end up being similar, with proximate points in latent space, even though to a human the regions may appear to be significantly different. There may be certain aspects of those regions that are similar in at least some ways, and that may be associated with certain behaviors or relationships that are shared across a cluster. When membership to such a cluster is determined, these learned behaviors or relationships can then be applied, where previously those similarities may not have been apparent or readily discoverable. [0092] Such functionality also can help to improve efficiency and reduce the cost of generating or updating maps or other environment representations. As mentioned previously, it can be beneficial in at least one embodiment to identify various map domains, such as Operational Design Domains (ODD), as well as to maximize the coverage of these ODDs when planning. In at least one embodiment, ODDs such as road intersections and other map structures in documents can be represented using a domain-specific language, such as RTL. As discussed in detail elsewhere herein, a large language model (LLM) can take input (e.g., sensor data or other observations) relating to a location or region, and generate one or more embeddings for each document. Such embeddings generated using an LLM, and expressed in a language such as RTL, can capture the semantic and topological information of local regions, domains, and/or maps. These semantic-aware embeddings can also support arithmetic operations, for example, which allows for the identification of patterns, relationships, and other aspects of the underlying map data. The ability to perform unsupervised clustering of these embeddings (or feature vectors, etc.) allows similar ODD instances to be identified as belonging to the same cluster. The clusters can thus correspond to road network structures, or other such structures or designs, that share similar characteristics, yielding unsupervised labeling of ODDs. [0093] In at least one embodiment, vector map features representing maps can be exported into a topology graph. The topology graphs can express topology at various levels, such as at the lane level or at the level of individual road segments. The topology graph can be exported and then split into a variety of sub-graphs. The sub-graphs can be selected or determined using any appropriate segmentation or selection criteria. The individual sub-graphs can, in turn, be converted into documents in the RTL language that is descriptive of that sub-graph, including semantic, topology, and geographic data for the region associated with the sub-graph. These sub-graphs can be used for a variety of purposes, such as to train a large language model. In at least one embodiment, the learned embeddings from the LLM can be stored in a repository, such as a latent space or a vector database, allowing for fast retrieval of similar sub-graphs. If a set of arbitrary map features is received as input, these features can be converted to an RTL document and the embeddings retrieved from the LLM. The embeddings can then be sued to query the vector database or other repository in order to receive identification of a set of similar sub-graphs as search results. [0114] An LLM 516 trained with an RTL (or other DSL) corpus built from a database 512 of map ground truth data can be queried to correct features output from a machine learning (ML) automation pipeline. The output of an LLM 516—such as by using a writer and/or parser component or module 520—can be mapped back to the extracted features. A difference (e.g., diff) operation can then be performed with respect to inferred landmarks from an automation component 518, for example, to perform any appropriate corrections to generate a map graph 522. An example use case is to infer the road topology (e.g., edges) from an incomplete set of nodes (e.g., landmarks) with potential applications in, for example and without limitation, tooling, quality assurance (QA), and automation. In some embodiments, document embeddings may be indexed in a vector database, or n-dimensional latent space (where n can represent a number of extracted features or feature types), and the index can then be used to cluster similar intersections—thus allowing the unsupervised labeling and retrieval of operation design domains (ODDs) (e.g., features or landmarks).”).
Kotaru et al., Manjunath et al., and Laprise et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. in combination with Manjunath et al. to incorporate the teachings of Laprise et al. of the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison which provides the benefit of improving the robustness and reliability of the system, and can help to ensure that the processed data has proper semantic meaning and is well-formed ([0109] of Laprise et al.).
As to independent claim 11, Kotaru et al. teaches:
11. A system (see ¶ [0003]: “Disclosed is a conversational artificial intelligence (AI) system which features a question-and-answer interface with advanced capabilities for providing accurate and relevant responses related to wireless communication specifications in the 5G mobile networking space. The system includes three elements, including a domain-specific database, a context extractor, and a feedback mechanism…”) comprising:
a computer processor (see ¶ [0003] above and further ¶ [0169]: “A further example includes a computer-readable storage device storing computer-executable instructions, which when executed by at least one processor in a computing device, synthesize information from wireless communication specifications using a conversational artificial intelligence (AI) system employable by a user, the computer-readable storage device including instructions executable by the at least one processor for: building a domain-specific database of wireless communication specifications; providing a base foundation large language model (LLM) for identifying relevant information from the domain-specific database in response to queries from the user made through a user interface (UI) to the conversational AI system; generating tokenized text from the domain-specific database from which word embedding vectors are generated; training the LLM using the word embedding vectors from the domain-specific database; generating tokenized text for a query received from the user at the UI from which word embedding vectors are generated; using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities; providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM; and creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.”);
a data repository in communication with the computer processor (see ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “…the computer-readable storage device including instructions executable by the at least one processor…”) and storing:
a query comprising a natural language statement having (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028 and 0054] citations as in claim 1, above. More specifically: “[0054] …When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.”, and ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “queries from the user”),
, (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citations as in claim 1, above. More specifically: ¶ [0054]: “… When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.” and ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.” and ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “queries from the user”),
a query vector comprising query data structure storing a semantic meaning of the query, in a multi-dimensional dense vector space, (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citations as in claim 1, above. More specifically: “¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs. OpenAI embedding model openai-textembedding-ada-002 is used to transform each of the text samples into word embedding vectors. The resulting vector representation for each sample is stored in an index for efficient querying. The resulting domain-specific database of domain-specific word embedding vectors consists of 86 million tokens.” and ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “generating tokenized text from the domain-specific database from which word embedding vectors are generated”),
a lookup table (see ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “domain-specific database”),
a found entry in the lookup table and the target entry in the lookup table (see ¶ [0003 and 0169] citations as in limitations above. More specifically ¶ [0169]: “word embedding vectors for the domain-specific database to identify matches having semantic similarities”), and
a lookup vector comprising a lookup data structure storing the plurality of entries of the lookup table and encoded semantic meanings of the plurality of entries, a large language model which, when applied by the processor to the query (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citations as in claim 1, above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.” and ¶ [0003 and 0169] citations as in limitations above and further ¶ [0138] citation as in limitation above and further: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”),
generates the query vector (see ¶ [0003 and 0169] citations as in limitations above and further ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above.),
a semantic matching algorithm which, when applied by the processor to both the query vector and the lookup vector, compares the query vector to the lookup vector and returns, as a result of comparing, the found entry in the lookup table (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138] citation as in limitations above.”); and
a lookup algorithm which, when applied by the processor to the lookup table using the found entry, looks up the target entry in the lookup table and returns the target entry (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.”, ¶ [0068] “Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors”, ¶ [0138]: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.” and ¶ [0139]: “Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.”)
the found entry comprising a closest semantic match to the query vector (see ¶ [0067 and 0138] citations as in limitation(s) above, more specifically: ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and further ¶ [0068 and 0139]: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt).
However, Kotaru et al. does not explicitly teach, but Manjunath et al. does teach:
a query comprising a natural language statement having (see ¶ [0055 and 0079] citations as in claim 1, above. More specifically: “[0055] In some embodiments, where a direct match is not found in the list of entities based on a search of the latent or semantic space for a similar embedding, a large language model (LLM) may be queried with the text from the query, the closest entities found, and/or additional information, and the LLM may be used to help determine a closest entity. For example, the LLM may be queried for synonyms or slang terms of an item, topping, size, etc. in the query, and the synonyms or slang terms may be converted to an embedding, and a search may be performed in the embedding space to determine if any entities match any of the synonyms or slang terms. In this way, an embedding may be found for one of the alternative synonyms or slang words, and the corresponding entity may be associated with the query. As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”),
wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table (see ¶ [0055] citation as in limitation above, more specifically: “…As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”),
wherein the large language model comprises a transformer-based large language model pre-trained on sentence datasets (see ¶ [0061]: “Referring back to the example of FIG. 1, in addition to, or alternatively from, the response component 124 using the templates to generate the responses, in other examples, the response component 124 may use one or more models, such as one or more language models (e.g., LLMs), to generate the responses. For instance, as described herein, the model(s) may include, but is not limited to, a Bidirectional Encoder Representations from Transformers (BERT) model, a Generative Pre-Trained Transformer 2 (GPT-2) model, a XLNet model, and/or any other type of language model. When using such a model(s), the response component 124 may use the similarity data 122, the intent/entity data 112, the text data 108, and/or the templates data 126 to generate the responses. ”)
Kotaru et al. and Manjunathet al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Manjunath et al. of receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table; and the language model comprises a transformer-based large language model pre-trained on sentence datasets which provides the benefit of being able to interpret the query from the user, even if the entity is not an exact match to a stored entity ([abstract] of Manjunath et al.).
However, Kotaru et al. in combination with Manjunath et al. do not explicitly teach, but Laprise et al. does teach:
wherein the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison (see ¶ [0035-0037, and 0051] and ¶ [0091-0093 and 0114] citations as in claim 1, above.).
Kotaru et al., Manjunath et al., and Laprise et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. in combination with Manjunath et al. to incorporate the teachings of Laprise et al. of wherein the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison which provides the benefit of improving the robustness and reliability of the system, and can help to ensure that the processed data has proper semantic meaning and is well-formed ([0109] of Laprise et al.).
Regarding claims 2 and 12, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
Kotaru et al. further teaches:
2 and 12. The method/system of claims 1 and 11, wherein [the semantic matching algorithm (claim 12)] comparing the query vector to the lookup vector (see ¶ [0066-0067 and 0138] citations as in claim 1 above. More specifically: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”) comprises identifying the found entry in the lookup vector as having a least semantic distance to the query vector, relative to other entries in the lookup table (see ¶ [0067 and 0138] citations as in claim 1 and limitation(s) above and further ¶ [0068 and 0139]: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors. [0139] Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.” and ¶ [0089]: “2) Similarity metrics: Three different similarity metrics are considered: cosine similarity, L2 distances and L1 distances…”).
Regarding claims 6 and 16, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
Kotaru et al. further teaches:
6. The method of claim 1, further comprising:
applying, prior to applying the semantic matching algorithm, the large language model to the lookup table to generate the lookup vector (see ¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs. OpenAI embedding model openai-textembedding-ada-002 is used to transform each of the text samples into word embedding vectors. The resulting vector representation for each sample is stored in an index for efficient querying. The resulting domain-specific database of domain-specific word embedding vectors consists of 86 million tokens.” and ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”).
16. The system of claim 11, wherein the large language model, when applied by the processor to the lookup table prior to applying the semantic matching algorithm, generates the lookup vector (see ¶ [0064 and 0066] citations as in claim 6, above.).
Regarding claims 7 and 17, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
Kotaru et al. further teaches:
7. The method of claim 1, further comprising:
receiving, prior to applying the semantic matching algorithm, a new entry to a new lookup table (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.”, ¶ [0068] “Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors”, ¶ [0138]: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.” and ¶ [0139]: “Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.”); and
applying, prior to applying the semantic matching algorithm, the large language model to the new lookup table to generate the lookup vector (see ¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs. OpenAI embedding model openai-textembedding-ada-002 is used to transform each of the text samples into word embedding vectors. The resulting vector representation for each sample is stored in an index for efficient querying. The resulting domain-specific database of domain-specific word embedding vectors consists of 86 million tokens.”),
wherein the new lookup table is the lookup table when looking up the target entry (see ¶ [0064] citation above.).
17. The system of claim 11, wherein the data repository further stores a new lookup table (see ¶ [0066-0068 and 0138-0139] citations as in claim 7 above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used…”),
the large language model, when applied by the processor to the new lookup table prior to applying the semantic matching algorithm, generates the lookup vector (see further ¶ [0064] citation as in claim 17 above.), and
the new lookup table is the lookup table when the lookup algorithm returns the target entry (see further ¶ [0064] citation as in claim 17 above.).
Regarding claims 9 and 18, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
However, Kotaru et al. in combination with Manjunath et al. do not explicitly teach, but Laprise et al. does teach:
9 and 18. The method/system of claims 1 and 11, wherein the large language model is programmed to map phrases to a multi-dimensional dense vector space suitable for a computer to perform vector similarity comparisons (see ¶ [0035-0037, and 0051] and ¶ [0091-0093 and 0114] citations as in claim 1, above.).
Kotaru et al. and Laprise et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Laprise et al. of wherein the large language model is programmed to map phrases to a multi-dimensional dense vector space suitable for a computer to perform vector similarity comparisons which provides the benefit of improving the robustness and reliability of the system, and can help to ensure that the processed data has proper semantic meaning and is well-formed ([0109] of Laprise et al.).
Regarding claims 10 and 19, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
Kotaru et al. further teaches:
10. The method of claim 1, wherein returning comprises providing the target entry to a data processing algorithm programmed to process the target entry to generate a secondary result (see ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” ¶ [0068]: “Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors.”, ¶ [0138]: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”, and ¶ [0139]: “Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.”).
19. The system of claim 11, wherein the system further comprises:
a data processing algorithm which, when applied by the processor to the target entry, processes the target entry to generate a secondary result (see ¶ [0067-0068 and 0138-039] citations as in claim 10 above.).
Claim 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1), Laprise et al. (US 20240419907 A1) and Lindrup et al. (US 20250046330 A1).
As to independent claim 20, Kotaru et al. teaches:
20. A method (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above…” and ¶ [0135]: “FIG. 16 is a flowchart 1600 of an illustrative method that may be performed when implementing the present conversational AI system in which a foundation model is adapted for information synthesis of wireless communication specifications…”) comprising:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028]: “FIG. 18 is a block diagram of an illustrative server or computing device that may be used at least in part to implement the present conversational AI system;” and ¶ [0054]: “FIG. 3 shows an illustrative architecture of the present conversational AI system 300. Prior to delving into the specifics, the following provides a concise summary of the operational procedure of the system. When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.”),
(see Fig. 2 (“what is numerology in 5G”) and ¶ [0028 and 0054] citation(s) as in limitation above, more specifically: “FIG. 18 is a block diagram of an illustrative server or computing device that may be used at least in part to implement the present conversational AI system;” and ¶ [0054]: “… When a user types in a query in a message bar 205 as shown in FIG. 2, the system appends the query with relevant text samples from the domain-specific database and feeds the foundation model with the resulting combined prompt as the input.” and further ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”);
executing, by the processor, a large language model to a lookup table to generate a lookup vector (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, 0064, 0066-0067 and 0137-0138] citation(s) as in claim 1, above and further ¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs. OpenAI embedding model openai-textembedding-ada-002 is used to transform each of the text samples into word embedding vectors. The resulting vector representation for each sample is stored in an index for efficient querying. The resulting domain-specific database of domain-specific word embedding vectors consists of 86 million tokens.”),
wherein the lookup vector comprises a lookup data structure storing a plurality of semantic meanings of a plurality of entries of the lookup table and encoded semantic meanings of the plurality of entries (see ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, and ¶ [0138]: “Block 1625 includes generating tokenized text for a query received from the user at the UI from which word embedding vectors are generated. Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”);
executing, by the processor and after applying the large language model to the lookup table, a large language model on the query to generate a query vector in a multi-dimensional dense vector space (see ¶ [0064, 0066 and 0138] citations as claim 1, above, more specifically: Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citations as in claim 1, above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.” and ¶ [0003 and 0169] citations as in limitations above and further ¶ [0138] citation as in limitation above and further: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.” and further ¶ [0064]: “As discussed above, the text needs to be tokenized and converted to word embedding vectors for the consumption by LLMs…”, ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0137]: “Block 1610 includes providing a base foundation large language model (LLM) for identifying relevant information from the domain-specific database in response to queries from the user made through a user interface (UI) to the conversational AI system …” and ¶ [0138]: “Block 1625 includes generating tokenized text for a query received from the user at the UI from which word embedding vectors are generated. ”), wherein:
the query vector comprises a query data structure storing a semantic meaning of the query (see ¶ [0138] citation as in limitation above and further: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”);
executing, by the server controller, a semantic matching algorithm to both the query vector and the lookup vector (see Fig. 2 (“what is numerology in 5G”) and ¶ [0028, 0054, and 0066] citations as in claim 1 and further ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.”, ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138] citation as in limitations above. More specifically: “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”), and wherein the semantic matching algorithm (see ¶ [0066-0067 and 0138] citations as in limitation(s) above.) further performs:
comparing the query vector to the lookup vector and returning a plurality of semantic distances between the query vector and a plurality of entries in the lookup table (see ¶ [0066-0067 and 0138] citations as in limitation(s) above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.” ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138] : “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.”),
However, Kotaru et al. does not explicitly teach, but Manjunath et al. does teach:
receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry (see ¶ [0055]: “In some embodiments, where a direct match is not found in the list of entities based on a search of the latent or semantic space for a similar embedding, a large language model (LLM) may be queried with the text from the query, the closest entities found, and/or additional information, and the LLM may be used to help determine a closest entity. For example, the LLM may be queried for synonyms or slang terms of an item, topping, size, etc. in the query, and the synonyms or slang terms may be converted to an embedding, and a search may be performed in the embedding space to determine if any entities match any of the synonyms or slang terms. In this way, an embedding may be found for one of the alternative synonyms or slang words, and the corresponding entity may be associated with the query. As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”
and ¶ [0079]: “The CPU(s) 1006 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 1000 to perform one or more of the methods and/or processes described herein. The CPU(s) 1006 may each include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) that are capable of handling a multitude of software threads simultaneously. The CPU(s) 1006 may include any type of processor, and may include different types of processors depending on the type of computing device 1000 implemented (e.g., processors with fewer cores for mobile devices and processors with more cores for servers). ”);
(see ¶ [0055] citation as in limitation above, more specifically: “…As an example, where a user request “pop,” synonyms for pop may be returned by the LLM, such as “soda” or “beverage.” Once soda or beverage are returned, embeddings may be generated for these terms, and the embedding space may be searched to find that the entity is “beverages.” As such, the proper entity may be determined even where an initial similarity for “pop” was not identified in the embedding space. In such examples, the LLM may be queried each time there is not a direct match, or the LLM may be queried each time the dissimilarity is greater than some threshold in the embedding space.”);
the language model comprises a transformer-based large language model pre-trained on sentence datasets (see ¶ [0061]: “Referring back to the example of FIG. 1, in addition to, or alternatively from, the response component 124 using the templates to generate the responses, in other examples, the response component 124 may use one or more models, such as one or more language models (e.g., LLMs), to generate the responses. For instance, as described herein, the model(s) may include, but is not limited to, a Bidirectional Encoder Representations from Transformers (BERT) model, a Generative Pre-Trained Transformer 2 (GPT-2) model, a XLNet model, and/or any other type of language model. When using such a model(s), the response component 124 may use the similarity data 122, the intent/entity data 112, the text data 108, and/or the templates data 126 to generate the responses. ”),
Kotaru et al. and Manjunath et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Manjunath et al. of receiving, at a server controller executable by a processor, a query comprising a natural language statement having an indirect reference to a target entry, wherein the indirect reference comprises information that does not directly identify the target entry, but which has a first semantic meaning comparable to semantic meanings of a plurality of entries in a lookup table; and the language model comprises a transformer-based large language model pre-trained on sentence datasets which provides the benefit of being able to interpret the query from the user, even if the entity is not an exact match to a stored entity ([abstract] of Manjunath et al.).
However, Kotaru et al. in combination with Manjunath et al. do not explicitly teach, but Laprise et al. does teach:
the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison (see ¶ [0035-0037, and 0051] and ¶ [0091-0093 and 0114] citations as in claim 1, above.).
Kotaru et al., Manjunath et al., and Laprise et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. in combination with Manjunath et al. to incorporate the teachings of Laprise et al. of wherein the large language model is programmed to map phrases to the multi-dimensional dense vector space for vector similarity comparison which provides the benefit of improving the robustness and reliability of the system, and can help to ensure that the processed data has proper semantic meaning and is well-formed ([0109] of Laprise et al.).
However, Kotaru et al. in combination with Manjunath et al. and Laprise et al. does not explicitly teach, but Lindrup et al. does teach:
comparing the plurality of semantic distances to a threshold value (see ¶ [0069, 0074, 0088, and 0091]: “[0069] … the word embeddings from the first sound signal, is compared with the word embeddings for each of the sound source signals and if two word embeddings from each of the two signals to be compared are determined to be close (i.e. that some similarity metric is higher than some threshold) then this increases a similarity counter with one and hereby the sound source signal that the user is most likely to be paying attention to can be determined as the signal having the highest similarity counter score. [0074] One Language model capable of carrying out the tasks above is the Natural Language API from Google. Alternatives include e.g. Amazon comprehend, Microsoft Text Analytics, Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer 3 (GPT-3), which can understand semantics and are able to predict meaningful words and sentence continuations. [0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”),
adding a set of entries, from the plurality of entries, to a list of candidate entries when a corresponding semantic distance in the plurality of semantic distances satisfies the threshold value (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above. More specifically: “[0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”), and
transmitting the list of candidate entries to a remote user device (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above and further ¶ [0149]: “…further on to the loudspeaker 406 that provides the audio output corresponding to the output signal.”);
receiving, at the server controller, a user a selection of one of the candidate entries as being a found entry in the lookup table (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above. More specifically: [0091] “According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure…” and further ¶ [0006]: “Within the present context an audio device system may comprise a single audio device (a so called monaural audio device system) or comprise two audio devices, one for each ear of the user (a so called binaural audio device system). […] the audio device system may also include a remote microphone system (which generally can also be considered a computing device) comprising additional microphones and/or may even include a remote server providing abundant processing resources and generally these additional devices will also include link means adapted to operationally connect to the various other devices of the audio device system.”);
looking up, by the server controller, using the found entry in the lookup table, the target entry in the lookup table (see ¶ [0006, 0069, 0074, 0088, and 0091] citations as in limitation above . More specifically: [0091] “According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure…”),
the found entry comprising a closest semantic match to the query vector (see ¶ [0006, 0069, 0074, 0088, and 0091] citations as in limitation above and further ¶ [0089]: “According to another embodiment the sound source signal having the highest score of at least one of the semantic similarity measure and the syntactic similarity measure is selected as the output signal.”); and
returning, by the server controller, the target entry (see ¶ [0006, 0069, 0074, 0088, and 0091] citations as in limitation above and further ¶ [0149]: “…further on to the loudspeaker 406 that provides the audio output corresponding to the output signal.”).
Kotaru et al., Manjunath et al., Laprise et al. and Lindrup et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. in combination with Manjunath et al., Laprise et al. to incorporate the teachings of Lindrup et al. of comparing the plurality of semantic distances to a threshold value, adding a set of entries, from the plurality of entries, to a list of candidate entries when a corresponding semantic distance in the plurality of semantic distances satisfies the threshold value, and transmitting the list of candidate entries to a remote user device; receiving, at the server controller, a user a selection of one of the candidate entries as being a found entry in the lookup table; looking up, by the server controller, using the found entry in the lookup table, the target entry in the lookup table, the found entry comprising a closest semantic match to the query vector; and returning, by the server controller, the target entry which provides the benefit of enhancing conversation signals ([0145] of Lindrup et al.).
Claims 3 and 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and further in view of Laprise et al. (US 20240419907 A1) as applied to claims 1 and 11 above, and further in view of Lindrup et al. (US 20250046330 A1).
Regarding claims 3 and 13, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
3 and 13. The method of claims 1 and 11, wherein [the semantic matching algorithm (claim 13)] comparing the query vector to the lookup vector (see ¶ [0066-0067 and 0138] citations as in claim 1 above. More specifically: ¶ [0066]: “When the user inputs a query into the present conversational AI system, the query is pre-processed, tokenized, and transformed into a word embedding vector, using methods similar to those described above. The query vector is then compared to each of the word embedding vectors in the domain-specific database to identify the samples in the database that are closest to the query semantically.” ¶ [0067]: “To match the user query to relevant samples in the database, a similarity metric is used. The most common similarity metric used in natural language processing (NLP) is cosine similarity. Cosine similarity measures the angle between two vectors in a high dimensional space. If two vectors are very similar, their cosine similarity will be close to 1. If they are very different, their cosine similarity will be close to 0. Other similarity metrics have been evaluated, as discussed below.” and ¶ [0138] : “…Block 1630 includes using a context extractor to augment the query with context from the domain-specific database by comparing word embedding vectors for the query to word embedding vectors for the domain-specific database to identify matches having semantic similarities.” and further ¶ [0089]: “ 2) Similarity metrics: Three different similarity metrics are considered: cosine similarity, L2 distances and L1 distances…”) comprises:
identifying the found entry in the lookup vector as having a semantic distance to the query vector (see ¶ [0066-0067, 0089 and 0138] citations as in limitation and/or claim 1 above. “similarity metrics” More specifically: ¶ [0089]: “ 2) Similarity metrics: Three different similarity metrics are considered: cosine similarity, L2 distances and L1 distances…”);
However, Kotaru et al. does not explicitly teach, but Lindrup et al. does teach:
comparing the semantic distance to a threshold value (see ¶ [0069, 0074, 0088, and 0091]: “[0074] One Language model capable of carrying out the tasks above is the Natural Language API from Google. Alternatives include e.g. Amazon comprehend, Microsoft Text Analytics, Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer 3 (GPT-3), which can understand semantics and are able to predict meaningful words and sentence continuations. [0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”); and
returning the found entry when the semantic distance satisfies the threshold value (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above. More specifically: “[0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”).
Kotaru et al. and Lindrup et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Lindrup et al. of comparing the semantic distance to a threshold value and returning the found entry when the semantic distance satisfies the threshold value which provides the benefit of enhancing conversation signals ([0145] of Lindrup et al.).
Claims 4 and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and further in view of Laprise et al. (US 20240419907 A1) as applied to claims 1 and 11 above, and further in view of Lindrup et al. (US 20250046330 A1) and Kim et al. (US 20210150155 A1).
Regarding claims 4 and 14, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
Kotaru et al. further teaches:
4. The method of claim 1, wherein [the semantic matching algorithm (claim 14)] comparing the query vector to the lookup vector (see ¶ [0066-0067 and 0138] citations as in claim 1 above, and more specifically the citations re-incorporated in claims 3 and 13, above.) comprises [the semantic matching algorithm (claim 14)]:
identifying the found entry in the lookup vector as having a semantic distance to the query vector (see ¶ [0066-0067, 0089 and 0138] citations as in limitation and/or claim 1 above. “similarity metrics” More specifically: ¶ [0089]: “ 2) Similarity metrics: Three different similarity metrics are considered: cosine similarity, L2 distances and L1 distances…”);
However, Kotaru et al. does not explicitly teach, but Lindrup et al. does teach:
comparing the semantic distance to a first threshold value (see ¶ [0069, 0074, 0088, and 0091]: “[0069] … the word embeddings from the first sound signal, is compared with the word embeddings for each of the sound source signals and if two word embeddings from each of the two signals to be compared are determined to be close (i.e. that some similarity metric is higher than some threshold) then this increases a similarity counter with one and hereby the sound source signal that the user is most likely to be paying attention to can be determined as the signal having the highest similarity counter score. [0074] One Language model capable of carrying out the tasks above is the Natural Language API from Google. Alternatives include e.g. Amazon comprehend, Microsoft Text Analytics, Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer 3 (GPT-3), which can understand semantics and are able to predict meaningful words and sentence continuations. [0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”);
comparing, (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above.);
adding, (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above. More specifically: “[0088] According to one embodiment the sound source signal having the word embedding similarity measure that is most similar with the word embedding similarity measure of the first sound signal is selected as the output signal. [0091] According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure, the syntactic similarity measure, a sound pressure level score reflecting the strength of the signal, a previous participant score reflecting whether the speaker representing the sound source signal has previously participated in the conversation that the user is paying attention to and having a speech onset within said predetermined duration after speech ending of the first sound signal.”);
transmitting, to a user device, the list of candidate entries (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above and further ¶ [0149]: “…further on to the loudspeaker 406 that provides the audio output corresponding to the output signal.”); and
receiving, from the user device, a selection of the found entry from the list of candidate entries (see ¶ [0069, 0074, 0088, and 0091] citations as in limitation above . More specifically: [0091] “According to another embodiment the sound source signal having a highest combined score is selected as output signal, wherein the combined score is obtained by combining at least some of: the word embedding similarity measure score, the semantic similarity measure…”).
Kotaru et al. and Lindrup et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Lindrup et al. of comparing the semantic distance to a first threshold value, comparing the semantic distance to a second threshold value; adding the found entry to a list of candidate entries comprising additional entries in the lookup vector; transmitting, to a user device, the list of candidate entries; and receiving, from the user device, a selection of the found entry from the list of candidate entries which provides the benefit of enhancing conversation signals ([0145] of Lindrup et al.).
However, Kotaru et al. in combination with Lindrup et al. do not explicitly teach, but Kim et al. does teach:
comparing/adding, responsive to the semantic distance satisfying the second threshold value (see ¶ [0005]: “In one general aspect, a processor-implemented method includes determining a first similarity between an input sentence of a user and a select first database query sentence, and dependent on a determination that the first similarity fails to meet a first threshold, determining a second similarity between a portion of the input sentence, less than all of the input sentence, and a select second database query sentence, and in response to the second similarity meeting a second threshold, outputting a response sentence corresponding to the second database query sentence as a response to the input sentence.”)
Kotaru et al., Lindrup er al. and Kim et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. in combination with Lindrup et al. to incorporate the teachings of Kim et al. of comparing/adding, responsive to the semantic distance satisfying the second threshold value which provides the benefit of generating most appropriate responses ([0058] of Kim et al.).
Claims 5 and 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and further in view of Laprise et al. (US 20240419907 A1) as applied to claims 1 and 11 above, and further in view of Bender et al. (US 20210294829 A1).
Regarding claims 5 and 15, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claims 1 and 11, above.
5 and 15. The method/system of claims 1 and 11,
wherein [the semantic matching algorithm (claim 15)] looking up the target entry (see ¶ [0067-0068 and 0138-0139] citations as in limitation(s) above. More specifically: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors.”) comprises:
looking up, using the found entry, a plurality of second entries in the lookup table (see ¶ [0067-0068 and 0138-0139] citations as in limitation(s) above. More specifically: “[0068] Using the similarity metric, the NLP model can rank the samples in the database according to their similarity to the user query. The most similar samples are then used to create a prompt. The main challenge is efficient retrieval of relevant vectors from the database. FAISS (Facebook AI Similarity Search) library is used to efficiently compute the similarity between the query vector and all the stored vectors.”
Here, the Examiner notes that the most similar samples from all the stored vectors would read on the found entry and the plurality of second entries in the lookup table.),
wherein the target entry is among the plurality of second entries (see ¶ [0067-0068 and 0138-0139] citations as in limitation(s) above.);
transmitting, to a user device, the plurality of second entries (see ¶ [0067-0068 and 0138-0139] citations as in limitation(s) above. More specifically: “[0139] Block 1635 includes providing a conversational response to the user's query through the UI, the response being generated based on tailored prompts to the LLM. Block 1640 includes creating an issue for resolution by expert feedback, the issue including the user query, the response, and the context, wherein the resolved issue is incorporated into the domain-specific database.” and further ¶ [0070]: “The present conversational AI system incorporates few-shot learning techniques by creating a prompt template with few pre-defined example queries, relevant context obtained from specifications and reference responses. A prompt is then generated, following the template illustrated in Table 2, with the { query } variable assigned to the new user query and the { context } variable assigned to the top-ranked samples in the database according to the similarity metric with the user query. The LangChain library is used to create the prompt according to the template and the OpenAI text-davinci-003 model is used as the foundation LLM.”
Here, the Examiner notes that the created prompt using the similar samples would read on the target entry in the lookup table being returned.); and
However, Kotaru et al. does not explicitly teach, but Bender et al. does teach:
receiving, from the user device, a selection of the target entry in the lookup table (see Figs. 4 and 5A-B and ¶ starting at Col. 3, line 4: “(8) In some implementations, one or more of the processors are remote from the client device and providing the responses of the selected set for display in the interface includes transmitting content to the client device to cause the client device to visually render the responses of the selected set in the interface a selectable manner. […] (11) In some implementations, determining that the second response satisfies the difference criterion relative to the first response based on comparing the first embedding to the second embedding includes: calculating a measure between the first embedding and the second embedding, and determining that the measure satisfies a threshold. In some of those implementations, the measure is a cosine similarity measure, a dot product, or a Euclidian distance. (12) In some implementations, the candidate set of responses are ranked and selecting the first response is based on it being the most highly ranked. In some of those implementations, the method further includes: identifying an additional response, of the responses of the candidate set, that is the next most highly ranked response following the first response; generating an additional embedding over the neural network response encoder model based on applying the additional response as input to the neural network response encoder model; determining that the additional response fails to satisfy the difference criterion relative to the first response based on comparing the first embedding and the additional embedding; and omitting the additional response from the selected set based on the second response failing to satisfy the difference criterion.”).
Kotaru et al. and Bender et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Bender et al. of receiving, from the user device, a selection of the target entry in the lookup table which provides the benefit of increasing the likelihood that one of the provided responses is sufficient to convey the essence of an intended response of the user (Col. 2, line 27 of Bender et al.).
Claim 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Kotaru et al. (US 20240330589 A1) further in view of Manjunath et al. (US 20240370690 A1) and further in view of Laprise et al. (US 20240419907 A1) as applied to claim 1 above, and further in view of Bayless et al. (US 20250112878 A1).
Regarding claim 8, Kotaru et al. in combination with Manjunath et al., and Laprise et al. teach the limitations as in claim 1, above.
However, Kotaru et al. does not explicitly teach, but Bayless et al. does teach:
8. The method of claim 1, wherein the large language model comprises a transformer-based large language model that is pre-trained on sentence data sets (see ¶ [0045]: “…LLMs 118 are trained on large sets or corpuses of text data to generate human-like textual responses to prompts. LLMs 118 are generally trained in two stages, pre-training and fine-tuning. During the pre-training stage, LLMs 118 are trained on massive datasets of unlabeled text data (or “unsupervised learning”) where transformers allow the LLMs 118 to process and learn the patterns and relationships between words…”).
Kotaru et al. and Bayless et al. are considered to be analogous to the claimed invention because they are in the same field of endeavor in data processing (i.e., conversation data). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Kotaru et al. to incorporate the teachings of Bayless et al. of wherein the large language model comprises a transformer-based large language model that is pre-trained on sentence data sets which provides the benefit of generating human-like textual responses to prompts ([0045] of Bayless et al.).
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Keisha Y Castillo-Torres whose telephone number is (571)272-3975. The examiner can normally be reached Monday - Friday, 9:00 am - 4:00 pm (EST).
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Pierre-Louis Desir can be reached at (571)272-7799. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
Keisha Y. Castillo-Torres
Examiner
Art Unit 2659
/Keisha Y. Castillo-Torres/Examiner, Art Unit 2659
/PIERRE LOUIS DESIR/Supervisory Patent Examiner, Art Unit 2659