Notice of Pre-AIA or AIA Status
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 102
2. In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
3. Claims 1 and 12-13 are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Qin (US 2024/0346256).
Regarding Claim 1:
Qin discloses a method comprising :
a) receiving an artificial intelligence request, wherein the artificial intelligence request comprises a payload (Qin: ¶[0003], ¶[0049] discloses receiving a user query as the input to the retrieval augmented response generation system. The received query corresponds to the claimed artificial intelligence request, and the content of the query is the claimed payload);
b) retrieving, from a context database, a set of exemplars corresponding to the artificial intelligence request (Qin: ¶[0036], ¶[0019] and ¶[0052] discloses datasets 112 that are expressly databases containing augmentation information. The system compares an encoding of the received query to encoding corresponding to stored augmentation information, determines a subset that matches a condition and then retrieves pieces of augmentation information corresponding to that subset. Those retrieved pieces function as the claimed set of exemplars corresponding to the AI request because they are selected based on the query and retrieved from a stored database to be insert into the prompt for the model);
c) generating a first artificial intelligence prompt, wherein the first artificial intelligence prompt comprises:
i) the set of exemplars corresponding to the artificial intelligence request (Qin: ¶[0046] ¶[0022] discloses generating an augmented prompt that includes the retrieved augmentation information as content. Because the retrieved augmentation information is the set selected/retrieved for the query, including it in the augmented prompt satisfies this portion); and
ii) the payload from the artificial intelligence request (Qin: ¶[0022], ¶[0046] discloses that the augmented prompt includes the original query as the question portion of the prompt); and
d) obtaining a first artificial intelligence output by performing acts comprising providing the first artificial intelligence prompt to a first trained machine learning model (Qin: ¶[0003] and ¶[0047]-[0054] discloses providing the augmented prompt to LLM 214 and receiving the resulting response).
Regarding Claim 12:
Qin further discloses a non-transitory computer readable medium having stored thereon instructions for performing the method of claim 1 (Qin: ¶[0069] discloses a computer readable medium having instructions to perform the method disclosed in claim 1).
Regarding Claim 13:
Qin further discloses a system comprising a computer comprising a non-transitory computer readable medium storing instructions operable to, when executed, configure the computer to perform the method of claim 1 (Qin: ¶[0069] discloses a system with memory having instructions to perform the method disclosed in claim 1).
Claim Rejections - 35 USC § 103
4. In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
5. Claim 2 is rejected under 35 U.S.C. 103 as being unpatentable over Qin (US 2024/0346256) in view of Freitag (US 2021/0019373).
Regarding Claim 2:
Qin further discloses the method of claim 1, wherein the method comprises:
a) generating a second artificial intelligence prompt, wherein the second artificial intelligence prompt comprises the first artificial intelligence output and the payload from the artificial intelligence request (Qin: ¶[0046] discloses a prompt generator creates an augmented prompt and then provides it to an LLM. This functions as a new/second prompt generated by the system. ¶[0019] discloses the prompt may include the original query, contextual information for answering the query, the retrieved augmentation information and request to answer the original query. The “payload from the artificial intelligence request” reads naturally as the request data bundled with the query, context and retrieved supporting content);
Qin does not disclose:
b) obtaining a second artificial intelligence output by performing acts comprising providing the second artificial intelligence prompt to a second trained model; and
c) determining a final artificial intelligence output based on the first artificial intelligence output and the second artificial intelligence output.
However, discloses these limitations:
b) obtaining a second artificial intelligence output by performing acts comprising providing the second artificial intelligence prompt to a second trained model (Freitag: ¶[0025], ¶[0047]-[0048] the second output is edited text 110, which is still the translation corresponding to the original payload, but corrected); and
c) determining a final artificial intelligence output based on the first artificial intelligence output and the second artificial intelligence output (Freitag: ¶[0024]-[0025] and ¶[0049] discloses edited text output which functions as the final output).
Qin and Freitag are both combinable because they both disclose content pertinent to one another. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Freitag teaches that applying an additional trained model to evaluate and score supporting information and then merge/rank results improves answer quality and confidence. Qin provides the payload and prompt construction while Beamon provides the multi-model/multi-response framework. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose using multiple models to generate alternative responses and ensemble style ranking selection to improve quality and factuality. The motivation for doing so is “ a large number of training instances can be automatically generated from a variety of diverse resources, which can improve the accuracy and/or robustness of edited translated text generated using an APE model trained on such training instances.” as discloses in ¶[0010] of Freitag.
6. Claims 9-10 are rejected under 35 U.S.C. 103 as being unpatentable over Qin in view of Freitag and further in view of Huang (US 2017/0169015).
Regarding Claim 9:
The combination of Qin and Freitag further discloses the method of claim 2, wherein:
a) the second artificial intelligence output comprises:
i) text in the second language corresponding to the payload from the artificial intelligence request (Qin: ¶[0046] discloses a prompt generator creates an augmented prompt and then provides it to an LLM. This functions as a augmented/second language corresponding to the payload/input prompt);
The combination of Qin and Freitag does not disclose the crossed out limitation above. However, Huang discloses:
ii) a first confidence, wherein the first confidence is confidence in the text in the second language comprised by the first artificial intelligence output as accurately translating the payload from the artificial intelligence request from the first language to the second language (Huang: ¶[0013] discloses a machine translation engine, and that confidence scores can be used to select which translation to use and determine whether alternate translations need to be generated); and
iii) a second confidence, wherein the second confidence is confidence in the text in the second language comprised by the second artificial intelligence output as accurately translating the payload from the artificial intelligence request from the first language to the second language (Huang: ¶[0030] discloses scoring another translation (i.e., another candidate output), also discloses scoring different translations of the same source content item and selecting final translation portions. Second AI output is held analogous to “another translation” and it likewise has a confidence score from the confidence models); and
b) determining the final artificial intelligence output based on the first artificial intelligence output and the second artificial intelligence output comprises determining text for the final artificial intelligence output selected from:
i) the text in the second language comprised by the first artificial intelligence output (Huang: ¶[0030] discloses translation sorter 352 receives scores from one or more translations and selects the translation that has the highest confidence score as the best translation); and
ii) the text in the second language comprised by the second artificial intelligence output based on the first confidence and the second confidence (Huang: ¶[0030] discloses translation sorter 352 receives scores from one or more translations and selects the translation that has the highest confidence score as the best translation).
It would have been obvious to one of ordinary skill in the art to modify the combination of Qin and Freitag, which discloses a multi stage artificial intelligence output system, to further compute confidence values for each candidate output and select the final output based on those confidence values, because Huang expressly teaches generating and using translation confidence scores to improve translation output selection and downstream handling. Huang discloses the confidence scoring model to score translation and produce a confidence score for translation and even selecting among alternatives by choosing a combination of the higher scoring phrases as the best translation which overall demonstrates that “confidence scoring system can improve machine translations by providing confidence scores that can help make determinations” as disclosed in ¶[0013] of Huang.
Regarding Claim 10:
The combination of Qin and Freitag further discloses the method of claim 2, wherein the final artificial intelligence output comprises:
a) text in the second language corresponding to the payload from the artificial intelligence request (Qin: Qin: ¶[0046] discloses a prompt is generated from the original input and a augmented prompt/second text (second language) is produced); and
The combination of Qin and Freitag does not disclose b) a confidence in the text comprised by the final artificial intelligence output as accurately translating the payload from the artificial intelligence request from the first language to the second language.
However, Huang discloses:
b) a confidence in the text comprised by the final artificial intelligence output as accurately translating the payload from the artificial intelligence request from the first language to the second language (Huang: ¶[0010]-[0011]discloses that the final output includes a confidence in the translated texts accuracy).
It would have been obvious to one of ordinary skill in the art to modify the combination of Qin and Freitag, which discloses a multi stage artificial intelligence output system, to further compute confidence values for each candidate output and select the final output based on those confidence values, because Huang expressly teaches generating and using translation confidence scores to improve translation output selection and downstream handling. Huang discloses the confidence scoring model to score translation and produce a confidence score for translation and even selecting among alternatives by choosing a combination of the higher scoring phrases as the best translation which overall demonstrates that “confidence scoring system can improve machine translations by providing confidence scores that can help make determinations” as disclosed in ¶[0013] of Huang.
7. Claims 3-6 are rejected under 35 U.S.C. 103 as being unpatentable over Qin in view of Kuhn (US 2011/0093254).
Regarding Claim 3:
Qin further discloses the method of claim 1, wherein:
a) the payload from the artificial intelligence request comprises a document to be translated from a first language into a second language (Qin: ¶[0039] discloses LLMs being used for augmentation from first language to second (augmented) language);
c) the first artificial intelligence output comprises text in the second language corresponding to the payload from the artificial intelligence request (Qin: ¶[0001], ¶[0047], the system is used for the input payload of the text/document to be modified).
Qin does not explicitly disclose:
b) each exemplar from the set of exemplars corresponding to the artificial intelligence request comprises:
i) text in the first language; and
ii) corresponding text in the second language.
However, Kuhn does disclose:
b) each exemplar from the set of exemplars corresponding to the artificial intelligence request comprises:
i) text in the first language (Kuhn: ¶[0050] discloses a database (stored exemplars) as bilingual sentence pairs where each pair includes a source language sentence and its corresponding target language translation); and
ii) corresponding text in the second language (Kuhn: ¶[0050] discloses a database (stored exemplars) as bilingual sentence pairs where each pair includes a source language sentence and its corresponding target language translation).
Qin and Kuhn are both combinable because they both disclose content pertinent to one another. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Kuhn teaches maintaining and using a database of bilingual exemplars. The difference between the claimed subject matter is that Qin does not expressly require the exemplars to be bilingual pairs, whereas Kuhn expressly teaches exemplars in a bilingual paired form. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose using a bilingual paired dataset. The motivation for doing so is “Many human translators already use a translation memory (TM) to increase their productivity” as discloses in ¶[0002] of Kuhn.
Regarding Claim 4:
The proposed combination of Qin in view of Kuhn further discloses the method of claim 3, wherein the context database comprises a superset of exemplars, wherein each exemplar from the superset of exemplars comprises text in the first language and corresponding text in the second language (Kuhn: ¶[0050] discloses a database (stored exemplars) as bilingual sentence pairs where each pair includes a source language sentence and its corresponding target language translation).
Qin and Kuhn are both combinable because they both disclose content pertinent to one another. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Kuhn teaches maintaining and using a database of bilingual exemplars. The difference between the claimed subject matter is that Qin does not expressly require the exemplars to be bilingual pairs, whereas Kuhn expressly teaches exemplars in a bilingual paired form. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose using a bilingual paired dataset. The motivation for doing so is “Many human translators already use a translation memory (TM) to increase their productivity” as discloses in ¶[0002] of Kuhn.
Regarding Claim 5:
The proposed combination of Qin in view of Kuhn further discloses the method of claim 4, wherein the method comprises:
a) generating a first set of embeddings, wherein the first set of embeddings comprises embeddings for words included in the payload of the artificial intelligence request (Qin: ¶[0042] discloses the encoder tokenizes the query text string and generate embeddings for each token);
b) selecting the set of exemplars corresponding to the artificial intelligence request from the superset of exemplars based on, for each exemplar from the set of exemplars corresponding to the artificial intelligence request, a distance between:
i) the first set of embeddings (Qin: ¶[0041]-[0045] selects a subset from a larger collection by computing a distance (cosine similarity) between the embeddings derived from embeddings from the input and embeddings derived from each stored items text); and
ii) a set of embeddings which comprises embeddings for words included in the text in the first language from that exemplar embeddings (Qin: ¶[0041]-[0045] selects a subset from a larger collection by computing a distance (cosine similarity) between the embeddings derived from embeddings from the input and embeddings derived from each stored items text, this similarity functions as the distance used to select the corresponding exemplars from the superset).
Regarding Claim 6:
The proposed combination of Qin in view of Kuhn further discloses the method of claim 4 wherein:
a) the artificial intelligence request comprises the payload and a value for a target parameter (Qin: ¶[0039] discloses an AI system receiving a request that includes the user’s query that include the users query plus an associated parameter value carried as contextual information. This contextual value is a target parameter value received with the input payload);
b) each exemplar from the superset of exemplars comprised by the context database has a corresponding value for the target parameter (Qin: ¶[0040] discloses that each stored augmentation information may include metadata such as a product identifier); and
c) the method comprises selecting the set of exemplars corresponding to the artificial intelligence request from the superset of exemplars based on, for each exemplar from the set of exemplars corresponding to the artificial intelligence request, identifying the value for the target parameter corresponding to that exemplar as matching the value for the target parameter comprised by the artificial intelligence request (Qin: ¶[0040]-[0043] discloses the requests the proper augmentation information previously stored. Each exemplar’s target parameter value is included in the exemplar text string and encoded into the first feature vector. Each exemplars target (e.g., identifier/metadata) is included in the exemplar text string and encoded into the second feature vector. The system then selects/retrieves exemplars by comparing the vectors and retrieving those most similar which is exactly the examples whose encoded parameter value matches because the parameter value is embedded on both sides).
8. Claims 7-8 are rejected under 35 U.S.C. 103 as being unpatentable over Qin in view of Kuhn, and further in view of Dodelson (US 2014/0193796).
Regarding Claim 7:
The proposed combination of Qin in view of Kuhn further discloses the method of claim 6, except wherein:
a) the target parameter is reading level;
b) the first artificial intelligence output has a reading level matching the value of the target parameter comprised by the artificial intelligence request; and
c) the payload comprised by the artificial intelligence request has a reading level which does not match the value of the target parameter comprised by the artificial intelligence request.
However, Dodelson discloses:
a) the target parameter is reading level (Dodelson: ¶[0009]-[0012] discloses system assesses user skill level including reading level, using LEXILE and assigns a LEXILE reading score);
b) the first artificial intelligence output has a reading level matching the value of the target parameter comprised by the artificial intelligence request (Dodelson: ¶[0052]-[0054] discloses the system produces multiple versions of unmodified content, where each aligned version is aligned to a specific skill level); and
The proposed combination of Qin and Kuhn do not disclose
c) the payload comprised by the artificial intelligence request has a reading level which does not match the value of the target parameter comprised by the artificial intelligence request.
However, Dodelson discloses:
c) the payload comprised by the artificial intelligence request has a reading level which does not match the value of the target parameter comprised by the artificial intelligence request (Dodelson: ¶[0052]-[0054] discloses the system first obtains unmodified content, it may determine the unmodified content is appropriate for a certain reading level, then generate an aligned version for a lower reading comprehension. The payload = the unmodified content, which does not match the target reading level, before the system rewrites it into the matching aligned version).
Qin and Kuhn in view of Dodelson are both combinable because each disclose content pertinent to one another. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Kuhn teaches maintaining and using a database of bilingual exemplars. Dodelson discloses a readability based text adaption system. The difference between the claimed subject matter is that Qin does not expressly disclose a readability comparison. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose this readability feature. The motivation for doing so is “The 100 system matches a version of the aligned content to a user by matching specific areas of learning where the user exhibits a need for improvement, as assessed by the system 100 in step 208.” as discloses in ¶[0054] of Dodelson.
Regarding Claim 8:
The proposed combination of Qin, Kuhn and Dodelson further discloses the method of claim 7, except wherein the first language and the second language are the same language (Dodelson: ¶[0052]-[0054] discloses the transformation is occurring within the same language, while changing reading level).
Qin and Kuhn in view of Dodelson are both combinable because each disclose content pertinent to one another. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Kuhn teaches maintaining and using a database of bilingual exemplars. Dodelson discloses a readability based text adaption system. The difference between the claimed subject matter is that Qin does not expressly disclose a readability comparison. It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose this readability feature. The motivation for doing so is “The 100 system matches a version of the aligned content to a user by matching specific areas of learning where the user exhibits a need for improvement, as assessed by the system 100 in step 208.” as discloses in ¶[0054] of Dodelson.
9. Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Qin in view of Marcu (US 2014/0149102).
Regarding Claim 11:
Qin further discloses the method of claim 1, except wherein the method comprises:
a) receiving an approved translation, wherein the approved translation comprises text in the second language corresponding to the payload from the artificial intelligence request; and
b) updating the context database by adding a new exemplar comprising:
i) the payload from the artificial intelligence request; and
ii) the approved translation.
However, Marcu discloses:
a) receiving an approved translation, wherein the approved translation comprises text in the second language corresponding to the payload from the artificial intelligence request (Marcu: ¶[0029] and ¶[0032] discloses feedback includes corrections from humans providing accuracy ratings and a correction unit that captures corrected translation content. ¶[0025]-[0031] discloses a translation/correction that is target-language text corresponding to the source text (payload)); and
b) updating the context database by adding a new exemplar comprising:
i) the payload from the artificial intelligence request (Marcu: ¶[0031]-[0033] feedback is stored for processing including the source unit); and
ii) the approved translation (Marcu: ¶[0031]-[0033] feedback is stored for processing including the target/corrected translation).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to disclose manual approving of translation and storing these results into Qin. Qin teaches retrieving segmented response generation where a system retrieves augmentation information and generates an augmented prompt including the request payload and provides that prompt to an LLM to obtain an output. Marcu discloses translator (human) feedback, which Qin does not explicitly disclose. The suggestion/motivation for doing so is translator provided feedback improves generated translation as disclosed by Marcu in ¶[0003].
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to IAN SCOTT MCLEAN whose telephone number is (703)756-4599. The examiner can normally be reached "Monday - Friday 8:00-5:00 EST, off Every 2nd Friday".
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, Hai Phan can be reached at (571) 272-6338. 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.
/IAN SCOTT MCLEAN/Examiner, Art Unit 2654
/HAI PHAN/Supervisory Patent Examiner, Art Unit 2654