DETAILED ACTION
This office action is in response to Applicant’s submission filed on 1/16/2025. Claims 1-29 are pending in the application of which Claims 1, 13, 20 and 27 are independent and have been examined.
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 .
Information Disclosure Statement
The information disclosure statement(s)(IDS) submitted on 1/16/2025, and 11/10/2025 have been considered by the examiner.
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- 3, 7-14, 16-21, and 23-26 are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter without significantly more. The claims as whole, considering all claim elements both individually and in combination, do not amount to significantly more than an abstract idea.
The group of independent claims 1, 13, and 20 disclose the same inventive concept with slightly different claim language. In this analysis, claim 1 is analyzed as a representative of the other independent claims. The independent claim 1 recites: “ … accessing a query; determining, from a corpus of content comprising a plurality of clauses from a plurality of passages, one or more passages that are potentially relevant to the query, wherein the one or more passages comprise a subset of three or more clauses; encoding the subset of three or more clauses to determine a clause-to-query relevance score for each clause, wherein a particular clause-to-query relevance score of at least one clause of the subset of three or more clauses is dependent on content of at least one other clause of the subset of three or more clauses; based at least in part on one or more clause-to-query relevance scores of one or more clauses of the subset of three or more clauses, removing the one or more clauses from the subset of three or more clauses to generate a revised subset of clauses; generating a prompt to a natural language processor, the prompt comprising: instructions based at least in part on the query, and at least part of the revised subset of clauses as context for generating a response by the natural language processor; prompting the natural language processor with the prompt; and accessing and storing a result generated based at least in part on the natural language processor executing the prompt.
The aforementioned limitations, under its broadest reasonable interpretation, cover performance of the limitation in the mind but for the recitation of generic computer components. That is, other than reciting “natural language processor”, and “processor”, nothing in the claim element precludes the step from practically being performed in the mind. For example, but for the “natural language processor”, and “processor” language, “a “accessing a query”, “determining”, from a corpus of content comprising a plurality of clauses from a plurality of passages, one or more passages that are potentially relevant to the query, wherein the one or more passages comprise a subset of three or more clauses; encoding the subset of three or more clauses to “determine” a clause-to-query relevance score for each clause, wherein a particular clause-to-query relevance score of at least one clause of the subset of three or more clauses is dependent on content of at least one other clause of the subset of three or more clauses; based at least in part on one or more clause-to-query relevance scores of one or more clauses of the subset of three or more clauses, “removing” the one or more clauses from the subset of three or more clauses to “generate” a revised subset of clauses; “generating” a prompt to a natural language processor, the prompt comprising: instructions based at least in part on the query, and at least part of the revised subset of clauses as context for “generating” a response by the natural language processor; prompting the natural language processor with the prompt; and “accessing” and storing a result generated based at least in part on the natural language processor executing the prompt.
In the context of this claim encompasses the human receives a query or a question and in order to respond to the question at hand, he will consult with resource material which relates to the query/question and pick a few passage that are closest to the query context. At the same, the requirement for the passage to be adequate is to make sure at least three sentences are relevant to the query. Next, he assesses the relevancy confidence/score of each sentence to the query (judgment call) a clause-to-query score. Next, the lowest score is dropped and the remainder constitute “subset of clauses”. Using a model (generic processor) is used to request to provide an answer based on the query and the subset of clauses. Human can also do the same task. Next the answer can be documented on a piece of paper.
Generating a prompt to a natural language processor, the prompt comprising: instructions based at least in part on the query, and at least part of the revised subset of clauses as context for generating a response by the natural language processor; prompting the natural language processor with the prompt. The following steps which deal with generating a prompt is a matter of design by the user/developer. The conversational natural language processor (i.e. conversational AI) is doing what the model is programmed to do which is to output a response based on a given context. Therefore, the natural language processor is nothing but generic tool being used to apply the abstract idea.
All of these steps can be performed in the mind and/or using a pen and paper. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claim recites an abstract idea.
This judicial exception is not integrated into a practical application. In particular, the claim only recites additional elements - using a “natural language processor”, and “processor” to perform all of the above-mentioned steps. The use of a “natural language processor”, and “processor” is recited at a high-level of generality (i.e., as a generic computer device performing a generic computer function) such that it amounts no more than mere instructions to apply the exception using a generic computer component. Accordingly, this additional element does not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea. The claim is directed to an abstract idea. The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception. The only element mentioned is the usage of a “conversational AI”, which due to lack of specificity can be considered as a generic processor. As discussed above with respect to integration of the abstract idea into a practical application, the additional element of a processor is merely for the purpose of data gathering and/or insignificant extra-solution activity that amount to no more than mere instructions to apply the exception using a generic computer component. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept. The claim is not patent eligible.
The dependent Claims do not add limitations that could help the Claim as a whole to amount to significantly more than the Abstract idea identified for the Independent Claim:
Similarly, dependent claims 2- 3, 6-12, 14, 16-19, 21, and 23-26 are also not patent eligible as they include additional steps that are directed towards an abstract idea as they can be practically performed in the mind without being integrated into a practical application, or including any additional elements sufficient to amount to significantly more than the judicial exception.
Claim Rejections - 35 USC § 103
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.
Claims 1, 3, 5-7, 13, 15 - 16, 20, and 22- 23 are rejected under 35 U.S.C. 103 as being unpatentable over Kanuga et al. (US 20250094717A1)(herein " Kanuga"), and in further view of Singal et al. (US 20190155913 A1)(herein " Singal ").
Regarding claims 1, 13 and 20 Kanuga teaches [A computer-implemented method comprising: - claim 1], [A system comprising: one or more processors; one or more non-transitory computer-readable media storing instructions, which, when executed by the system, cause the system to perform a set of actions comprising: - claim 13], and [One or more non-transitory computer-readable media storing instructions which, when executed, cause performance of a set of actions comprising: - claim 20] (Kanuga, Par. 0009:” … one or more non-transitory computer-readable media are provided for storing instructions which, when executed by one or more processors, cause a system to perform part or all of one or more methods disclosed herein.”)
accessing a query; (Kanuga, Par. 0064:” … the retriever 214A is configured to identify and retrieve chunks that are similar to the query.”, and Par. 0066:” … For example, the query “What is my 401k contribution limit?” is mapped to a “semantic search” knowledge task type. …”).
determining, from a corpus of content comprising a plurality of clauses from a plurality of passages, one or more passages that are potentially relevant to the query, wherein the one or more passages comprise a subset of three or more clauses; (Kanuga, Par. 0006:”… accessing the text portion [passage] comprises accessing a document and splitting the document into a plurality of text portions. In some aspects, accessing the text portion further comprises selecting and retrieving a subset of the plurality of text portions.”, and Par. 0004:”… accessing a text portion [passage]; identifying a plurality of sentences [plurality of clauses] in the text portion; embedding each of the plurality of sentences in the text portion to generate a respective plurality of text sentence embeddings; providing, to a language model, the text portion or a derivative thereof and a query; receiving, from the language model, a response to the query based on the text portion; identifying a plurality of sentences in the response; embedding the plurality of sentences in the response to generate a plurality of response embeddings; comparing each of the plurality of response embeddings to each of the plurality of sentence embeddings to generate a similarity score [relevancy] for each sentence embedding-response embedding pair; ..”, and Par. 0101:” … can configure how many sentences from the context documents should be mapped to each sentence in the response. For example, the user could configure to receive x=1, 2, 3, or more sentences as a reference. The selected sentences could be those that surround the highest-ranked sentence or the top x ranked sentences.”)
based at least in part on one or more clause-to-query relevance scores of one or more clauses of the subset of three or more clauses, removing the one or more clauses from the subset of three or more clauses to generate a revised subset of clauses; (Kanuga, Par. 0101:” … can configure how many sentences from the context documents should be mapped to each sentence in the response. For example, the user could configure to receive x=1, 2, 3, or more sentences as a reference. The selected sentences could be those that surround the highest-ranked sentence or the top x ranked sentences.”, and Par. 0090:” … the top n document chunks 438 are reranked by the ranker 440, and the top k most relevant sentence groups 442 are retrieved at step 502.”) Note: Top k most relevant sentence reads on generating a revised subset of clauses.
generating a prompt to a natural language processor, the prompt comprising: instructions based at least in part on the query, and (Kanuga, Par. 0067:” … Based on the original or rewritten query, the execution engine 214 generates a prompt. The prompt may include data from various sources (knowledge, API, dialog history, etc.) and relevant information from the context and memory store 206.”, and Par. 0068:” At step 9, the prompt is sent to the output pipeline 216. The main role of the output pipeline 216 is to synthesize responses for providing to the user 202. These responses may be in the format of a Conversation Message Model (CMM) and output by the response LLM 216B as rich multi-modal responses. For example, a prompt is generated based on the knowledge and conversational data and provided to the response LLM 216B. The response LLM is a language model such as ChatGPT (e.g., GPT-4), Falcon 40B Instruct, Cohere Command, or the like that is configured to answer a query based on corresponding knowledge information. The response LLM 216B generates output which may be provided directly to the user 202, …”).
at least part of the revised subset of clauses as context for generating a response by the natural language processor; (Kanuga, Par. 0084:” … the retrieved top n document chunks 438 are provided as input to the ranker 440. At step 441, the ranker 440 splits the top n document chunks 438 into sentence groups and reranks the sentence groups, returning the top k most relevant sentence groups 442. K is a configurable integer.”, and Par. 0085:”… the top k most relevant sentence groups 442 (or top n most similar document chunks 438 if steps 439-441 are omitted) are provided to the reader LLM 446. The contextualized user query is also provided to the reader LLM at step 444. At step 447, the reader LLM 446 outputs a final answer 448 to the user query 416.”)
prompting the natural language processor with the prompt; and (Kanuga, Par. 0068:” … a prompt is generated based on the knowledge and conversational data and provided to the response LLM 216B. The response LLM is a language model such as ChatGPT (e.g., GPT-4), Falcon 40B Instruct, Cohere Command, or the like that is configured to answer a query based on corresponding knowledge information.”)
accessing and storing a result generated based at least in part on the natural language processor executing the prompt. (Kanuga, Par. 0068:” … The responses are output to the user at step 10.”, and Par. 0082:” … the CQR model 424 is a model such as an LLM that is configured to use previous conversation history 421 to add contextual information to a query when needed, generating a contextualized user query 428. At step 426, the CQR model outputs the contextualized user query 428. At step 430, the contextualized user query 428 is embedded to generate a query embedding 432.”) Note: LLM that is configured to use previous conversation history reads on accessing and storing a result generated based on the execution of the prompt.
Kanuga does not teach, however, Singal teaches encoding the subset of three or more clauses to determine a clause-to-query relevance score for each clause, wherein a particular clause-to-query relevance score of at least one clause of the subset of three or more clauses is dependent on content of at least one other clause of the subset of three or more clauses; (Singal, Par. 0062:’ … features are extracted for every query-sentence pair in the document (410), which are then sent to a pre-trained model, in order to rank the sentences in terms of relevance to the query. The trained model then returns a ranking score for each pair (412). Using these ranking scores, the most relevant sentence(s) are returned to the user and displayed (414).”, and Par. 0073:”… for a particular sentence, a preceding sentence and/or a subsequent sentence may be used as context. More generally, any one or more adjacent sentence may be used, such as sentences within a same paragraph as a selected sentence. In these implementations, it is assumed that, for a particular query, a relevant sentence is more likely to be surrounded by other sentences that are also somewhat relevant with respect to the query.”) Note: ranking score for each pair of query sentence represents “clause-to-query relevance score”.
Singal is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga further in view of Singal to encode the subset of three or more clauses to determine a clause-to-query relevance score for each clause, wherein a particular clause-to-query relevance score of at least one clause of the subset of three or more clauses is dependent on content of at least one other clause of the subset of three or more clauses. Motivation to do so would provide sentences that are relevant to the query, even when the sentences do not include any of the exact terms of the query [Singal, Par. 0017].
Regarding claim 3, Kanuga, as modified above, teaches the method of claim 1.
Kanuga, as modified above, further teaches wherein determining, from the corpus of content, the one or more passages comprises determining a passage-to-query similarity between vector embeddings of the plurality of passages and a vector embedding of the query. (Kanuga, Par. 0081:” … At step 404, the documents are parsed and chunked to generate parsed and chunked documents 406. … At step 408, the chunked objects are embedded to generate chunk embeddings 410 [passage embedding].”, and Par. 0083:”… the chunk embeddings 410 and the query embedding 432, respectively, are provided to the retriever 435. The retriever 435 retrieves, at step 436, a subset of the chunk embeddings 410 that are most similar to the query embedding 432. As shown in FIG. 4, the top n most similar document chunks 438 are retrieved. The retriever 435 may, for example, compare the embeddings via cosine similarity to identify the top n most similar document chunks 438. N is a configurable integer.”)
Regarding claims of 5, 15, and 22, Kanuga, as modified above, teaches the method, the system, and media of claims 1, 13, and 20, respectively.
Kanuga, as modified above, does not teach, however, Singal further teaches training a model to identify which clauses of subsets of clauses are most relevant to respective queries to generate a trained clause selection model at least in part by: (Singal, Par. 0078:” … In unsupervised training, the network learns to identify a structure or pattern in the provided input.”, and Par. 0062:”… features are extracted for every query-sentence pair in the document (410), which are then sent to a pre-trained model, in order to rank the sentences in terms of relevance to the query. The trained model then returns a ranking score for each pair (412). Using these ranking scores, the most relevant sentence(s) are returned to the user and displayed (414).”)
prompting a large language model with a plurality of prompts that request selection of most relevant clauses; and (Singal, Par. 0067:” … Based on the original or rewritten query, the execution engine 214 generates a prompt. The prompt may include data from various sources (knowledge, API, dialog history, etc.) and relevant information from the context and memory store 206.”, and Par. 0068:” … a prompt is generated based on the knowledge and conversational data and provided to the response LLM 216B. The response LLM is a language model such as ChatGPT (e.g., GPT-4), Falcon 40B Instruct, Cohere Command, or the like that is configured to answer a query based on corresponding knowledge information.”, and Par. 0085:”… the top k most relevant sentence groups 442 (or top n most similar document chunks 438 if steps 439-441 are omitted) are provided to the reader LLM 446.")
assigning weights to features of clauses based at least in part on results from prompting with the plurality of prompts; (Singal, Par. 0016:” Each sentence or other specified grammatical unit may then be ranked using one or more machine learning techniques. For example, a model may be trained on the feature set used to obtain the types of features just referenced, so that the sentences may be ranked accordingly. Then, the highest-ranked sentence(s) may be returned as search results, …”, and Par. 0057:” … For example, when the specified grammatical unit is a sentence, the ranking module 128 may rank each query/sentence pair, using the feature values provided by the feature extraction module 126, e.g., using a trained model that is trained using the particular features for which values were obtained.”) Note: clause/sentence ranked according to the feature which reads on assigning weights to the features of clauses.
wherein encoding the subset of three or more clauses uses the trained clause selection model. (Singal, Par. 0036:” … to rank the sentences, or other grammatical units selected in conjunction with the received query, using a sentence ranking model.”, and Par. 0073:” … For example, for a particular sentence, a preceding sentence and/or a subsequent sentence may be used as context. More generally, any one or more adjacent sentence may be used, such as sentences within a same paragraph as a selected sentence. In these implementations, it is assumed that, for a particular query, a relevant sentence is more likely to be surrounded by other sentences that are also somewhat relevant with respect to the query.”)
Regarding claim 6, Kanuga, as modified above, teaches the method of claim 5.
Kanuga, as modified above, does not teach, however, Singal further teaches wherein the trained clause selection model is trained to select subsets of clauses that are most relevant to the respective queries relative to a total amount of information represented among each subset of the subsets of clauses for a particular passage. (Singal, Par. 0073:” … For example, for a particular sentence, a preceding sentence and/or a subsequent sentence may be used as context. More generally, any one or more adjacent sentence may be used, such as sentences within a same paragraph as a selected sentence. In these implementations, it is assumed that, for a particular query, a relevant sentence is more likely to be surrounded by other sentences that are also somewhat relevant with respect to the query.”, and Par. 0085:”… the top k most relevant sentence groups 442 (or top n most similar document chunks 438 if steps 439-441 are omitted) are provided to the reader LLM 446.") Note: Selecting based on the surrounding sentences in a paragraph reads on total amount of information, as the “one or more adjacent sentence” could encompass the entire paragraph.
Regarding claims of 7, 16, and 23, Kanuga, as modified above, teaches the method, the system, and media of claims 1, 13, and 20, respectively.
Kanuga, as modified above, further teaches wherein determining the one or more passages is based at least in part on one or more passage-to-query similarity scores determined using an initial passage selection model, the computer-implemented method further comprising: (Kanuga, Par. 0081:” … At step 404, the documents are parsed and chunked to generate parsed and chunked documents 406. … At step 408, the chunked objects are embedded to generate chunk embeddings 410 [passage embedding].”, and Par. 0083:”… the chunk embeddings 410 and the query embedding 432, respectively, are provided to the retriever 435. The retriever 435 retrieves, at step 436, a subset of the chunk embeddings 410 that are most similar to the query embedding 432. As shown in FIG. 4, the top n most similar document chunks 438 are retrieved. The retriever 435 may, for example, compare the embeddings via cosine similarity to identify the top n most similar document chunks 438. N is a configurable integer.”) Note: chunk (passage) embeddings that are most similar to the query embedding represents “passage-to-query similarity score”
adjusting the one or more passage-to-query similarity scores using another passage selection model; (Kanuga, Par. 0090:”… one or more selection mechanisms can be applied to identify and retrieve one or more text portions [passage] that are relevant to a query. For example, a retriever 435 can compare [similarity] chunk embeddings 410 corresponding to the parsed and chunked documents 406 to a query embedding 432 corresponding to the user query 416 (which may first be contextualized to produce the contextualized user query 430). The retriever then retrieves a subset of the parsed and chunked documents 406, such as the top n document chunks 438. Alternatively, or additionally, the top n document chunks 438 are reranked by the ranker 440, and the top k most relevant sentence groups 442 are retrieved at step 502.”) Note: reranking document chunks reads on “adjusting the one or more passage-to-query similarity score”.
wherein the clause selection model is trained to prune clauses separately from the other passage selection model that is trained to determine most relevant passages, and (Kanuga, Par 0084:” At step 439, the retrieved top n document chunks 438 are provided as input to the ranker 440. At step 441, the ranker 440 splits the top n document chunks 438 into sentence groups and reranks the sentence groups, returning the top k most relevant sentence groups 442. K is a configurable integer. In some embodiments, the ranker 440 uses a different ranking mechanism than the retriever 435, such as a classifier-based approach.”) Note: “reranking the top k most relevant sentence” reads on pruning clauses.
wherein the other passage selection model incorporates more features than the initial passage selection model. (Kanuga, Par. 0081:” … At step 404, the documents are parsed and chunked [passage] to generate parsed and chunked documents 406. … At step 408, the chunked objects are embedded to generate chunk embeddings 410.”, and Par. 0083:”… The retriever 435 retrieves, at step 436, a subset of the chunk embeddings 410 that are most similar to the query embedding 432. As shown in FIG. 4, the top n most similar document chunks 438 are retrieved. The retriever 435 may, for example, compare the embeddings via cosine similarity to identify the top n most similar document chunks 438. N is a configurable integer.”) Note: the “top n most similar document chunk” reads on passage with more features than initial passage.
Kanuga, as modified above, does not teach, however, Singal further teaches determining the one or more clause-to-query relevance scores for the one or more clauses using a clause selection model; (Singal, Par. 0062:’ … features are extracted for every query-sentence pair in the document (410), which are then sent to a pre-trained model, in order to rank the sentences in terms of relevance to the query. The trained model then returns a ranking score for each pair (412). Using these ranking scores, the most relevant sentence(s) are returned to the user and displayed (414).”, and Par. 0073:”… it is assumed that, for a particular query, a relevant sentence is more likely to be surrounded by other sentences that are also somewhat relevant with respect to the query.”) Note: the score represents the clause-to-query relevance score.
Claims 2, 14, and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Xia et al. (US11003720 B1)(herein "Xia").
Regarding claims of 2, 14, and 21, Kanuga, as modified above, teaches the method, the system, and media of claims 1, 13, and 20, respectively.
Kanuga, as modified above, does not teach, however, Xia teaches wherein encoding the subset of three or more clauses comprises, for each clause of the subset of three or more clauses: determining a token-to-query relevance score of each token in the clause based at least in part on a token-to-query similarity between a token embedding of the token and a query embedding of the query; (Xia, claim 4:” … obtaining word [token] embeddings of respective texts of messages in the second set of relevant messages and obtaining a word embedding of the search query, wherein each word embedding is a vector in a multi-dimensional space; using the word embeddings of the texts of messages in the second set of relevant messages and the word embedding of the search query to calculate a respective relevance score between the search query and each message in the second set of relevant messages, wherein the relevance scores generated by the third sub-stage is a measure of whether words [token] of the respective messages are used in a same way as words of the search query; and using the relevance scores as a boost factor for the scores generated by the second stage for messages in the second set of relevant messages.”, and Col. 6, ll. 20-25:” a given message can include a string of text. The string of text is converted into a vector in multi-dimensional space. This vector can then be input to the neural network as a feature. Using the text embeddings as a model feature allows the model to identify the similarity of two texts and to infer relevance between them.”)
determining the clause-to-query relevance score of the clause based at least in part on the token-to-query relevance score of each token in the clause; (Xia, claim 4:” … obtaining word [token] embeddings of respective texts of messages [clause] in the second set of relevant messages and obtaining a word embedding of the search query, wherein each word embedding is a vector in a multi-dimensional space; using the word embeddings of the texts of messages in the second set of relevant messages and the word embedding of the search query to calculate a respective relevance score between the search query and each message in the second set of relevant messages, wherein the relevance scores generated by the third sub-stage is a measure of whether words [token] of the respective messages are used in a same way as words of the search query; and using the relevance scores as a boost factor for the scores generated by the second stage for messages in the second set of relevant messages.”) Note: relevant score is calculated based on query and each message where token (word) of the respective message (clause).
wherein the token embedding for at least one token of the at least one clause is based at least in part on content of the at least one clause and the at least one other clause of the subset of three or more clauses. (Xia, Col. 6, ll. 40-52:” … uses word [token] embeddings of the query and the message text to calculate a relevance score. … In particular, the score generated by the third sub-stage 201 is a measure of whether the words [tokens] of the messages [clauses] are used in the same way as the words [tokens]of the search query. For example, the search term “cat” can mean the animal, but a message including “cat” could also refer to “central African time.” Identifying the context [other clauses] of the respective texts [clauses] can be used to boost messages having a matching context or that are using words in the same sense as the search query.”)
Xia is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Xia to determine a token-to-query relevance score of each token in the clause based at least in part on a token-to-query similarity between a token embedding of the token and a query embedding of the query; determining the clause-to-query relevance score of the clause based at least in part on the token-to-query relevance score of each token in the clause; wherein the token embedding for at least one token of the at least one clause is based at least in part on content of the at least one clause and the at least one other clause of the subset of three or more clauses. Motivation to do so would present search results in relevance order while improving search result accuracy (Xia, Col. 2, ll. 27-34).
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Agrawal et al. (US20220327287 A1)(herein " Agrawal").
Regarding claim 4, Kanuga, as modified above, teaches the method of claim 1.
Kanuga, as modified above, further teaches splitting a plurality of documents into a plurality of passages comprising consecutive sentences; and (Kanuga, Par. 0081:” … At step 404, the documents are parsed and chunked to generate parsed and chunked documents 406. ... At step 408, the chunked objects are embedded to generate chunk embeddings 410.”, and Par. 0091:” At step 510, a plurality of sentences are identified in the text portion. Each text portion may contain multiple sentences. The computing system may parse the text portion to split it into sentences.”)
Kanuga, as modified above, does not teach, however, Agrawal teaches training a model to identify which passages of the plurality of passages are most relevant to respective queries to generate a trained passage selection model; (Agrawal, Par. 0063:’ … the training system 420 (of FIG. 4) can train a model for use by the passage-selecting network 402 based on a set of training examples provided by a search engine's search log. Each such training example may include a query submitted by a user together with a passage in a document that has been identified by the search engine as being relevant to the query. In addition, or alternatively, at least some of the training examples may include manually-created pairings of queries and passages. The training system 420 trains the model to maximize the likelihood [most relevant] that the model will correctly identify the relevance of a query embedding (produced based on a query) and an identified passage.”)
wherein determining the one or more passages uses the trained passage selection model. (Agrawal, Par. 0094:” … where the base models are trained to perform one or more language-modeling tasks …”, and Par. 0128:” … The computing system includes a neural network that includes a first neural network (e.g., the passage-selecting network 402) trained using a first set of training examples (e.g., the first set 424), …”, and claim 14:” … the first neural network being configured to select a subset of textual passages for use in generating questions regarding the electronic document, …”).
Agrawal is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Agrawal to train a model to identify which passages of the plurality of passages are most relevant to respective queries to generate a trained passage selection model; wherein determining the one or more passages uses the trained passage selection model. Motivation to do so would allow search and retrieval systems to surface highly specific paragraphs independently, ensuring that users receive direct answers to micro-intents rather than broad documents.
Claims 8, 17, and 24 are rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Nallamothu et al. (US20250390677A1)(herein " Nallamothu"), and Fox et al. (US 20210390127 A)(herein “Fox”).
Regarding claims of 8, 17, and 24, Kanuga, as modified above, teaches the method, the system, and media of claims 1, 13, and 20, respectively.
Kanuga, as modified above, teaches clause-to-query relevance scores [claim 1 rejection], and teaches passage-to-query similarity scores [claim 7 rejection].
Kanuga, as modified above, does not teach, however, Nallamothu teaches wherein determining the one or more passages is based at least in part on one or more passage-to-query similarity scores, the computer-implemented method further comprising determining the one or more [[clause-to-query relevance scores]] for the one or more clauses and adjusting the one or more [[passage-to-query similarity scores]] using a combined machine learning model; however, Nallamothu in Par. 0052 teaches … utilizing the trained machine-learning model, the determined matches and the similarity score(s) to adjust one or more parameters of the similarity assessment to improve matching accuracy.
Nallamothu is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Nallamothu to determine one or more [[clause-to-query relevance scores]] for the one or more clauses and adjusting the one or more [[passage-to-query similarity scores]] using a combined machine learning model. Motivation to do so would allow performing advanced analysis on a document content, and improving the accuracy and efficiency of document processing (Nallamothu, Par. 0004 & 0019).
Kanuga, as modified above, does not teach, however, Fox teaches wherein removing the one or more clauses from the subset of three or more clauses to generate the revised subset of clauses is based at least in part on the one or more clause-to-query relevance scores. (Fox, Par. 0170:” … include loading document data 150, performing keyword extraction using a keyword extraction algorithm or method like RAKE, scoring the sentences based on keyword statistics, and re-ranking the sentences using MMR. The summary of a deposition should be concise and should not have redundant sentences included in it. The keyword recognition component 138 can use TF-IDF scores for each sentence. Then cosine similarity is used to calculate the similarity between two candidate sentences, and between the query and the candidate sentence; these help to remove the redundancy [clause]. If the similarity score between two candidate sentences is high, then the sentence which is more relevant to the query is ranked higher, and the other sentence is given a very low score.”, and Par. 0171:” … The algorithm iteratively compares each candidate sentence with the query, and adds the sentence to R based on the similarity score. It also computes the similarity between the sentence and other sentences in R; if two sentences have a high similarity score, the sentence most similar to the query is selected and the other sentence is discarded.”) Note: similarity between the candidate sentences and the query represents “clause-to-query relevance scores”.
Fox is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Fox to wherein removing the one or more clauses from the subset of three or more clauses to generate the revised subset of clauses is based at least in part on the one or more clause-to-query relevance scores. Motivation to do so would allow minimizing redundancy and leads to query-relevant summarization (Fox, Par. 0172).
Claims 9, 10, 18, and 25 are rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Fox.
Regarding claims of 9, 18, and 25, Kanuga, as modified above, teaches the method, the system, and media of claims 1, 13, and 20, respectively.
Kanuga, as modified above, further teaches wherein determining, from the corpus of content, the one or more passages results in an initial ranking of two or more passages, (Kanuga, Par. 0064:” … the retriever 214A is configured to rank the indexed chunks [passage] and select some number of highest-ranked chunks to be retrieved.”)
the computer-implemented method further comprising concurrently re-ranking the two or more passages and [[selecting the one or more clauses for removal]]; (Kanuga, Par. 0090:” … the top n document chunks 438 are reranked by the ranker 440, and the top k most relevant sentence groups 442 are retrieved at step 502.”)
wherein the natural language processor comprises a large language model. (Kanuga, Par. 0042:” … As part of generating these responses, digital assistant 106 may perform natural language generation (NLG) using one or more Large Language Models (LLMs).”)
Kanuga, as modified above, does not teach, however Fox teaches [[the computer-implemented method further comprising concurrently re-ranking the two or more passages and]] selecting the one or more clauses for removal; (Fox, Par. 0171:” … It also computes the similarity between the sentence and other sentences in R; if two sentences have a high similarity score, the sentence most similar to the query is selected and the other sentence is discarded.”)
Fox is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Fox to select the one or more clauses for removal. Motivation to do so would allow minimizing redundancy and leads to query-relevant summarization (Fox, Par. 0172).
Regarding claim 10, Kanuga, as modified above, teaches the method of claim 9.
Kanuga, as modified above, does not teach, however, Fox teaches selecting the one or more clauses for removal based on which subsets of clauses are most relevant to the query relative to a total amount of information represented among each subset of the subsets of clauses for a particular passage. (Fox, Par. 0171:” … It also computes the similarity between the sentence and other sentences in R; if two sentences have a high similarity score, the sentence most similar to the query [relevant] is selected and the other sentence is discarded.”)
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, Singal, and Xia, in further view of Debnath et al. (US 20250061118 A1)(herein " Debnath").
Regarding claim 11, Kanuga, as modified above, teaches the method of claim 2.
Kanuga, as modified above, does not teach, however, Debnath teaches causing display of an option to adjust a threshold for selecting clauses for removal, (Debnath, Par. 0030:” After retrieving the context document 106, the state is determined with context and query embedding. The state is used as input to the Q* table to find the recommended action. Based on the action, the threshold for context reduction is computed and top-k sentences are selected. Then the reduced context version is produced by discarding sentences outside limits established by the top-k sentences and reducing any sentences other than the top-k sentences that remain.”, and Par. 0041:”… the computing device 600 may also include one or more peripheral devices 660. … For example, in some embodiments, the peripheral devices 660 may include a display, touch screen, graphics circuitry, …, interface devices, and/or peripheral devices.”) Note: top-k reads on threshold, and selecting “k” adjust a threshold for sentences/clauses removal.
receiving a selection of the option to adjust the threshold, resulting in an adjusted threshold; (Debnath, Par. 0004:” … identifying a context document relating to a query. A number of sentences of the context document to preserve is determined. The sentences of the context document are ranked according to respective similarities between the sentences and the query. A reduced context is generated that preserves the determined number of highest ranked sentences of the context document and eliminates other sentences from the context document. The query is executed with a language model, including the reduced context in a prompt, to generate a response.”) Note: “reduced context is generated” is the outcome of adjusted threshold.
wherein the one or more clauses are selected for removal based at least in part on the adjusted threshold. (Debnath, Par. 0004:” … identifying a context document relating to a query. A number of sentences of the context document to preserve is determined. The sentences of the context document are ranked according to respective similarities between the sentences and the query. A reduced context is generated that preserves the determined number of highest ranked sentences of the context document and eliminates other sentences from the context document. The query is executed with a language model, including the reduced context in a prompt, to generate a response.”, and Par. 0030:” … the reduced context version is produced by discarding sentences outside limits established by the top-k sentences and reducing any sentences other than the top-k sentences that remain.”)
Debnath is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Debnath to cause display of an option to adjust a threshold for selecting clauses for removal, receiving a selection of the option to adjust the threshold, resulting in an adjusted threshold; wherein the one or more clauses are selected for removal based at least in part on the adjusted threshold. Motivation to do so would dynamically control over the balance between model accuracy and computational efficiency.
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Boston et al. (US 20160180438 A1)(herein "Boston"), and Fox.
Regarding claim 12, Kanuga, as modified above, teaches the method of claim 1.
Kanuga, as modified above, does not teach, however, Boston teaches wherein a first clause of the subset of three or more clauses is a title of a first passage of the plurality of passages, a second clause of the subset of three or more clauses is in a body of the first passage of the plurality of passages, and a third clause of the subset of three or more clauses is also in the body of the first passage of the plurality of passages; (Boston, Par. 0031:” … a passage may be said to entail a hypothesis if the truth of the passage makes the truth of the hypothesis likely. Thus, in some embodiments, as described further below, a passage determined to entail a hypothesis derived from the user's question may be identified as providing supporting evidence for an answer to the question, …”, and Par. 0076:” … In some embodiments in which unstructured documents are indexed in units such as individual natural language sentences, as discussed above, individual indexed units (e.g., document sentences) may be scored for relevance to the search query, and high-scoring units may be retrieved individually for further analysis. In some such embodiments, as discussed above, annotation of index units (e.g., body text sentences) with text from corresponding section headings and document titles may increase the relevance score of a passage by simulating proximity to terms that appear in the heading and/or title to which that passage belongs. …. In some embodiments, passages may be identified and delineated by combining adjacent sentences which are indexed individually, and the relevance score of the passage as a whole may be computed by combining the relevance scores of the sentences it spans.”) Note: sentence that contain heading and/or title reads on first clause, and the corresponding passage reads on “first passage”. Similarly, the body text sentences read on second and third clause.
Boston is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Boston to wherein a first clause of the subset of three or more clauses is a title of a first passage of the plurality of passages, a second clause of the subset of three or more clauses is in a body of the first passage of the plurality of passages, and a third clause of the subset of three or more clauses is also in the body of the first passage of the plurality of passages. Motivation to do so would provide for the extracted answer information as contributing to the best and accuracy of answer to the user's question (Boston, Par. 0077 & 81).
Kanuga, as modified above, does not teach, however, Fox teaches wherein the third clause is removed from the subset of three or more clauses to generate the revised subset of clauses. (Fox, Par. 0171:” … It also computes the similarity between the sentence and other sentences in R; if two sentences have a high similarity score, the sentence most similar to the query is selected and the other sentence is discarded.”)
Fox is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Fox to wherein the third clause is removed from the subset of three or more clauses to generate the revised subset of clauses. Motivation to do so would allow minimizing redundancy and leads to query-relevant summarization (Fox, Par. 0172).
Claims 19, and 26 are rejected under 35 U.S.C. 103 as being unpatentable over Kanuga, and Singal, and in further view of Debnath.
Regarding claims of 19, and 26, Kanuga, as modified above, teaches the system, and media of claims 13, and 20, respectively.
Kanuga, as modified above, does not teach, however, Debnath teaches causing display of an option to adjust a threshold for selecting clauses for removal, (Debnath, Par. 0030:” After retrieving the context document 106, the state is determined with context and query embedding. The state is used as input to the Q* table to find the recommended action. Based on the action, the threshold for context reduction is computed and top-k sentences are selected. Then the reduced context version is produced by discarding sentences outside limits established by the top-k sentences and reducing any sentences other than the top-k sentences that remain.”, and Par. 0041:”… the computing device 600 may also include one or more peripheral devices 660. … For example, in some embodiments, the peripheral devices 660 may include a display, touch screen, graphics circuitry, …, interface devices, and/or peripheral devices.”) Note: top-k reads on threshold, and selecting “k” adjust a threshold for sentences/clauses removal.
receiving a selection of the option to adjust the threshold, resulting in an adjusted threshold; (Debnath, Par. 0004:” … identifying a context document relating to a query. A number of sentences of the context document to preserve is determined. The sentences of the context document are ranked according to respective similarities between the sentences and the query. A reduced context is generated that preserves the determined number of highest ranked sentences of the context document and eliminates other sentences from the context document. The query is executed with a language model, including the reduced context in a prompt, to generate a response.”) Note: “reduced context is generated” is the outcome of adjusted threshold.
wherein the one or more clauses are selected for removal based at least in part on the adjusted threshold. (Debnath, Par. 0004:” … identifying a context document relating to a query. A number of sentences of the context document to preserve is determined. The sentences of the context document are ranked according to respective similarities between the sentences and the query. A reduced context is generated that preserves the determined number of highest ranked sentences of the context document and eliminates other sentences from the context document. The query is executed with a language model, including the reduced context in a prompt, to generate a response.”, and Par. 0030:” … the reduced context version is produced by discarding sentences outside limits established by the top-k sentences and reducing any sentences other than the top-k sentences that remain.”)
Debnath is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Kanuga, as modified above, further in view of Debnath to cause display of an option to adjust a threshold for selecting clauses for removal, receiving a selection of the option to adjust the threshold, resulting in an adjusted threshold; wherein the one or more clauses are selected for removal based at least in part on the adjusted threshold. Motivation to do so would dynamically control over the balance between model accuracy and computational efficiency.
Claim 27 is rejected under 35 U.S.C. 103 as being unpatentable over Boston (US 20160180438 A1), and in further view of Singal (US 20190155913 A1), and Gofman et al. (US 20260017496 A1)(herein “Gofman”).
Regarding claim 27, Boston teaches A computer-implemented method for optimizing clauses retrieved from a data store to determine an answer to a query received from a user, the computer-implemented method comprising: processing the query and the first set of clauses to produce a subset of clauses with a context pruning subsystem that concurrently: prunes the first set of clauses to produce a subset of the first set of clauses, and (Boston, Par. 0062:” … document/passage analyzer 130 may divide natural language texts into units, with each unit having a separate entry and annotations in the document's index. ... In some embodiments, natural language texts may be divided into units of sentences [clauses], and each sentence may have its own entry in a document index. In some embodiments, as discussed further below, this may allow text passages to be identified in support of answers to users' questions by combining adjacent sentences individually indexed and determined to provide relevant evidence for the question's answer. … may evaluate the relevance of indexed natural language text to a user's question by searching and/or scoring individual text units [clause] corresponding to index entries (e.g., individual sentences) along with their associated annotations in the index.”, and Par. 0084:”… perform entailment analysis on retrieved passages, and may reject [prune] any passages that do not entail a question hypothesis. For example, a passage about a different film that states, “You won't find any tall blue aliens in this movie,” may have been retrieved as being highly relevant to the search query because it contains many of the search keywords, but it would not entail the question constraint hypothesis, and it would not provide strong support for the other film as being the answer to the user's question. … evidence scorer 180 may score passages based at least in part on the degree to which they entail one or more question hypotheses, and may rank passages based on their scores. In some embodiments, evidence scorer 180 may disregard [prune] passages that score below a suitably defined threshold, or may retain only the N-best-scoring passages for any suitably defined value of N, or may prune low-scoring passages according to any other suitable criteria.”) Note: “retain only the N-best-scoring” reads on “produce a subset of the first set of clauses”.
generates a re-rank score for at least one passage comprising the subset of the first set of clauses by relevance to answering the query; (Boston, Par. 0084:” … evidence scorer 180 may score passages based at least in part on the degree to which they entail one or more question hypotheses, and may rank passages based on their scores. In some embodiments, evidence scorer 180 may disregard [prune] passages that score below a suitably defined threshold, or may retain only the N-best-scoring passages for any suitably defined value of N, or may prune low-scoring passages according to any other suitable criteria.”, and Par. 0109:”… one or more supporting natural language text passages for one or more answer items (e.g., the top-scoring passage for an answer item, a ranked list of a suitable number of top-scoring passages for the same answer item, etc.) may also be provided to answer builder 190.”) Note: A “ranked list of top-scoring passages” reads on re-ranked score passage.
rankings of the remaining subsets of the candidate clauses. (Boston, Par. 0084:” … evidence scorer 180 may score passages based at least in part on the degree to which they entail one or more question hypotheses, and may rank passages based on their scores. In some embodiments, evidence scorer 180 may disregard [prune] passages that score below a suitably defined threshold, or may retain only the N-best-scoring passages for any suitably defined value of N, or may prune low-scoring passages according to any other suitable criteria.”, and Par. 0109:”… one or more supporting natural language text passages for one or more answer items (e.g., the top-scoring passage for an answer item, a ranked list of a suitable number of top-scoring passages for the same answer item, etc.) may also be provided to answer builder 190.”) Note: disregard passages read on “remaining subsets". Furthermore, the candidate clauses, is mapped above.
Boston, does not teach, however, Singal teaches retrieving from the data store a first set of clauses that are ranked by relevance to answering the query; (Singal, Par. 0002:”Computers, and associated data stores, are capable of providing access to large numbers of potentially lengthy documents and other content.”, and Par. 0006:” … receiving a query with respect to at least one document, identifying sentences [first set of clauses] within the at least one document, and extracting a plurality of features characterizing a relevance of each sentence of the sentences to the query. The sentences may be ranked based on the features, and the sentences may be visually designated within the document, in an order corresponding to the ranking.”)
Singal is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Boston further in view of Singal to retrieve from the data store a first set of clauses that are ranked by relevance to answering the query. Motivation to do so would provide sentences that are relevant to the query, even when the sentences do not include any of the exact terms of the query [Singal, Par. 0017].
Boston, as modified above, does not teach, however, Gofman teaches prompting a natural language processor to provide the answer to the query, (Gofman, Par. 0069:”… for each generated synthetic query, use the LLM 208 to generate a response to the synthetic query from the related document, … configured to instruct the LLM 208 to generate a response to a synthetic query from its related document by providing the LLM 208 with an extraction prompt that comprises the query, the related document and instructions to generate a concise response to the query from the related document.”, and Par. 0070:”… You are given a query and a supporting document, please extract an answer from the document.“)
wherein the prompting uses the query and the subset of clauses as included context for providing the answer to the query; (Gofman, Par. 0069:”… for each generated synthetic query, use the LLM 208 to generate a response to the synthetic query from the related document [context], … configured to instruct the LLM 208 to generate a response to a synthetic query from its related document [context] by providing the LLM 208 with an extraction prompt that comprises the query, the related document [context] and instructions to generate a concise response to the query from the related document.“)
wherein [[the context pruning subsystem]] is trained for pruning at least some clauses using a cross-encoder that encodes queries and candidate clauses of a training set and (Gofman, Par. 0096:” … the training dataset may be used, by a training module 240, to train or fine-tune a reranker model 242. … The term “reranker model” is used to mean a specialized machine learning model designed to rank documents/passages [candidate clauses] based on their relevance to a query, such as, but not limited to, a cross-encoder model. Some reranker models may calculate a relevance score for a query-document pair, and the relevance scores can be used to rank a set of documents.”) Note: the context pruning subsystem was mapped above.
generates output comprising: [[remaining subsets]] of the [[candidate clauses]] for the [[queries of the training set]], wherein the remaining subsets exclude the at least some clauses that were pruned; and (Boston, Par. 0002:” the output would typically be a list of one or more web pages whose texts contain the keywords “capital” and “Liechtenstein.”, and Par. 0084:” … evidence scorer 180 may score passages based at least in part on the degree to which they entail one or more question hypotheses, and may rank passages based on their scores. In some embodiments, evidence scorer 180 may disregard [prune] passages that score below a suitably defined threshold, or may retain only the N-best-scoring passages for any suitably defined value of N, or may prune low-scoring passages according to any other suitable criteria.”, and Par. 0109:”… one or more supporting natural language text passages for one or more answer items (e.g., the top-scoring passage for an answer item, a ranked list of a suitable number of top-scoring passages for the same answer item, etc.) may also be provided to answer builder 190.”) Note: disregard passages or may retain only the N-best-scoring passages read on “remaining subsets exclude the at least some clauses that were pruned”, or equivalently "remaining subsets". Furthermore, the double bracketed limitations were mapped above.
Gofman is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Boston further in view of Gofman to prompt a natural language processor to provide the answer to the query, wherein the prompting uses the query and the subset of clauses as included context for providing the answer to the query; wherein [[the context pruning subsystem]] is trained for pruning at least some clauses using a cross-encoder that encodes queries and candidate clauses of a training set and generates output comprising. Motivation to do so would enhance system scalability to support larger sets of documents [Gofman, Par. 0100].
Claims 28, and 29 are rejected under 35 U.S.C. 103 as being unpatentable over Boston, Singal, and Gofman, and in further view of Thorsley et al. (US20220374766A1 )(herein " Thorsley").
Regarding claim 28, Boston, as modified above, teaches the method of claim 27.
Boston, as modified above, does not teach, however, Singal further teaches receiving the query from a user interface; and (Singal, Par. 0013:” … provide a user interface within the application to enable users to submit a query, …”)
outputting the answer to the user interface; (Singal, Par. 0013:” … The user interface displays search results by, e.g., displaying each grammatical unit within the search results, either separately from, and/or highlighted or identified within, the searched document(s).”, and Par. 0014:” … the highest-ranked sentence(s) may be returned as search results, using the user interface specified above.”)
Boston, as modified above, does not teach, however, Gofman further teaches wherein the rankings of the remaining subsets of the candidate clauses are generated using re-rank scores; and (Gofman, Par. 0003:” … compute a score for each of the retrieved documents that indicates the relevance of the document to the query. The scores can then be used to reorder the documents retrieved in the first phase by relevance to the query. “, and Par. 0081:” … A score is then assigned to each document based on the outcome of the pairwise rankings. The scores assigned to the documents are then used to rank the documents. “, and Par. 0096:” … The term “reranker model” is used to mean a specialized machine learning model designed to rank documents/passages based on their relevance to a query, such as, but not limited to, a cross-encoder model. Some reranker models may calculate a relevance score for a query-document pair, and the relevance scores can be used to rank a set of documents.” Note: the remaining subsets, the candidate clauses were mapped before.
wherein the context pruning subsystem is stored in a memory and the natural language processor is a large language model. (Gofman teaches large language model (see Par. 0005, 0022, 0024, 0035, etc.). Also teaches memory (see Par. 0005, 0022, 0103, 0104, etc.)
Boston, as modified above, does not teach, however, Thorsley teaches wherein the at lease some clauses are pruned using a binary mask; (Thorsley, Par. 0045:” In threshold token pruning, a threshold-based token pruning approach prunes tokens if the token has a token importance score less than a threshold denoted by θ(l)∈ [Image Omitted] . Specifically, a pruning strategy may be defined by imposing a binary mask M(l)(·): {1, . . . , n}→{0,1} that indicates whether a token should be kept or pruned:”, and Par. 0046:” In other words, a token may be pruned if the token has a token importance score greater than a predetermined threshold, and evaluating the token for pruning may be done with a simple comparison operator without a top-k calculation. Furthermore, in some embodiments, once a token is pruned, a pruned token may be excluded from calculations in all succeeding layers, thereby gradually reducing the computation complexity in bottom layers (i.e., cascade token pruning). In other embodiments, tokens may be pruned if they have an importance score equal to or less than the predetermined threshold.”)
PNG
media_image1.png
60
282
media_image1.png
Greyscale
Thorsley is considered to be analogous to the claimed invention because it is in the same field of endeavor. 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 Boston further in view of Thorsley to wherein the at lease some clauses are pruned using a binary mask. Motivation to do so would allow for pruning to performed to remove tokens that are irrelevant to a task prior to performing the corresponding calculation (Thorsley, Par. 0025).
Regarding claim 29, Boston, as modified above, teaches the method of claim 28.
Boston, as modified above, does not teach, however, Singal further teaches executable logic for receiving input via an interactive input device, and wherein the interactive input device is a display, a keyboard, a mouse, a microphone, a speaker, or any combination thereof. (Singal, Par. 0039:” … may be entered by the user by way of an associated keyboard, or using any other suitable text entry technique.”, and Par. 0085:”… may be performed by one or more programmable processors executing a computer program to perform functions by operating on input data and generating output.”, and Par. 0087:” To provide for interaction with a user, implementations may be implemented on a computer having a display device, … displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer.”)
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure. Boytsov et al. (US20220253447A1) teaches in ABS:” The controller may be configured to receive a query and document, tokenize the query into a sequence of query tokens and tokenize the document into a sequence of document tokens, generate a matrix of token pairs for each of the query and the document tokens, retrieve for each entry in the matrix of token pairs, a precomputed similarity score produced by a neural conditional translation probability network, wherein the neural network has been trained in a ranking task using a corpus of paired queries and respective relevant documents, produce a ranking score for each document with respect to each query via a product-of-sum aggregation of each of the similarity scores for the respective query; and output the document and associated ranking score of the document.”
Examiner's Note: Examiner has cited particular columns and line numbers and/or paragraph numbers in the references applied to the claims above for the convenience of the applicant. Although the specified citations are representative of the teachings of the art and are applied to specific limitations within the individual claim, other passages and figures may apply as well. It is respectfully requested from the applicant in preparing responses, to fully consider the references in entirety as potentially teaching all or part of the claimed invention, as well as the context of the passage as taught by the prior art or disclosed by the Examiner.
In the case of amending the Claimed invention, Applicant is respectfully requested to indicate the portion(s) of the specification which dictate(s) the structure relied on for proper interpretation and also to verify and ascertain the metes and bounds of the claimed invention.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DARIOUSH AGAHI whose telephone number is (408)918-7689. The examiner can normally be reached Monday - Thursday and alternate Fridays, 7:30-4:30 PT.
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, Bhavesh Mehta can be reached on 571-272-7453. 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.
DARIOUSH AGAHI, P.E.
Primary Examiner
/DARIOUSH AGAHI/Primary Examiner, Art Unit 2656
Prosecution Insights