DETAILED ACTION
Claims 1 – 20 are pending.
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 .
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the Applicant regards as his invention.
Claims 1 – 8 and 15 – 20 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the Applicant), regards as the invention.
With regard to independent claim 1 (fifth limitation) and claim 15 (sixth limitation), the claims recite ‘finetuning the language model and the graph neural network … by maximizing the loss function.’ The limitation here provides ‘maximizing’ the loss function,’ which is contrary to what is commonly understood in the art for a technique like that being performed here. As understood and well-known in the art, a ‘loss function’ or ‘cost function’ gets minimized rather than maximized, so that the error encountered in the technique can be greatly reduced. This is an optimization technique to improve the system. Maximizing the loss, would be anti-thetical to the aim of optimising the system. The Specification does not define this ‘maximizing’ in any other way than what would be linguistically understood as a maximization. The Specification in [0016] provides ‘apply the loss function for optimizing the language model and the graph neural network’ which would be in line with ‘minimizing’ the loss function, rather than ‘maximizing’ it for the purpose of optimizing the language model and the graph neural network.
For the purpose of this office Action, unless it can be properly explained by the Applicant that the ‘maximizing’ is indeed what is intended, the Examiner will interpret this as ‘minimizing’ the loss function.
Claims 1 and 15 are hereby rejected under 35 U.S.C. 112(b).
Dependent claims 2 – 8 and 16 – 20 are also hereby rejected under 35 U.S.C. 112(b) for failing to rectify the issue raised regarding their respective base claims.
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.
Claims 1, 2, 4, 8, 15, 16 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Yebes Torres et al. (US 2024/0096125 Al: hereafter — Yebes Torres) in view of Mendelson (US 2025/0378100 A1) further in view of CHENG et al. (US 2022/0301173 A1: hereafter — Cheng).
For claim 1, Yebes Torres discloses a computer-implemented method, executed on a computing device, comprising:
transforming a plurality of content portions into a plurality of embeddings [[using a language model]] (Yebes Torres: [0077] — extracting features from text segments (the text segments being the claimed plurality of content portions) to generate embeddings);
generating a graph with nodes representing respective embeddings and an edge between a pair of nodes representing a similarity distance between the respective embeddings [[that is less than or equal to a predefined threshold]] (Yebes Torres: [0096] — a graph generator based on each text segment (text segments that are represented by embeddings), the nodes being the text segment [embeddings] and the edges representing interactions between text segments (a pair of nodes) that is based on a proximity of neighbouring text segments to a given text segment [embedding]);
generating a category prediction for each content portion by processing the graph using a graph neural network (Yebes Torres: [0030] — a graph neural network (GNN); [0105] — segment tagger circuitry structured to classify text segments; FIG. 8 Steps 806 → 812 — generating a graph to represent the content portions, leading to a classification of the text segments).
The reference of Yebes Torres provides teaching for transforming content portions into embeddings and generating a graph based on the embeddings so as to make a category prediction for the content portions. This however differs from the claimed invention in that the claimed invention fails to provide teaching for obtaining the plurality of embeddings for the plurality of content portions using a language model, as well as finetuning the language model and the graph neural network for automatically tagging content portions.
The reference of Mendelson is however now introduced to teach this as:
transforming a plurality of content portions into a plurality of embeddings using a language model (Mendelson: [0012], [0022] — using a large language model on a particular [edge] text to generate an embedding vector representation);
determining a loss function using a plurality of predefined categories and the category predicted for each content portion (Mendelson: [0021] — outputting one or more classifications made by the GNN model regarding the node of the graph (the nodes being the embeddings, thereby teaching making classifications/categories for each content portion); [0031] — the presence of a loss functions for the embeddings based on plurality of known groupings);
finetuning the language model and the graph neural network for automatically tagging content portions with a category by maximizing the loss function (Mendelson: [0023] — fine-tuning the GNN and the LLM for the purpose of making predictions; FIG. 4, [0006] — improving predictions made by the GNN; [0031] — minimising a loss function for the purpose of making one or more classifications, thereby optimising the technique).
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to modify the teaching of Yebes Torres which provides transforming content portions into embeddings and generating a graph based on the embeddings, by incorporating the known teaching of Mendelson which provides transforming the content portions into embedding through the use of a large language model for the purpose of fine-tuning the language model and the graph neural network, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result of presenting a particular method for obtaining vector embeddings, being the use of a language model which is routine and well-known in the art, while ensuring the LM and the GNN being used are well-trained to be able to address newly-encountered content portions so that an overall classification of content may be determined. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415-421, 82 USPQ2d 1385, 1395-97 (2007).
The combination of Yebes Torres in view of Mendelson provides teaching for generating a graph with nodes for embeddings and edges between pairs of nodes representing a similarity distance between embeddings. This however differs from the claimed invention in that the claimed invention further provides teaching for generating the graph with edges such that the similarity distance is less than or equal to a predefined threshold.
This is however not new to the art as the reference of Cheng is now introduced to teach this as:
generating a graph with nodes representing respective embeddings and an edge between a pair of nodes representing a similarity distance between the respective embeddings that is less than or equal to a predefined threshold (Cheng: [0011], [0018], [0048] — generating a graph whereby an edge exists between nodes with based on a Euclidean distance for similarity being below a threshold).
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to incorporate the known teaching of Cheng which creates an edge between nodes based on a similarity distance being below a threshold, to improve upon the teaching of the combination of Yebes Torres in view of Mendelson which provides the generation of a graph with node representing respective embeddings, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result of ensuring connections between nodes that have been determined to have the closest relationships, thereby building a graph that shows proper relatedness rather than a graph where the relationships between embeddings are not close enough. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415–421, 82 USPQ2d 1385, 1395-97 (2007).
For claim 2, claim 1 is incorporated and the combination of Yebes Torres in view of Mendelson further in view of Cheng provides teaching for the computer-implemented method, further comprising:
processing a target content portion for automatically tagging with a new category by generating an embedding using the finetuned language model (Mendelson: [0031] — the LLM here is trained to output embeddings of input text (input text here taken as the target content) and the text is descriptive of one or more items that could be associated with a node in the graph); [0032] — a graph may be built out of the computing action (input) data 206 (indicating the presence of content portions within the input data); [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0022] — generating text embeddings of input text based on the large language model);
adding a new node to the graph representing the embedding of the target content portion (Mendelson: [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0036] — the GNN being used to generate one or more classifications of the respective entities represented in the graph); and
generating a plurality of category predictions for the target content portion using the graph neural network and the graph (Mendelson: [0036] — the GNN being used to generate one or more classifications of the respective entities represented in the graph; [0031] — making groupings for the embeddings).
For claim 4, claim 1 is incorporated and the combination of Yebes Torres in view of Mendelson further in view of Cheng discloses the computer-implemented method, wherein determining the loss function includes determining a cross-entropy score between the plurality of predefined categories and the category predicted for each content portion (Yebes Torres: [0119] — for a segment tagger, using a cross-entropy loss function, to present several possible outcomes for the segment).
For claim 8, claim 1 is incorporated and the combination of Yebes Torres in view of Mendelson further in view of Cheng discloses the computer-implemented method, wherein generating the category prediction for each content portion includes generating an adjacency matrix representative of the graph (Cheng: [0010], [0011] — generating a graph while making use of an adjacency matrix).
For claim 15, Yebes Torres discloses a computer program product residing on a non-transitory computer readable medium having a plurality of instructions stored thereon which (Yebes Torres: [0210] — non-transitory machine-readable storage medium), when executed by a processor (Yebes Torres: [0210] — execution by a processor), cause the processor to perform operations comprising:
transforming a plurality of content portions into a plurality of embeddings [[using a language model]] (Yebes Torres: [0077] — extracting features from text segments (the text segments being the claimed plurality of content portions) to generate embeddings);
generating a graph with nodes representing respective embeddings and an edge between a pair of nodes representing a similarity distance between the respective embeddings [[that is less than or equal to a predefined threshold]] (Yebes Torres: [0096] — a graph generator based on each text segment (text segments that are represented by embeddings), the nodes being the text segment [embeddings] and the edges representing interactions between text segments (a pair of nodes) that is based on a proximity of neighbouring text segments to a given text segment [embedding]);
generating a category prediction for each content portion by processing the adjacency matrix using a graph neural network (Yebes Torres: [0030] — a graph neural network (GNN); [0105] — segment tagger circuitry structured to classify text segments; FIG. 8 Steps 806 → 812 — generating a graph to represent the content portions, leading to a classification of the text segments).
The reference of Yebes Torres provides teaching for transforming content portions into embeddings and generating a graph based on the embeddings so as to make a category prediction for the content portions. This however differs from the claimed invention in that the claimed invention fails to provide teaching for obtaining the plurality of embeddings for the plurality of content portions using a language model, as well as finetuning the language model and the graph neural network for automatically tagging content portions.
The reference of Mendelson is however now introduced to teach this as:
transforming a plurality of content portions into a plurality of embeddings using a language model (Mendelson: [0012] — using a large language model on particular [edge] text to generate an embedding vector representation);
determining a loss function using a plurality of predefined categories and the category predicted for each content portion (Mendelson: [0021] — outputting one or more classifications made by the GNN model regarding the node of the graph (the nodes being the embeddings, thereby teaching making classifications/categories for each content portion); [0031] — the presence of a loss functions for the embeddings based on plurality of known groupings);
finetuning the language model and the graph neural network by maximizing the loss function (Mendelson: [0023] — fine-tuning the GNN and the LLM for the purpose of making predictions; FIG. 4, [0006] — improving predictions made by the GNN; [0031] — minimising a loss function for the purpose of making one or more classifications, thereby optimising the technique); and
processing a target content portion for automatically tagging with a new category by generating an embedding using the finetuned language model (Mendelson: [0031] — the LLM here is trained to output embeddings of input text (input text here taken as the target content) and the text is descriptive of one or more items that could be associated with a node in the graph); [0032] — a graph may be built out of the computing action (input) data 206 (indicating the presence of content portions within the input data); [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0022] — generating text embeddings of input text based on the large language model).
The same motivation for combination for incorporating the Mendelson reference as applied to claim 1 is applicable here still.
The combination of Yebes Torres in view of Mendelson provides teaching for generating a graph with nodes for embeddings and edges between pairs of nodes representing a similarity distance between embeddings. This however differs from the claimed invention in that the claimed invention further provides teaching for generating the graph with edges such that the similarity distance is less than or equal to a predefined threshold.
This is however not new to the art as the reference of Cheng is now introduced to teach this as:
generating a graph with nodes representing respective embeddings and an edge between a pair of nodes representing a similarity distance between the respective embeddings that is less than or equal to a predefined threshold (Cheng: [0011], [0018], [0048] — generating a graph whereby an edge exists between nodes with based on a Euclidean distance for similarity being below a threshold);
generating an adjacency matrix representative of the graph (Cheng: [0010], [0011] — generating a graph while making use of an adjacency matrix).
The same motivation for combination for incorporating the Cheng reference as applied to claim 1 is applicable here still.
For claim 16, claim 15 is incorporated and the combination of Yebes Torres in view of Mendelson further in view of Cheng discloses the computer program product, wherein determining the loss function includes determining a cross-entropy score between the plurality of predefined categories and the category predicted for each content portion (Yebes Torres: [0119] — for a segment tagger, using a cross-entropy loss function, to present several possible outcomes for the segment).
For claim 20, claim 15 is incorporated and the combination of Yebes Torres in view of Mendelson further in view of Cheng discloses the computer program product, wherein the operations further comprise:
adding a new node to the graph representing the embedding of the target content portion (Mendelson: [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0036] — the GNN being used to generate one or more classifications of the respective entities represented in the graph); and
generating a plurality of category predictions for the target content portion using the graph neural network and the graph (Mendelson: [0036] — the GNN being used to generate one or more classifications of the respective entities represented in the graph; [0031] — making groupings for the embeddings).
Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Yebes Torres (US 2024/0096125 Al) in view of Mendelson (US 2025/0378100 A1) further in view of Cheng (US 2022/0301173 A1) as applied to claim 2, and further in view of Dong et al. (US 11,947,912 B1: hereafter — Dong).
For claim 3, claim 2 is incorporated but the combination of Yebes Torres in view of Mendelson further in view of Chang fails to disclose the limitations of this claim, for which the reference of Dong is now introduced to teach as the computer-implemented method, further comprising:
processing a query against the plurality of content portions by processing tokens of the query against a plurality of category predictions generated for the plurality of content portions (Dong: FIG. 7 Step 710 — first text data from a natural language input (the input is the query); Step 730 — second data which is the grouping of text; Col 2 line 66 – Col 3 line 21 — generating spans which are the several content portions, suing tokens of the natural language input whereby the spans are given their particular categories; Col 3 lines 57–61 — obtaining predicted named entities; Col 12 lines 49–53 — entity embedding for predicted category representation); and
providing a query result from the plurality of content portions using the plurality of category predictions generated for the plurality of content portions (Dong: FIG. 7 Step 760, Col 14 lines 4–15 — generating the output data (the query result) based on all the predicted NER tags for the groupings of the text).
The combination of Yebes Torres in view of Mendelson further in view of Cheng provides teaching for generating a category prediction for each content portion, but differs from the claimed invention in that the claimed invention further provides processing a query against the plurality of content portions so as to provide a query result from the plurality of content portions. This isn’t new to the art as the reference of Dong is seen to teach above.
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to combine the known teaching of Dong which tokenises an input query against a plurality of category predictions to be able to obtain query results from the plurality of content portions, with the teaching of the combination of Yebes Torres in view of Mendelson further in view of Cheng which teaches generating a category prediction for each content portion, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result of generating a response to a query whereby the response makes use of terms used in the query, making the response directly relatable to the query. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415–421, 82 USPQ2d 1385, 1395-97 (2007).
Claims 5 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Yebes Torres (US 2024/0096125 Al) in view of Mendelson (US 2025/0378100 A1) further in view of Cheng (US 2022/0301173 A1) as applied to claims 4 and 16, and further in view of Chican et al. (US 2024/0070971 A1: hereafter — Chican).
For claim 5, claim 1 is incorporated and the combination of Yebes Torres in view of Mendelson provides teaching for the computer-implemented method, wherein determining the loss function includes [[determining a soft silhouette score]] using the plurality of embeddings and the category predicted for each content portion (Mendelson: [0021] — outputting one or more classifications made by the GNN model regarding the node of the graph (the nodes being the embeddings, thereby teaching making classifications/categories for each content portion); [0031] — the presence of a loss functions for the embeddings base on plurality of known groupings).
This combination fails to teach of the loss function being a soft-silhouette score. This is however not new to the art as the reference of Chican is introduced to teach of a soft-silhouette loss, as the computer-implemented method, wherein determining the loss function includes determining a soft silhouette score using the plurality of embeddings and the category predicted for each content portion (Chican: [0024] — a loss function; [0104], [0222] — a soft silhouette loss).
The combination of Yebes Torres in view of Mendelson further in view of Cheng provides teaching for determining a loss function for the plurality of embeddings, but differs from the claimed invention in that the claimed invention further provides that the loss function includes a soft-silhouette score. This is however not new to the art as the reference of Chican is seen to teach above.
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to improve upon the loss function calculation as provided by the teaching of the combination of Yebes Torres in view of Mendelson further in view of Cheng, by incorporating the known teaching of Chican which provides that the loss function includes a soft-silhouette loss, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result that a soft silhouette guides learned representations towards forming compact and well-separated clusters, suitable for deciding on how well an embedding would fit into a classification. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415–421, 82 USPQ2d 1385, 1395-97 (2007).
For claim 17, claim 16 is incorporated and the combination of Yebes Torres in view of Mendelson provides teaching for the computer program product, wherein determining the loss function includes [[determining a soft silhouette score]] using the plurality of embeddings and the category predicted for each content portion (Mendelson: [0021] — outputting one or more classifications made by the GNN model regarding the node of the graph (the nodes being the embeddings, thereby teaching making classifications/categories for each content portion); [0031] — the presence of a loss functions for the embeddings base on plurality of known groupings).
This combination fails to teach of the loss function being a soft-silhouette score. This is however not new to the art as, just as applied to claim 5 above, the reference of Chican is introduced to teach of a soft-silhouette loss, as:
the computer program product, wherein determining the loss function includes determining a soft silhouette score using the plurality of embeddings and the category predicted for each content portion (Chican: [0024] — a loss function; [0104], [0222] — a soft silhouette loss).
The same motivation for combination for incorporating the Chican reference as applied to claim 5 is applicable here still.
Claims 6 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Yebes Torres (US 2024/0096125 Al) in view of Mendelson (US 2025/0378100 A1) further in view of Cheng (US 2022/0301173 A1) as applied to claims 1 and 15, and further in view of GHOSHAL et al. (US 2022/0012268 A1: hereafter — Ghoshal).
For claim 6, claim 1 is incorporated but the combination of Yebes Torres in view of Mendelson further in view of Cheng fails to disclose the limitations of this claim, for which the reference of Ghoshal is now introduced to teach this as the computer-implemented method, wherein the plurality of predefined categories include a plurality of user-defined categories for the plurality of content portions (Ghoshal: [0006], [0007] — when new documents come into existence and new taxonomies are needed, there can be user-defined categories for placing new content into accurate categories).
The combination of Yebes Torres in view of Mendelson further in view of Cheng provides teaching for classifying content portions into categories, but differs from the claimed invention in that the claimed invention further provides teaching for the categories being user-defined categories. This isn’t new to the art as the reference of Ghoshal is seen to teach above.
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to improve upon the teaching of the combination of Yebes Torres in view of Mendelson further in view of Cheng which provides teaching for classifying content portions into categories, by incorporating the known teaching of the reference of Ghoshal which provides user-defined categories, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result of ensuring that there are available categories to classifying content portions which currently do not have categories. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415–421, 82 USPQ2d 1385, 1395-97 (2007).
For claim 18, claim 15 is incorporated and as applied to claim 6 above, the combination of Yebes Torres in view of Mendelson further in view of Cheng and further in view of Ghoshal discloses the computer program product, wherein the plurality of predefined categories include a plurality of user-defined categories for the plurality of content portions (Ghoshal: [0006], [0007] — when new documents come into existence and new taxonomies are needed, there can be user-defined categories for placing new content into accurate categories).
The same motivation for combination for incorporating the Ghoshal reference as applied to claim 6 is applicable here still.
Claims 7 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Yebes Torres (US 2024/0096125 Al) in view of Mendelson (US 2025/0378100 A1) further in view of Cheng (US 2022/0301173 A1) as applied to claims 1 and 15, and further in view of Terry et al. (US 20190129933 A1: hereafter — Terry).
For claim 7, claim 1 is incorporated but the combination of Yebes Torres in view of Mendelson further in view of Cheng fails to disclose the limitations of this claim, for which the reference of Terry is now introduced to teach this as the computer-implemented method, wherein the plurality of predefined categories include a plurality of predefined categories generated by a generative artificial intelligence (AI) model for the plurality of content portions (Terry: [0033], [0063] — document classification; [0088] — AI-generated document classifications).
The combination of Yebes Torres in view of Mendelson further in view of Cheng provides teaching for classifying content portions into categories, but differs from the claimed invention in that the claimed invention further provides teaching for the categories being generated by a generative AI model. This isn’t new to the art as the reference of Terry is seen to teach above.
Hence, before the effective filing date of the claimed invention, one of ordinary skill in the art would have found it obvious to improve upon the teaching of the combination of Yebes Torres in view of Mendelson further in view of Cheng which provides teaching for classifying content portions into categories, by incorporating the known teaching of the reference of Terry which provides an AI model which generates categories user-defined categories, to thereby come up with the claimed invention. The combination of both prior art elements would have provided the predictable result of ensuring that there are available categories to classifying content portions which currently do not have categories, without needing the presence of a human for such category generation, making it an automated process. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 415–421, 82 USPQ2d 1385, 1395-97 (2007).
For claim 19, claim 15 is incorporated and as applied to claim 7 above, the combination of Yebes Torres in view of Mendelson further in view of Cheng and further in view of Terry discloses the computer program, wherein the plurality of predefined categories include a plurality of predefined categories generated by a generative artificial intelligence (AI) model for the plurality of content portions (Terry: [0033], [0063] — document classification; [0088] — AI-generated document classifications).
The same motivation for combination for incorporating the Terry reference as applied to claim 7 is applicable here still.
Claims 9, 10, 11, 12, 13 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Mendelson (US 2025/0378100 A1) in view of Cheng (US 2022/0301173 A1).
For claim 9, Mendelson discloses a computing system comprising:
a memory (Mendelson: [0075] — computer-readable storge medium); and
a processor (Mendelson: [0075] — execution by a processor) configured to: process a target content portion for automatically tagging with a category by generating an embedding using a finetuned language model (Mendelson: [0031] — the LLM here is trained to output embeddings of input text (input text here taken as the target content) and the text is descriptive of one or more items that could be associated with a node in the graph); [0032] — a graph may be built out of the computing action (input) data 206 (indicating the presence of content portions within the input data); [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0022] — generating text embeddings of input text based on the large language model), to add a new node to a graph with nodes representing respective embeddings (Mendelson: [0054] — newly received information as new entities may be added to available data sets and input into an LLM and an updated graph may be obtained, such that a small graph may be constructed around the new data point represented by the new entity; [0036] — the GNN being used to generate one or more classifications of the respective entities represented in the graph) and an edge between a pair of nodes representing a similarity distance between the respective embeddings (Mendelson: [0031] — relationships between node entities may be represented by an edge) [[that is less than or equal to a predefined threshold]], and to generate a plurality of category predictions for the target content portion using the graph neural network and the graph (Mendelson: [0052] — making multiple predictions/classifications for nodes, making predictions for each of the entities (indicating the making of predictions for the target content) by making use of the GNN model (the presence of the nodes indicates the use of the graph)).
The reference of Mendelson provides teaching for generating a graph having nodes and edges, but fails to teach that the edges are based on a similarity distance between respective embeddings meeting a threshold. This however isn’t new to the art as the reference of Cheng is now introduced to teach this as:
an edge between a pair of nodes representing a similarity distance between the respective embeddings that is less than or equal to a predefined threshold (Cheng: [0011], [0018], [0048] — generating a graph whereby an edge exists between nodes with based on a Euclidean distance for similarity being below a threshold).
The same motivation for combination which introduced the Cheng reference as applied to claim 1 is applicable here still.
For claim 10, claim 9 is incorporated and the combination of Mendelson in view of Cheng discloses the computing system, wherein the processor is further configured to: training the finetuned language model (Mendelson: [0023] — fine-tuning the GNN and the LLM for the purpose of making predictions; FIG. 4, [0006] — improving predictions made by the GNN; [0031] — minimising a loss function for the purpose of making one or more classifications, thereby optimising the technique).
For claim 11, claim 10 is incorporated and the combination of Mendelson in view of Cheng discloses the computing system, wherein training the finetuned language model includes transforming a plurality of content portions into a plurality of embeddings using a language model (Mendelson: [0012], [0022] — using a large language model on a particular [edge] text to generate an embedding vector representation).
For claim 12, claim 11 is incorporated and the combination of Mendelson in view of Cheng discloses the computing system, wherein training the finetuned language model includes generating an adjacency matrix representative of the graph (Cheng: [0010], [0011] — generating a graph while making use of an adjacency matrix).
For claim 13, claim 11 is incorporated and the combination of Mendelson in view of Cheng discloses the computing system, wherein training the finetuned language model includes generating a category prediction for each content portion by processing the adjacency matrix using a graph neural network (Cheng: [0010] — an adjacency matrix is used in graphing a graph convolutional neural being used for the purpose of predicting an instance label for each node (using this with Mendelson [0023] which provides that the GNN is used to fine-tune the LLM, this in turn being used to further fine-tune the GNN)).
For claim 14, claim 13 is incorporated and the combination of Mendelson in view of Cheng discloses the system, wherein training the finetuned language model includes determining a loss function using a plurality of predefined categories and the category predicted for each content portion (Mendelson: [0021] — outputting one or more classifications made by the GNN model regarding the node of the graph (the nodes being the embeddings, thereby teaching making classifications/categories for each content portion); [0031] — the presence of a loss functions for the embeddings based on plurality of known groupings).
Conclusion
The prior art made of record and not relied upon is considered pertinent to Applicant’s disclosure.
Renders et al. (US 2021/0349954 A1) provides teaching for the edges of a graph being based on a k-nearest neighbour graph based on similarity between the embeddings ([0070], [0129], [0172]).
Geng et al. (US 2023/0153531 A1) provides teaching for a set of embeddings being generated for each token based on one or more transformer-encoder language models (e.g., a pre-trained ALBERT model) [0025].
Any inquiry concerning this communication or earlier communications from the Examiner should be directed to OLUWADAMILOLA M. OGUNBIYI whose telephone number is (571)272-4708. The Examiner can normally be reached Monday – Thursday (8:00 AM – 5:30 PM Eastern Standard Time).
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, PARAS D. SHAH can be reached at (571) 270-1650. 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.
/OLUWADAMILOLA M OGUNBIYI/Examiner, Art Unit 2653
/Paras D Shah/Supervisory Patent Examiner, Art Unit 2653
02/13/2026