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 .
DETAILED ACTION
This Office Action is sent in response to Application’s Communication received on 10/10/2023 for application number 18/378491. The Office hereby acknowledges receipt of the following and placed of record in file: Specification, Drawing, Abstract, Oath/Declaration, and Claims.
Claims (1-18), 19 and 20 are presented for examination.
Examiner Notes
Examiner acknowledges that the cited Whang Korean Publication does not qualify as Prior Art against the present Application. Therefore, the examiner is issuing Non-Final rejection based on new rejection.
Information Disclosure Statement
The information disclosure statements (IDS) submitted on 10/10/2023 was filed prior to current Office Action. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
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 therefore, subject to the conditions and requirements of this title.
Claim 19 is rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (i.e., a law of nature, a natural phenomenon, or an abstract idea) without significantly more.
Step 1: Claim 19 is drawn to a method each of which is within the four statutory categories (e.g., a process, a machine).
Step 2A - Prong One: In prong one of step 2A, the claims are analyzed to evaluate whether they recite a judicial exception.
Claim 19.
A method for embedding a knowledge graph, the method being performed by at least one processor and comprising:
obtaining a knowledge instance including a mask element, the mask element being a numeric entity;
generating, through an embedder, an instance embedding of the knowledge instance;
and
predicting a value of the mask element by inputting, to a predictor comprising a regression layer, the instance embedding.
Applicant’s limitation of “predicting a value of the mask element” by supplying an instance embedding to a predictor “comprising a regression layer,” reasonably interpreted in light of the specification, recites estimating an unknown numeric value from a numerical representation through regression. This is a mathematical calculation. Paragraph [0080] describes embeddings as vector representations, and paragraph [0131] expressly describes the regression layer as outputting the predicted numeric value based on an embedding. The identified exception is the numerical estimation operation; the conclusion does not rest merely on the presence of artificial intelligence, a neural network, or an unspecified algorithm.
Step 2A Prong 2:
Considered as a whole, claim 19 obtains information with a missing numeric entry, creates its numerical representation, and calculates the missing value. The obtaining step supplies the data for the calculation; the embedder supplies its numerical input; and the processor carries out those operations. The claim ends with the predicted value and identifies no particular use of that value beyond the knowledge graph context. It does not recite the particular base-and-qualifier aggregation, contextual information flow, type encoding, or training arrangement described elsewhere in the specification. On this construction, the additional limitations provide the computational environment and inputs for the mathematical operation without integrating it into a specifically claimed technological improvement (Applicant’s claim 19; ¶¶[0069]–[0070], [0080], [0096]–[0104], [0130]–[0137]).
Applicant’s asserted improvement has been considered. Paragraphs [0005], [0008], [0137], and [0145] describe handling numeric knowledge-graph values and improved embedding accuracy. An improvement to machine-learning functionality can itself be a practical application; external physical action is not required. The concern here is that claim 19 does not specify how an improved embedding or regression mechanism achieves that result beyond the broad sequence identified above.
Step 2B: The claim does not provide an inventive concept (significantly more than the abstract idea). The claim is ineligible.
Individually and in the claimed ordered combination, the additional processor implementation and data acquisition do not supply an inventive concept beyond executing the representation-and-regression calculation. Applicant’s own disclosure permits implementation on any device having a computing function and identifies CPUs, GPUs, and other well-known processors for executing program arithmetic; these passages support the finding as to the general-purpose computing implementation. They are not treated as an admission that every disclosed neural architecture is conventional. The claimed order supplies a representation to the regression operation but adds no further technological implementation to that calculation. This rationale does not infer conventionality merely from the existence of Shomer, Chen, or Tay in the prior-art discussion (Applicant’s specification, ¶¶[0069]–[0070], [0156]–[0158], [0162]; MPEP §§2106.05(d), 2106.07(a)).
The Examiner has therefore determined that the elements, or combination of additional elements, do not integrate the abstract idea into a practical application. Accordingly, claim 19 is proposed for rejection under 35 U.S.C. §101 as directed to mathematical calculation without additional elements amounting to significantly more, on the construction and evidence stated above.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102 of this title, 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 1-6, 16-17 and 19-20 are rejected under AIA 35 U.S.C. 103(a) as being unpatentable over of Harry Shomer et al. Learning Representations for Hyper-Relational Knowledge Graphs, 30 August 2022 (hereinafter Shomer) in view Chen et al. Sanxing Chen, Xiaodong Liu, Jianfeng Gao, Jian Jiao, Ruofei Zhang, and Yangfeng Ji. HittER: Hierarchical Transformers for Knowledge Graph Embeddings. EMNLP 2021, pp. 10395–10407. First verified publication used here: November 2021 (hereinafter Chen).
Regarding claim 1, Shomer teaches A method for embedding a knowledge graph, the method being performed by at least one processor (Shomer, §4 and Appendix A, pp. 4 and 10; Wherein Shomer implements a knowledge graph embedding model using PyTorch Geometric on a Tesla V100 GPU. The expressly identified GPU supplies the claimed processor; the embedding operation is the purpose of the disclosed framework) and comprising obtaining a basic knowledge instance and a qualifier instance associated with the basic knowledge instance (Shomer, §3.2 and Fig. 1, pp. 1 and 3–4; Shomer identifies a base triple and associated qualifier relation entity pairs. The triple corresponds to claimed basic knowledge instance, and an associated pair corresponds to claimed qualifier instance) wherein at least one of the basic knowledge instance or the qualifier instance comprises a mask element (Shomer, §§4 and 4.2.1, Fig. 2, pp. 3–5; Yu, §2.3, p. 3; Shomer forms examples with a masked base entity or a masked qualifier entity. Either placement satisfies claimed alternative “at least one” requirement) Generating, through an embedder, a first instance embedding corresponding to the basic knowledge instance (Shomer, §4.1.2, Eq. (12), p. 5, Shomer concatenates the embeddings to generate a vector representing the base triple. The base triple corresponds to the claimed basic knowledge instance, and the resulting vector corresponds to the claimed first instance embedding. The concatenation and projection operations perform the claimed embedding generation)
Shomer does not teach a second instance embedding corresponding to the qualifier instance; and performing a prediction of the mask element by inputting, to a predictor, an instance embedding associated with the mask element, wherein the instance embedding corresponds to the first instance embedding or the second instance embedding.
However, in the analogous art of knowledge graph representation learning and completion, Chen teaches a second instance embedding corresponding to the qualifier instance (FIG. 2, §2.2, pp. 10397–10398, wherein Chen separately compresses each entity–relation pair before contextualization. Chen’s bottom entity Transformer packs each entity relation pair into one vector, which is supplied to the upper context Transformer. This supports generating an embedding for a pair) and performing a prediction of the mask element by inputting, to a predictor, an instance embedding associated with the mask element, wherein the instance embedding corresponds to the first instance embedding or the second instance embedding (FIG. 2, §2.3, pp. 10397–10398 wherein Chen masks or perturbs the source entity and uses the corresponding context transformer output through a classification layer to recover that entity. This supports predicting a masked element from contextual output associated with its input group).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine Shomer with Chen by incorporating Chen’s grouped-input contextualization and associated-output prediction into Shomer’s hyper-relational embedding method. Specifically, the modified method would retain Shomer’s base-triple vector; generate one vector for each qualifier pair using Chen’s pair-composition approach; contextualize the group vectors together; and supply the vector of the group containing the mask to the prediction head.
Regarding claim 2, Shomer as modified by Chen teaches updating the embedder and the predictor (Shomer, §4.2.1, p. 5; Shomer trains its embedding and prediction architecture by minimizing prediction losses for masked entities. Training adjusts the trainable parameters used to generate embeddings and predict the masked entities, thereby teaching the claimed updating of the embedder and the predictor) based on a difference between a result of the prediction and a correct answer about the mask element (Shomer, §4.2.1, Eq. (14), p. 5; Shomer calculates a binary cross entropy loss for predicting masked base entities and an additional binary cross-entropy loss for predicting masked qualifier entities. Each loss compares the prediction with correct entity label. Equation (14) combines these losses into the training objective. Minimizing the discrepancy between the predicted result and the correct answer for the masked element, thereby teaching the claimed difference-based updating)
Regarding claim 3, Shomer as modified by Chen teaches wherein the embedder comprises a context encoder configured to wherein the embedder comprises a context encoder (§2.2, Fig. 2, pp. 10397–10398, wherein Chen teaches a hierarchical embedding model with an upper context transformer that processes representations produced by the lower entity transformer. This supports the context-encoder component) perform an analysis of an association between a first embedding of the basic knowledge instance and a second embedding of the qualifier instance (§§4.1–4.2, Eq (12), p.5 wherein Shomer generates representation of an entire base triple), (§2.2, Fig. 2, pp. 10397–10398 wherein Chen processes grouped entity relations representations using its context transformer) generate the first instance embedding and the second instance embedding based on a result of the analysis of the association between the first embedding of the basic knowledge instance and the second embedding of the qualifier instance (§2.2 and Fig. 2, pp. 10397–10398, wherein Chen teaches the context block returns updated vectors at the respective input-group positions. Retaining the base position and the qualifier position supplies the two output instance embeddings).
Regarding claim 4, Shomer as modified by Chen teaches wherein the context encoder is implemented based on a self-attention module (§2.2, Fig. 2 and §3.3, pp. 10397–10400, Chen’s bidirectional Transformer blocks use multi-head attention to contextualize grouped inputs. Employing that context block supplies the required self-attention module)
Regarding claim 5, Shomer as modified by Chen teaches wherein the generating of the first instance embedding and the second instance embedding comprises: reflecting type information for distinguishing the basic knowledge instance and the qualifier instance from each other in the first embedding and the second embedding (Chen, §2.2, p. 10398; Shomer, §§3.2 and 4.1, pp. 3–4; Chen assigns different type embeddings to the intermediate source representation and intermediate neighbor representations before contextualization. Chen applies that disclosed role-tagging technique to the base-instance vector and qualifier-instance vectors already produced by Shomer) analyzing the association between the first embedding and the second embedding in which the type information is reflected through the context encoder (§2.2, pp. 10397–10398. Chen teaches the type-marked intermediate vectors enter the context Transformer. In the modified architecture, attention therefore operates on the base and qualifier vectors after their different role encodings have been included. This provides required ordering of typing before association analysis).
Regarding claim 6, Shomer as modified by Chen teaches a basic aggregator configured to aggregate first element embeddings of the basic knowledge instance to generate the first embedding (§4.1.2, Eq. (12), p. 5.; Shomer concatenates the base triple’s entity, relation, and entity embeddings and applies a linear projection. That operation aggregates the constituent elements into a base-triple vector. Shomer places this vector at the base-instance input to the context encoder.) an auxiliary aggregator configured to aggregate second element embeddings of the qualifier instance to generate the second embedding (Shomer, §4.1.1, Eq. (8), p. 4; Chen, §2.2, p. 10397.; The relation and entity of each qualifier are combined using Chen’s pair-composition block and retained as a separate group vector. Shomer’s module called “qualifier aggregator” instead updates qualifier entities from surrounding facts)
Regarding claim 16, Shomer as modified by Chen teaches wherein the predictor comprises a first predictor associated with a first prediction task and a second predictor associated with a second prediction task (Shomer, §4.2.1, Eq. (14), p. 5; Applicant specification ¶[0141], for construction only. Shomer trains base-entity prediction and qualifier-entity prediction using the same decoder, the two task-conditioned invocations of the Shomer prediction head can supply the first and second predictors) wherein, based on at least one of a type of the mask element, a format of a value of the mask element, and a type of a knowledge instance comprising the mask element, the first prediction task and the second prediction task are distinguished from each other . (§4.2.1, p. 5; Shomer distinguishes prediction in the base triple from prediction in a qualifier. This corresponds to the alternative based on the type of knowledge instance containing the mask)
Regarding claim 17, Shomer as modified by Chen teaches wherein the first predictor and the second predictor are configured to share at least one weight parameter with each other (§§4.2–4.2.1 and Fig. 2, pp. 3 and 5; wherein Shomer describes prediction of masked head or tail entities in a base triple and an auxiliary task predicting masked qualifier entities. The decoder operating on masked based entity corresponds to the first prediction function, while the decoder operating on masked qualifier entity corresponds to the second prediction function. These tasks are distinguished by whether the masked entity belongs to the basic knowledge instance or the qualifier instance. Shomer describes common decoder that processes the masked statement, extracts the transformer output at the masked position, and applies a fully connected prediction layer. Both base-entity and qualifier-entity prediction use this decoding procedure).
Regarding claim 19, the claim is similar in scope to claim 1 therefore the claims are rejected under similar rationale.
Regarding claim 20, the claim is similar in scope to claim 1 therefore the claims are rejected under similar rationale.
Claims 7-14 are rejected under AIA 35 U.S.C. 103(a) as being unpatentable over of Harry Shomer et al. Learning Representations for Hyper-Relational Knowledge Graphs, 30 August 2022 (hereinafter Shomer) in view Chen et al. Sanxing Chen, Xiaodong Liu, Jianfeng Gao, Jian Jiao, Ruofei Zhang, and Yangfeng Ji. HittER: Hierarchical Transformers for Knowledge Graph Embeddings. EMNLP 2021, pp. 10395–10407. First verified publication used here: November 2021 (hereinafter Chen) and further in view of Wang et al. Link Prediction on N-ary Relational Facts: A Graph-based Approach, 2021 (hereinafter Wang).
Regarding claim 7, Shomer as modified by Chen teaches performing, through the embedder, an analysis of an association among embeddings of the basic knowledge instance, the first qualifier instance, and the second qualifier instance (Shomer, §4.1; Chen, §2.2 and Fig. 2; In Shomer, the input to the context block contains the base group and both qualifier groups. Self-attention permits each group to incorporate information from the others. The retained group vectors, rather than a single undifferentiated qualifier sum, provide the claimed three-way analysis) Generating the first instance embedding and the second instance embedding based on a result of the analysis of the association (Chen, §2.2, pp. 10397–10398; Wherein the updated vectors at the base and first-qualifier positions depend on the contextualization that also includes the second qualifier).
Shomer and Chen do not teach wherein the generating of the first instance embedding and the second instance embedding comprises: obtaining a second qualifier instance associated with the basic knowledge instance.
However, in the analogous art of knowledge graph representation learning and completion, Wang teaches wherein the generating of the first instance embedding and the second instance embedding comprises: obtaining a second qualifier instance associated with the basic knowledge instance (§2, p. 397; Wang disclosures permit a set of multiple qualifier or auxiliary pairs associated with one base triple. Selecting two members of that set supplies first and second qualifier instances).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine the Shomer and Chen method with Wang by incorporating Wang’s relation-position masking and relation-vocabulary classification into the grouped-instance prediction method, for the purpose of completing facts whose missing item is a relation rather than an entity. Shomer’s express discussion of adding that component provides a specific reason to combine, while Wang provides the operational training and output mechanism.
Regarding claim 8, Shomer as modified by Chen and Wang teaches wherein the predicting of the mask element comprises inputting, to the predictor, element embeddings of a knowledge instance comprising the mask element (Wang, §3.2 and Fig. 2, p. 398; Shomer, §4.2, p. 5; Chen, Fig. 2. Wang supplies element embeddings to its attention-based prediction pipeline. However, claim also inherits claim 1’s instance-embedding input. Shomer’s predictor receives the grouped context vector).
Regarding claim 9, Shomer as modified by Chen and Wang teaches reflecting, in the element embeddings, information for distinguishing types of elements from each other (Chen, §2.1, p. 10396; Wang, §3.1, pp. 397–398. Chen adds different learned type embeddings to the entity, relation, and special-token inputs. That is direct support for distinguishing element roles in their representations. Wang’s typed edges are related evidence) Inputting, to the predictor, the element embeddings in which the information is reflected (Chen, §2.1 and Fig. 2, pp. 10396–10397; Wang, §3.2, p. 398. Chen processes typed element inputs with a transformer used for prediction. Applying that typing to the element-input branch would meet this added ordering).
Regarding claim 10, Shomer as modified by Chen and Wang teaches Type information of the knowledge instance corresponding to the instance embedding is reflected in the instance embedding associated with the mask element. (Chen, §2.2, p. 10398. Chen types of intermediate grouped representations. Chen supports distinguishing base and qualifier roles at the context-encoder input) wherein the instance embedding in which the type information is reflected is input to the predictor (§§2.1–2.3, p. 10398-10396; wherein Chen uses the instance embedding as an input to BERT).
Regarding claim 11, Shomer as modified by Chen and Wang teaches wherein the predictor comprises a context reflector and a prediction layer (§3.2, Fig. 2, p. 398; §3.3, Eq. (4), p. 399; Wang supplies attention layers followed by a classification layer. These provide contextual processing and prediction as two functional stages) generating a plurality of embeddings by reflecting, through the context reflector, the instance embedding associated with the mask element in the element embeddings (Wang, §3.2, pp. 398–399; Chen, §2.2 and Fig. 2, pp. 10397–10398; Wang updates element vectors using other element vectors. Chen contextualizes grouped vectors using other grouped vectors) predicting the mask element by inputting, to the prediction layer, an embedding corresponding to the mask element among the plurality of embeddings (§3.3, Eq. (4), p. 399; Wang selects the hidden representation of the mask token and sends it through the prediction layer. This corresponds to selecting the mask-associated member of several output embeddings).
Regarding claim 12, Shomer as modified by Chen and Wang teaches wherein the context reflector is implemented based on a self-attention module (Wang, §3.2, Fig. 2, pp. 398–399; Wang uses fully connected, multi-head self-attention to contextualize its input vertices. This supports self-attention technology).
Regarding claim 13, Shomer as modified by Chen and Wang teaches wherein the mask element is a discrete entity (Shomer, §4.2.1, p. 5; Chen, §2.3, p. 10398; The masked base/qualifier entity in Shomer, and the perturbed source entity in Chen, are selected from a discrete entity vocabulary. The proposed Shomer architecture therefore supports the discrete-entity alternative) wherein the predictor comprises a classification layer configured to output a predicted probability distribution for predefined entities (Chen, §§2.1–2.3, pp. 10396–10398; Wang, §3.3, Eq. (4), p. 399; Chen scores the entity candidates and normalizes their scores with softmax. Wang explicitly outputs a distribution over the entity vocabulary. These disclosures support an actual probability distribution).
Regarding claim 14, Shomer as modified by Chen and Wang teaches wherein the mask element is a relation (Wang, §2, Definition 2, p. 397; §3.3, p. 399. Wang expressly permits masking the main relation or an auxiliary attribute relation, not only entities. The combination of Chen and Wang extends the Shomer architecture’s masking task to such relation positions) wherein the predictor comprises a classification layer configured to output a predicted probability distribution for predefined relations (Wang, §3.3, Eq. (4), p. 399. Wang selects the relation vocabulary when the missing element is a relation and applies softmax over candidate relations. The combination of Shomer, Chen and Wang incorporate this task-specific output space in the proposed grouped predictor).
Claim 15 is rejected under AIA 35 U.S.C. 103(a) as being unpatentable over of Harry Shomer et al. Learning Representations for Hyper-Relational Knowledge Graphs, 30 August 2022 (hereinafter Shomer) in view Chen et al. Sanxing Chen, Xiaodong Liu, Jianfeng Gao, Jian Jiao, Ruofei Zhang, and Yangfeng Ji. HittER: Hierarchical Transformers for Knowledge Graph Embeddings. EMNLP 2021, pp. 10395–10407. First verified publication used here: November 2021 (hereinafter Chen) and further in view of Tay et al. Multi-Task Neural Network for Non-discrete Attribute Prediction in Knowledge Graphs, 2027 (hereinafter Tay).
Regarding claim 15, Shomer and Chen do not teach wherein the mask element is a numeric entity, and wherein the predictor comprises a regression layer configured to output a predicted numeric value.
However, in the analogous art of knowledge graph representation learning and completion, Tay teaches teach wherein the mask element is a numeric entity (Tay, §3.1 and §4.5, pp. 4 and 8. Tay teaches the numeric-slot construction and mapping) and wherein the predictor comprises a regression layer configured to output a predicted numeric value (Tay, §3.4, Eqs. (8)-(13), p. 5. Tay Applies the regression-output mapping to the predictor of Shomer. Tay identifies the additional modification and motivation).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine the Shomer and Chen method with Tay by incorporating Tay’s regression output into Wang’s masked prediction method for numeric targets, to complete missing numerical attributes rather than choose categorical labels (Tay, §1.1.2). The same numeric-output substitution can be applied to Shomer.
Claim 18 is rejected under AIA 35 U.S.C. 103(a) as being unpatentable over of Harry Shomer et al. Learning Representations for Hyper-Relational Knowledge Graphs, 30 August 2022 (hereinafter Shomer) in view Chen et al. Sanxing Chen, Xiaodong Liu, Jianfeng Gao, Jian Jiao, Ruofei Zhang, and Yangfeng Ji. HittER: Hierarchical Transformers for Knowledge Graph Embeddings. EMNLP 2021, pp. 10395–10407. First verified publication used here: November 2021 (hereinafter Chen) and further in view of Wang et al. Link Prediction on N-ary Relational Facts: A Graph-based Approach, 2021 (hereinafter Wang) and further in view of Wen Zhang et al. Iteratively Learning Embeddings and Rules for Knowledge Graph Reasoning 2019 (hereinafter Zhang).
Regarding claim 18, Shomer as modified by Chen and Wang teaches generating a complete knowledge instance based on a result of the prediction in a knowledge instance comprising the mask element (Shomer, §4.2, p. 5; Wang, §2, p. 397; The predicted candidate fills the unknown position of the existing fact to produce a completed record. This is a concrete use of the prediction result and scoring candidates).
Shomer, Chen and Wang do not teach inserting the generated complete knowledge instance into the knowledge graph.
However, in the analogous art of knowledge graph representation learning and completion, Zhang teaches teach inserting the generated complete knowledge instance into the knowledge graph (Introduction, p. 2; §3.3, p. 6. Zhang describes injecting newly inferred triples into the KG for subsequent embedding learning. Zhang applies that insertion step to accepted completed records from Shomer).
It would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention to combine the Shomer, Chen and Wang method with Zhang by incorporating Zhang’s insertion of inferred triples after filling a masked field, for the purpose of making accepted completed facts available for subsequent graph use and embedding learning.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HASSAN MRABI whose telephone number is (571)272-8875. The examiner can normally be reached on Monday-Friday, 7:30am-5pm. Alt, Friday, EST.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Viker Lamardo can be reached on 571-270-5871. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/HASSAN MRABI/Examiner, Art Unit 2144
