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 .
Status of Claims
Claims 1 – 20 are pending and examined herein.
Claims 6, 16 are rejected under 35 U.S.C. 112(b).
Claims 1 – 20 are rejected under 35 U.S.C. 101.
Claims 1 – 20 are rejected under 35 U.S.C. 103.
Information Disclosure Statement
The information disclosure statement filed 0 fails to comply with the provisions of 37 CFR 1.97, 1.98 and MPEP § 609 because 116662342 CN foreign reference only includes abstract and a legible copy of the cited foreign patent document was not provided. Also, "Going Digital: A survey on Digitalization and Large Scale Data Analytics in Healthcare", "Enabling scalable AI for Digital Health: Interoperability, consent and ethics support" were not considered because a legible copy of the cited non patent literature publication, or the portion listed in IDS was not provided. It has been placed in the application file, but the information referred to therein has not been considered as to the merits. Applicant is advised that the date of any re-submission of any item of information contained in this information disclosure statement or the submission of any missing element(s) will be the date of submission for purposes of determining compliance with the requirements based on the time of filing the statement, including all certification requirements for statements under 37 CFR 1.97(e). See MPEP § 609.05(a).
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 6, 16 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.
Claims 6 and 16 recites the limitation “a plurality of counter fitted entities”, “the plurality of entities”, and “a plurality of retrofitted entities”. None of their respective independent claims previously introduce “a plurality of entities”. It is unclear whether “the plurality of entities” refers to the previously recited plurality of counterfitted entities or to a seprater plurality of entities. There is lack of antecedent basis for this limitation of the claims. For the examination purposes, “the plurality of entities” would refer back to previously introduced “a plurality of counterfitted entities”.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1 - 20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
MPEP § 2109(III) sets out steps for evaluating whether a claim is drawn to patent-eligible subject matter. The analysis of claims 1 – 20, in accordance with these steps, follows.
Step 1 Analysis:
Step 1 is to determine whether the claim is directed to a statutory category (process, machine, manufacture, or composition of matter.
Claims 1 – 13 are directed to a method, meaning that it is directed to the statutory category of process. Claims 14 – 19 are directed to a system, which is the statutory category of machine. Claim 20 is directed to one or more non-transitory computer-readable storage media, which can be an article of manufacture.
Step 2A Prong One, Step 2A Prong Two, and Step 2B Analysis:
Step 2A Prong One asks if the claim recites a judicial exception (abstract idea, law of nature, or natural phenomenon). If the claim recites a judicial exception, analysis proceeds to Step 2A Prong Two, which asks if the claim recites additional elements that integrate the abstract idea into a practical application. If the claim does not integrate the judicial exception, analysis proceeds to Step 2B, which asks if the claim amounts to significantly more than the judicial exception. If the claim does not amount to significantly more than the judicial exception, the claim is not eligible subject matter under 35 U.S.C. 101.
Regarding claim 1, the following claim elements are abstract ideas:
identifying … one or more candidate node paths and one or more node relations …(This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components or by a human using a pen and paper.)
generating … one or more context-relationship ranking predictions based on the one or more candidate node paths, the one or more node relations, and the one or more context embeddings; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components or by a human using a pen and paper.)
generating … one or more subgraph data objects from the one or more knowledge graph data objects based on the one or more context-relationship ranking predictions; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components or by a human using a pen and paper.)
and generating … one or more answer outputs for the query input based on the one or more subgraph data objects. (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components or by a human using a pen and paper.)
The following claim elements are additional elements which, taken alone or in combination with the other additional elements, do not integrate the judicial exception into a practical application nor amount to significantly more than the judicial exception:
,by one or more processors, (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
generating … one or more context embeddings based on a contextual representation data object that is associated with a query input; (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
… that are associated with one or more knowledge graph data objects; (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
, by the one or more processors and using a predictive machine learning model, (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 2, the rejection of claim 1 is incorporated herein. Further, claim 2 recites the following abstract idea:
determining a plurality of similar contexts for a plurality of entities or topics based on one or more feature similarities. (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components.)
Claim 2 does not recite additional element
Regarding claim 3, the rejection of claim 2 is incorporated herein. Further, claim 3 recites the following additional elements:
generating one or more retrofitted entities from a first portion of the plurality of entities or topics based on one or more inclusion features; (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
and generating one or more counter-fitted entities from a second portion of the plurality of entities or topics based on one or more exclusion features. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 4, the rejection of claim 2 is incorporated herein. Further, claim 4 recites the following abstract idea:
determining one or more hop relations based on a plurality of shared documents; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components.)
associating the query input with one or more documents, one or more entities, or one or more topics based on the one or more hop relations; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components.)
Claim 4 recites following additional elements:
generating one or more retrofitted entities or one or more counter-fitted entities based on the associations. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 5, the rejection of claim 1 is incorporated herein. Further, claim 5 recites the following abstract idea:
determining a plurality of positive constraints for a plurality of entities based on a cost function that is associated with a maximum distance measurement between the plurality of entities and a plurality of retrofitted entities. (This is merely reciting mathematical relationship, which is mathematical concept.)
Regarding claim 6, the rejection of claim 1 is incorporated herein. Further, claim 6 recites the following abstract idea:
determining a plurality of negative constraints for a plurality of counter-fitted entities based on a cost function that is associated with a minimum distance measurement between the plurality of entities and a plurality of retrofitted entities. (This is merely reciting mathematical relationship, which is mathematical concept.)
Regarding claim 7, the rejection of claim 1 is incorporated herein. Further, claim 7 recites the following additional element:
wherein the predictive machine learning model comprises a supervised machine learning model. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 8, the rejection of claim 1 is incorporated herein. Further, claim 8 recites the following abstract idea:
comparing one or more knowledge graph embeddings that are associated with one or more nodes of the one or more candidate node paths with the one or more context embeddings; (This is merely reciting mathematical relationship, which is mathematical concept.)
and determining similarity between the one or more nodes and the one or more context embeddings based on the comparison. (This is merely reciting mathematical relationship, which is mathematical concept.)
Claim 8 does not recite additional element
Regarding claim 9, the rejection of claim 1 is incorporated herein. Further, claim 9 recites the following additional element:
generating one or more parameters based on training data that comprises a plurality of training queries, a plurality of training node paths, a plurality of training node relations, and a plurality of contextual path labels. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 10, the rejection of claim 9 is incorporated herein. Further, claim 10 recites the following abstract idea:
generating one or more validation context-relationship ranking predictions for one or more validation node paths based on the one or more parameters; (This is practical to perform in the human mind under its broadest reasonable interpretation aside from the recitation of generic computer components.)
generating one or more similarity scores by comparing the one or more validation context-relationship ranking predictions with respective one or more actual rankings for the one or more validation node paths; (This is merely reciting mathematical relationship, which is mathematical concept.)
Claim 10 recites following additional element
and fine-tuning the predictive machine learning model by generating one or more updated parameters based on the one or more similarity scores. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 11, the rejection of claim 1 is incorporated herein. Further, claim 11 recites the following additional element:
generating… an output sequence of node relations based on an input sequence (This is mere data gathering and outputting, an insignificant extra solution activity, which is a well-understood, routine conventional activity. It does not integrate the judicial exception into a practical application. See MPEP § 2106.05(d). Therefore, this does not amount to significantly more than the judicial exception.)
, using a variational autoencoder machine learning model, … that comprises one or more candidate entities, one or more candidate topics, and one or more candidate node relations that are associated with the one or more knowledge graph data objects. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 12, the rejection of claim 11 is incorporated herein. Further, claim 12 recites the following additional element:
generating, using a bidirectional recurrent neural network, one or more embeddings of the one or more knowledge graph data objects based on the input sequence. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Regarding claim 13, the rejection of claim 1 is incorporated herein. Further, claim 13 recites the following additional element:
wherein the one or more knowledge graph data objects comprise (i) a plurality of nodes associated with a plurality of topics, a plurality of entities, or a plurality of documents and (ii) a plurality of edges between the plurality of nodes. (These are mere instructions to apply abstract idea on a generic computer. See MPEP § 2106.05(f). Therefore, this does not amount to significantly more than the judicial exception.)
Claims 14, 20 recite substantially similar subject matter to claim 1 respectively and are rejected with the same rationale, mutatis mutandis.
Claims 15 – 19 recite substantially similar subject matter to claims 5-6, 8 – 10 respectively and are rejected with the same rationale, mutatis mutandis.
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.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claims 1-3, 5-7, 9, 11-16, 18, 20 are rejected under 35 U.S.C. 103 as being unpatentable over Zhang et al. (NPL: “Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering”) in view of Lo et al. (NPL: “Contextual Path Retrieval: A Contextual Entity Relation Embedding-based Approach”).
Regarding Claim 1, Zhang teaches
identifying, by the one or more processors, one or more candidate node paths and one or more node relations that are associated with one or more knowledge graph data objects; (Pg. 5775 – 5776 of Zhang states “Path expanding starts from a topic entity and follows a sequential decision process. Here a path is defined as a sequence of relations (r1,··· ,r|p|), since a question usually implies the intermediate relations excluding the entities. Suppose a partial path p(t) = (r1,··· ,rt) has been retrieved at time t, a tree can be induced from p(t) by filling in the intermediate entities along the path, i.e., T(t) = (eq,r1,E1,··· ,rt,Et). Each Et is an entity set as a head entity and a relation can usually derive multiple tail entities. Then we select the next relation from the union of the neighboring relations of Et. The relevance of each relation r to the question q is measured by the dot product between their embeddings, i.e., s(q, r) = f(q)⊤h(r), (4) where both f and h are instantiated by RoBERTa (Liu et al., 2019). Specifically, we input the question or the name of r into RoBERTa and take its [CLS] token as the output embedding. According to the assumption (Chen et al., 2019b; He et al., 2021; Qiu et al., 2020a; Zhou et al., 2018) that expanding relations at different time steps should attend to specific parts of a query, we update the embedding of the question by simply concatenating the original question with the historical expanded relations in p(t) as the input of RoBERTa, i.e., f(q(t)) = RoBERTa([q;r1;··· ;rt]), (5) Thus s(q,r) is changed to s(q(t),r) = f(q(t))⊤h(r). Then the probability of a relation r being expanded can be formalized as: p(r|q(t)) = 1 1 + exp(s(q(t),END) − s(q(t),r)) , (6) where END is a virtual relation named as “END”. The score s(q(t),END) represents the threshold of the relevance score.” Path in Zhang is defined as a sequence of relations expanded from a topic entity in the knowledge base via top K beam search over neighboring relations.)
generating, by the one or more processors and using a predictive machine learning model, one or more context-relationship ranking predictions based on the one or more candidate node paths, the one or more node relations, and the one or more context embeddings; (Pg. 5775 – 5776 of Zhang states “Path expanding starts from a topic entity and follows a sequential decision process. Here a path is defined as a sequence of relations (r1,··· ,r|p|), since a question usually implies the intermediate relations excluding the entities. Suppose a partial path p(t) = (r1,··· ,rt) has been retrieved at time t, a tree can be induced from p(t) by filling in the intermediate entities along the path, i.e., T(t) = (eq,r1,E1,··· ,rt,Et). Each Et is an entity set as a head entity and a relation can usually derive multiple tail entities. Then we select the next relation from the union of the neighboring relations of Et. The relevance of each relation r to the question q is measured by the dot product between their embeddings, i.e., s(q, r) = f(q)⊤h(r), (4) where both f and h are instantiated by RoBERTa (Liu et al., 2019). Specifically, we input the question or the name of r into RoBERTa and take its [CLS] token as the output embedding. According to the assumption (Chen et al., 2019b; He et al., 2021; Qiu et al., 2020a; Zhou et al., 2018) that expanding relations at different time steps should attend to specific parts of a query, we update the embedding of the question by simply concatenating the original question with the historical expanded relations in p(t) as the input of RoBERTa, i.e., f(q(t)) = RoBERTa([q;r1;··· ;rt]), (5) Thus s(q,r) is changed to s(q(t),r) = f(q(t))⊤h(r). Then the probability of a relation r being expanded can be formalized as: p(r|q(t)) = 1 1 + exp(s(q(t),END) − s(q(t),r)) , (6) where END is a virtual relation named as “END”. The score s(q(t),END) represents the threshold of the relevance score.” Zhang’s equations show relation relevance scored as the dot product of the context derived question embedding and the relation embedding.)
generating, by the one or more processors, one or more subgraph data objects from the one or more knowledge graph data objects based on the one or more context-relationship ranking predictions; (Pg. 5776 of Zhang states “Since the top-1 relevant path cannot be guaranteed to be right, we perform a top-K beam search at each time to get K paths. From each topic entity, we obtain K paths which result in nK paths in total by n topic entities. nK paths correspond to nK instantiated trees. We take the union of top-K trees from one topic entity into a single subgraph, and then merge the same entities from different subgraphs to induce the final subgraph. This can reduce the subgraph size, i.e., the answer reasoning space, as the subgraphs from different topic entities can be viewed as the constraints of each other. Specifically, from the n subgraphs of the n topic entities, we find the same entities and merge them. From these merged entities, we trace back in each subgraph to the root (i.e., a topic entity) and trace forward to the leaves. Then we only keep the entities and relations along the tracing paths of all the trees to form the final subgraph.” Zhang described top k ranked trees per topic entity union, duplicate entities merged, and tracing paths retained as the final subgraph.)
and generating, by the one or more processors, one or more answer outputs for the query input based on the one or more subgraph data objects. (Pg. 5775 of Zhang states “knowledge base (KB) G organizes the factual information as a set of triples, i.e., G = {(e, r, e′)|e, e′ ∈ E,r ∈ R}, where E and R denote the entity set and the relation set respectively. Given a factoid question q, KBQA is to figure out the answers Aq to the question q from the entity set E of G. The entities mentioned in q are topic entities denoted by Eq = {eq}, which are assumed to be given. This paper considers the complex questions where the answer entities are multi-hops away from the topic entities, called multi-hop KBQA. Probabilistic Formalization of KBQA. Given a question q and one of its answers a ∈ Aq, we formalize the KBQA problem as maximizing the probability distribution p(a|G,q). Instead of directly reasoning on G, we retrieve a subgraph G ⊆ G and infer a on G. Since G is unknown, we treat it as a latent variable and rewrite p(a|G,q) as: p(a|G,q) = pϕ(a|q,G)pθ(G|q). (1) G In the above equation, the target distribution p(a|G,q) is jointly modeled by a subgraph retriever pθ(G|q) and an answer reasoner pϕ(a|q,G). The subgraph retriever pθ defines a prior distribution over a latent subgraph G conditioned on a question q, while the answer reasoner pϕ predicts the likelihood of the answer a given G and q.” The reasoner produces the answer from the retrieved subgraph G.)
However, Zhang does not explicitly teach that
generating, by one or more processors, one or more context embeddings based on a contextual representation data object that is associated with a query input;
Lo teaches that
generating, by one or more processors, one or more context embeddings based on a contextual representation data object that is associated with a query input; (Pg. 7 of Lo states “Definition 2. Contextual Path Retrieval Problem (CPR): Given a knowledge graph G, a query ⟨eh,et,d⟩ consists of a context document d ∈ D and two entities eh and et mentioned in d, we want to retrieve fromG the contextual paths between eh and et. For example, in Figure 1, Alfonso Cuarón and Children of Men are two entities in a context document containing several sentences. The CPR problem is to retrieve the contextual path(s) from knowledge graph that explains the connection between Alfonso Cuarón and Children of Men. In the figure, the path ⟨Al f onso Cuar ´on,director,Children of Men⟩ is a candidate contextual path. We divide the CPR task into twosubtasks: (a) construction of candidate paths connectingeh and et and (b) determination of contextual paths among the candidates.” Pg. 8 of Lo states “In the training phase, three major components are trained, namely, context encoder, path encoder, and path ranker. First, the context encoder is trained to return context representations zcx,m and zcx,mth of entities eh and et, respectively, for each query qm = ⟨eh,et,d⟩∈Q. The context encoder may combine the embeddings of other words and/or entities in the relevant section of the context document as it generates the two context representations. The second component path encoder returns a path representation for each path found in G. In particular, it generates the representations {zpt,ml} for the candidate paths {pmi } of the query qm. While paths are sequences of entities and relations, path encoder turns them into vector representation forms that can be matched against representation of the query context. During the training phase, both context encoder and path encoder are learned from the queries with input entity pairs and their labelled contextual and non-contextual paths.” Lo defines a query as a triple of head entity, tail entity, and context document. Then the context encoder generates context representations for the query entities from that document anchored structure.)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from Zhang with Lo. Zhang teaches ranking knowledge graph paths based on a query and forming a subgraph from the ranked paths. Lo teaches encoding query context and candidate paths using context fused entity embeddings and path representations. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Lo with Zhang so that document context information could be considered when ranking the candidate paths and select paths more relevant to the query context. The combination would have been predictable to improve the relevance of the retrieved subgraph by accounting for contextual similarities and differences when scoring candidate path.
Regarding claim 2, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
determining a plurality of similar contexts for a plurality of entities or topics based on one or more feature similarities (Pg. 11 of Lo states “Constraints tell us how entity embeddings should be updated. In this section, we introduce all constraints used in the retrofitting cost function: C(ze , ze ) = CNA + SDA + NCR +VSP + RP. To obtain a context-fused entity embedding method that leverages knowledge from its context, we propose two in-context constraints (CNA and SDA), one out-context constraint (NCR), and two regularization terms (VSP and RP). This is the only context encoder that does not differentiate the setting of context range of WD and CW, as the constraints are retrieved from both in and out-context. First, the in-context constraints capture the co-occurrence(s) of an entity e with a query entity e within some context range. With retrofitting, we adjust the embeddings of e and e, i.e.,ze and ze , to incorporate their context similarity. Through the retrofitting process, the entity will gradually move its representation towards the in-context constraints in the vector space to obtain more semantic similarity with them. We define two kinds of in-context constraints based on how close the entities are to e in the context document.” Lo determines similar contexts for entities based on cooccurrence within a defined context window or shared context document. That cooccurrence determination is the claimed feature similarity and the resulting entity grouping is the claimed similar context. )
Regarding claim 3, the rejection of claim 2 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
generating one or more retrofitted entities from a first portion of the plurality of entities or topics based on one or more inclusion features; (Pg. 11 of Lo states “First, the in-context constraints capture the co-occurrence(s) of an entity e′ with a query entity e within some context range. With retrofitting, we adjust the embeddings of e and e′, i.e., ze and ze′ , to incorporate their context similarity. Through the retrofitting process, the entity will gradually move its representation towards the in-context constraints in the vector space to obtain more semantic similarity with them. We define two kinds of in-context constraints based on how close the entities are to e in the context document. Close Neighbor Attract (CNA):ze is retrofitted with entity e′ ∈ ECW d,e . For a query entity e and context document d, we defined context window entity set ECW d,e to be the set of entities that appear in the context window of e in d.”)
and generating one or more counter-fitted entities from a second portion of the plurality of entities or topics based on one or more exclusion features. (Pg. 12 of Lo states “The out-context constraints, however, are entities that we do not want the target entity e to be close to in the vector space. In the case of context encoding in CPR,such out-context constraints are negative context, that is, entities that do not appear in the context of e. ze will learn to repel themselves from the these entities during the retrofitting(counter-fitting) process. Negative Context Repel (NCR): ze is counter-fitted with entity e† ∈ ENC e ,where ENC e is the set of entities that have relation with e but do not appear in the context document d containing e.”)
Regarding claim 5, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
determining a plurality of positive constraints for a plurality of entities based on a cost function that is associated with a maximum distance measurement between the plurality of entities and a plurality of retrofitted entities. (Pg. 11 – 12 of Lo states “Close Neighbor Attract (CNA):ze is retrofitted with entity e′ ∈ ECW d,e . For a query entity e and context document d, we defined context window entity set ECW d,e to be the set of entities that appear in the context window of e in d. For example, in the following context window of the query entity Daniel Craig: “British actor Daniel Craig has been confirmed as the man to follow Pierce Brosnan as the sixth James Bond.” CNA includes ECW Daniel Craig = {Pierce Brosnan, James Bond}. Let d(·) be any kind of distance measurement and τ(x) ≜ max(0,x), CAN thus derives the new embeddings of e and e′,i.e.,z′ e and z′e′,respectively, by minimizing the cost function: CNA(z′e) = τ (d(z′e,z′e′) − γ ),e′∈ECWe where γ is the ideal maximum distance between e and e′ ∈ ECW. Here, we empirically set γ =0.”)
Regarding claim 6, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
determining a plurality of negative constraints for a plurality of counter-fitted entities based on a cost function that is associated with a minimum distance measurement between the plurality of entities and a plurality of retrofitted entities. (Pg. 12 of Lo states “The out-context constraints, however, are entities that we do not want the target entity e to be close to in the vector space. In the case of context encoding in CPR,such out-context constraints are negative context, that is, entities that do not appear in the context of e. ze will learn to repel themselves from the these entities during the retrofitting(counter-fitting) process. Negative Context Repel (NCR): ze is counter-fitted with entity e† ∈ ENC e ,where ENC e is the set of entities that have relation with e but do not appear in the context document d containing e. From the previous example, entity Daniel Craig’s NCR includes ENC Daniel Craig = {Knives Out, Logan Lucky, ...}. Intending to push e away from e†s, AR seeks to minimize the following cost function: NCR(z′e) = τ (δ −d(z′e,z′e†)),e†∈ENC e where δ is the ideal minimum distance between e and e† ∈ ENCe. Here, we empirically set δ =1.”)
Regarding claim 7, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
wherein the predictive machine learning model comprises a supervised machine learning model. (Pg. 5785 of Zhang states “Supervised Training. WebQSP provides the relation chains corresponding to each (question, answer) pair. For each question, we use each relation chain from each topic entity to the answer as the ground truth path. In this way, we obtain 3,098 (question, path) instances which can be decomposed into 5,394 (question, relation) instances in total for supervised training. The learning rate for supervised training is set as 5e-5. An epoch takes about 5 minutes and the loss function converges within 10 epochs on WebQSP/CWQ.”)
Regarding claim 9, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
training the predictive machine learning model by generating one or more parameters based on training data that comprises a plurality of training queries, a plurality of training node paths, a plurality of training node relations, and a plurality of contextual path labels. (Pg. 5776 – 5777 of Zhang states “Since the ground truth subgraphs are not easy to be obtained, we resort to the weak supervision signals constructed from the (q,a) pairs. Specifically, from each topic entity of a question, we retrieve all the shortest paths to each answer as the supervision signals, as paths are easier to be obtained than graphs. Since maximizing the log likelihood of a path equals to P|p| t=1 log p(rt|q(t)) according to Eq. (7), we can maximize the probabilities of all the intermediate relations in a path. To achieve the goal, we decompose a path p = (r1, · · · , r|p|) into |p| + 1 (question, relation) instances, including ([q], r1), ([q; r1], r2), ..., ([q; r1; r2; · · · ; r|p|−1], r|p|), and an additional END instance ([q; r1; r2; · · · ; r|p|],END), and optimize the probability of each instance. We replace the observed relation at each time step with other sampled relations as the negative instances to optimize the probability of the observed ones.” Zhang decomposes each supervision path into (question, relation) training instances, replacing the observed relation with sampled negatives to form negative instances)
Regarding claim 11, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
generating, using a variational autoencoder machine learning model, an output sequence of node relations based on an input sequence that comprises one or more candidate entities, one or more candidate topics, and one or more candidate node relations that are associated with the one or more knowledge graph data objects. (Pg. 14 of Lo states “For path encoding, we propose a PathVAE to encode paths into their latent representations that preserved as much semantics as possible for reconstructing the path. While many sequential embedding models exist, we choose to utilize a variational autoencoder structure for (i) VAE is unsupervised, so the learning of PathVAE could be separated from other components; (ii) the PathVAE is inductive, so new paths that are unseen from the training data can still be encoded. Unlike entities, path is a sequence of entities and relations. It is essential to preserve the ordering of the entities/relations as we generate the path representation. We first break up a path pm = ⟨eh,r1,e2,r2,...,r|pm|−1,et⟩ into a sequence of elements and feed each element to a Long Short-Term Memory (LSTM), one at a time [19]. One could also choose to use other sequential encoding methods such as Bi-LSTM or transformer to replace the LSTM layer. As each element is a single entity or relation, PathVAE encoderqϕ takes each entity or relation embedding generated by a base embedding model and obtains the overall path embedding after processing the entire sequence of path elements using LSTM.” Pg. 5775 of Zhang states “A knowledge base (KB) G organizes the factual information as a set of triples, i.e., G = {(e, r, e′)|e, e′ ∈ E, r ∈ R}, where E and R denote the entity set and the relation set respectively. Given a factoid question q, KBQA is to figure out the answers Aq to the question q from the entity set E of G. The entities mentioned in q are topic entities denoted by Eq = {eq}, which are assumed to be given. This paper considers the complex questions where the answer entities are multi-hops away from the topic entities, called multi-hop KBQA… Path expanding starts from a topic entity and follows a sequential decision process. Here a path is defined as a sequence of relations (r1, · · · , r|p|), since a question usually implies the intermediate relations excluding the entities. Suppose a partial path p(t) = (r1, · · · , rt) has been retrieved at time t, a tree can be induced from p(t) by filling in the intermediate entities along the path, i.e., T(t) = (eq, r1,E1, · · · , rt,Et). Each Et is an entity set as a head entity and a relation can usually derive multiple tail entities. Then we select the next relation from the union of the neighboring relations of Et.” Zhang’s topic entity anchored candidate path is encoded with Lo’s PathVAE and the decoder reconstructions Zhang’s relation path output.)
Regarding claim 12, the rejection of claim 11 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
generating, using a bidirectional recurrent neural network, one or more embeddings of the one or more knowledge graph data objects based on the input sequence. (Pg. 13 of Lo states “BERT is the state-of-the-art representation learning method, which utilizes transformer to learn a bi-directional language model [16]. Each token is embedded with respect to the surrounding tokens in the context. Thus, a word will have different representations when it appears in different contexts. The context of eh(et) is encoded using BERT to be zcx,mh (zcx,mt).We use the BERT-small model with three layers, and the dimension size of 512.” Pg. 14 of Lo states “While many sequential embedding models exist, we choose to utilize a variational autoencoder structure for (i) VAE is unsupervised, so the learning of PathVAE could be separated from other components; (ii) the PathVAE is inductive, so new paths that are unseen from the training data can still be encoded… One could also choose to use other sequential encoding methods such as Bi-LSTM or transformer to replace the LSTM layer.”)
Regarding claim 13, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
wherein the one or more knowledge graph data objects comprise (i) a plurality of nodes associated with a plurality of topics, a plurality of entities, or a plurality of documents and (ii) a plurality of edges between the plurality of nodes. (Pg. 5775 of Zhang states “A knowledge base (KB) G organizes the factual information as a set of triples, i.e., G = {(e, r, e′)|e, e′ ∈ E, r ∈ R}, where E and R denote the entity set and the relation set respectively. Given a factoid question q, KBQA is to figure out the answers Aq to the question q from the entity set E of G. The entities mentioned in q are topic entities denoted by Eq = {eq}, which are assumed to be given. This paper considers the complex questions where the answer entities are multi-hops away from the topic entities, called multi-hop KBQA” Zhang defines the knowledge base as triples where it comprises entity set and relation set. The entities are nodes and the relations are edges of knowledge graph. )
Claims 14, 20 recite substantially similar subject matter to claim 1 respectively and are rejected with the same rationale, mutatis mutandis.
Claims 15 – 16, 18 recite substantially similar subject matter to claims 5-6, 9 respectively and are rejected with the same rationale, mutatis mutandis.
Claims 4 are rejected under 35 U.S.C. 103 as being unpatentable over Zhang et al. (NPL: “Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering”) in view of Lo et al. (NPL: “Contextual Path Retrieval: A Contextual Entity Relation Embedding-based Approach”), further in view of Gramatica (U.S. Pub. 20140337306).
Regarding claim 4, the rejection of claim 2 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
associating the query input with one or more documents, one or more entities, or one or more topics; (Pg. 7 of Lo states “Contextual Path Retrieval Problem (CPR): Given a knowledge graph G, a query <eh, et ,d> consists of a context document d ∈ D and two entities eh and et mentioned in d, we want to retrieve from G the contextual paths between eh and et .”)
and generating one or more retrofitted entities or one or more counter-fitted entities based on the associations. (Pg. 11 – 12 of Lo states “Close Neighbor Attract (CNA): ze is retrofitted with entity e∈ ECW d,e . For a query entity e and context document d, we defined context window entity set ECWd,e to be the set of entities that appear in the context window of e in d… Same Document Attract (SDA): ze is retrofitted with entity e∈ ESD d,e . We defined same document entity set ESD d,e to be the set of entities that appear in the context document d of the query entity e and they do not exist in ECW d,e . Using the same example in CNA, towards ESD Daniel Craig = Ed – ECW Daniel Craig. Here, Ed is the set of entities included in d. Similar to CNA, SDA is designed to adjust the embeddings of e and e ∈ ESD d,e by minimizing the cost function… Negative Context Repel (NCR): ze is counter-fitted with entity e† ∈ ENC e , where ENC e is the set of entities that have relation with e but do not appear in the context document d containing e.”)
However, the combination does not explicitly teach that
determining one or more hop relations based on a plurality of shared documents;
associating the query input … based on the one or more hop relations
Gramatica teaches that
determining one or more hop relations based on a plurality of shared documents; ([0074] of Gramatica states “Specifically, paths—i.e. indirect relations among concepts—can be interpreted as complex inferences never explicitly expressed in any source sentence or document. If, for example, the relation “A implies B” is identified from a source and the relation “B implies C” in identified from another source, i.e. Source 1: A→B - - - Source 2: B→C” [0076] of Gramatica states “Even if the inference A→C, i.e. A implies C, is never mentioned in any source paper, the system according to the present embodiment will be able to build it from its constituent parts, leveraging on the graph representation of interlinked concepts. The graph produced by the two sentences has in fact the simple form A→B→C and A→C is one acceptable connection, i.e. a correlation between A and C. Obviously, in a complex data structure the chain of inferences are much longer than three elements and in this case the system will produce an inference path X→Y. In a further step according to the present embodiment, the search machine may rank these paths according to a measures of relevance based on link weights.” Gramatica teaches deriving a multihop relation from multiple source documents that share an intermediate concept. Also, the path across the information nodes involves at least two data objects.)
associating the query input … based on the one or more hop relations ([0011] of Gramatica states “establishing an information network with a plurality of information nodes and links between the information nodes, said information nodes being related to said information units and said links being related to said detected correlations and weighted accordingly; analyzing a link connectivity state of said weighted information network to find a path across multiple information nodes and provide inferences across a multitude of nodes, the nodes being input by a query searched by a user. As a result, a connection may be identified between two or many (end points) information nodes via a path of other intermediate information nodes that was previously unknown to the user.” Combined with Lo, the query entities or document are associated with the additional documents and entities reaches through Gramatica’s multihop relation.)
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from Gramatica with Zhang and Lo. Zhang teaches ranking knowledge graph paths based on a query and forming a subgraph from the ranked paths. Lo teaches encoding query context and candidate paths using context fused entity embeddings and path representations. Gramatica teaches forming multihop paths from relationships identified in separate source documents. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Gramatica into the combination of Lo and Zhang to identify indirect relationships from information distributed across multiple documents, including relationships not expressly stated in any single document. It would have been predictable combination to provide additional multihop relationships for associating a query with relevant entities, topics, and documents.
Claims 8, 17 are rejected under 35 U.S.C. 103 as being unpatentable over Zhang et al. (NPL: “Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering”) in view of Lo et al. (NPL: “Contextual Path Retrieval: A Contextual Entity Relation Embedding-based Approach”), further in view of Wu et al. (U.S. Pub. 20220108188).
Regarding claim 8, the rejection of claim 1 is incorporated herein. Furthermore, the combination of Zhang and Lo does not explicitly teach that
comparing one or more knowledge graph embeddings that are associated with one or more nodes of the one or more candidate node paths with the one or more context embeddings;
and determining similarity between the one or more nodes and the one or more context embeddings based on the comparison.
Wu teaches that
comparing one or more knowledge graph embeddings that are associated with one or more nodes of the one or more candidate node paths with the one or more context embeddings; ([0049] of Wu states “For example, the KG subgraph encoder component 502 can employ one or more bidirectional message passing neural networks (“MPNNs”) to encode both the nodes and edges of the one or more KG subgraphs… For instance, each node representation “sv” in a given KG subgraph “gi” can be updated in accordance with Equations 6 and 7 below. s′ v├ =W 2├s v ├+Σu∈N ├ (v) s u·ρ├(e u,v) (6) s′ v┤ =W 2┤s v ┤+Σu∈N ┤ (v) s u·ρ┤(e u,v) (7) Where “W2├” and “W2┤” can be learnable parameters. Similarly, “N├(v)” and “N┤(v)” can be nodes in the given KG subgraph that neighbor node “v”. Further, (u, v)∈Ei can be a directed edge of the given KG subgraph. Additionally, ρ(⋅) can be a single-layer perceptron with rectified linear unit (“ReLU”) activation function. The neural network embedding “ri” for a given KG subgraph “gi” can be computed by the KG subgraph encoder component 502 in accordance with Equation 8 below.
PNG
media_image1.png
22
307
media_image1.png
Greyscale
Where “φ(⋅)” can be a single-layer perceptron, “Ni” can be the number nodes of the given KG subgraph, and “s′n” can be the final node embedding of the node “n”.” [0051] of Wu states “For example, the match component 602 can calculate a matching score between the question graph embedding “rQ” and the KG subgraph embedding “ri”. For instance, each KG subgraph can be characterized by one or more respective KG subgraph embeddings, and each KG subgraph embedding can be compared with the question graph embedding;” Wu teaches initializing and updating each node representation in a KG subgraph and computing the KG subgraph embedding from the final node embeddings. Then it compares that node derived KG subgraph embedding with a question graph embedding.)
and determining similarity between the one or more nodes and the one or more context embeddings based on the comparison. ([0051] of Wu states “The match scores can characterize an amount of similarity between the respective KG subgraph and the question graph. In various embodiments, the match component 602 can execute a similarity algorithm to computer the match scores. For example, the match component 602 can execute a cosine similarity algorithm to compare the neural network embeddings in accordance with Equation 9 below. {circumflex over (t)}=cos(r Q ,r i)(9)” [0052] of Wu states “The match component 602 can determine that a given KG subgraph embedding matches the question graph embedding based on the associate match score being greater than or equal to a defined threshold (e.g., defined via the one or more input devices 106). Further, the match component 602 can rank the KG subgraphs associated with the matched KG subgraph embeddings based on the match score. For example, amongst the KG subgraph embeddings determined by the match component 602 to match the question graph embedding, the match component 602 can rank the KG subgraph embeddings based on their match score. For instance, as the match score increases, an amount of similarity between the associate KG subgraph and question graph can also increase.” Wu teaches calculating a match score between the node derived KG subgraph embedding and the question graph embedding using cosine similarity. It also teaches that the score characterizes the amount of similarity and could be used to rank the KG subgraphs. )
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from Wu with Zhang and Lo. Zhang teaches ranking knowledge graph paths based on a query and forming a subgraph from the ranked paths. Lo teaches encoding query context and candidate paths using context fused entity embeddings and path representations. Wu teaches comparing question graph and knowledge graph subgraph embeddings using cosine similarity and ranking the subgraphs based on the resulting match scores. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Wu into the combination of Lo and Zhang to provide a numerical measure of the similarity between the query representation and each candidate subgraph for use in ranking the candidate subgraphs. It would have been predictable combination to allow consistent and precise ranking with respect to the similarity of query context.
Claim 17 recites substantially similar subject matter to claim 8 respectively and is rejected with the same rationale, mutatis mutandis.
Claims 10, 19 are rejected under 35 U.S.C. 103 as being unpatentable over Zhang et al. (NPL: “Subgraph Retrieval Enhanced Model for Multi-hop Knowledge Base Question Answering”) in view of Lo et al. (NPL: “Contextual Path Retrieval: A Contextual Entity Relation Embedding-based Approach”), further in view of Hiranandani et al. (U.S. Pub. 20180121988).
Regarding claim 10, the rejection of claim 9 is incorporated herein. Furthermore, the combination of Zhang and Lo teaches that
generating one or more … context-relationship ranking predictions for one or more … node paths based on the one or more parameters; (Pg. 5775 – 5776 of Zhang states “Path expanding starts from a topic entity and follows a sequential decision process. Here a path is defined as a sequence of relations (r1,··· ,r|p|), since a question usually implies the intermediate relations excluding the entities. Suppose a partial path p(t) = (r1,··· ,rt) has been retrieved at time t, a tree can be induced from p(t) by filling in the intermediate entities along the path, i.e., T(t) = (eq,r1,E1,··· ,rt,Et). Each Et is an entity set as a head entity and a relation can usually derive multiple tail entities. Then we select the next relation from the union of the neighboring relations of Et. The relevance of each relation r to the question q is measured by the dot product between their embeddings, i.e., s(q, r) = f(q)⊤h(r), (4) where both f and h are instantiated by RoBERTa (Liu et al., 2019). Specifically, we input the question or the name of r into RoBERTa and take its [CLS] token as the output embedding. According to the assumption (Chen et al., 2019b; He et al., 2021; Qiu et al., 2020a; Zhou et al., 2018) that expanding relations at different time steps should attend to specific parts of a query, we update the embedding of the question by simply concatenating the original question with the historical expanded relations in p(t) as the input of RoBERTa, i.e., f(q(t)) = RoBERTa([q;r1;··· ;rt]), (5) Thus s(q,r) is changed to s(q(t),r) = f(q(t))⊤h(r). Then the probability of a relation r being expanded can be formalized as: p(r|q(t)) = 1 1 + exp(s(q(t),END) − s(q(t),r)) , (6) where END is a virtual relation named as “END”. The score s(q(t),END) represents the threshold of the relevance score.” Zhang teaches generating context relationship ranking predictions for node paths based on learned parameters. )
generating … context-relationship ranking predictions with respective one or more actual rankings for the one or more … node paths; (Pg. 5776 – 5777 of Zhang states “Since the ground truth subgraphs are not easy to be obtained, we resort to the weak supervision signals constructed from the (q,a) pairs. Specifically, from each topic entity of a question, we retrieve all the shortest paths to each answer as the supervision signals, as paths are easier to be obtained than graphs. Since maximizing the log likelihood of a path equals to P|p| t=1 log p(rt|q(t)) according to Eq. (7), we can maximize the probabilities of all the intermediate relations in a path. To achieve the goal, we decompose a path p = (r1, · · · , r|p|) into |p| + 1 (question, relation) instances, including ([q], r1), ([q; r1], r2), ..., ([q; r1; r2; · · · ; r|p|−1], r|p|), and an additional END instance ([q; r1; r2; · · · ; r|p|],END), and optimize the probability of each instance. We replace the observed relation at each time step with other sampled relations as the negative instances to optimize the probability of the observed ones.” Zhang teaches generating predicted ranking for knowledge graph node paths and identifies shortest paths to the correct answers as supervision signals. )
and fine-tuning the predictive machine learning model by generating one or more updated parameters … (Pg. 5777 of Zhang states “End-to-end training is an alternative to fine-tune the separately trained retriever and the reasoner jointly. The main idea is to leverage the feedback from the reasoner to guide the path expansion of the retriever. To enable this, we optimize the posterior pθ,ϕ(G|q, a) instead of the prior pθ(G|q), since the former one contains the additional likelihood pϕ(a|q, pk) which exactly reflects the feedback from the reasoner. We do not directly optimize the posterior pθ,ϕ(G|q, a), because G is induced from nK paths, making it unknown which path should receive the feedback from the likelihood computed on the whole G. Instead, we approximate p(G|q, a) by the sum of the probabilities of the nK paths and rewrite the posterior of each path by Bayes’ rule (Sachan et al., 2021), i.e.,
PNG
media_image2.png
78
215
media_image2.png
Greyscale
where pθ(pk|q) is the prior distribution of the kth path that can be estimated by Eq. (7), and pϕ(a|q, pk) is the likelihood of the answer a given the k-th path. Essentially, pϕ(a|q, pk) estimates the answer a on the single tree induced by the k-th path instead of the fused subgraph by nK paths. As a result, the reasoning likelihood on each tree can be reflected to the corresponding path that induces the tree. The reasoner for estimating pϕ(a|q, pk) is the same as that for calculating pϕ(a|q, G). In summary, the whole objective function for each training instance (q, a, G) is formalized as:
PNG
media_image3.png
96
228
media_image3.png
Greyscale
where the stop-gradient operation SG is to stop updating the parameters ϕ.” Zhang teaches the actual fine tuning operation and update of the knowledge graph path retriever parameters.)
validation context-relationship, validation node paths
one or more similarity scores by comparing the one or more validation… actual rankings for the one or more validation node paths… based on the one or more similarity scores.
Hiranandani teaches that
validation context-relationship, validation node paths ([0143] of Hiranandani states “After getting the ground truth ranking for each list by the above rank aggregation method, as a result, we had a total of Figure US20180121988A1-20180503-P00001 pairs of model comparisons. By performing 4:1:1 split on this dataset for training, validation and testing, we rank Rank-SVM on the training data. Validation data was used to achieve an optimal cost parameter required in the rank-SVM. And the test set was used to report the accuracy of the model.” [0144] of Hiranandani states “
PNG
media_image4.png
298
308
media_image4.png
Greyscale
”. Hiranandani teaches dividing ranking examples into training, validation, and testing sets then applying the trained ranking model to the validation data.)
one or more similarity scores by comparing the one or more validation… actual rankings for the one or more validation node paths… based on the one or more similarity scores. ([0141] of Hiranandani states “One can use any standard approach of Rank Aggregation to get the ground truth. Here, an “average the ranks and then rank the averages” method was used to get the ground truth ranking. In this algorithm, one determines average of the rankings of the models from different participants-annotated ranked lists and then rank the averages to get the final ground truth ranking.” [0143] of Hiranandani states “After getting the ground truth ranking for each list by the above rank aggregation method, as a result, we had a total of Figure US20180121988A1-20180503-P00001 pairs of model comparisons. By performing 4:1:1 split on this dataset for training, validation and testing, we rank Rank-SVM on the training data. Validation data was used to achieve an optimal cost parameter required in the rank-SVM. And the test set was used to report the accuracy of the model.” [0144] of Hiranandani states “
PNG
media_image4.png
298
308
media_image4.png
Greyscale
Finally, we determined w1=0.19, and w2=1.66, with cost=3, an accuracy of about 72.22% over validation set and an accuracy of 55.56% on the test set.”. Hiranandani teaches evaluating different parameter configurations using validation accuracy and using the validation results to determine the optimal Rank SVM cost and corresponding weights. Combining Hiranandani’s validation technique to Zhang, Zhang supplies the predicted rankings and node paths while Hiranandani supplies the comparison against actual rankings and the resulting validation accuracy score. The validation accuracy corresponds to the claimed similarity score. )
It would have been obvious to one with ordinary skill in the art before the effective filing date of the invention to combine the teachings from Hiranandani with Zhang and Lo. Zhang teaches ranking knowledge graph paths based on a query and forming a subgraph from the ranked paths. Lo teaches encoding query context and candidate paths using context fused entity embeddings and path representations. Hiranandani teaches validating a ranking model by comparing results on held out validation data with ground truth rankings and using the resulting validation accuracy to select an optimal model parameter. One with the ordinary skill in the art would have been motivated to incorporate the teachings of Hiranandani into the combination of Lo and Zhang to evaluate the model’s path ranking predictions on validation data and select the parameter configuration based on the validation results. It would have been predictable combination to improve prediction accuracy and reduce possible errors.
Claim 19 recites substantially similar subject matter to claim 10 respectively and is rejected with the same rationale, mutatis mutandis.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BYUNGKWON HAN whose telephone number is (571)272-5294. The examiner can normally be reached M-F: 9:00AM-6PM PST.
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, Li B Zhen can be reached at (571)272-3768. 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.
/BYUNGKWON HAN/ Examiner, Art Unit 2121
/Li B. Zhen/ Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights