DETAILED ACTION
This final Office action is responsive to Applicant’s amendment filed August 11, 2025. Claims 1, 13, 14, and 20 have been amended. Claims 1-20 are presented for examination.
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 .
Response to Arguments
Applicant's arguments filed August 11, 2025 have been fully considered but they are not persuasive.
Preliminarily, it is noted that the previously-pending claim objection of claim 13 and previously-pending rejection of claim 19 under 35 U.S.C. § 112(b) are withdrawn in response to Applicant’s amendments to claims 13 and 19.
Regarding the rejection under 35 U.S.C. § 101, Applicant argues the following:
…None of the claims can be practically performed in the human mind. Claim 1 requires a deep learning model with at least a thousand dimensions. Claim 1 further requires that the deep learning model provide word embedding for a concept, along with concept characteristics associated with the concept. Claim 1 requires that a centroid be determined in a high-dimensional word embeddings space. Euclidean distances and cosine similarities are also involved in determining quantifiable relationships between the centroid and the concept characteristics in the high-dimensional word embeddings space. A prominence index for each concept characteristic is then determined based on the Euclidean distances and cosine similarities. The prominence index is used for ranking the concept characteristics. (Page 10 of Applicant’s response)
The Examiner maintains that a human user could indeed perform calculations to determine Euclidean distances and cosine similarities. While the deep learning model including at least a thousand dimensions would not practically be performed in the human mind, the details of the deep learning model could incorporate mathematical concepts based on a broadest reasonable interpretation.
Applicant submits that the Examiner has not identified any relevant sub-categories for organizing human activity. The Examiner respectfully disagrees. As explained in the rejection, the evaluated process is related to “systems and methods for using high-dimensionality embedding centrality with prominence index for aligning and identifying qualities or characteristics associated with concepts” (Spec: ¶ 1), which (under its broadest reasonable interpretation) may include the concept being related to a job description including job titles and/or roles and skills and the process may be used to evaluate a candidate profile, as evidenced by claims 6-11, 13, and 18-20, thereby exemplifying the organization of human activity.
Applicant argues:
In contrast to the assertion in the Office Action, the claims are integrated into a practical application. The Specification at § [0003] describes a specific problem being solved in the art, i.e., the problem of explainability in machine learning systems. Given a deep learning model and an input to the model (i.e., a concept provided to the model), the claims provide a way of explaining what characteristics the deep learning model has associated with the concept. The claims provide a ranking of most relevant characteristics. Therefore, the claims transform a black box deep learning model into a model that is more explainable based on the ranking of characteristics associated with the concept… (Page 11 of Applicant’s response)
Ranking concept characteristics can be performed by a human and is part of the algorithm used to organize human activity (as explained above and in the rejection). The underlying algorithms of machine learning models incorporate mathematical concepts. The machine learning details are only generally implemented at a high level by generic processing elements. Performing a ranking with machine learning simply adjusts the underlying algorithms. The nature of how machine learning is performed is not an improvement from a technological perspective within the scope of the claims. At best, it might present an improvement in the algorithms used; however, the algorithms presented in the claims are directed to the details of the abstract ideas, which is insufficient to present a practical application or inventive concept under the § 101 guidelines.
Regarding the art rejections, Applicant argues that Inamdar does not address the claim limitations because “Inamdar does not solve a problem associated with explainability of machine learning models. Inamdar provides a candidate recommendation system and does not provide any insights associated to how a deep learning model organizes a concept.” (Page 13 of Applicant’s response) The claims do not present specific details regarding the “explainability of machine learning models” beyond ranking concepts based on relevance. As explained in the rejection and as seen in ¶¶ 17-18, 119 of Inamdar, Inamdar recognizes that the relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally. Machine learning models may initially be trained using non-industry relevant terms (Inamdar: ¶ 21) and then the machine learning model substitutes more appropriate terms as part of the training process (Inamdar: ¶¶ 42-43). Retraining is carried out continuously to improve performance and to incorporate more industry-relevant training data (Inamdar: ¶ 52). Model effectiveness is assessed for each model based on each model’s ranking effectiveness (Inamdar: ¶¶ 53-54). The most effective models (i.e., higher ranking models) are identified at a point when the machine learning model is deemed to have been sufficiently trained (Inamdar: ¶ 54). In other words, a trained model is deemed to be more relevant and to use the more relevant data (including evaluated words) for a given industry when sufficient accuracy is achieved. The best words relevant to a given industry reflect a highest ranking concept being the most relevant to the given industry. Identifying one iteration/version of a model being trained as being more effective than another iteration/version of the model is an example of implicitly ranking the various iterations/versions of the model to find a relatively highest ranking model and corresponding concepts that are most relevant in the industry at hand.
The rejection has also been revised to include the Van Hoang reference to help address the limitation regarding “each vector including at least a thousand dimensions.”
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 non-statutory subject matter.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The claimed invention is directed to “systems and methods for using high-dimensionality embedding centrality with prominence index for aligning and identifying qualities or characteristics associated with concepts” (Spec: ¶ 1) without significantly more.
Step
Analysis
1: Statutory Category?
Yes – The claims fall within at least one of the four categories of patent eligible subject matter. Process (claims 14-19), Apparatus (claims 1-13, 20)
2A – Prong 1: Judicial Exception Recited?
Yes – Aside from the additional elements identified in Step 2A – Prong 2 below, the claims recite:
[Claims 1, 14] A system/method for determining most relevant characteristics of a concept that performs the following:
provide a model;
receive a concept, the concept including text for probing the model;
determine a word embedding representing the concept using the model;
determine a centroid for the concept, based at least in part on concept characteristics associated with the concept, the centroid representing an average position of the concept characteristics in a high-dimensional word embeddings space;
determine Euclidean distances for each of the concept characteristics with respect to the centroid;
determine cosine similarities for each of the concept characteristics with respect to the centroid;
determine a prominence index for each of the concept characteristics based on the determined Euclidean distances and the determined cosine similarities; and
provide a ranking of the concept characteristics based on the prominence index, the ranking including a highest ranking concept characteristic being the most relevant concept characteristic.
[Claims 3, 16] receive at least one of the concept characteristics.
[Claims 4, 17] wherein a first concept characteristic and a second concept characteristic are received;
determine that the second concept characteristic is an outlier for the concept.
[Claim 5] wherein the concept is received.
[Claims 6, 18] wherein the concept is a job title or role, and wherein the concept characteristics include skills associated with the job title or role.
[Claims 7, 19] wherein the concept and at least one of the concept characteristics are received via a job posting.
[Claim 8] provide an identified concept characteristic absent from the job posting.
[Claim 9] provide an identified concept characteristic within the job posting that should be removed.
[Claims 10, 20] receive a candidate profile including one or more skills; and
determine a fitment score for the candidate profile based on the ranking of the concept characteristics.
[Claim 11] receive a candidate profile including one or more skills; and
determine a skill gap between the candidate profile and the job title or role.
[Claim 12] wherein the prominence indexes are determined based on harmonic mean of the normalized Euclidean distances and the cosine similarities.
[Claim 13] A system for optimizing skill fitment in job descriptions and that performs the following:
provide a model;
receive the job description, the job description including job titles and/or roles and skills;
determine word embeddings representing the job titles, the roles, and/or the skills using the model;
determine a centroid for each job title or role, based at least in part on the skills associated with the job titles or roles, the centroid representing an average position of skills in a high-dimensional word embeddings space;
determine Euclidean distances for each of the skills with respect to each respective centroid;
determine cosine similarities for each of the skills with respect to each respective centroid;
determine a prominence index for each of the skills based on respective Euclidean distances and respective cosine similarities; and
provide a ranking of the skills for each of the job titles or roles based on the prominence index, the ranking for each of the job titles or roles including a highest ranking skill being a most relevant skill.
Aside from the additional elements, the aforementioned claim details exemplify the abstract idea(s) of a mental process (since the details include concepts performed in the human mind, including an observation, evaluation, judgment, and/or opinion). As explained in MPEP § 2106.04(a)(1)(III), “[t]he courts consider a mental process (thinking) that ‘can be performed in the human mind, or by a human using a pen and paper’ to be an abstract idea. CyberSource Corp. v. Retail Decisions, Inc., 654 F.3d 1366, 1372, 99 USPQ2d 1690, 1695 (Fed. Cir. 2011).” The limitations reproduced above, as drafted, are a process that, under its broadest reasonable interpretation, covers performance of the limitations in the mind but for the recitation of generic computer components. That is, other than reciting the additional elements identified in Step 2A – Prong 2 below, nothing in the claim elements precludes the steps from practically being performed in the mind and/or by a human using a pen and paper. For example, but for the recitations of generic computer and other processing components (identified in Step 2A – Prong 2 below), the respectively recited steps/functions of the claims, as drafted and set forth above, are a process that, under its broadest reasonable interpretation, covers performance of the limitations in the mind and/or with the use of pen and paper. A human user could evaluate a model and received text to determine words representing a concept, determine a centroid for the concept, determine Euclidean distances for concept characteristics, determine cosine similarities for the concept characteristics, determine a prominence index, provide a ranking of the concept characteristics based on the prominence index, determine that a concept characteristic is an outlier, evaluate concepts related to job postings, etc. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind (and/or with pen and paper) but for the recitation of generic computer components, then it falls within the “Mental Processes” grouping of abstract ideas. Accordingly, the claims recite an abstract idea.
Aside from the additional elements, the aforementioned claim details exemplify a method of organizing human activity (since the details include examples of commercial or legal interactions, including advertising, marketing or sales activities or behaviors, and/or business relations and managing personal behavior or relationships or interactions between people, including social activities, teaching, and following rules or instructions). More specifically, the evaluated process is related to “systems and methods for using high-dimensionality embedding centrality with prominence index for aligning and identifying qualities or characteristics associated with concepts” (Spec: ¶ 1), which (under its broadest reasonable interpretation) may include the concept being related to a job description including job titles and/or roles and skills and the process may be used to evaluate a candidate profile, as evidenced by claims 6-11, 13, and 18-20 (i.e., organizing human activity); therefore, aside from the recitations of generic computer and other processing components (identified in Step 2A – Prong 2 below), the limitations identified in the more detailed claim listing above encompass the abstract idea of organizing human activity.
Providing a deep learning model representing words as vectors, each vector including at least a thousand dimensions; receiving a concept, the including text for probing the deep learning model; determining a centroid representing an average position of concept characteristics, determining Euclidean distances for each of the concept characteristics with respect to the centroid; determining cosine similarities for each of the concept characteristics with respect to the centroid; determining a prominence index for each of the concept characteristics based on the prominence index, the ranking including a highest ranking concept characteristic being the most relevant concept characteristic (claims 1, 13, 14) are examples of mathematical concepts. Claims 2 and 15 recite wherein the deep learning model includes GloVe, Word2Vec, or BERT and these models are understood to incorporate mathematical concepts.
2A – Prong 2: Integrated into a Practical Application?
No – The judicial exception(s) is/are not integrated into a practical application.
Claims 1-12 and 20 present a system comprising a processor and a non-transitory computer readable medium storing instruction such that when the instructions are executed by the processor, the system is configured to perform the recited operations. Claim 1 also recites that a ranking is provided to a client device.
Claim 3 receives at least one of the concept characteristics from the client device.
Claim 4 recites wherein a first concept characteristic and a second concept characteristic are received from the client device.
Claim 5 recites wherein the concept is received from the client device or a database.
Claim 8 provides an identified concept characteristic absent from the job posting to the client device.
Claim 9 provides an identified concept characteristic within the job posting that should be removed to the client device.
Claim 13 presents a system comprising a processor and a non-transitory computer readable medium storing instruction such that when the instructions are executed by the processor, the system is configured to perform the recited operations. Claim 13 also recites that a ranking is provided to a client device.
Claim 14 recites that a ranking is provided to a client device.
Claim 16 receives at least one of the concept characteristics from the client device.
Claim 17 recites wherein a first concept characteristic and a second concept characteristic are received from the client device.
The claims as a whole merely describe how to generally “apply” the abstract idea(s) in a computer environment. The claimed processing elements are recited at a high level of generality and are merely invoked as a tool to perform the abstract idea(s). Simply implementing the abstract idea(s) on a general-purpose processor is not a practical application of the abstract idea(s); Applicant’s specification discloses that the invention may be implemented using general-purpose processing elements and other generic components (Spec: ¶¶ 17-22, 67-70).
The use of a processor/processing elements (e.g., as recited in all of the claims) facilitates generic processor operations. The use of a memory or machine-readable media with executable instructions facilitates generic processor operations.
The additional elements are recited at a high-level of generality (i.e., as generic processing elements performing generic computer functions) such that the incorporation of the additional processing elements amounts to no more than mere instructions to apply the judicial exception(s) using generic computer components. There is no indication in the Specification that the steps/functions of the claims require any inventive programming or necessitate any specialized or other inventive computer components (i.e., the steps/functions of the claims may be implemented using capabilities of general-purpose computer components). Accordingly, the additional elements do not integrate the abstract ideas into a practical application because they do not impose any meaningful limits on practicing the abstract idea. The claims are directed to an abstract idea(s).
The processing components presented in the claims simply utilize the capabilities of a general-purpose computer and are, thus, merely tools to implement the abstract idea(s). As seen in MPEP § 2106.05(a)(I) and § 2106.05(f)(2), the court found that accelerating a process when the increased speed solely comes from the capabilities of a general-purpose computer is not sufficient to show an improvement in computer-functionality and it amounts to a mere invocation of computers or machinery as a tool to perform an existing process (see FairWarning IP, LLC v. Iatric Sys., 839 F.3d 1089, 1095, 120 USPQ2d 1293, 1296 (Fed. Cir. 2016)).
Regarding references to machine learning, claims 1, 13, and 14 recite providing a deep learning model representing words as vectors, each vector including at least a thousand dimensions; and receiving a concept, the including text for probing the deep learning model. Claims 2 and 15 recite wherein the deep learning model includes GloVe, Word2Vec, or BERT. Considering that the implementation of the deep learning model is performed using processing elements, such an implementation is presented as a generic recitation of machine learning in the claims and as a general link to technology. The machine learning-based processing elements are simply tools to generally automate the underlying process that could be performed by a human. It is further noted that, as described in Applicant’s Specification, the machine learning operations are generic machine learning operations (Spec: ¶ 30 – “At step 202, the server 102 provides a deep learning model (e.g., the language model 110). The language model 110 can be GloVe, Word2Vec, or BERT.” In other words, models commonly known in the art may be used.). The Specification presents no assertion that there is any improvement in the automated machine learning process itself. Such a generic recitation of machine learning, as recited in the claims, is little more than automating an analogous process that can be performed by a human.
There is no transformation or reduction of a particular article to a different state or thing recited in the claims.
Additionally, even when considering the operations of the additional elements as an ordered combination, the ordered combination does not amount to significantly more than what is present in the claims when each operation is considered separately.
2B: Claim(s) Provide(s) an Inventive Concept?
No – The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception(s). As discussed above with respect to integration of the abstract idea(s) into a practical application, the use of the additional elements to perform the steps identified in Step 2A – Prong 1 above amounts to no more than mere instructions to apply the exceptions using a generic computer component(s). Mere instructions to apply an exception using a generic computer component(s) cannot provide an inventive concept. The claims are not patent eligible.
As explained above, there is nothing in the claims as a whole that adds significantly more to the abstract idea(s).
Even if the use of GloVe, Word2Vec, and BERT models is seen as more than mathematical concepts generally implemented by generic processing elements, evidence showing that GloVe, Word2Vec, and BERT models are well-understood, routine, and conventional is provided below. In support of the Examiner’s assertion (above) that GloVe, Word2Vec, or BERT are known models, Shen et al. (Y. Shen and J. Liu, "Comparison of Text Sentiment Analysis based on Bert and Word2vec," 2021 IEEE 3rd International Conference on Frontiers Technology of Information and Computer (ICFTIC), Greenville, SC, USA, 2021, pp. 144-147, doi: 10.1109/ICFTIC54370.2021.9647258) states, “Most of the existing sentiment classification models use Word2Vec, GloVe, etc. to obtain the word vector representation of the text. But these methods ignore the context of words. In response to this problem, a neural network model based on the combination of BERT (bidirectional encoder representations from transformers) pre-trained language model and BLSTM (bidirectional long short-term memory network) and attention mechanism is proposed for text sentiment analysis in this paper.” (Shen: Abstract)
Similarly, Gandhi et al. (US 2019/0325029) states, “Using the word vectors, an embodiment of the method applies certain convolutions and data processing to determine a representation of a document (or documents) in d-dimensions (as described with reference to FIG. 3; note that one could use different semantic representations for words than the one described herein (co-occurrence based embeddings), such as word2vec, GLoVE, BERT, etc).” (Gandhi: ¶ 81)
Michelson et al. (US 2020/0402672) states, “A feature value of a word-embedding feature of an individual attribute value may specify a representation of the individual attribute value as determined through one or more word-embedding models. The word-embedding models may include language modeling and/or feature learning techniques where words or phrases may be mapped to vectors of real numbers. The word-embedding models may include one or more of Tomas Mikolov's Word2vec, Stanford University's GloVe, AllenNLP's Elmo, Google’s BERT, and/or other techniques.” (Michelson: ¶ 35)
Truong et al. (US 10,853,385) states, “Embedder 238 may include machine learning models configured to embed data. Embedder 238 may include algorithms to return one or more vector representations of showing relationships between raw data, including, for example, a word2vec method, a GloVe (Global Vector) method, a transformer method (e.g., a Bidirectional Encoder Representations from Transformer (BERT) method), an Embeddings from Language Models (ELMo) method, PCA, and/or any other method to obtain vector representation for data. More generally, embedding may include implementing algorithms (e.g., models) to transform data into an n-dimensional space, where the number of dimensions (n) may vary. Dimensions may be based on relationships among input data. Embedder 238 may include one or more embedding layers (i.e., one or more embedding algorithms that embed data in series and/or in parallel). An embedding network layer may include a natural language processing model, a binary classification model, a convolutional neural network model, a deep learning model, a transformer model (e.g., a BERT model), an ELMo representation model, and/or any other model configured to embed data.” (Truong: col. 10: 3-24)
Anisimov et al. (US 2020/0364539) states, “The present technology contemplates that is it possible to calibrate raw data for use by the neural networks of the training machine learning models such that it includes the word-level lexical information and/or image-level significance information into the training machine learning model by word-embedding and image embedding-systems. Word embeddings are tensor representations of words in a chosen text corpus, that translate semantic affinity of the words into a euclidean space. Word embeddings can be extracted with suitable natural language processing (NLP) solutions such as word2vec, GloVe and BERT that gained a massive popularity in recent years.” (Anisimov: ¶ 213)
Pablo et al. (US 2020/0356627) states, “FIG. 4 provides an illustration 310 that shows an example of extracting semantic data from a corpus of training materials 302. In this example the semantic information is extracted into a semantic embedding function 306. In machine learning semantic analysis is the task of building structures that approximate concepts from a large set of documents. Various semantic analysis tools are available in machine learning. One particular approach is to transform each word into a vector. Transforming the text used into a specific numerical vector may be accomplished by various methods. For example using probabilistic relationships between words in the documents, as is found in tools such as Word2vec by Tomas Mikolov at Google, “GloVe: Global Vectors for Word Representation” by Jeffrey Pennington, et al., etc. Examples of other tools that can create embedding functions include ELMo, ULMFit, BERT and others. The embedding functions these tools create allow the words that exist in the documents to be represented in a mathematical vector were each dimension of the vector has been found to correlate to some statistical relationship to other words in the corpus of documents, e.g. semantic vector space.” (Pablo: ¶ 49)
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, 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 1-7, 10-11, and 13-20 are rejected under 35 U.S.C. 103 as being unpatentable over Inamdar et al. (US 2021/0097472) in view of Van Hoang et al. (US 2018/0247271).
[Claim 1] Inamdar discloses a system for determining most relevant characteristics of a concept, the system comprising a processor and a non-transitory computer readable medium storing instructions such that when the instructions are executed by the processor (¶¶ 137-138 – processor, memory, software; ¶¶ 77-80 – Similarities between candidate documents and objective documents related to job descriptions are measured as scores.), the system is configured to:
provide a deep learning model (¶ 132 – “The machine learning model can include, for example, one or more of a decision tree, an artificial neural network, a support vector machine, a clustering process, a Bayesian network, a reinforcement learning model, naïve Bayes classification, a genetic algorithm, a rule-based model, a self-organized map, and an ensemble method, such as a random forest classifier or a gradient boosting decision tree.”), the deep learning model representing words as vectors (¶ 36 – “In another example, a word embedding approach, such as Word2Vec, or a document embedding approach, such as Doc2Vec can be used. In Word2Vec, a neural network with an input layer, in which each node represents a term, is trained on proximate word pairs within a document to provide a classifier that identifies words likely to appear in proximity to one another. The weights for the links between an input node representing a given word and the hidden layer can then be used to characterize the content of the document, including semantic and syntactic relatedness between the words. Document embedding is an extension of the word embedding approach. In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences.”);
receive a concept, the concept including text for probing the deep learning model (¶ 36 – “In another example, a word embedding approach, such as Word2Vec, or a document embedding approach, such as Doc2Vec can be used. In Word2Vec, a neural network with an input layer, in which each node represents a term, is trained on proximate word pairs within a document to provide a classifier that identifies words likely to appear in proximity to one another. The weights for the links between an input node representing a given word and the hidden layer can then be used to characterize the content of the document, including semantic and syntactic relatedness between the words. Document embedding is an extension of the word embedding approach. In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences.”; ¶¶ 17-18, 119 – Relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally.);
determine a word embedding representing the concept using the deep learning model (¶ 36 – “In another example, a word embedding approach, such as Word2Vec, or a document embedding approach, such as Doc2Vec can be used. In Word2Vec, a neural network with an input layer, in which each node represents a term, is trained on proximate word pairs within a document to provide a classifier that identifies words likely to appear in proximity to one another. The weights for the links between an input node representing a given word and the hidden layer can then be used to characterize the content of the document, including semantic and syntactic relatedness between the words. Document embedding is an extension of the word embedding approach. In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences.”);
determine a centroid for the concept, based at least in party on concept characteristics associated with the concept, the centroid representing an average position of concept characteristics in a high-dimensional word embeddings space (¶¶ 17-18, 119 – Relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally.; ¶ 36 – “In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences. In some examples, other approaches can be utilized and as such the above approaches are not exclusive.”; ¶ 47 – “In some examples, the ranking can be based on a weighted mean of the title and job parameters, and can be carried out via the sumquery command.”;
PNG
media_image1.png
164
266
media_image1.png
Greyscale
);
determine Euclidean distances for each of the concept characteristics with respect to the centroid (¶¶ 17-18, 119 – Relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally.; ¶ 36 – “In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences. In some examples, other approaches can be utilized and as such the above approaches are not exclusive.”; ¶ 47 – “In some examples, the ranking can be based on a weighted mean of the title and job parameters, and can be carried out via the sumquery command.”;
PNG
media_image1.png
164
266
media_image1.png
Greyscale
);
determine cosine similarities for each of the concept characteristics with respect to the centroid (¶¶ 17-18, 119 – Relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally.; ¶ 36 – “In document embedding, context from each paragraph (or appropriate text) is included as an input to the model, and link weights associated with these inputs is generated for each paragraph as part of the training process, representing the specific context of that paragraph. This can be used in combination with the training data associated with the individual words to generate a document vector for the document that captures embedding representations averaged across occurring words and word sequences. In some examples, other approaches can be utilized and as such the above approaches are not exclusive.”; ¶ 47 – “In some examples, the ranking can be based on a weighted mean of the title and job parameters, and can be carried out via the sumquery command.”;
PNG
media_image1.png
164
266
media_image1.png
Greyscale
);
determine a prominence index for each of the concept characteristics based on the determined Euclidean distances and the determined cosine similarities (¶¶ 125-130 – Candidate documents are ranked based on the distances of their corresponding candidate vectors to objective vectors corresponding to objective documents using one or more distance metrics, such as cosine similarity, Euclidean distance, etc.); and
provide, to a client device, a ranking of the concept characteristics based on the prominence index, the ranking including a highest ranked concept characteristic being the most relevant concept characteristic (¶¶ 17-18, 119 – Relevance of training data and models is dependent on the particular industry. Deep learning models are used and the best information for each respective industry is relied upon ideally. The most relevant data for each respective industry is used to determine the best rankings. Machine learning models may initially be trained using non-industry relevant terms (Inamdar: ¶ 21) and then the machine learning model substitutes more appropriate terms as part of the training process (Inamdar: ¶¶ 42-43). Retraining is carried out continuously to improve performance and to incorporate more industry-relevant training data (Inamdar: ¶ 52). Model effectiveness is assessed for each model based on each model’s ranking effectiveness (Inamdar: ¶¶ 53-54). The most effective models (i.e., higher ranking models) are identified at a point when the machine learning model is deemed to have been sufficiently trained (Inamdar: ¶ 54). In other words, a trained model is deemed to be more relevant and to use the more relevant data (including evaluated words) for a given industry when sufficient accuracy is achieved. The best words relevant to a given industry reflect a highest ranking concept being the most relevant to the given industry. Identifying one iteration/version of a model being trained as being more effective than another iteration/version of the model is an example of implicitly ranking the various iterations/versions of the model to find a relatively highest ranking model and corresponding concepts that are most relevant in the industry at hand.; ¶¶ 125-130 – Candidate documents are ranked based on the distances of their corresponding candidate vectors to objective vectors corresponding to objective documents using one or more distance metrics, such as cosine similarity, Euclidean distance, etc.; ¶¶ 20, 47 -- Job-related metrics, such as job titles are evaluated.).
Inamdar does not explicitly disclose “each vector including at least a thousand dimensions.” Like Inamdar, Van Hoang uses deep learning (such as Word2Vec) to vectorize job posting to identify relevant words and phrases related to job titles (Van Hoang: ¶ 50, 55-56, 94). Van Hoang explicitly states, “Although Word2Vec uses a training corpus to determine the word cosine distance between words, such a training corpus can be obtained from the hundreds of thousands of job postings of the job posting database 122.” (Van Hoang: ¶ 55) The Examiner submits that it would have been obvious to one of ordinary skill in the art before the effective filing date of Applicant’s invention to modify Inamdar such that each vector includes at least a thousand dimensions so that a larger collection of data may be used to glean a good sample size of truly representative data for jobs and relevant skills in each respective industry.
[Claim 2] Inamdar discloses wherein the deep learning model includes GloVe, Word2Vec, or BERT (¶ 36 – “In another example, a word embedding approach, such as Word2Vec, or a document embedding approach, such as Doc2Vec can be used. In Word2Vec, a neural network with an input layer, in which each node represents a term, is trained on proximate word pairs within a document to provide a classifier that identifies words likely to appear in proximity to one another.”).
[Claim 3] Inamdar discloses that the system is further configured to receive at least one of the concept characteristics from the client device (¶ 33 – “In further examples, the search engine server 106 can be programmed to employ an indexer 120. The indexer 120 can be programmed to read and index the candidate data to provide a candidate index. Thus, the indexer 120 can be programmed to read and index the resume data 112 and the job description data 114 to provide a resume index 122 and a job description index 124. The resume data 112 and the job description data 114 may be indexed based on indexing schema data (not shown in FIG. 1) that can be user definable (e.g., via the input device 116). In some examples, the resume index 122 and the job description index 124 can be stored local to the candidate recommendation system 102 while, in other examples, may be stored at a remote location (e.g., a remote database). In further examples, the indexer 120 can be programmed to read and index other types of candidate data, such as the employee data and the project data and generate corresponding indexes (not shown in FIG. 1) to enable searching and generating of a ranked candidate list, such as the ranked job list, the ranked project list or the ranked employee list.”; ¶ 30 – “The input device 116 can be any type of device capable of supporting a communications interface to the candidate recommendation system 102. Exemplary input devices 116 can include a server, a mobile device, a mobile computer, a tablet, etc. The input device 116 can be connected to the candidate recommendation system 102 using a provided network (e.g., via common internet protocols), such as a wired or wireless network. Example networks can include an Internet, an intranet, a WiFi network, a WiMAX network, a mobile telephone network, and combinations thereof. The input device 116 can be configured to enable a user to interact with the candidate recommendation system 102 via a local interface (e.g., a web browser, software application, etc.) to execute one or more searches for relevant candidate information (e.g., a list of candidates).”).
[Claim 4] Inamdar discloses wherein a first concept characteristic and a second concept characteristic are received from the client device (¶ 33 – “In further examples, the search engine server 106 can be programmed to employ an indexer 120. The indexer 120 can be programmed to read and index the candidate data to provide a candidate index. Thus, the indexer 120 can be programmed to read and index the resume data 112 and the job description data 114 to provide a resume index 122 and a job description index 124. The resume data 112 and the job description data 114 may be indexed based on indexing schema data (not shown in FIG. 1) that can be user definable (e.g., via the input device 116). In some examples, the resume index 122 and the job description index 124 can be stored local to the candidate recommendation system 102 while, in other examples, may be stored at a remote location (e.g., a remote database). In further examples, the indexer 120 can be programmed to read and index other types of candidate data, such as the employee data and the project data and generate corresponding indexes (not shown in FIG. 1) to enable searching and generating of a ranked candidate list, such as the ranked job list, the ranked project list or the ranked employee list.”; ¶ 30 – “The input device 116 can be any type of device capable of supporting a communications interface to the candidate recommendation system 102. Exemplary input devices 116 can include a server, a mobile device, a mobile computer, a tablet, etc. The input device 116 can be connected to the candidate recommendation system 102 using a provided network (e.g., via common internet protocols), such as a wired or wireless network. Example networks can include an Internet, an intranet, a WiFi network, a WiMAX network, a mobile telephone network, and combinations thereof. The input device 116 can be configured to enable a user to interact with the candidate recommendation system 102 via a local interface (e.g., a web browser, software application, etc.) to execute one or more searches for relevant candidate information (e.g., a list of candidates).”), the system being further configured to: determine that the second concept characteristic is an outlier for the concept (¶ 100 – “ In some examples, the set of best-ranked resumes can include between fifty and one-hundred resumes. In these examples, the feature generator 231 can be programmed to determine, for each candidate document, a feature vector representing the candidate document and the objective document. Each feature in the feature vector can be generated or derived from the candidate resume data 214 and the corresponding data for the job description.” In other words, resumes that do not fall into the set of best-ranked resumes are examples of outliers.).
[Claim 5] Inamdar discloses wherein the concept is received from the client device or a database (¶ 33 – “In further examples, the search engine server 106 can be programmed to employ an indexer 120. The indexer 120 can be programmed to read and index the candidate data to provide a candidate index. Thus, the indexer 120 can be programmed to read and index the resume data 112 and the job description data 114 to provide a resume index 122 and a job description index 124. The resume data 112 and the job description data 114 may be indexed based on indexing schema data (not shown in FIG. 1) that can be user definable (e.g., via the input device 116). In some examples, the resume index 122 and the job description index 124 can be stored local to the candidate recommendation system 102 while, in other examples, may be stored at a remote location (e.g., a remote database). In further examples, the indexer 120 can be programmed to read and index other types of candidate data, such as the employee data and the project data and generate corresponding indexes (not shown in FIG. 1) to enable searching and generating of a ranked candidate list, such as the ranked job list, the ranked project list or the ranked employee list.”; ¶ 30 – “The input device 116 can be any type of device capable of supporting a communications interface to the candidate recommendation system 102. Exemplary input devices 116 can include a server, a mobile device, a mobile computer, a tablet, etc. The input device 116 can be connected to the candidate recommendation system 102 using a provided network (e.g., via common internet protocols), such as a wired or wireless network. Example networks can include an Internet, an intranet, a WiFi network, a WiMAX network, a mobile telephone network, and combinations thereof. The input device 116 can be configured to enable a user to interact with the candidate recommendation system 102 via a local interface (e.g., a web browser, software application, etc.) to execute one or more searches for relevant candidate information (e.g., a list of candidates).”).
[Claim 6] Inamdar discloses wherein the concept is a job title or role, and wherein the concept characteristics include skills associated with the job title or role (¶ 43 – “The coarse search query parser 128 may be programmed to feed the obtained words into the baseline ML model 130 to generate the candidate search parameter vector (e.g., a fixed-length numerical vector) to represent at least one candidate search parameter (e.g., the job description information and the job title information).”; ¶ 44 – “In further examples, the coarse search query parser 128 may be programmed to compare each candidate search parameter vector and each candidate vector to determine whether if any document (e.g., resume) includes one or more words that match one or more words of the search query request (e.g., the job description information or the job title information).”; ¶ 83 – “Thus, the requisition code can have a defined file format and can include (or be representative of) a job description and a job title. In additional or alternative examples, the candidate recommendation system 202 can be programmed to receive resume code as the search query request. The resume code can include the job experiences, education, skills, accomplishments, etc.”).
[Claim 7] Inamdar discloses wherein the concept and at least one of the concept characteristics are received via a job posting (¶ 16 – “Candidates are matched to “objectives”, which can be a job, position, project, or contract to which an individual or corporate entity is matched, or an individual or corporate entity to which an appropriate job, contract, or position is matched. In some implementations, the candidate recommendation system will be designed and trained to match a specific type of candidates (e.g., job applications) to a specific type of objective (e.g., job openings). In this implementation, information about the objective may be provided to the system when a candidate ranking is desired, although it will be appreciated that the information could instead be loaded and indexed prior to a candidate ranking request. In another implementation, a single system contains candidates of varying types (e.g., job candidates and job openings). In such a system, the objective is the candidate for which a query is submitted, and the objective will be matched to candidates of a different type.”).
[Claim 10] Inamdar discloses that the system is further configured to receive a candid