DETAILED ACTION
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 Amendment
(Submitted on 12/4/2025)
The applicant has amended the claims 1, 6, 9 and 15 in which the claim 1, and claim 6 is amended from features from claim 2 and claim 7, claim 9 is amended with features of claim 10 and claim 15 is amended with features of claim 16. The claims 2,7, 10 and 16 are CANCELED. The applicant has provided the following arguments and the examiner has provided the responses to each argument.
In regard to 101 rejections:
- On Pages 11-12, the applicant argues that that amended claim 1 should be eligible under Steps 2A and 2B, as the adding specific limitations can provide improvements within the context of beyond conventional, routine as the claim provides a specific process carried out to training a tag recommendation model. As described in the present application, the claimed process of amended independent claim 1 provides the following beneficial effects of “ ERNIE is used to represent the training materials semantically, which can make the representations of features of the training materials more accurate. By training the double-layer neural network structure, the coverage of the materials is increased, thereby improving the accuracy of the obtained interest tags" and "By splicing the training behavior vectors and the training service vectors in some embodiments of the disclosure, final training semantic vectors of fixed or reasonable length are obtained, which is beneficial to improve the generalization capability of the neural network model." The applicant argues that therefore the claim language from amended independent claim 1 demonstrates that the presently claimed subject matter is directed to a technical problem that is unique to training a tag recommendation model. The specific steps of the claimed process solve such unique technical problems and are not well understood, routine, or conventional.
Examiner’s Response:
Training a tag recommendation model is not inherently unique or unconventional. The key
is to demonstrate the novelty in the technical problem solving (how the model performs).
That is to say simply stating that the ML performs a task faster, efficiently or more
accurately is not sufficient to overcome 101 rejections. The applicant need to demonstrate
that the claim has captured the result and provide overall technical impact on the
recommendation system. The applicant need to be more specific in terms of overall
improvement to the computer technology so as for any POSITA to understand to the
overall computer improvement for practical applications. As such, the claim merely
suggesting an improvement to the model itself which is generic computer function. The
examiner submits same arguments to claims 9 and 15.
- On Page 12, the applicant argues that that amended claim 6 should be eligible under Steps 2A and 2B, as the adding specific limitations can provide improvements within the context of beyond conventional, routine as the claim 6 provides a process “ Through the method for obtaining a tag provided by the disclosure, the users interest tags can be accurately obtained, so that relevant materials can be recommended accurately" and that “ the tag vectors are parsed by using sigmoid as the activation function and that the interest tags corresponding to the features are obtained from the features in the tag vectors, and the interest tags of the user are determined from the obtained interest tags.". The applicant argument is claim 6 demonstrates that the presently claimed subject matter is directed to a technical problem that is unique to training a tag recommendation model. The specific steps of the claimed process solve such unique technical problems and are not well understood, routine, or conventional.
Examiner’s Response:
As noted before, Training a tag recommendation model is not inherently unique or unconventional. In regard to applicant’s argument specific to parsing of tag vectors, the parsing process is very conventional and routine perhaps well known to a POSITA that involve tokenization which is well understood in the NLP applications and the examiner submits that might be again just improvement to the model itself. That is to say simply stating that the ML performs a task faster, efficiently or more accurately is not sufficient to overcome 101 rejections. The applicant need to demonstrate that the claim has captured the result and provide overall technical impact on the recommendation system. The applicant need to be more specific in terms of overall improvement to the computer technology so as any POSITA to understand the overall computer improvement for practical applications. Thus, as such, the claim 6 merely suggesting an improvement to the model itself which is generic computer function.
In Conclusion, the amendment to claims 1, 6, 9 and 15 has not supported to overcome 101 rejections and the examiner reaffirms the 101 rejections on claims 1, 3-6, 8-9, 11-15 and 17-20. Claims 2, 7 , 10 and 16 have been CANCELED by the applicant.
In regard to 103 rejections:
- On Page 16 -17, the applicant summarizes the claim 1 as amended six limitations designated as features AJ, BJ, CJ, DJ, EJ, and FJ.
The examiner notes that limitations [Feature EJ] were drawn from claim 2 under FJ with respect to vector lengths and averaging are drawn from claims 2 (claims 10 and 16).
The examiner submits that the primary reference “Weiss” teaches the “fully connected neural network” as a “autoencoder” in claims 4, 6, 12 and 18.
Applicant Argument #1 (Page 17):
- The applicant argues with reference to features A and B that Weiss fails to explicitly disclose " in response to receiving an instruction," and Wang pertains to knowledge data collection and not to training materials for tag recommendation.
Examiner Response
The examiner “totally and respectfully disagrees” with the argument. The examiner notes that the applicant refers to WANG teaching of [0084], [0085], [0007], [0141], [0142] and [0124] for the arguments. The examiner first interprets that training materials for tag recommendation include data set that includes resources and user profiles with tag to build relationship between the tage and the resources. Reference WANG discloses in [0093] “ In the natural language processing system shown in FIG. 1, the user equipment may receive an instruction of the user. For example, the user equipment may receive a piece of text entered by the user, and then initiate a request to the data processing device, so that the data processing device executes a natural language processing application (for example, text classification, text inference, named entity recognition, or translation) on the piece of text obtained by the user equipment, to obtain a processing result (for example, a classification result, an inference result, a named entity recognition result, or a translation result) of a corresponding natural language processing application for the piece of text. For example, the user equipment may receive a piece of Chinese text entered by the user, and then initiate a request to the data processing device, so that the data processing device performs entity classification on the piece of Chinese text, to obtain an entity classification result for the piece of Chinese text. For example, the user equipment may receive a piece of Chinese text entered by the user, and then initiate a request to the data processing device, so that the data processing device translates the piece of Chinese text into English, to obtain an English translation for the piece of Chinese text” and further discloses in
[0018]” In a target processing model training process, the parameters of the original processing model and/or the parameters of the original fusion model are adjusted based on the first knowledge data and the training text, to obtain the target processing model and/or the target fusion model. This improves a capability of understanding natural language by the target processing model and/or the target fusion model, and improves accuracy of the processing result of the target processing model”.
Applicant Argument #2 (Page 18):
- The applicant argues on feature C that CAO does not disclose the semantic vector training and that CAO's vectors are independent and are not integrated with any semantic vector. Further, the applicant argues that in contrast, amended claim 1 specifies aggregating social network information into a pre-existing "training semantic vector" derived from a previous step. This step is entirely absent in CAO. The applicant further argues that CAO performs "feature creation," not the "aggregating into" as recited in amended claim 1. As shown in paragraph [0085] of CAO, the system "creates an overall feature matrix" by combining path counts from meta-paths and uses it as the training feature vector. The applicant refer to [0084] and [0085] for the arguments.
Examiner Response
The examiner “totally and respectfully disagrees”. As shown in paragraph [0085] of CAO, the system "creates an overall feature matrix" by combining path counts from meta-paths and uses it as the training feature vector”. The reference CAO teaches in [0092]” As used herein, a meta-path corresponds to a type of path within the network schema, containing a certain sequence of link types. For example, in FIG. 7, a meta-path
PNG
media_image1.png
50
337
media_image1.png
Greyscale
denotes a composite relationship from tweets to venues. The semantic meaning of this meta-path is that the tweet and the venue share common words via tips. The link type “contain.sup.−1” represents the inverted relation of “contain”. The tweet and venues connected through the meta-path can be regarded as being more likely to be linked than those without such correlations”. Further CAO teaches in [0093] “ Different meta-paths usually represent different relationships among linked nodes with different semantic meanings. For example, the meta-path
PNG
media_image2.png
58
398
media_image2.png
Greyscale
denotes that the tweet was posted by a Twitter user who is a mayor of the venue in Foursquare, while the meta-path
PNG
media_image3.png
68
376
media_image3.png
Greyscale
indicates the tweet was posted by a Twitter user whose friend checks in at the venue. In this way, relationships between tweets and venues can be described by different meta-paths with different semantics. Thus, four types of meta-paths as shown in FIG. 7 are extracted and summarized in FIG. 8. Further, the examiner submits that Cao teaches in [0015] “ In some implementations, identifying for the new social message corresponding meta-paths to the particular venue includes: obtaining a social graph as a social network schema based on types of entities and relationships extracted from a collection of messages and the collection of venues, wherein each type of entities is represented as a type of node in the social network schema and the relationships between the entities are represented as different types of links; and based on the social graph, content of the new social message and/or a user writing the new social message and/or social friends of the user, identifying for the new social message corresponding meta-paths connecting the new social message to the particular venue, wherein each of the corresponding meta-paths represents a type of path within the social network, containing a certain sequence of link types.
Applicant Argument #3 (Page 18):
- The applicant argues on 1.2 Feature C that CAO does not teach the “aggregation into a training semantic vector” as it does not teach the key input of “training semantic vector” as defined in claim 1 and the reference teaches directly “compute features” based on raw “social message” and “meta-paths” and the feature generation process is performed independently and does not take any form of “training semantic vector” as input. The applicant has made references CAO’s [0084], [0085] for the arguments. The applicant argues that none of the references make up for this deficiency.
Examiner’s Response
The examiner “totally and respectfully disagree” with the argument. WANG teaches in [0084]: "Using the plurality of social message and venue pairs, the server 104 computes features based on meta-paths and geo-coordinate information” and in [0085] teaches "the path counts for different meta-paths are combined to create an overall feature matrix and the overall feature matrix is represented as the training feature vector.". The examiner teaching in [0085]: "first encoding the respective training social message in the pair as a label". The reference CAO teaches in [0092]” As used herein, a meta-path corresponds to a type of path within the network schema, containing a certain sequence of link types. For example, in FIG. 7, a meta-path
PNG
media_image1.png
50
337
media_image1.png
Greyscale
denotes a composite relationship from tweets to venues. The semantic meaning of this meta-path is that the tweet and the venue share common words via tips. The link type “contain.sup.−1” represents the inverted relation of “contain”. The tweet and venues connected through the meta-path can be regarded as being more likely to be linked than those without such correlations”. Further CAO teaches in [0093] “ Different meta-paths usually represent different relationships among linked nodes with different semantic meanings. For example, the meta-path
PNG
media_image2.png
58
398
media_image2.png
Greyscale
denotes that the tweet was posted by a Twitter user who is a mayor of the venue in Foursquare, while the meta-path
PNG
media_image3.png
68
376
media_image3.png
Greyscale
indicates the tweet was posted by a Twitter user whose friend checks in at the venue. In this way, relationships between tweets and venues can be described by different meta-paths with different semantics. Thus, four types of meta-paths as shown in FIG. 7 are extracted and summarized in FIG. 8. Further, the examiner submits that Cao teaches in [0015] “ In some implementations, identifying for the new social message corresponding meta-paths to the particular venue includes: obtaining a social graph as a social network schema based on types of entities and relationships extracted from a collection of messages and the collection of venues, wherein each type of entities is represented as a type of node in the social network schema and the relationships between the entities are represented as different types of links; and based on the social graph, content of the new social message and/or a user writing the new social message and/or social friends of the user, identifying for the new social message corresponding meta-paths connecting the new social message to the particular venue, wherein each of the corresponding meta-paths represents a type of path within the social network, containing a certain sequence of link types.
Applicant Argument #4(Page 19):
- The applicant argues that the Semantic- entities and relationships essentially involves extracting and constructing a new set of features from social data. This fundamentally differs in technical means and concept from the "aggregation" operation recited in amended claim 1, which involves injecting or fusing social network information into an existing semantic vector carrier. Therefore, Feature C cannot be reached by combining Weiss, Wang, and CAO.
Examiner’s Response
The examiner “totally and respectfully disagrees” on the combination argument. First, the examiner interprets that fusing the social network involves integrating user attributes, social relationships as a vector representation.
Wiess in [0007] teaches “ the present invention includes systems, methods, circuits, and associated computer executable code for deep learning based natural language understanding, wherein: (1) a word tokenization and spelling correction model/machine may generate corrected word sets outputs based on respective character strings inputs; and/or (2) a word semantics derivation model/machine may generate semantically tagged sentences outputs based on respective word sets inputs” . The reference CAO teaches in [0092]” As used herein, a meta-path corresponds to a type of path within the network schema, containing a certain sequence of link types. For example, in FIG. 7, a meta-path
PNG
media_image1.png
50
337
media_image1.png
Greyscale
denotes a composite relationship from tweets to venues. The semantic meaning of this meta-path is that the tweet and the venue share common words via tips. The link type “contain.sup.−1” represents the inverted relation of “contain”. The tweet and venues connected through the meta-path can be regarded as being more likely to be linked than those without such correlations”. Perhaps it may be known to the POSITA that the meta-path that provides a composite relationship can represent social network information into a semantic vector carrier where meta-path capture semantic relationships. Further WANG discloses in [0007] “ using the target text vector and the target knowledge vector based on a target fusion model, to obtain a fused target text vector and a fused target knowledge vector; and processing the fused target text vector and/or the fused target knowledge vector based on a target processing model, to obtain a processing result corresponding to a target task” and teaches in [0008]” uses the obtained fused target text vector and/or the obtained fused target knowledge vector as input data for the target processing model”.
Applicant Argument #5 (Page 20):
- The applicant argues on feature D that reference YE teaches 2-layer NN for service evaluation and are based on quality data and not interest tags. The applicant refers to [0118] and [0027-0029] of YE. Further, the applicant argues YE teaches a completely different problem of monitoring communication networks and does not teach “tag recommendation model “ of claim which the process examines aspect of tag and help understand user interests and develop algorithms based on the user’s tagging behavior.
Examiner’s Response
The examiner “totally and respectfully disagrees” with the argument. The examiner first submits that perhaps known to the POSITA that attention mechanism in NLP significantly enhance the ability to simulate biological behavior by allowing models to focus on relevant parts of the input data. They enable dynamic weighting of input tokens based on their relevance to the current task, overcoming the limitations of fixed-size context vectors in models . This capability mirrors human cognitive processes, as it allows models to prioritize information that is most informative for the current output step, thus improving performance in tasks like translation and summarization. The examiner submits that WANG [0115] “ The attention mechanism simulates an internal process of biological observation behavior, and is a mechanism that aligns internal experience with external feeling to increase observation precision of some regions. The mechanism can quickly select high-value information from a large amount of information by using limited attention resources. The attention mechanism is widely used in natural language processing tasks, especially machine translation, because the attention mechanism can quickly extract an important feature of sparse data. A self-attention mechanism (self-attention mechanism) is an improvement of the attention mechanism. The self-attention mechanism reduces dependence on external information and is better at capturing an internal correlation of data or features.
Applicant Argument #6 (Page 21):
- The applicant argues that reference “YE” within the context of double layer neural network does not disclose a tag recommendation model as it is used for service quality evaluation and argues that it is teaching a completely different problem. The applicant made references to YE teachings in [0118], [0027]-[0029], [0049]-[0053] in support of the arguments.
Examiner’s Response
The examiner “totally and respectfully “disagrees” with the argument. Reference YE discloses in [0027] “determining the tag based on the quality monitoring “ and further discloses in [0028]” determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable” and further in [0029] discloses ” calculating a value of the evaluation indicator using the quality monitoring data, and determining the value of the evaluation indicator as a tag” with further teaching in [0053]” a training module, configured to train a deep neural network model using the training set to obtain a service quality evaluation model”. In this context, the determination of the tag on quality monitoring relates to recommendation model with evaluation indicator as “interest tag”.
Applicant Argument #7 (Page 21):
- The applicant argues that with respect to features E and F, reference Weiss does not disclose “ representing the behavior training materials as training behavior vectors of “different length”. The specific context of argument is “vector” with specific references to [0074]-[0083]. Specifically the argument is that Weiss vectors and is limited to character sequences and structural annotations with no mention of vector representations.
Examiner’s Response:
The examiner totally “disagrees” with the argument. First, the examiner interprets that the training behavior vectors of different involves handling sequences that vary in length due to variability in the number of tokens or characters in the input data. Reference Weiss teaches in [0039]”According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams” and “ An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications”. Further reference Weiss discloses in [0047] According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input.
Applicant Argument #8 (Page 21):
- The applicant argues that YE’s "tag" and the "interest tag" in amended claim 1 are different in nature and function suggesting that "tag" in YE is an "evaluation indicator value" calculated from service quality monitoring data (as described in paragraph [0029], serving as a continuous, quantitative regression target. In contrast, the "interest tag" in amended claim 1 is a discrete, categorical recommendation target. Therefore, feature D cannot be reached by combing Weiss, WANG, CAO and YE.
Examiner’s Response
The examiner “totally and respectfully “disagrees” with the argument. The examiner first interprets that “ an interest tag” is a tool to capture the user needs and interest to help build a deep relationship and personal interactions by categorizing the tags. Reference YE discloses in [0094]” The embodiments of the present disclosure provide a training method for service quality evaluation models. The method may be applicable to the network frame illustrated in FIG. 1. The network frame may include service nodes, monitoring nodes, and a model training node. The service node may be a node in a CDN service system that provides services to users” and further discloses in [0110] “ The tag may intuitively reflect the service quality, and for different service types, the indicators used to evaluate service quality may be different. According to the service type to which the service quality evaluation model is applicable, a corresponding evaluation indicator may be used to determine the tag. The step of determining the tag based on the quality monitoring data may include: determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable,
Applicant Argument #9 (Page 22):
- The applicant argues that reference WANG does not disclose “ representing the behavior training of fixed length training service vector”. The applicant argues that attributes may be of fixed-length but it does teach representing “service training materials “ as vectors. The specific context of argument is “vector” with specific references to [0141] and [0145]. Once again, the argument says the WANG focus on knowledge representation and not on fusing the behavior vectors.
Examiner’s Response:
The examiner totally “disagrees” with the argument. The examiner refers to the same counter argument as before in regard to knowledge representation. Perhaps known to the POSITA, the examiner first interprets that the behavior service training service is providing support for users to improve their behavior. Perhaps known to the POSTA, the examiner first interprets that representation of behavior training in fixed-length involves converting a converting a complex data into a sequence of numbers or a feature vector. WANG discloses in [0135]” the LM may also be understood as a probability model used to calculate a probability of a sentence. In other words, the language model is a probability distribution of a natural language text sequence, and the probability distribution represents a possibility of existence of text with a specific sequence and a specific length. In short, the language model predicts a next word based on a context. Because there is no need to manually tag a corpus, the language model can learn rich semantic knowledge from an unlimited large-scale corpus. WANG discloses in [0122] “ By using NLP and components of the NLP, a very large amount of text data can be managed or a lot of automated tasks can be performed, and various problems can be resolved, such as automatic summarization (automatic summarization), machine translation (machine translation, MT), named entity recognition (named entity recognition, NER), relation extraction (relation extraction, RE), information extraction (information extraction, IE), sentiment analysis, speech recognition (speech recognition), question answering (question answering), and topic segmentation”. A sentiment analysis is a “behavior analysis”.
Applicant Argument #10 (Page 22):
- The applicant argues that reference WANG does not disclose “averaging” or “fusing operation” and therefore does not disclose “ obtaining the training semantic vectors by averaging the training behavior vectors that are averaged with the training service vectors” as required by amended claim 1.
Examiner’s Response
The examiner “ totally “disagrees” with the argument. The reference WANG teaches in [0008] “According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data” and discloses in [0050]” process the training text to obtain a first text vector; fuse the first text vector and the first knowledge vector based on an original fusion model” and” adjust parameters of the original fusion model based on the first task result and the second task result, to obtain the target fusion model” and discloses in [0112] “ A recurrent neural network (recurrent neural network, RNN) is used to process sequence data” and discloses in [210] “ optionally, if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector” and further discloses in [0298]” The knowledge aggregator #5 may have a complex multilayer network structure, for example, a multilayer self-attention mechanism network structure, a multilayer perceptron network structure, or a recurrent neural network structure, and may simply weight and average the encoded text sequence and the encoded knowledge sequence.
Applicant Argument #11 (Page 23):
- The applicant argues with reference to Claim 6 that the prior art references do not teach parsing of the tage vector and hence the references cannot be combined.
Examiner’s Response
The examiner “totally and respectfully” disagrees with the arguments. The applicant argument is with respect to the amended claim 6. The parsing is taught by reference “Kasai” as it is added in the new grounds of rejection of “Weiss” WANG, CAO and “Kasai”.
In Conclusion, the examiner’s rebuttal for the applicant’s the argument of the applicant of claim 1 is same as arguments for claims 9 and 15 and the examiner reaffirms the 103 rejections on claims 1, 3-6, 8-9, 11-15 and 17-20 and MOVE the application to FINAL REJECTION under 103. Claims 2, 7 , 10 and 16 have been CANCELED by the applicant.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of
matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the
conditions and requirements of this title.
Claims 1, 3-6, 8-9, 11-15 and 17-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1:
According to the first part of the analysis, in the instant case, claim 1 and claim 6 is directed to a method claim, claim 9 is directed to an electronic device comprising a processor and a memory, and claim 15 is directed a storage to execute the storage contents. Thus, each of the claims falls within one of the four statutory categories (i.e. process, machine, manufacture, or composition of matter).
In regard to claim 1. (Currently Amended)
Step 2A Prong 1:
“obtaining training semantic vectors that comprise the interest tags by representing
features of the training materials using a semantic enhanced representation frame;” is a mental
step of vector data representation.
“obtaining training encoding vectors by aggregating social networks into the training
semantic vectors;” a mental step of vector data aggregation.
“ representing the behavior training materials as training behavior vectors of different
lengths” a mental step of vector data representation.
“and representing the service training materials as fixed-length training service vectors
in the semantic enhanced representation frame” is a mental step of vector data representation.
“ and obtain the training semantic vectors by averaging the training behavior vectors and fusing
the training behavior vectors that are averaged with the training service vectors” is a mental
step of vector data representation.
Additional Elements
Step 2A Prong 2:
“ A method for training a tag recommendation model, comprising:” recited in the preamble do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h.
“ collecting training materials that comprise interest tags “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ A method for training a tag recommendation model, comprising:” recited in the preamble do is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ collecting training materials that comprise interest tags “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim.
These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining training encoding vectors by aggregating social networks into the training semantic vectors;” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 3: (Original)
Step 2A Prong 1:
“ wherein obtaining the training encoding vectors by aggregating the social networks
into the training semantic vectors, comprises:” “ is a mental step of data aggregation.
“ determining intimacy values between any two of the social networks; “is a mental step of data
comparison.
“ the intimacy values as values of elements in a matrix, “ is a mental step of data comparison.
“and generating an adjacency matrix based on the values of the elements; in response to that a
sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the
elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency
matrix is greater than weights assigned to other elements; mental step of matrix data
representation.
“ and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.”
is a mental step of matrix data representation.
Step 2A Prong 2: no additional elements.
Step 2B: no additional elements.
In regard to claim 4: (Original)
Step 2A Prong 2:
“ wherein obtaining the tag recommendation model by training the double- layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtaining new training encoding vectors by inputting the training encoding vectors into a forward network; “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein obtaining the tag recommendation model by training the double- layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises: “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ obtaining new training encoding vectors by inputting the training encoding vectors into a forward network; “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 5: (Original)
Step 2A Prong 1:
“ and determining first interest tags corresponding to the interest tags in the training tag
vectors, calculating a ratio of the first interest tags to the interest tags,” is a mental step of data
identification.
Additional Elements
Step 2A Prong 2:
“ wherein obtaining the tag recommendation model by determining the training tag
vectors as the independent variables, and the outputs as the interest tags, comprises: “
do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein obtaining the tag recommendation model by determining the training tag
vectors as the independent variables, and the outputs as the interest tags, comprises: “
is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
do not integrate the judicial exception into a practical application. These additional elements
“ determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 6: (Currently Amended)
Step 2A Prong 1:
“ obtaining training semantic vectors that comprise the interest tags by representing
features of the training materials using a semantic enhanced representation frame;” is a mental
step of vector data representation.
“obtaining training encoding vectors by aggregating social networks into the training
semantic vectors;” a mental step of vector data aggregation.
“ and obtaining the training semantic vectors that comprise the interest tags by representing the
features of the training materials using the semantic enhanced representation frame” mental
step of data identification.
“ representing the behavior training materials as training behavior vectors of different
lengths” a mental step of vector data representation.
“and representing the service training materials as fixed-length training service vectors
in the semantic enhanced representation frame” is a mental step of vector data representation.
“ and obtain the training semantic vectors by averaging the training behavior vectors and fusing
the training behavior vectors that are averaged with the training service vectors” is a mental
step of vector data representation.
Step 2A Prong 2:
“ A method for obtaining a tag, comprising: obtaining corresponding materials in response to receiving an instruction for obtaining an interest tag;” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“and obtaining the interest tags by inputting the encoding vectors into a pre-trained tag recommendation model.” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” ” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ A method for obtaining a tag, comprising: obtaining corresponding materials in response to receiving an instruction for obtaining an interest tag;” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“and obtaining the interest tags by inputting the encoding vectors into a pre-trained tag recommendation model.” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” does not amount to significantly more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
In regard to claim 8: (Currently Amended)
Step 2A Prong 2:
“ wherein parsing the tag vectors, and outputting the interest tags based on the probability threshold value of the tag recommendation model, comprises: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining a plurality of tags by parsing the tag vectors based on an activation function in the tag recommendation model” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and determining tags whose occurrence probability is greater than or equal to the probability threshold value in the plurality of tags as the interest tags” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein parsing the tag vectors, and outputting the interest tags based on the probability threshold value of the tag recommendation model, comprises: “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“obtaining a plurality of tags by parsing the tag vectors based on an activation function in the tag recommendation model” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ and determining tags whose occurrence probability is greater than or equal to the probability threshold value in the plurality of tags as the interest tags” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 9: (Currently Amended)
Step 2A Prong 1:
“ obtaining training semantic vectors that comprise the interest tags by representing
features of the training materials using a semantic enhanced representation frame;” is a mental
step of vector data representation.
“obtaining training encoding vectors by aggregating social networks into the training
semantic vectors;” a mental step of vector data aggregation.
“ and obtaining the training semantic vectors that comprise the interest tags by representing the
features of the training materials using the semantic enhanced representation frame” mental
step of data identification.
“ representing the behavior training materials as training behavior vectors of different
lengths” a mental step of vector data representation.
“and representing the service training materials as fixed-length training service vectors
in the semantic enhanced representation frame” is a mental step of vector data representation.
“ and obtain the training semantic vectors by averaging the training behavior vectors and fusing
the training behavior vectors that are averaged with the training service vectors” is a mental
step of vector data representation.
Additional Elements
Step 2A Prong 2:
“ An electronic device, comprising: a processor; and a memory communicatively coupled to the processor; wherein the memory is configured to store instructions executable by the processor, and the processor is configured to: “ do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ collecting training materials that comprise interest tags “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” ” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ An electronic device, comprising: a processor; and a memory communicatively coupled to the processor; wherein the memory is configured to store instructions executable by the processor, and the processor is configured to: “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ collecting training materials that comprise interest tags “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim.
These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” ”is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
In regard to claim 11: (Original)
Step 2A Prong 2:
“ wherein, the processor is further configured to: determine intimacy values between any two of the social networks” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein, the processor is further configured to: determine intimacy values between any two of the social networks” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 12: (Original)
Step 2A Prong 1:
“ determine the intimacy values as values of elements in a matrix, and generating an adjacency
matrix based on the values of the elements” is a mental step of data comparison.
“in response to that a sum of weights of elements in each row of the adjacency matrix is one,
assign weights to the elements, wherein a weight assigned to each of elements arranged
diagonally in the adjacency matrix is greater than weights assigned to other elements” is a
mental step of vector data representation.
“and obtain a training semantic vector corresponding to each element in the adjacency matrix,”
is a mental step of vector data representation.
“and obtain the training encoding vectors by calculating a product of the training semantic
vector and a value of each element after assigning by a graph convolutional network.” is a
mental step of vector data representation.
Additional Elements
Step 2A Prong 2:
“wherein the processor is further configured to: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtain new training encoding vectors by inputting the training encoding vectors into a forward network;” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtain training tag vectors by inputting the new training encoding vectors into a fully-connected network” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and obtain the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“wherein the processor is further configured to: “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ obtain new training encoding vectors by inputting the training encoding vectors into a forward network;” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ obtain training tag vectors by inputting the new training encoding vectors into a fully-connected network” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ and obtain the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 13: (Original)
Step 2A Prong 1:
“ obtain interest tags in the training tag vectors by parsing the training tag vectors by an activation function” is a mental process.
“ and determine first interest tags corresponding to the interest tags in the training tag vectors,”
“ calculating a ratio of the first interest tags to the interest tags” is a mental step of data
comparison.
Additional Elements
Step 2A Prong 2:
“ wherein the processor is further configured to:” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein the processor is further configured to:” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 14: (Original)
Step 2A Prong 2:
“ An electronic device, comprising: a processor; and a memory communicatively coupled to the processor” ” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“wherein the memory is configured to store instructions executable by the processor, and the processor is configured to perform the method “ do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ An electronic device, comprising: a processor; and a memory communicatively coupled to the processor” ” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“wherein the memory is configured to store instructions executable by the processor, and the processor is configured to perform the method “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 15: (Currently Amended)
Step 2A Prong 1:
“ obtaining training semantic vectors that comprise the interest tags by representing
features of the training materials using a semantic enhanced representation frame;” is a mental
step of vector data representation.
“obtaining training encoding vectors by aggregating social networks into the training
semantic vectors;” a mental step of vector data aggregation.
“ and obtaining the training semantic vectors that comprise the interest tags by representing the
features of the training materials using the semantic enhanced representation frame” mental
step of data identification.
“ representing the behavior training materials as training behavior vectors of different
lengths” a mental step of vector data representation.
“and representing the service training materials as fixed-length training service vectors
in the semantic enhanced representation frame” is a mental step of vector data representation.
“ and obtain the training semantic vectors by averaging the training behavior vectors and fusing
the training behavior vectors that are averaged with the training service vectors” is a mental
step of vector data representation.
Additional Elements
Step 2A Prong 2:
“A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement a method for training a tag recommendation model, the method comprising: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ collecting training materials that comprise interest tags “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” ” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement a method for training a tag recommendation model, the method comprising: “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ collecting training materials that comprise interest tags “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ wherein the training materials comprise behavior training materials and service training materials” ” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 17: (Original)
Step 2A Prong 1:
“ determining the intimacy values as values of elements in a matrix, and generating an
adjacency matrix based on the values of the elements;” is a mental step of data comparison.
“in response to that a sum of weights of elements in each row of the adjacency matrix is one,
assigning weights to the elements, wherein a weight assigned to each of elements arranged
diagonally in the adjacency matrix is greater than weights assigned to other elements” is a
mental step of vector data representation.
“ and obtaining a training semantic vector corresponding to each element in the adjacency
matrix, and obtaining the training encoding vectors by calculating a product of the training
semantic vector and a value of each element “is a mental step of vector data representation.
Additional Elements
Step 2A Prong 2:
“wherein obtaining the training encoding vectors by aggregating the social networks into the training semantic vectors, comprises: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ determining intimacy values between any two of the social networks; “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ after assigning by a graph convolutional network.” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“wherein obtaining the training encoding vectors by aggregating the social networks into the training semantic vectors, comprises: “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ determining intimacy values between any two of the social networks; “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ after assigning by a graph convolutional network.” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 18: (Original)
Step 2A Prong 2:
“ wherein obtaining the tag recommendation model by training the double-layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“obtaining new training encoding vectors by inputting the training encoding vectors into a forward network; “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network; “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“ and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tag” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein obtaining the tag recommendation model by training the double-layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises: “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“obtaining new training encoding vectors by inputting the training encoding vectors into a forward network; “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
do not integrate the judicial exception into a practical application. These additional elements
“ obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network; “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“ and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tag” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 19: (Original)
Step 2A Prong 1:
“obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function” is a mental process.
“and determining first interest tags corresponding to the interest tags in the training tag
vectors” is a mental step of data identification.
“ calculating a ratio of the first interest tags to the interest tag” is a mental step of data
calculation.
Additional Elements
Step 2A Prong 2:
“ wherein obtaining the tag recommendation model by determining the training tag vectors as the independent variables, and the outputs as the interest tags, comprises: “do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
“determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
“ wherein obtaining the tag recommendation model by determining the training tag vectors as the independent variables, and the outputs as the interest tags, comprises: “is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
“determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value” is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
In regard to claim 20: (Original)
Step 2A Prong 2:
A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement the method “ do not integrate the judicial exception into a practical application. These additional elements are merely directed to using a computer as a tool to perform an abstract idea. See MPEP 2106.05(h).
Step 2B:
A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement the method “ is directed to a generic computer function and does not amount to significantly more than the judicial exception in the claim. See MPEP 2106.05(h).
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.
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, 4, 9, 12, 15 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over
Tal Weiss et.al (hereinafter Weiss) US 2016/0350655 A1,
in view of Yasheng WANG et.al(hereinafter WANG) US 2022/0147715 A1,
in view of BOKAI CAO et.al (hereinafter CAO) US 2016/0275401 A1,
in view of Tangzhi YE (hereinafter YE) US 2021/0027170 A1.
In regard to claim 1: (Currently Amended)
Weiss discloses:
- A method for training a tag recommendation model, comprising:
collecting training materials that comprise interest tags
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
(BRI: A training material is a training data set)
- using the training encoding vectors as inputs and the interest tags as outputs.
In [0013] :
According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input, to a recurrent, or a convolutional, neural network language model,
In [0042]:
According to some embodiments, the variable length ‘character level input streams’ may be Unicode character streams, and the system may further comprise a UTF-8 encoder for applying UTF-8 encoding to the Unicode character streams prior to their inputting to the neural network
In [0039]:
An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an auto-encoder can be structured as a “double-layer” neural network by using separate layers for the encoder and decoder each consisting of multiple layers)
In [0090]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog related metadata for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- representing the behavior training materials as training behavior vectors of different lengths,
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’; (2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications.
In [0047]:
According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input.
In [0074]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog structure logic for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the enumeration of questions asked as part of learned dialog words inputs from its corpus/corpora. Questions asked by the system as part of a dialog may be fed as inputs, along with other dialog words inputs, to the model, enabling model learning of discourse-based language logical and structural characteristics/constraints. According to some embodiments, enumerated questions may be coupled-to/grouped-with/paired-with relevant corresponding answers and/or responses.
(BRI: A Q and A is a form of a training behavior)
In [0075]:
For example, if the dialog corpus from which the system learns looks like this:
[0076] User: Flight to Chicago
[0077] Intelligent Agent: When?
[0078] User: Tomorrow
[0079] Questions asked by the system are fed into the model as well, so that the input that the model learns looks like this:[
0080] <us> Flight to Chicago </us>
[0081] <as> When? </as>
[0082] <us> Tomorrow </us>
In [0083] :
Where <us> means User Start of Sentence; </us> means User End of Sentence; <as> means Intelligent Agent Start of Sentence; and </as> means Intelligent Agent End of Sentence.
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’
(BRI: Q and A is a variable length character string)
Weiss does not explicitly disclose:
in response to receiving an instruction for collecting training materials;
- obtaining training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame;
- wherein the training materials comprise behavior training materials and service training materials;
and obtaining the training semantic vectors that comprise the interest tags by representing the features of the training materials using the semantic enhanced representation frame, comprises:
- and representing the service training materials as fixed-length training service vectors, in the semantic enhanced representation frame;
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
However, WANG discloses:
in response to receiving an instruction for collecting training materials;
- obtaining training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame;
In [0007]:
a text processing method is provided, including: obtaining target knowledge data, where the target knowledge data includes a first named entity, a second named entity, and an association between the first named entity and the second named entity;
in [0141]:
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation.
In [0142]:
Entity: The entity refers to an object that is distinguishable and exists independently, for example, a person, a city, a plant, or a commodity. Everything in the world is constituted by concrete objects, which refer to entities, for example, “China”, “United States”, and “Japan”. The entity is a most basic element in the knowledge graph. There are different relations between different entities.
in [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
In [0024] :
By setting, in the training text, the second knowledge identifier used to indicate the named entity, the original processing model can be guided to inject knowledge and semantic information into the second knowledge identifier, and the model can be guided to focus on the named entity indicated by the second identifier or extract a local knowledge feature.
In [0033]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0034] :
After the first text vector and the first knowledge vector are fused, the first text vector is fused with knowledge information, and the first knowledge vector is fused with semantic background information.
- wherein the training materials comprise behavior training materials and service training materials;
In [0141] :
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation. An association between two entities is described by using a relation, for example, a relation between Beijing and China. For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight. Currently, the knowledge graph has been widely used to refer to various large-scale knowledge bases (knowledge bases).
- and obtaining the training semantic vectors that comprise the interest tags by representing the features of the training materials using the semantic enhanced representation frame, comprises:
In [0013]:
By setting, in the to-be-processed text, the first knowledge identifier used to indicate the named entity, the target processing model can be guided to inject knowledge and semantic information into the first knowledge identifier, and the model can be guided to focus on the named entity indicated by the first identifier or extract a local knowledge feature.
In [0014]:
in a possible implementation, the fused target text vector includes at least a part of information in the target knowledge data, and the fused target knowledge vector includes semantic background information of the to-be-processed text.
In [0019]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0118]:
The generation model mainly generates samples (samples) with a same distribution from training data, and estimates a joint probability distribution of input x and a category label y. The discrimination model determines whether the input is real data or data generated by the generation model, that is, estimates a conditional probability distribution that a sample belongs to a specific category.
In [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
- and representing the service training materials as fixed-length training service vectors, in the semantic enhanced representation frame;
In [0144]
Content: The content is usually used as names, descriptions, and interpretations of entities and semantic categories, and may be expressed by text, images, and audio/videos.
In [0145] :
Attribute (value) (property): The attribute points to an attribute value of an entity from the entity. Different attribute types correspond to edges of different types of attributes. The attribute value refers to a value of an attribute specified by an object. For example, “area”, “population”, and “capital” are several different attributes of the entity “China”. The attribute value mainly refers to the value of the attribute specified by the object. For example, a value of the area attribute specified by “China” is “9.6 million square kilometers”.
In [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight
(BRI: the attributes are of fixed length)
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
In [0008] According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data,
In [0040]:
With reference to the second aspect, in a possible implementation, the original fusion model is any one of the following models: a multilayer self-attention model, a multilayer perceptron model, a recurrent neural network model, a weight model, a convolutional neural network model, a generative adversarial network model, and a reinforcement learning neural network model.
In [0050]:
With reference to the third aspect, in a possible implementation, the processor is further configured to: obtain first knowledge data, where the first knowledge data includes a third named entity, a fourth named entity, and an association between the third named entity and the fourth named entity, and the target knowledge data includes the first knowledge data; process the first knowledge data to obtain a first knowledge vector, where the first knowledge vector includes a vector corresponding to the third named entity, a vector corresponding to the fourth named entity, and a vector corresponding to the association between the third named entity and the fourth named entity; obtain training text and a first task result that corresponds to the training text and the target task, where the training text includes one or more named entities, and the one or more named entities include the third named entity; process the training text to obtain a first text vector; fuse the first text vector and the first knowledge vector based on an original fusion model, to obtain a fused first text vector and a fused first knowledge vector; process the fused first text vector and/or the fused first knowledge vector based on an original processing model, to obtain a second task result; and adjust parameters of the original processing model based on the first task result and the second task result, to obtain the target processing model; and/or adjust parameters of the original fusion model based on the first task result and the second task result, to obtain the target fusion model.
In [0298]:
The knowledge aggregator #5 may have a complex multilayer network structure, for example, a multilayer self-attention mechanism network structure, a multilayer perceptron network structure, or a recurrent neural network structure, and may simply weight and average the encoded text sequence and the encoded knowledge sequence.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
Weiss and WANG do not explicitly disclose:
- obtaining training encoding vectors by aggregating social networks into the training semantic vectors;
However, CAO discloses:
- obtaining training encoding vectors by aggregating social networks into the training semantic vectors;
In [0084]:
Using the plurality of social message and venue pairs, the server 104 computes (406) features based on meta-paths and geo-coordinate information. In some implementations, meta-paths are used to compute the features and the computed features include measures of geo features.
In [0085]:
the computation (406) is performed for a pair in the plurality of social message and venue pairs, first encoding the respective training social message in the pair as a label
Having encoded the label, the server 104 further identifies for the respective training social message corresponding training meta-paths to the respective venue in the pair.
In [0085]:
In some implementations, the path counts for different meta-paths are combined to create an overall feature matrix and the overall feature matrix is represented as the training feature vector.
In [0093]:
Different meta-paths usually represent different relation-ships among linked nodes with different semantic meanings.
In [0014]:
identifying for the respective training social message corresponding training meta-paths to the respective venue in the pair; encoding the corresponding training meta-paths to a corresponding training feature vector, wherein each element of the corresponding training feature vector includes a measure based on a respective type of the respective training social message connected to the respective venue in the pair; and giving the encoded labels and training feature vectors to the classifier for training.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG and CAO.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
One of ordinary skill would have motivation to combine Weiss, WANG and CAO that can concatenate different types of meta-path features to achieve significant improvement over accuracy ([CAO [0132])
Weiss, WANG and CAO do not explicitly disclose:
- and obtaining the tag recommendation model by training a double-layer neural network structure
However, YE discloses:
- and obtaining the tag recommendation model by training a double-layer neural network structure
In [0027] :
Optionally, the step of determining the tag based on the quality monitoring data includes: in [0028]:
determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable; and
in [0029]:
calculating a value of the evaluation indicator using the quality monitoring data, and determining the value of the evaluation indicator as a tag.
In [0048]:
a training apparatus for service quality evaluation models is provided, and the apparatus includes:
in [0049:
a collection module, configured to collect machine performance data, network characteristic data, and quality monitoring data of a service node according to a fixed cycle;
[0051]:
the processing module; further configured to determine a tag based on the quality monitoring data;
in [0052]:
the processing module, further configured to build a training set using the characteristic value and the tag; a
and [0053] a training module, configured to train a deep neural network model using the training set to obtain a service quality evaluation model.
(BRI: the determination of the tag on quality monitoring relates to recommendation model)
In [0118]:
Optionally, the embodiments of the present disclosure may adopt an LSTM neural network model with a double-layer structure. In the following, an LSTM neural network model with a double-layer structure is provided as an example to illustrate the training process of the model.
In [0094]:
The embodiments of the present disclosure provide a training method for service quality evaluation models. The method may be applicable to the network frame illustrated in FIG. 1. The network frame may include service nodes, monitoring nodes, and a model training node. The service node may be a node in a CDN service system that provides services to users:
in [0110]
The tag may intuitively reflect the service quality, and for different service types, the indicators used to evaluate service quality may be different. According to the service type to which the service quality evaluation model is applicable, a corresponding evaluation indicator may be used to determine the tag. The step of determining the tag based on the quality monitoring data may include: determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable.
(BRI: the that reflects the service quality within the context of services to “users” is a “interest tag”).
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, CAO and YE.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network
One of ordinary skill would have motivation to combine Weiss , WANG, CAO and YE that can
improve the efficiency of the service quality evaluation, but also reduce the operating costs.
(YE [0084]).
In regard to claim 4: (Original)
Weiss discloses:
wherein obtaining the tag recommendation model by training the double- layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises:
- and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags.
In [0007] :
The present invention includes systems, methods, circuits, and associated computer executable code for deep learning based natural language understanding, wherein: (1) a word tokenization and spelling correction model/machine may generate corrected word sets outputs based on respective character strings inputs; and/or (2) a word semantics derivation model/machine may generate semantically tagged sentences outputs based on respective word sets inputs.
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data,
In [0010]:
Both the supervised and unsupervised training phases may be repeated (in whole or in part ) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0090]:
According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network
In [0010]; According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an autoencoder is a type of fully connected NN that provides encoder and decoder using fully connected layer in both encode and decoder)
Weiss does not explicitly disclose:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
However, WANG discloses:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
In [0210] Optionally, if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector. Encoding the first knowledge data may be understood as converting the first knowledge data into a vector, for example, encoding the structured knowledge is converting the structured knowledge into a vector.
0008] According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data, and uses the obtained fused target text vector and/or the obtained fused target knowledge vector as input data for the target processing model
In [0120]:
In a training process, a neural network may correct values of parameters in an initial neural network model by using an error back propagation (back propagation, BP) algorithm, so that a reconstruction error loss of the neural network model becomes increasingly smaller. Specifically, an input signal is forward transferred until an error loss occurs in output
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
In regard to claim 9: (Currently Amended)
Weiss discloses:
- An electronic device, comprising: a processor; and a memory communicatively coupled to the processor; wherein the memory is configured to store instructions executable by the processor, and the processor is configured to:
In [0024], in [0028]
- collect training materials that comprise interest tags
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
- using the training encoding vectors as inputs and the interest tags as outputs.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an auto-encoder can be structured as a “double-layer” neural network by using separate layers for the encoder and decoder each consisting of multiple layers)
In [0062]:
an unsupervised, or weakly supervised, learning process, executed by a word semantics derivation model of a system for Deep Learning may include: (1) receiving a set of one or more words (e.g. ‘words input’, ‘word set’, ‘sentence’); (2) entering the word set, as an input set, to a sequence classifying, deep multi-layered recurrent, and/or recursive, neural network based, word semantics derivation model, wherein the word semantics derivation model is adapted for: (i) weakly supervising the model learning by providing a substantially small amount of ‘right’ semantic taggings as learning examples to the model; (ii) assigning markup language semantic tags to at least some, and/or a subset, of the words; (3) repeating stages (1) and (2) one or more additional times, while utilizing stochastic gradient descend for learning ‘correct’ semantic tagging, and improving following taggings' outputs.
In [0090]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog related metadata for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- represent the behavior training materials as training behavior vectors of different lengths,
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’; (2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications.
In [0047]:
According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input.
In [0074]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog structure logic for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the enumeration of questions asked as part of learned dialog words inputs from its corpus/corpora. Questions asked by the system as part of a dialog may be fed as inputs, along with other dialog words inputs, to the model, enabling model learning of discourse-based language logical and structural characteristics/constraints. According to some embodiments, enumerated questions may be coupled-to/grouped-with/paired-with relevant corresponding answers and/or responses.
(BRI: A Q and A is a form of a training behavior)
In [0075]:
For example, if the dialog corpus from which the system learns looks like this:
[0076] User: Flight to Chicago
[0077] Intelligent Agent: When?
[0078] User: Tomorrow
[0079] Questions asked by the system are fed into the model as well, so that the input that the model learns looks like this:[
0080] <us> Flight to Chicago </us>
[0081] <as> When? </as>
[0082] <us> Tomorrow </us>
In [0083] :
Where <us> means User Start of Sentence; </us> means User End of Sentence; <as> means Intelligent Agent Start of Sentence; and </as> means Intelligent Agent End of Sentence.
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’
(BRI: Q and A is a variable length character string)
Weiss does not explicitly disclose:
- obtain training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame;
- wherein the training materials comprise behavior training materials and service training materials;
However, WANG discloses:
- obtain training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame;
In [0007]:
a text processing method is provided, including: obtaining target knowledge data, where the target knowledge data includes a first named entity, a second named entity, and an association between the first named entity and the second named entity;
in [0141]:
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation.
In [0142]:
Entity: The entity refers to an object that is distinguishable and exists independently, for example, a person, a city, a plant, or a commodity. Everything in the world is constituted by concrete objects, which refer to entities, for example, “China”, “United States”, and “Japan”. The entity is a most basic element in the knowledge graph. There are different relations between different entities.
in [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
In [0024] :
By setting, in the training text, the second knowledge identifier used to indicate the named entity, the original processing model can be guided to inject knowledge and semantic information into the second knowledge identifier, and the model can be guided to focus on the named entity indicated by the second identifier or extract a local knowledge feature.
In [0033]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0034] :
After the first text vector and the first knowledge vector are fused, the first text vector is fused with knowledge information, and the first knowledge vector is fused with semantic background information.
- wherein the training materials comprise behavior training materials and service training materials;
In [0141] :
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation. An association between two entities is described by using a relation, for example, a relation between Beijing and China. For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight. Currently, the knowledge graph has been widely used to refer to various large-scale knowledge bases (knowledge bases).
- and representing the service training materials as fixed-length training service vectors, in the semantic enhanced representation frame;
In [0144]
Content: The content is usually used as names, descriptions, and interpretations of entities and semantic categories, and may be expressed by text, images, and audio/videos.
In [0145] :
Attribute (value) (property): The attribute points to an attribute value of an entity from the entity. Different attribute types correspond to edges of different types of attributes. The attribute value refers to a value of an attribute specified by an object. For example, “area”, “population”, and “capital” are several different attributes of the entity “China”. The attribute value mainly refers to the value of the attribute specified by the object. For example, a value of the area attribute specified by “China” is “9.6 million square kilometers”.
In [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight
(BRI: the attributes are of fixed length)
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
In [0144]
Content: The content is usually used as names, descriptions, and interpretations of entities and semantic categories, and may be expressed by text, images, and audio/videos.
In [0145] :
Attribute (value) (property): The attribute points to an attribute value of an entity from the entity. Different attribute types correspond to edges of different types of attributes. The attribute value refers to a value of an attribute specified by an object. For example, “area”, “population”, and “capital” are several different attributes of the entity “China”. The attribute value mainly refers to the value of the attribute specified by the object. For example, a value of the area attribute specified by “China” is “9.6 million square kilometers”.
In [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight
(BRI: the attributes are of fixed length)
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
In [0008] According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data,
In [0040]:
With reference to the second aspect, in a possible implementation, the original fusion model is any one of the following models: a multilayer self-attention model, a multilayer perceptron model, a recurrent neural network model, a weight model, a convolutional neural network model, a generative adversarial network model, and a reinforcement learning neural network model.
In [0050]:
With reference to the third aspect, in a possible implementation, the processor is further configured to: obtain first knowledge data, where the first knowledge data includes a third named entity, a fourth named entity, and an association between the third named entity and the fourth named entity, and the target knowledge data includes the first knowledge data; process the first knowledge data to obtain a first knowledge vector, where the first knowledge vector includes a vector corresponding to the third named entity, a vector corresponding to the fourth named entity, and a vector corresponding to the association between the third named entity and the fourth named entity; obtain training text and a first task result that corresponds to the training text and the target task, where the training text includes one or more named entities, and the one or more named entities include the third named entity; process the training text to obtain a first text vector; fuse the first text vector and the first knowledge vector based on an original fusion model, to obtain a fused first text vector and a fused first knowledge vector; process the fused first text vector and/or the fused first knowledge vector based on an original processing model, to obtain a second task result; and adjust parameters of the original processing model based on the first task result and the second task result, to obtain the target processing model; and/or adjust parameters of the original fusion model based on the first task result and the second task result, to obtain the target fusion model.
In [0298]:
The knowledge aggregator #5 may have a complex multilayer network structure, for example, a multilayer self-attention mechanism network structure, a multilayer perceptron network structure, or a recurrent neural network structure, and may simply weight and average the encoded text sequence and the encoded knowledge sequence.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
Weiss and WANG do not explicitly disclose:
- obtain training encoding vectors by aggregating social networks into the training semantic vectors;
However, CAO discloses:
- obtain training encoding vectors by aggregating social networks into the training semantic vectors;
In [0084]:
Using the plurality of social message and venue pairs, the server 104 computes (406) features based on meta-paths and geo-coordinate information. In some implementations, meta-paths are used to compute the features and the computed features include measures of geo features.
In [0085]:
the computation (406) is performed for a pair in the plurality of social message and venue pairs, first encoding the respective training social message in the pair as a label
Having encoded the label, the server 104 further identifies for the respective training social message corresponding training meta-paths to the respective venue in the pair.
In [0085]:
In some implementations, the path counts for different meta-paths are combined to create an overall feature matrix and the overall feature matrix is represented as the training feature vector.
In [0093]:
Different meta-paths usually represent different relation-ships among linked nodes with different semantic meanings.
In [0014]:
identifying for the respective training social message corresponding training meta-paths to the respective venue in the pair; encoding the corresponding training meta-paths to a corresponding training feature vector, wherein each element of the corresponding training feature vector includes a measure based on a respective type of the respective training social message connected to the respective venue in the pair; and giving the encoded labels and training feature vectors to the classifier for training.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG and CAO.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
One of ordinary skill would have motivation to combine Weiss, WANG and CAO that can concatenate different types of meta-path features to achieve significant improvement over accuracy ([CAO [0132])
Weiss, WANG and CAO do not explicitly disclose:
- and obtaining the tag recommendation model by training a double-layer neural network
However, YE discloses:
- and obtaining the tag recommendation model by training a double-layer neural network
In [0027] :
Optionally, the step of determining the tag based on the quality monitoring data includes: in [0028]:
determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable; and
in [0029]:
calculating a value of the evaluation indicator using the quality monitoring data, and determining the value of the evaluation indicator as a tag.
In [0048]:
a training apparatus for service quality evaluation models is provided, and the apparatus includes:
in [0049:
a collection module, configured to collect machine performance data, network characteristic data, and quality monitoring data of a service node according to a fixed cycle;
[0051]:
the processing module; further configured to determine a tag based on the quality monitoring data;
in [0052]:
the processing module, further configured to build a training set using the characteristic value and the tag; a
and [0053] a training module, configured to train a deep neural network model using the training set to obtain a service quality evaluation model.
(BRI: the determination of the tag on quality monitoring relates to recommendation model)
In [0118]:
Optionally, the embodiments of the present disclosure may adopt an LSTM neural network model with a double-layer structure. In the following, an LSTM neural network model with a double-layer structure is provided as an example to illustrate the training process of the model.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, CAO and YE.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network
One of ordinary skill would have motivation to combine Weiss , WANG, CAO and YE that can
improve the efficiency of the service quality evaluation, but also reduce the operating costs.
(YE [0084]).
In regard to claim 12: (Original)
Weiss discloses:
wherein obtaining the tag recommendation model by training the double- layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises:
- and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags.
In [0007] :
The present invention includes systems, methods, circuits, and associated computer executable code for deep learning based natural language understanding, wherein: (1) a word tokenization and spelling correction model/machine may generate corrected word sets outputs based on respective character strings inputs; and/or (2) a word semantics derivation model/machine may generate semantically tagged sentences outputs based on respective word sets inputs.
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data,
In [0010]:
Both the supervised and unsupervised training phases may be repeated (in whole or in part ) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0090]:
According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network
In [0010]; According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an autoencoder is a type of fully connected NN that provides encoder and decoder using fully connected layer in both encode and decoder)
Weiss does not explicitly disclose:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
However, WANG discloses:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
In [0210] Optionally, if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector. Encoding the first knowledge data may be understood as converting the first knowledge data into a vector, for example, encoding the structured knowledge is converting the structured knowledge into a vector.
In [0008]:
According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data, and uses the obtained fused target text vector and/or the obtained fused target knowledge vector as input data for the target processing model
In [0120]:
In a training process, a neural network may correct values of parameters in an initial neural network model by using an error back propagation (back propagation, BP) algorithm, so that a reconstruction error loss of the neural network model becomes increasingly smaller. Specifically, an input signal is forward transferred until an error loss occurs in output
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
In regard to claim 15: (Currently Amended)
Weiss discloses:
- A non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to implement a method for training a tag recommendation model, the method comprising:
In [0028], in [0029]
(BRI: all semiconductor computer readable is non-transitory)
- collecting training materials that comprise interest tags in response to receiving an instruction for collecting training materials;
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
(BRI: A training material is a training data set)
- using the training encoding vectors as inputs and the interest tags as outputs.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an auto-encoder can be structured as a “double-layer” neural network by using separate layers for the encoder and decoder each consisting of multiple layers)
In [0062]:
an unsupervised, or weakly supervised, learning process, executed by a word semantics derivation model of a system for Deep Learning may include: (1) receiving a set of one or more words (e.g. ‘words input’, ‘word set’, ‘sentence’); (2) entering the word set, as an input set, to a sequence classifying, deep multi-layered recurrent, and/or recursive, neural network based, word semantics derivation model, wherein the word semantics derivation model is adapted for: (i) weakly supervising the model learning by providing a substantially small amount of ‘right’ semantic taggings as learning examples to the model; (ii) assigning markup language semantic tags to at least some, and/or a subset, of the words; (3) repeating stages (1) and (2) one or more additional times, while utilizing stochastic gradient descend for learning ‘correct’ semantic tagging, and improving following taggings' outputs.
In [0090]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog related metadata for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- representing the behavior training materials as training behavior vectors of different lengths,
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’; (2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications.
In [0047]:
According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input.
In [0074]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog structure logic for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the enumeration of questions asked as part of learned dialog words inputs from its corpus/corpora. Questions asked by the system as part of a dialog may be fed as inputs, along with other dialog words inputs, to the model, enabling model learning of discourse-based language logical and structural characteristics/constraints. According to some embodiments, enumerated questions may be coupled-to/grouped-with/paired-with relevant corresponding answers and/or responses.
(BRI: A Q and A is a form of a training behavior)
In [0075]:
For example, if the dialog corpus from which the system learns looks like this:
[0076] User: Flight to Chicago
[0077] Intelligent Agent: When?
[0078] User: Tomorrow
[0079] Questions asked by the system are fed into the model as well, so that the input that the model learns looks like this:[
0080] <us> Flight to Chicago </us>
[0081] <as> When? </as>
[0082] <us> Tomorrow </us>
In [0083] :
Where <us> means User Start of Sentence; </us> means User End of Sentence; <as> means Intelligent Agent Start of Sentence; and </as> means Intelligent Agent End of Sentence.
In [0039]:
According to some embodiments of the present invention, a neural network based system for spell correction and tokenization of natural language, may comprise: (1) An artificial neural network architecture, to generate variable length ‘character level output streams’ for system fed variable length ‘character level input streams’
(BRI: Q and A is a variable length character string)
However, WANG discloses:
- obtaining training semantic vectors that comprise the interest tags by representing features of the training materials using a semantic enhanced representation frame;
In [0007]:
a text processing method is provided, including: obtaining target knowledge data, where the target knowledge data includes a first named entity, a second named entity, and an association between the first named entity and the second named entity;
in [0141]:
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation.
In [0142]:
Entity: The entity refers to an object that is distinguishable and exists independently, for example, a person, a city, a plant, or a commodity. Everything in the world is constituted by concrete objects, which refer to entities, for example, “China”, “United States”, and “Japan”. The entity is a most basic element in the knowledge graph. There are different relations between different entities.
in [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
In [0024] :
By setting, in the training text, the second knowledge identifier used to indicate the named entity, the original processing model can be guided to inject knowledge and semantic information into the second knowledge identifier, and the model can be guided to focus on the named entity indicated by the second identifier or extract a local knowledge feature.
In [0033]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0034] :
After the first text vector and the first knowledge vector are fused, the first text vector is fused with knowledge information, and the first knowledge vector is fused with semantic background information.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
- wherein the training materials comprise behavior training materials and service training materials;
In [0141] :
The knowledge graph describes various entities or concepts and relations between the entities or concepts in the real world, and forms a huge semantic network diagram, where a node represents an entity or a concept, and an edge is constituted by an attribute or a relation. An association between two entities is described by using a relation, for example, a relation between Beijing and China. For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight. Currently, the knowledge graph has been widely used to refer to various large-scale knowledge bases (knowledge bases).
- and obtaining the training semantic vectors that comprise the interest tags by representing the features of the training materials using the semantic enhanced representation frame, comprises:
In [0013]:
By setting, in the to-be-processed text, the first knowledge identifier used to indicate the named entity, the target processing model can be guided to inject knowledge and semantic information into the first knowledge identifier, and the model can be guided to focus on the named entity indicated by the first identifier or extract a local knowledge feature.
In [0014]:
in a possible implementation, the fused target text vector includes at least a part of information in the target knowledge data, and the fused target knowledge vector includes semantic background information of the to-be-processed text.
In [0019]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0118]:
The generation model mainly generates samples (samples) with a same distribution from training data, and estimates a joint probability distribution of input x and a category label y. The discrimination model determines whether the input is real data or data generated by the generation model, that is, estimates a conditional probability distribution that a sample belongs to a specific category.
In [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
- and representing the service training materials as fixed-length training service vectors, in the semantic enhanced representation frame;
In [0144]
Content: The content is usually used as names, descriptions, and interpretations of entities and semantic categories, and may be expressed by text, images, and audio/videos.
In [0145] :
Attribute (value) (property): The attribute points to an attribute value of an entity from the entity. Different attribute types correspond to edges of different types of attributes. The attribute value refers to a value of an attribute specified by an object. For example, “area”, “population”, and “capital” are several different attributes of the entity “China”. The attribute value mainly refers to the value of the attribute specified by the object. For example, a value of the area attribute specified by “China” is “9.6 million square kilometers”.
In [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight
(BRI: the attributes are of fixed length)
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
In [0144]
Content: The content is usually used as names, descriptions, and interpretations of entities and semantic categories, and may be expressed by text, images, and audio/videos.
In [0145] :
Attribute (value) (property): The attribute points to an attribute value of an entity from the entity. Different attribute types correspond to edges of different types of attributes. The attribute value refers to a value of an attribute specified by an object. For example, “area”, “population”, and “capital” are several different attributes of the entity “China”. The attribute value mainly refers to the value of the attribute specified by the object. For example, a value of the area attribute specified by “China” is “9.6 million square kilometers”.
In [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight
(BRI: the attributes are of fixed length)
- and obtaining the training semantic vectors by averaging the training behavior vectors and fusing the training behavior vectors that are averaged with the training service vectors.
In [0008] According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data,
In [0040]:
With reference to the second aspect, in a possible implementation, the original fusion model is any one of the following models: a multilayer self-attention model, a multilayer perceptron model, a recurrent neural network model, a weight model, a convolutional neural network model, a generative adversarial network model, and a reinforcement learning neural network model.
In [0050]:
With reference to the third aspect, in a possible implementation, the processor is further configured to: obtain first knowledge data, where the first knowledge data includes a third named entity, a fourth named entity, and an association between the third named entity and the fourth named entity, and the target knowledge data includes the first knowledge data; process the first knowledge data to obtain a first knowledge vector, where the first knowledge vector includes a vector corresponding to the third named entity, a vector corresponding to the fourth named entity, and a vector corresponding to the association between the third named entity and the fourth named entity; obtain training text and a first task result that corresponds to the training text and the target task, where the training text includes one or more named entities, and the one or more named entities include the third named entity; process the training text to obtain a first text vector; fuse the first text vector and the first knowledge vector based on an original fusion model, to obtain a fused first text vector and a fused first knowledge vector; process the fused first text vector and/or the fused first knowledge vector based on an original processing model, to obtain a second task result; and adjust parameters of the original processing model based on the first task result and the second task result, to obtain the target processing model; and/or adjust parameters of the original fusion model based on the first task result and the second task result, to obtain the target fusion model.
In [0298]:
The knowledge aggregator #5 may have a complex multilayer network structure, for example, a multilayer self-attention mechanism network structure, a multilayer perceptron network structure, or a recurrent neural network structure, and may simply weight and average the encoded text sequence and the encoded knowledge sequence.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
Weiss and WANG do not explicitly disclose:
- obtaining training encoding vectors by aggregating social networks into the training semantic vectors; and obtaining the tag recommendation model
However, CAO discloses:
- obtaining training encoding vectors by aggregating social networks into the training semantic vectors; and obtaining the tag recommendation model
In [0084]:
Using the plurality of social message and venue pairs, the server 104 computes (406) features based on meta-paths and geo-coordinate information. In some implementations, meta-paths are used to compute the features and the computed features include measures of geo features.
In [0085]:
the computation (406) is performed for a pair in the plurality of social message and venue pairs, first encoding the respective training social message in the pair as a label
Having encoded the label, the server 104 further identifies for the respective training social message corresponding training meta-paths to the respective venue in the pair.
In [0085]:
In some implementations, the path counts for different meta-paths are combined to create an overall feature matrix and the overall feature matrix is represented as the training feature vector.
In [0093]:
Different meta-paths usually represent different relation-ships among linked nodes with different semantic meanings.
In [0014]:
identifying for the respective training social message corresponding training meta-paths to the respective venue in the pair; encoding the corresponding training meta-paths to a corresponding training feature vector, wherein each element of the corresponding training feature vector includes a measure based on a respective type of the respective training social message connected to the respective venue in the pair; and giving the encoded labels and training feature vectors to the classifier for training.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG and CAO.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
One of ordinary skill would have motivation to combine Weiss, WANG and CAO that can concatenate different types of meta-path features to achieve significant improvement over accuracy ([CAO [0132])
Weiss, WANG and CAO do not explicitly disclose:
- and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs
However, YE discloses:
- and obtaining the tag recommendation model by training a double-layer neural network structure using the training encoding vectors as inputs and the interest tags as outputs
In [0027] :
Optionally, the step of determining the tag based on the quality monitoring data includes: in [0028]:
determining an evaluation indicator of service quality based on the service type to which the service quality evaluation model is applicable; and
in [0029]:
calculating a value of the evaluation indicator using the quality monitoring data, and determining the value of the evaluation indicator as a tag.
In [0048]:
a training apparatus for service quality evaluation models is provided, and the apparatus includes:
in [0049:
a collection module, configured to collect machine performance data, network characteristic data, and quality monitoring data of a service node according to a fixed cycle;
in [0051]:
the processing module; further configured to determine a tag based on the quality monitoring data;
in [0052]:
the processing module, further configured to build a training set using the characteristic value and the tag; a
and [0053] a training module, configured to train a deep neural network model using the training set to obtain a service quality evaluation model.
(BRI: the determination of the tag on quality monitoring relates to recommendation model)
In [0118]:
Optionally, the embodiments of the present disclosure may adopt an LSTM neural network model with a double-layer structure. In the following, an LSTM neural network model with a double-layer structure is provided as an example to illustrate the training process of the model.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG , CAO and YE.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network
One of ordinary skill would have motivation to combine Weiss , WANG and YE that can
improve the efficiency of the service quality evaluation, but also reduce the operating costs.
(YE [0084]).
In regard to claim 18: (Original)
Weiss discloses:
wherein obtaining the tag recommendation model by training the double- layer neural network structure using the training encoding vectors as the inputs and the interest tags as the outputs, comprises:
- and obtaining the tag recommendation model by determining the training tag vectors as independent variables, and outputs as the interest tags.
In [0007] :
The present invention includes systems, methods, circuits, and associated computer executable code for deep learning based natural language understanding, wherein: (1) a word tokenization and spelling correction model/machine may generate corrected word sets outputs based on respective character strings inputs; and/or (2) a word semantics derivation model/machine may generate semantically tagged sentences outputs based on respective word sets inputs.
In [0010]:
According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data,
In [0010]:
Both the supervised and unsupervised training phases may be repeated (in whole or in part ) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0090]:
According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- obtaining training tag vectors by inputting the new training encoding vectors into a fully-connected network
In [0010]; According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an autoencoder is a type of fully connected NN that provides encoder and decoder using fully connected layer in both encode and decoder)
Weiss does not explicitly disclose:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
However, WANG discloses:
- obtaining new training encoding vectors by inputting the training encoding vectors into a forward network;
In [0210] Optionally, if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector. Encoding the first knowledge data may be understood as converting the first knowledge data into a vector, for example, encoding the structured knowledge is converting the structured knowledge into a vector.
In [0008] According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data, and uses the obtained fused target text vector and/or the obtained fused target knowledge vector as input data for the target processing model
In [0120]:
In a training process, a neural network may correct values of parameters in an initial neural network model by using an error back propagation (back propagation, BP) algorithm, so that a reconstruction error loss of the neural network model becomes increasingly smaller. Specifically, an input signal is forward transferred until an error loss occurs in output
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss , and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
Claims 3, 11 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over
Tal Weiss et.al (hereinafter Weiss) US 2016/0350655 A1,
in view of Yasheng WANG(hereinafter WANG) US 2022/0147715 A1,
in view of BOKAI CAO et.al (hereinafter CAO) US 2016/0275401 A1,
in view of Tangzhi YE (hereinafter YE) US 2021/0027170 A1.
further in view of Hamid Hatami-Hanza (hereinafter Hatami) US 2015/0227559 A1.
In regard to claim 3: (Original)
Weiss, WANG, CAO and YE do not explicitly disclose:
- determining intimacy values between any two of the social networks;
- the intimacy values as values of elements in a matrix,
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
However, Hatami discloses:
- determining intimacy values between any two of the social networks;
In [0105]:
The methods and systems of the present invention and can be used for applications ranging from document classification, search engine document retrieval, news analysis, knowledge discovery and research trajectory optimization, question answering, computer conversation, spell checking, summarization, categorizations, categorization, clustering, distillation, automatic composition generation, genetics and genomics, signal and image processing, to novel applications in economical systems by evaluating a value for economical entities, crime investigation, financial applications such as financial decision making, credit checking, decision support systems, stock valuation, target advertising, and as well measuring the influence of a member in a social network, and/or any other problem that can be represented by graphs and for any group of entities with some kind of relations or association.
In [0022]:
In another aspect various measures of "association strength" are given from which the relations of ontological subjects of the composition can be revealed. Algorithms and formulations and calculation methods are given to evaluate such "association strength" according to various exemplary association aspects.
In [0023]:
According to another aspect of the present invention measures are given for evaluating the "relational association strengths" of the ontological subjects of different orders to each other or to one or more target ontological subject.
(BRI: the association strength is the intimacy value)
- the intimacy values as values of elements in a matrix,
In [0186]:
The association strength data structures usually in the form a matrix therefore is instrumental to build such cognitive networks for variety of tasks in general and for building neural nets in particular
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
In [0279] :
Referring to FIG. 3 now, it is to show that any composition of ontological subjects can in principal be represented by a graph which in this preferred embodiment shown as an asymmetric graph. The exemplified graph is corresponded to one of the exemplary "association strength matrix", i.e. an ASM, as representative of its adjacency matrix.
In [0283] :
The association strength matrix could be regarded as the adjacency matrix of any graphs such as social graphs or any network of any thing. For instance the graphs can be built representing the relations between the concepts and entities or any other desired set of OSs in a special area of science, market, industry or any "body of knowledge".
In [0134]:
The first type of a "value significance measure" is defined as a function of "Frequency of Occurrences" of OS.sub.i.sup.k is called here FO.sub.i.sup.k|l and can be given by:
PNG
media_image4.png
22
363
media_image4.png
Greyscale
wherein
F
O
i
k
|
l
i
s obtained by counting the occurrences of OSs of the particular order, e.g. counting the appearances of particular word in the text or counting its total occurrences in the partitions, or more conveniently be obtained from the
C
O
M
k
|
l
(the elements on the main diagonal of the
C
O
M
k
|
l
) or by using Eq. 4, or any other way of counting the occurrences of OS.sub.i.sup.k in the desired partitions of the composition.
In [0302]:
at lease one of the algorithms of calculating one of the measures in order to assign a value on the part or partitions of the compositions and based on the assigned value process one or more of the partitions or OSs of the particular order as an output in the form of a service or data. The output could be simply one or more tags or OS/s that the input composition can be characterized with, i.e. significant keywords of the composition.
In [0303]:
As another example, the output or outcome of the investigator of FIG. 10, could be to provide the partitions of the input composition which have exhibited intrinsic value significances of above a predetermined threshold. Another output could be the novel parts or the OSs of the compositions that scored a predetermined level of a particular type of novelty value significance
In [0109]:
Assuming we have an input composition of ontological subjects, e.g. an input text, the "Participation Matrix" (PM) is a matrix indicating the participation of one or more ontological subjects of particular order in one or more partitions of the composition. In other words in terms of our definitions, PM indicate the participation of one or more lower order OS into one or more OS of higher or the same order.
PNG
media_image5.png
17
358
media_image5.png
Greyscale
In [0122] :
For example the PMs, ASMs, OSM or co-occurrences of the ontological subjects etc. can be represented by a matrix, sparse matrix, table, database rows, no sql databases, JSON, dictionaries and the like which can be stored in various forms of data structures.
In [0186]:
data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net)
in [0186]:
The relatedness is measured by one or more of the above measures and partitions that exhibited an association strength value greater (or sometimes smaller) than a predetermined threshold to a particular OS, can be grouped or clustered together. Further these data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net).
In [0187]:
a number of partitions of the composition or the BOK that have exhibited a predetermined threshold of relative association strength or predetermined criteria of satisfying enough association strength to a target subject or to each other can be categorized or being clustered as group together
categorized or being clustered as group together.
in [0248:
In another aspect the novelty is observed in relation or combination with other OSs since novelty could occurs in a context and therefore in relation to other ontological subjects. The stand alone or the intrinsic "novelty value significance value" in this case is defined as sum of the novelty that an OS will have with a desired number of other OSs.
In [0249]:
These measures of novelty are intrinsic since it adds up all the pair-wise novelty values for each OS.sup.k so that a NVSM type 2 can be defined as:
PNG
media_image6.png
32
543
media_image6.png
Greyscale
wherein the pair-wise novelty measures are summed over the column (i.e. the j subscript).
In [0230] :
When there are multiple OSs of interest the pair-wise value significances can be used in combination and perhaps with various weight to achieve the same filtering effect for a set of OSs.
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
FIG. 3: shows one exemplary embodiment of a directed asymmetric network or graph corresponding to a composition of ontological subjects.
In [0179]:
using the above relational rasm one can conveniently find the most related partitions of a composition to one or more target OS for the desired goal of the investigation (e.g quick retrieval of documents, sentences, or paragraphs with high semantic relevancy).
In [0257] :
In accordance with another aspect of the methods of investigation of the compositions of ontological subject of the present invention, the participation matrix can, for instance, routinely being transformed to other types of objects or participation matrices by operating one or more vector or matrices on the PM. For example one can multiply the PM by a diagonal matrix (M by M) from the right side whose diagonal values are the reciprocal of the number of constituent OSs of order k in the partitions or the higher order OS of order l.
In [0254]:
Moreover all these matrices (e.g. such as PM, COM, ASM/s, RASM, RVSMs NVSM, RNVSMs etc.) can be regarded as an adjacency matrix for a corresponding graph wherein the matrix carry the data of the connectivity between the nodes or objects of the graph
In [0040]:
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG , CAO, YE and Hatami.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network.
Hatami teaches representing intimacy using matrix.
One of ordinary skill would have motivation to combine Weiss , WANG , CAO , YE and Hatami that can reduce the training iteration and resources for training using association strength (Hatami [0186])
In regard to claim 11: (Original)
Weiss, WANG, CAO and YE do not explicitly disclose:
- determining intimacy values between any two of the social networks;
- the intimacy values as values of elements in a matrix,
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
However, Hatami discloses:
- determining intimacy values between any two of the social networks;
In [0105]:
The methods and systems of the present invention and can be used for applications ranging from document classification, search engine document retrieval, news analysis, knowledge discovery and research trajectory optimization, question answering, computer conversation, spell checking, summarization, categorizations, categorization, clustering, distillation, automatic composition generation, genetics and genomics, signal and image processing, to novel applications in economical systems by evaluating a value for economical entities, crime investigation, financial applications such as financial decision making, credit checking, decision support systems, stock valuation, target advertising, and as well measuring the influence of a member in a social network, and/or any other problem that can be represented by graphs and for any group of entities with some kind of relations or association.
In [0022]:
In another aspect various measures of "association strength" are given from which the relations of ontological subjects of the composition can be revealed. Algorithms and formulations and calculation methods are given to evaluate such "association strength" according to various exemplary association aspects.
In [0023]:
According to another aspect of the present invention measures are given for evaluating the "relational association strengths" of the ontological subjects of different orders to each other or to one or more target ontological subject.
(BRI: the association strength is the intimacy value)
- the intimacy values as values of elements in a matrix,
In [0186]:
The association strength data structures usually in the form a matrix therefore is instrumental to build such cognitive networks for variety of tasks in general and for building neural nets in particular
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
In [0279] :
Referring to FIG. 3 now, it is to show that any composition of ontological subjects can in principal be represented by a graph which in this preferred embodiment shown as an asymmetric graph. The exemplified graph is corresponded to one of the exemplary "association strength matrix", i.e. an ASM, as representative of its adjacency matrix.
In [0283] :
The association strength matrix could be regarded as the adjacency matrix of any graphs such as social graphs or any network of any thing. For instance the graphs can be built representing the relations between the concepts and entities or any other desired set of OSs in a special area of science, market, industry or any "body of knowledge".
In [0134]:
The first type of a "value significance measure" is defined as a function of "Frequency of Occurrences" of OS.sub.i.sup.k is called here FO.sub.i.sup.k|l and can be given by:
PNG
media_image4.png
22
363
media_image4.png
Greyscale
wherein
F
O
i
k
|
l
i
s obtained by counting the occurrences of OSs of the particular order, e.g. counting the appearances of particular word in the text or counting its total occurrences in the partitions, or more conveniently be obtained from the
C
O
M
k
|
l
(the elements on the main diagonal of the
C
O
M
k
|
l
) or by using Eq. 4, or any other way of counting the occurrences of OS.sub.i.sup.k in the desired partitions of the composition.
In [0302]:
at lease one of the algorithms of calculating one of the measures in order to assign a value on the part or partitions of the compositions and based on the assigned value process one or more of the partitions or OSs of the particular order as an output in the form of a service or data. The output could be simply one or more tags or OS/s that the input composition can be characterized with, i.e. significant keywords of the composition.
In [0303]:
As another example, the output or outcome of the investigator of FIG. 10, could be to provide the partitions of the input composition which have exhibited intrinsic value significances of above a predetermined threshold. Another output could be the novel parts or the OSs of the compositions that scored a predetermined level of a particular type of novelty value significance
In [0109]:
Assuming we have an input composition of ontological subjects, e.g. an input text, the "Participation Matrix" (PM) is a matrix indicating the participation of one or more ontological subjects of particular order in one or more partitions of the composition. In other words in terms of our definitions, PM indicate the participation of one or more lower order OS into one or more OS of higher or the same order.
PNG
media_image5.png
17
358
media_image5.png
Greyscale
In [0122] :
For example the PMs, ASMs, OSM or co-occurrences of the ontological subjects etc. can be represented by a matrix, sparse matrix, table, database rows, no sql databases, JSON, dictionaries and the like which can be stored in various forms of data structures.
In [0186]:
data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net)
in [0186]:
The relatedness is measured by one or more of the above measures and partitions that exhibited an association strength value greater (or sometimes smaller) than a predetermined threshold to a particular OS, can be grouped or clustered together. Further these data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net).
In [0187]:
a number of partitions of the composition or the BOK that have exhibited a predetermined threshold of relative association strength or predetermined criteria of satisfying enough association strength to a target subject or to each other can be categorized or being clustered as group together
categorized or being clustered as group together.
in [0248:
In another aspect the novelty is observed in relation or combination with other OSs since novelty could occurs in a context and therefore in relation to other ontological subjects. The stand alone or the intrinsic "novelty value significance value" in this case is defined as sum of the novelty that an OS will have with a desired number of other OSs.
In [0249]:
These measures of novelty are intrinsic since it adds up all the pair-wise novelty values for each OS.sup.k so that a NVSM type 2 can be defined as:
PNG
media_image6.png
32
543
media_image6.png
Greyscale
wherein the pair-wise novelty measures are summed over the column (i.e. the j subscript).
In [0230] :
When there are multiple OSs of interest the pair-wise value significances can be used in combination and perhaps with various weight to achieve the same filtering effect for a set of OSs.
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
FIG. 3: shows one exemplary embodiment of a directed asymmetric network or graph corresponding to a composition of ontological subjects.
In [0179]:
using the above relational rasm one can conveniently find the most related partitions of a composition to one or more target OS for the desired goal of the investigation (e.g quick retrieval of documents, sentences, or paragraphs with high semantic relevancy).
In [0257] :
In accordance with another aspect of the methods of investigation of the compositions of ontological subject of the present invention, the participation matrix can, for instance, routinely being transformed to other types of objects or participation matrices by operating one or more vector or matrices on the PM. For example one can multiply the PM by a diagonal matrix (M by M) from the right side whose diagonal values are the reciprocal of the number of constituent OSs of order k in the partitions or the higher order OS of order l.
In [0254]:
Moreover all these matrices (e.g. such as PM, COM, ASM/s, RASM, RVSMs NVSM, RNVSMs etc.) can be regarded as an adjacency matrix for a corresponding graph wherein the matrix carry the data of the connectivity between the nodes or objects of the graph
In [0040]:
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG , CAO, YE and Hatami.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network.
Hatami teaches representing intimacy using matrix.
One of ordinary skill would have motivation to combine Weiss , WANG, CAO, YE and Hatami that can Reduce the training iteration and resources for training using association strength (Hatami [0186])
In regard to claim 17: (Original)
Weiss, WANG, CAO and YE do not explicitly disclose:
- determining intimacy values between any two of the social networks;
- the intimacy values as values of elements in a matrix,
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
However, Hatami discloses:
- determining intimacy values between any two of the social networks;
In [0105]:
The methods and systems of the present invention and can be used for applications ranging from document classification, search engine document retrieval, news analysis, knowledge discovery and research trajectory optimization, question answering, computer conversation, spell checking, summarization, categorizations, categorization, clustering, distillation, automatic composition generation, genetics and genomics, signal and image processing, to novel applications in economical systems by evaluating a value for economical entities, crime investigation, financial applications such as financial decision making, credit checking, decision support systems, stock valuation, target advertising, and as well measuring the influence of a member in a social network, and/or any other problem that can be represented by graphs and for any group of entities with some kind of relations or association.
In [0022]:
In another aspect various measures of "association strength" are given from which the relations of ontological subjects of the composition can be revealed. Algorithms and formulations and calculation methods are given to evaluate such "association strength" according to various exemplary association aspects.
In [0023]:
According to another aspect of the present invention measures are given for evaluating the "relational association strengths" of the ontological subjects of different orders to each other or to one or more target ontological subject.
(BRI: the association strength is the intimacy value)
- the intimacy values as values of elements in a matrix,
In [0186]:
The association strength data structures usually in the form a matrix therefore is instrumental to build such cognitive networks for variety of tasks in general and for building neural nets in particular
- and generating an adjacency matrix based on the values of the elements; in response to that a sum of weights of elements in each row of the adjacency matrix is one, assigning weights to the elements, wherein a weight assigned to each of elements arranged diagonally in the adjacency matrix is greater than weights assigned to other elements;
In [0279] :
Referring to FIG. 3 now, it is to show that any composition of ontological subjects can in principal be represented by a graph which in this preferred embodiment shown as an asymmetric graph. The exemplified graph is corresponded to one of the exemplary "association strength matrix", i.e. an ASM, as representative of its adjacency matrix.
In [0283] :
The association strength matrix could be regarded as the adjacency matrix of any graphs such as social graphs or any network of any thing. For instance the graphs can be built representing the relations between the concepts and entities or any other desired set of OSs in a special area of science, market, industry or any "body of knowledge".
In [0134]:
The first type of a "value significance measure" is defined as a function of "Frequency of Occurrences" of OS.sub.i.sup.k is called here FO.sub.i.sup.k|l and can be given by:
PNG
media_image4.png
22
363
media_image4.png
Greyscale
wherein
F
O
i
k
|
l
i
s obtained by counting the occurrences of OSs of the particular order, e.g. counting the appearances of particular word in the text or counting its total occurrences in the partitions, or more conveniently be obtained from the
C
O
M
k
|
l
(the elements on the main diagonal of the
C
O
M
k
|
l
) or by using Eq. 4, or any other way of counting the occurrences of OS.sub.i.sup.k in the desired partitions of the composition.
In [0302]:
at lease one of the algorithms of calculating one of the measures in order to assign a value on the part or partitions of the compositions and based on the assigned value process one or more of the partitions or OSs of the particular order as an output in the form of a service or data. The output could be simply one or more tags or OS/s that the input composition can be characterized with, i.e. significant keywords of the composition.
In [0303]:
As another example, the output or outcome of the investigator of FIG. 10, could be to provide the partitions of the input composition which have exhibited intrinsic value significances of above a predetermined threshold. Another output could be the novel parts or the OSs of the compositions that scored a predetermined level of a particular type of novelty value significance
In [0109]:
Assuming we have an input composition of ontological subjects, e.g. an input text, the "Participation Matrix" (PM) is a matrix indicating the participation of one or more ontological subjects of particular order in one or more partitions of the composition. In other words in terms of our definitions, PM indicate the participation of one or more lower order OS into one or more OS of higher or the same order.
PNG
media_image5.png
17
358
media_image5.png
Greyscale
In [0122] :
For example the PMs, ASMs, OSM or co-occurrences of the ontological subjects etc. can be represented by a matrix, sparse matrix, table, database rows, no sql databases, JSON, dictionaries and the like which can be stored in various forms of data structures.
In [0186]:
data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net)
in [0186]:
The relatedness is measured by one or more of the above measures and partitions that exhibited an association strength value greater (or sometimes smaller) than a predetermined threshold to a particular OS, can be grouped or clustered together. Further these data can be readily used to build a neural network type system (for learning, reasoning etc.) whose edge/connection weights can be obtained from the data of association strengths of the ontological subjects (e.g. the node of a neural net).
In [0187]:
a number of partitions of the composition or the BOK that have exhibited a predetermined threshold of relative association strength or predetermined criteria of satisfying enough association strength to a target subject or to each other can be categorized or being clustered as group together
categorized or being clustered as group together.
in [0248:
In another aspect the novelty is observed in relation or combination with other OSs since novelty could occurs in a context and therefore in relation to other ontological subjects. The stand alone or the intrinsic "novelty value significance value" in this case is defined as sum of the novelty that an OS will have with a desired number of other OSs.
In [0249]:
These measures of novelty are intrinsic since it adds up all the pair-wise novelty values for each OS.sup.k so that a NVSM type 2 can be defined as:
PNG
media_image6.png
32
543
media_image6.png
Greyscale
wherein the pair-wise novelty measures are summed over the column (i.e. the j subscript).
In [0230] :
When there are multiple OSs of interest the pair-wise value significances can be used in combination and perhaps with various weight to achieve the same filtering effect for a set of OSs.
- and obtaining a training semantic vector corresponding to each element in the adjacency matrix, and obtaining the training encoding vectors by calculating a product of the training semantic vector and a value of each element after assigning by a graph convolutional network.
FIG. 3: shows one exemplary embodiment of a directed asymmetric network or graph corresponding to a composition of ontological subjects.
In [0179]:
using the above relational rasm one can conveniently find the most related partitions of a composition to one or more target OS for the desired goal of the investigation (e.g quick retrieval of documents, sentences, or paragraphs with high semantic relevancy).
In [0257] :
In accordance with another aspect of the methods of investigation of the compositions of ontological subject of the present invention, the participation matrix can, for instance, routinely being transformed to other types of objects or participation matrices by operating one or more vector or matrices on the PM. For example one can multiply the PM by a diagonal matrix (M by M) from the right side whose diagonal values are the reciprocal of the number of constituent OSs of order k in the partitions or the higher order OS of order l.
In [0254]:
Moreover all these matrices (e.g. such as PM, COM, ASM/s, RASM, RVSMs NVSM, RNVSMs etc.) can be regarded as an adjacency matrix for a corresponding graph wherein the matrix carry the data of the connectivity between the nodes or objects of the graph
In [0040]:
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG , CAO, YE and Hatami.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network.
Hatami teaches representing intimacy using matrix.
One of ordinary skill would have motivation to combine Weiss , WANG, CAO and YE that can
Reduce the training iteration and resources for training using association strength (Hatami [0186])
Claims 6, 8, 14 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over
Tal Weiss et.al (hereinafter Weiss) US 2016/0350655 A1,
in view of Yasheng WANG(hereinafter WANG) US 2022/0147715 A1,
in view of BOKAI CAO et.al (hereinafter CAO) US 2016/0275401 A1.
in view of Jungo Kasao, et.al (hereinafter Kasai) End-to-end Graph-based TAG Parsing with Neural Network, Proceedings of NAACL-HLT 2018, pages 1181–1194.
In regard to claim 6: (Currently Amended)
Weiss discloses:
- A method for obtaining a tag, comprising:
obtaining corresponding materials in response to receiving an instruction for obtaining an interest tag;
in [0047]:
According to some embodiments, an unsupervised, or weakly supervised, learning process, executed by a word tokenization and spelling correction model of a system for Deep Learning may include: (1) receiving a string of one or more characters; (2) encoding and indexing the characters as a multi-value index; (3) embedding each character as a numbers vector; (4) entering a matrix of one or more character number vectors, as input, to a recurrent, or a convolutional, neural network language model, wherein the language model is adapted for: (i) parsing the data into words and tokenizing the words; (ii) correcting misspelled words; and/or (iii) auto-completing words;
In [0062]:
an unsupervised, or weakly supervised, learning process, executed by a word semantics derivation model of a system for Deep Learning may include: (1) receiving a set of one or more words (e.g. ‘words input’, ‘word set’, ‘sentence’); (2) entering the word set, as an input set, to a sequence classifying, deep multi-layered recurrent, and/or recursive, neural network based, word semantics derivation model, wherein the word semantics derivation model is adapted for: (i) weakly supervising the model learning by providing a substantially small amount of ‘right’ semantic taggings as learning examples to the model; (ii) assigning markup language semantic tags to at least some, and/or a subset, of the words; (3) repeating stages (1) and (2) one or more additional times, while utilizing stochastic gradient descend for learning ‘correct’ semantic tagging, and improving following taggings' outputs.
In [0090]:
According to some embodiments of the present invention, the word semantics derivation model may utilize dialog related metadata for, and/or as part of, words inputs tagging. According to some embodiments, the model may learn about dialog and discourse based on the context world of the ‘dialog's metadata’. According to some embodiments, dialog metadata types may include, but are not limited to, those pertaining to: the scope of the metadata and the hosting application's general context (e.g. traveling—flights, hotels etc.); specific contexts of the dialog, and/or utterance(s) within it, relevant to specific capabilities of the application's scope, such as the current interest of the user, derived, for example, from the application's page/function that the user is viewing/using; and/or positioning data such as latitude and longitude of the end user.
- obtaining tag vectors by inputting the new encoding vectors into a fully- connected network;
In [0010]; According to some embodiments of the present invention, the word semantics derivation model/machine training may be weakly-supervised (i.e. based on substantially small amounts of characterized data, followed by substantially large amounts of uncharacterized data) wherein ‘right’ examples (e.g. word sets semantically tagged correctly) constitute the initial input training data, thus yielding a supervised training phase for providing the model with initial ‘correct’ knowledge, which may be followed by an unsupervised training phase that is based on uncharacterized data inputs for training the model on the utilization of the accumulated knowledge for semantically tagging ‘incorrect’/un-encountered inputs. Both the supervised and unsupervised training phases may be repeated (in whole or in part) multiple times until sufficient accuracy and breadth of tagging are achieved.
In [0039]:
2) An auto-encoder (may also be referred to as a ‘Noise Insertion Module’) for injecting random character level modifications to variable length ‘character level input streams’, wherein the characters may include a space-between-token character; and/or (3) An unsupervised training mechanism for adjusting the neural network to learn correct variable length ‘character level output streams’, wherein correct variable length ‘character level output streams’ needs to be identical to respective original variable length ‘character level input streams’ prior to their random character level modifications
(BRI: an autoencoder is a type of fully connected NN that provides encoder and decoder using fully connected layer in both encode and decoder)
Weiss does not explicitly disclose:
- obtaining semantic vectors that comprise interest tags by representing features of the materials using a semantic enhanced representation frame;
However, WANG discloses:
- obtaining semantic vectors that comprise interest tags by representing features of the materials using a semantic enhanced representation frame;
In [0007]:
a text processing method is provided, including: obtaining target knowledge data, where the target knowledge data includes a first named entity, a second named entity, and an association between the first named entity and the second named entity;
in [0141]:
For an attribute of an entity, an “attribute-value pair” is used to describe an intrinsic characteristic, for example, a person has attributes such as age, height, and weight. Currently, the knowledge graph has been widely used to refer to various large-scale knowledge bases (knowledge bases).
In [0142]:
Entity: The entity refers to an object that is distinguishable and exists independently, for example, a person, a city, a plant, or a commodity. Everything in the world is constituted by concrete objects, which refer to entities, for example, “China”, “United States”, and “Japan”. The entity is a most basic element in the knowledge graph. There are different relations between different entities.
in [0124]:
Sequence labeling: A model needs to provide a classification category for each word in a sentence based on a context. For example, the sequence labeling is Chinese word segmentation, part-of-speech tagging, named entity recognition, or semantic role tagging.
(BRI: chinese word segmentation is directly related to part-of-speech (POS) tagging)
In [0024] :
By setting, in the training text, the second knowledge identifier used to indicate the named entity, the original processing model can be guided to inject knowledge and semantic information into the second knowledge identifier, and the model can be guided to focus on the named entity indicated by the second identifier or extract a local knowledge feature.
In [0033]:
in a possible implementation, the fused first text vector includes at least a part of information in the first knowledge data, and the fused first knowledge vector includes semantic background information of the training text.
In [0034] :
After the first text vector and the first knowledge vector are fused, the first text vector is fused with knowledge information, and the first knowledge vector is fused with semantic background information.
- and obtaining the interest tags by inputting the encoding vectors into a pre-trained tag recommendation model.
In [0025]:
With reference to the first aspect, in a possible implementation, the original fusion model is obtained through training based on the first knowledge data and preset pre-training text.
In [0131]:
Part-of-speech tagging (part-of-speech tagging): A part-of-speech (noun, verb, adjective, or the like) is assigned to each word in natural language text.
in [0210] :
if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector.
- obtaining new encoding vectors by inputting the encoding vectors into a forward network in the tag recommendation model;
In [0210] Optionally, if the first knowledge data includes structured knowledge, the first knowledge data may be encoded by using an existing knowledge encoding method (for example, translating embedding, TransE), and obtained encoded information is the first knowledge vector. Encoding the first knowledge data may be understood as converting the first knowledge data into a vector, for example, encoding the structured knowledge is converting the structured knowledge into a vector.
In [0008]:
According to the solution provided in this application, the target fusion model fuses the target text vector corresponding to the to-be-processed text and the target knowledge vector corresponding to the target knowledge data, and uses the obtained fused target text vector and/or the obtained fused target knowledge vector as input data for the target processing model
In [0120]:
In a training process, a neural network may correct values of parameters in an initial neural network model by using an error back propagation (back propagation, BP) algorithm, so that a reconstruction error loss of the neural network model becomes increasingly smaller. Specifically, an input signal is forward transferred until an error loss occurs in output
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss and WANG.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
One of ordinary skill would have motivation to combine Weiss and WANG that can improve the capability of understanding natural language by the target processing model and/or the target fusion model (WANG [0059]).
Weiss and WANG do not explicitly disclose:
- obtaining encoding vectors by aggregating social networks into the semantic vectors;
However, CAO discloses:
- obtaining encoding vectors by aggregating social networks into the semantic vectors;
In [0084]:
Using the plurality of social message and venue pairs, the server 104 computes (406) features based on meta-paths and geo-coordinate information. In some implementations, meta-paths are used to compute the features and the computed features include measures of geo features.
In [0085]:
the computation (406) is performed for a pair in the plurality of social message and venue pairs, first encoding the respective training social message in the pair as a label
Having encoded the label, the server 104 further identifies for the respective training social message corresponding training meta-paths to the respective venue in the pair.
In [0085]:
In some implementations, the path counts for different meta-paths are combined to create an overall feature matrix and the overall feature matrix is represented as the training feature vector.
In [0093]:
Different meta-paths usually represent different relation-ships among linked nodes with different semantic meanings.
In [0014]:
identifying for the respective training social message corresponding training meta-paths to the respective venue in the pair; encoding the corresponding training meta-paths to a corresponding training feature vector, wherein each element of the corresponding training feature vector includes a measure based on a respective type of the respective training social message connected to the respective venue in the pair; and giving the encoded labels and training feature vectors to the classifier for training.
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG and CAO.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
One of ordinary skill would have motivation to combine Weiss, WANG and CAO that can concatenate different types of meta-path features to achieve significant improvement over accuracy ([CAO [0132])
Weiss, Wang and CAO do not explicitly disclose:
- and parsing the tag vectors, and outputting the interest tags based on a probability threshold value of the tag recommendation model.
However, Kasai discloses:
- and parsing the tag vectors, and outputting the interest tags based on a probability threshold value of the tag recommendation model.
In [Abstract, Page 1181]:
We present a graph-based Tree Adjoining Grammar (TAG) parser that uses BiL-STMs, highway connections, and character-level CNNs.
In [2, Page 1182]:
TAG parsing can be decomposed into supertagging and parsing. Supertagging assigns to words elementary trees (supertags) chosen from a finite set, and parsing determines how these elementary trees can be combined to form a derivation tree that yield the observed sentence.
In [2.2.2, Page 1184]:
We first perform unlabeled arc-factored scoring using the final output vectors from the BiLSTMs, and then label the resulting arcs. Specifically, suppose that we score edges coming into the ith word in a sentence i.e. assigning scores to the potential parents of the ith word. Denote the final output vector from the BiLSTM for the kth word by hk and suppose that hk is d-dimensional. Then, we produce two vectors from two separate multilayer perceptrons (MLPs) with the ReLU activation:
PNG
media_image7.png
82
362
media_image7.png
Greyscale
where
h
k
a
r
c
-
d
e
p
and
h
k
a
r
c
-
h
e
a
d
are
d
a
r
c
dimensional vectors that represent the kth word as a dependent and a head respectively.
In [5.2, Page 1188]:
The unbounded dependency corpus (Rimell et al., 2009) specifically evaluates parsers on unbounded dependencies, which involve a constituent moved from its original position, where an unlimited number of clause boundaries can intervene. The corpus comprises 7 constructions: object extraction from a relative clause (ObRC), object extraction from a reduced relative clause (ObRed), subject extraction from a relative clause (SbRC), free relatives (Free), object wh-questions (ObQ), right node raising (RNR), and subject extraction from an embedded clause (SbEm)
(BRI: object wh-questions" can be considered an "interest tag" within the specific domain of linguistics, grammar, and language education)
In [A.3, Page 1192]:
Small Clauses The UDR corpus has inconsistency with regards to small clauses. UDR gives an analysis that a small clause contains a subject and a complement as in (nsubj, guy, liar) “the guy who I call a liar.” in the subject embedded constructions. However, in the object question and object free relative constructions, a small clause is analyzed as two arguments of the verb. For instance, UDR specifies (what, adopted, dobj) in “we adopted what I would term pseudo-capitalism.” To solve this problem we add an arc from the head of the matrix clause to the subject in a small clause with label 1
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG and CAO.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
Kasai teaches parsing of tag vector,
One of ordinary skill would have motivation to combine Weiss , WANG , CAO, and Kasai and Kasai that can use a join modeling (tagging and parsing) to improve the performance of the tasks (Kasai [I, Page 1182]).
In regard to claim 8: (Currently Amended)
Weiss, WANG, and CAO do not explicitly disclose:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
However, YE discloses:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
In [0114]:
A large number of training samples may be obtained using the collected raw data, and the training samples may be divided according to a preset division ratio to build a training set, a verification set, and a test set. For example, the division ratio may be 60%, 20%, and 20%. The training set may be used to train the model; the validation set may be used to validate the trained model and select a model with high accuracy; and the test set may be used to further test and optimize the model selected by the validation set.
Weiss , WANG, CAO, and YE do not explicitly disclose:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
However, Kasai discloses:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
In [Abstract, Page 1181]:
We present a graph-based Tree Adjoining Grammar (TAG) parser that uses BiL-STMs, highway connections, and character-level CNNs.
In [2, Page 1182]:
TAG parsing can be decomposed into supertagging and parsing. Supertagging assigns to words elementary trees (supertags) chosen from a finite set, and parsing determines how these elementary trees can be combined to form a derivation tree that yield the observed sentence.
In [2.2.2, Page 1184]:
We first perform unlabeled arc-factored scoring using the final output vectors from the BiLSTMs, and then label the resulting arcs. Specifically, suppose that we score edges coming into the ith word in a sentence i.e. assigning scores to the potential parents of the ith word. Denote the final output vector from the BiLSTM for the kth word by hk and suppose that hk is d-dimensional. Then, we produce two vectors from two separate multilayer perceptrons (MLPs) with the ReLU activation:
PNG
media_image7.png
82
362
media_image7.png
Greyscale
where
h
k
a
r
c
-
d
e
p
and
h
k
a
r
c
-
h
e
a
d
are
d
a
r
c
dimensional vectors that represent the kth word as a dependent and a head respectively.
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
In [2.2.2, Page 1184]:
Given the head prediction of each word in the sentence, we assign labeling scores using vectors obtained from two additional MLP with ReLU. For the kth word, we obtain:
PNG
media_image8.png
66
291
media_image8.png
Greyscale
Let pi be the index of the predicted head of the ith word, and r be the number of dependency relations in the dataset. Then, the probability distribution `i over the possible dependency relations of the arc pointing from the pi th word to the ith word is calculated by
PNG
media_image9.png
80
475
media_image9.png
Greyscale
PNG
media_image10.png
57
486
media_image10.png
Greyscale
(BRI: the dependency relations is the threshold as it may represents the cutoff within the probability for classifying the dependency)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, CAO, YE and Kasai.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches a ratio of the first interest tags to the interest tags,
Kasai teaches parsing and activation function.
One of ordinary skill would have motivation to combine Weiss , WANG , YE and Kasai that can use a join modeling (tagging and parsing) to improve the performance of the tasks
(Kasai [I, Page 2]).
In regard to claim 14: (Original)
Weiss discloses:
- An electronic device, comprising: a processor; and a memory communicatively coupled to the processor; wherein the memory is configured to store instructions executable by the processor, and the processor is configured to perform the method
In [0024], in [0028]
In regard to claim 20: (Original)
Weiss discloses:
- A non-transitory computer-readable storage medium having computer instructions stored thereon. wherein the computer instructions are configured to cause a computer to implement the method
In [0028], in [0029]:
(BRI: all semiconductor computer readable is non-transitory)
Claims 5, 13, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over
Tal Weiss et.al (hereinafter Weiss) US 2016/0350655 A1,
in view of Yasheng WANG(hereinafter WANG) US 2022/0147715 A1,
in view of BOKAI CAO et.al (hereinafter CAO) US 2016/0275401 A1,
in view of Tangzhi YE (hereinafter YE) US 2021/0027170 A1.
further in view of Jungo Kasao, et.al (hereinafter Kasai) End-to-end Graph-based TAG Parsing with Neural Network, Proceedings of NAACL-HLT 2018, pages 1181–1194.
In regard to claim 5: (Original)
Weiss, WANG and CAO do not explicitly disclose:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
However, YE discloses:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
In [0114]:
A large number of training samples may be obtained using the collected raw data, and the training samples may be divided according to a preset division ratio to build a training set, a verification set, and a test set. For example, the division ratio may be 60%, 20%, and 20%. The training set may be used to train the model; the validation set may be used to validate the trained model and select a model with high accuracy; and the test set may be used to further test and optimize the model selected by the validation set.
Weiss , WANG, CAO, and YE do not explicitly disclose:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
However, Kasai discloses:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
In [Abstract, Page 1181]:
We present a graph-based Tree Adjoining Grammar (TAG) parser that uses BiL-STMs, highway connections, and character-level CNNs.
In [2, Page 1182]:
TAG parsing can be decomposed into supertagging and parsing. Supertagging assigns to words elementary trees (supertags) chosen from a finite set, and parsing determines how these elementary trees can be combined to form a derivation tree that yield the observed sentence.
In [2.2.2, Page 1184]:
We first perform unlabeled arc-factored scoring using the final output vectors from the BiLSTMs, and then label the resulting arcs. Specifically, suppose that we score edges coming into the ith word in a sentence i.e. assigning scores to the potential parents of the ith word. Denote the final output vector from the BiLSTM for the kth word by hk and suppose that hk is d-dimensional. Then, we produce two vectors from two separate multilayer perceptrons (MLPs) with the ReLU activation:
PNG
media_image7.png
82
362
media_image7.png
Greyscale
where
h
k
a
r
c
-
d
e
p
and
h
k
a
r
c
-
h
e
a
d
are
d
a
r
c
dimensional vectors that represent the kth word as a dependent and a head respectively.
In [5.2, Page 1188]:
The unbounded dependency corpus (Rimell et al., 2009) specifically evaluates parsers on unbounded dependencies, which involve a constituent moved from its original position, where an unlimited number of clause boundaries can intervene. The corpus comprises 7 constructions: object extraction from a relative clause (ObRC), object extraction from a reduced relative clause (ObRed), subject extraction from a relative clause (SbRC), free relatives (Free), object wh-questions (ObQ), right node raising (RNR), and subject extraction from an embedded clause (SbEm)
(BRI: object wh-questions" can be considered an "interest tag" within the specific domain of linguistics, grammar, and language education)
In [A.3, Page 1192]:
Small Clauses The UDR corpus has inconsistency with regards to small clauses. UDR gives an analysis that a small clause contains a subject and a complement as in (nsubj, guy, liar) “the guy who I call a liar.” in the subject embedded constructions. However, in the object question and object free relative constructions, a small clause is analyzed as two arguments of the verb. For instance, UDR specifies (what, adopted, dobj) in “we adopted what I would term pseudo-capitalism.” To solve this problem we add an arc from the head of the matrix clause to the subject in a small clause with label 1
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
In [2.2.2, Page 1184]:
Given the head prediction of each word in the sentence, we assign labeling scores using vectors obtained from two additional MLP with ReLU. For the kth word, we obtain:
PNG
media_image8.png
66
291
media_image8.png
Greyscale
Let pi be the index of the predicted head of the ith word, and r be the number of dependency relations in the dataset. Then, the probability distribution `i over the possible dependency relations of the arc pointing from the pi th word to the ith word is calculated by
PNG
media_image9.png
80
475
media_image9.png
Greyscale
PNG
media_image10.png
57
486
media_image10.png
Greyscale
(BRI: the dependency relations is the threshold as it may represents the cutoff within the probability for classifying the dependency)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, YE, and Kasai.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network.
Kasai teaches parsing and activation function.
One of ordinary skill would have motivation to combine Weiss , WANG , CAO, YE, and Kasai that can use a join modeling (tagging and parsing) to improve the performance of the tasks (Kasai [I, Page 1182]).
In regard to claim 13 : (Original)
Weiss, WANG and CAO do not explicitly disclose:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
However, YE discloses:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
In [0114]:
A large number of training samples may be obtained using the collected raw data, and the training samples may be divided according to a preset division ratio to build a training set, a verification set, and a test set. For example, the division ratio may be 60%, 20%, and 20%. The training set may be used to train the model; the validation set may be used to validate the trained model and select a model with high accuracy; and the test set may be used to further test and optimize the model selected by the validation set.
Weiss , WANG, CAO, YE do not explicitly disclose:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
However, Kasai discloses:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
In [Abstract, Page 1181]:
We present a graph-based Tree Adjoining Grammar (TAG) parser that uses BiL-STMs, highway connections, and character-level CNNs.
In [2, Page 1182]:
TAG parsing can be decomposed into supertagging and parsing. Supertagging assigns to words elementary trees (supertags) chosen from a finite set, and parsing determines how these elementary trees can be combined to form a derivation tree that yield the observed sentence.
In [2.2.2, Page 1184]:
We first perform unlabeled arc-factored scoring using the final output vectors from the BiLSTMs, and then label the resulting arcs. Specifically, suppose that we score edges coming into the ith word in a sentence i.e. assigning scores to the potential parents of the ith word. Denote the final output vector from the BiLSTM for the kth word by hk and suppose that hk is d-dimensional. Then, we produce two vectors from two separate multilayer perceptrons (MLPs) with the ReLU activation:
PNG
media_image7.png
82
362
media_image7.png
Greyscale
where
h
k
a
r
c
-
d
e
p
and
h
k
a
r
c
-
h
e
a
d
are
d
a
r
c
dimensional vectors that represent the kth word as a dependent and a head respectively.
In [5.2, Page 1188]:
The unbounded dependency corpus (Rimell et al., 2009) specifically evaluates parsers on unbounded dependencies, which involve a constituent moved from its original position, where an unlimited number of clause boundaries can intervene. The corpus comprises 7 constructions: object extraction from a relative clause (ObRC), object extraction from a reduced relative clause (ObRed), subject extraction from a relative clause (SbRC), free relatives (Free), object wh-questions (ObQ), right node raising (RNR), and subject extraction from an embedded clause (SbEm)
(BRI: object wh-questions" can be considered an "interest tag" within the specific domain of linguistics, grammar, and language education)
In [A.3, Page 1192]:
Small Clauses The UDR corpus has inconsistency with regards to small clauses. UDR gives an analysis that a small clause contains a subject and a complement as in (nsubj, guy, liar) “the guy who I call a liar.” in the subject embedded constructions. However, in the object question and object free relative constructions, a small clause is analyzed as two arguments of the verb. For instance, UDR specifies (what, adopted, dobj) in “we adopted what I would term pseudo-capitalism.” To solve this problem we add an arc from the head of the matrix clause to the subject in a small clause with label 1
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
In [2.2.2, Page 1184]:
Given the head prediction of each word in the sentence, we assign labeling scores using vectors obtained from two additional MLP with ReLU. For the kth word, we obtain:
PNG
media_image8.png
66
291
media_image8.png
Greyscale
Let pi be the index of the predicted head of the ith word, and r be the number of dependency relations in the dataset. Then, the probability distribution `i over the possible dependency relations of the arc pointing from the pi th word to the ith word is calculated by
PNG
media_image9.png
80
475
media_image9.png
Greyscale
PNG
media_image10.png
57
486
media_image10.png
Greyscale
(BRI: the dependency relations is the threshold as it may represents the cutoff within the probability for classifying the dependency)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, YE and Kasai.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
CAO teaches aggregation of social networks.
YE teaches double-layer neural network.
Kasai teaches parsing and activation function.
One of ordinary skill would have motivation to combine Weiss , WANG , YE and Kasai that can use a join modeling (tagging and parsing) to improve the performance of the tasks
(Kasai [I, Page 1182]).
In regard to claim 19: (Original)
Weiss, WANG and CAO do not explicitly disclose:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
However, YE discloses:
- and determining first interest tags corresponding to the interest tags in the training tag vectors, calculating a ratio of the first interest tags to the interest tags,
In [0114]:
A large number of training samples may be obtained using the collected raw data, and the training samples may be divided according to a preset division ratio to build a training set, a verification set, and a test set. For example, the division ratio may be 60%, 20%, and 20%. The training set may be used to train the model; the validation set may be used to validate the trained model and select a model with high accuracy; and the test set may be used to further test and optimize the model selected by the validation set.
Weiss , WANG, CAO and YE do not explicitly disclose:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
However, Kasai discloses:
- obtaining interest tags in the training tag vectors by parsing the training tag vectors by an activation function;
In [Abstract, Page 1181]:
We present a graph-based Tree Adjoining Grammar (TAG) parser that uses BiL-STMs, highway connections, and character-level CNNs.
In [2, Page 1182]:
TAG parsing can be decomposed into supertagging and parsing. Supertagging assigns to words elementary trees (supertags) chosen from a finite set, and parsing determines how these elementary trees can be combined to form a derivation tree that yield the observed sentence.
In [2.2.2, Page 1184]:
We first perform unlabeled arc-factored scoring using the final output vectors from the BiLSTMs, and then label the resulting arcs. Specifically, suppose that we score edges coming into the ith word in a sentence i.e. assigning scores to the potential parents of the ith word. Denote the final output vector from the BiLSTM for the kth word by hk and suppose that hk is d-dimensional. Then, we produce two vectors from two separate multilayer perceptrons (MLPs) with the ReLU activation:
PNG
media_image7.png
82
362
media_image7.png
Greyscale
where
h
k
a
r
c
-
d
e
p
and
h
k
a
r
c
-
h
e
a
d
are
d
a
r
c
dimensional vectors that represent the kth word as a dependent and a head respectively.
In [5.2, Page 1188]:
The unbounded dependency corpus (Rimell et al., 2009) specifically evaluates parsers on unbounded dependencies, which involve a constituent moved from its original position, where an unlimited number of clause boundaries can intervene. The corpus comprises 7 constructions: object extraction from a relative clause (ObRC), object extraction from a reduced relative clause (ObRed), subject extraction from a relative clause (SbRC), free relatives (Free), object wh-questions (ObQ), right node raising (RNR), and subject extraction from an embedded clause (SbEm)
(BRI: object wh-questions" can be considered an "interest tag" within the specific domain of linguistics, grammar, and language education)
In [A.3, Page 1192]:
Small Clauses The UDR corpus has inconsistency with regards to small clauses. UDR gives an analysis that a small clause contains a subject and a complement as in (nsubj, guy, liar) “the guy who I call a liar.” in the subject embedded constructions. However, in the object question and object free relative constructions, a small clause is analyzed as two arguments of the verb. For instance, UDR specifies (what, adopted, dobj) in “we adopted what I would term pseudo-capitalism.” To solve this problem we add an arc from the head of the matrix clause to the subject in a small clause with label 1
- determining a probability threshold value of the tag recommendation model, and obtaining the tag recommendation model whose output tag probability value is greater than or equal to the probability threshold value.
In [2.2.2, Page 1184]:
Given the head prediction of each word in the sentence, we assign labeling scores using vectors obtained from two additional MLP with ReLU. For the kth word, we obtain:
PNG
media_image8.png
66
291
media_image8.png
Greyscale
Let pi be the index of the predicted head of the ith word, and r be the number of dependency relations in the dataset. Then, the probability distribution `i over the possible dependency relations of the arc pointing from the pi th word to the ith word is calculated by
PNG
media_image9.png
80
475
media_image9.png
Greyscale
PNG
media_image10.png
57
486
media_image10.png
Greyscale
(BRI: the dependency relations is the threshold as it may represents the cutoff within the probability for classifying the dependency)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Weiss, WANG, YE , and Kasai.
Weiss teaches collection of data and training of tags.
WANG teaches semantic representation.
Kasai teaches parsing and activation function.
One of ordinary skill would have motivation to combine Weiss , WANG , CAO, and Kasai that can use a join modeling (tagging and parsing) to improve the performance of the tasks
(Kasai [I, Page 2]).
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the
examiner should be directed to TIRUMALE KRISHNASWAMY RAMESH whose telephone number is (571)272-4605. The examiner can normally be reached by phone.
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 on phone (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.
/TIRUMALE K RAMESH/Examiner, Art Unit 2121
/Li B. Zhen/Supervisory Patent Examiner, Art Unit 2121
Prosecution Insights