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 .
Examiner’s Note
The Examiner encourages Applicant to schedule an interview to discuss issues related to, for example, the rejections noted below under 35 U.S.C § 112 and § 103.
Providing supporting paragraph(s) for each limitation of amended/new claim(s) in Remarks is strongly requested for clear and definite claim interpretations by Examiner.
Priority
Acknowledgment is made of applicant's claim for the present application filed on 01/09/2023.
Claim Objections
Claim(s) 1-18 is/are objected to because of the following informalities.
Claim(s) 1 is/are objected to because of the following informalities:
it appears that “a first and second output” (line 9) needs to read “a first output and a second output” or something else. Appropriate correction is required. In addition, claim(s) 9, 18 is/are objected to for the same reason.
it appears that “the first and second output” (line 11) needs to read “the first and second outputs” or something else. Appropriate correction is required. In addition, claim(s) 9, 18 is/are objected to for the same reason.
it appears that “the outputs” (line 12) needs to read “the first and second outputs” or something else. Appropriate correction is required. In addition, claim(s) 9, 18 is/are objected to for the same reason.
Claim(s) 3 is/are objected to because of the following informalities: it appears that “the sequence to sequence model” (line 4) needs to read “the sequence to sequence deep learning model” or something else. Appropriate correction is required. In addition, claim(s) 11 is/are objected to for the same reason.
Claim(s) 6 is/are objected to because of the following informalities: it appears that “the multi-task learning” (line 4) needs to read “the deep multi-task learning” or something else. Appropriate correction is required. In addition, claim(s) 14 is/are objected to for the same reason.
Claim(s) 8 is/are objected to because of the following informalities: it appears that “the output nodes” (line 4) needs to read “the at least some output nodes” or something else. Appropriate correction is required. In addition, claim(s) 16 is/are objected to for the same reason.
Claim(s) 18 is/are objected to because of the following informalities: it appears that “concatenating” (line 15) needs to read “concatenate” or something else. Appropriate correction is required.
Claim(s) 1, 3, 6, 8-9, 11, 14, 16, 18 each recite(s) limitations that raise issues of indefiniteness as set forth above, and their dependent claims are objected to at least based on their direct and/or indirect dependency from the claims listed above. Appropriate explanation and/or amendment is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim(s) 3-18 is/are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim(s) 3 recite(s) the limitation “the metadata” (the last line). There is insufficient antecedent basis for this limitation in the claim. It is not clear if it indicates “metadata” (claim 1, line 4) or “metadata” (claim 1, line 5), or something else. It appears it may need to read “metadata”, or something else. For the purposes of examination, “metadata” is used. In addition, claim(s) 11 is/are rejected for the same reason.
Claim(s) 5 recite(s) the limitation “the outputs” (line 5). There is insufficient antecedent basis for this limitation in the claim. It is not clear if it indicates “the outputs” (claim 1, line 12) or “outputs” (claim 5, line 2), or something else. It appears it may need to read “outputs”, or something else. For the purposes of examination, “outputs” is used. In addition, claim(s) 13 is/are rejected for the same reason.
Claim(s) 6 recite(s) the limitation “the task specific layer applies deep multi-task learning and comprises an input layer, a hidden layer and an output layer” (line 1), and its analogous claim 14 recites “the multi-task learning model performs deep multi-task learning and comprises an input layer, a hidden layer and an output layer”. In addition, Fig 1A shows “Input Layer 142”. It is not clear if “the task specific layer” of claim 6 has another separate input layer other than “Input Layer 142”. Thus, it appears that “the task specific layer” (claim 6) may need to read “the multi-task learning model” like its analogous claim 14, or something else. For the purposes of examination, “the multi-task learning model” is used for claim 6.
Claim(s) 7 recite(s) the limitation “the metadata classification” (line 4). There is insufficient antecedent basis for this limitation in the claim. It is not clear if it indicates “metadata classification” (claim 1, line 1) or one of “separate metadata classification” (claim 1, line 18) or one of “different metadata classification” (claim 1, line 21), or something else. It appears it may need to read “a metadata classification”, or something else. For the purposes of examination, “a metadata classification” is used. In addition, claim(s) 15 is/are rejected for the same reason.
Claim(s) 9 recite(s) the limitation “the metadata classifications” (5th last line). There is insufficient antecedent basis for this limitation in the claim. It is not clear if it indicates “separates metadata classifications” (9th last line) or “different metadata classifications” (6th last line), or something else. It appears it may need to read “metadata classifications”, or something else. For the purposes of examination, “data elements” is used. In addition, “the metadata classifications” (claim 9, 3rd last line) is/are rejected for the same reason. In addition, claim(s) 17 (line 2), 18 (two times) is/are rejected for the same reason.
Claim(s) 11 recite(s) the limitation “the data elements” (line 5). There is insufficient antecedent basis for this limitation in the claim. It is not clear if it indicates “data elements” (claim 9, line 2) or “data elements” (claim 11, line 2), or something else. It appears it may need to read “data elements”, or something else. For the purposes of examination, “data elements” is used.
Claim(s) 13 recite(s) the limitation “the optimization layer” (line 5). There is insufficient antecedent basis for this limitation in the claim. It is not clear what it is referring to. It appears that it needs to read “an optimization layer” or something else. For the purposes of examination, “an optimization layer” is used.
Claim(s) 3, 5, 6-7, 9, 11, 13, 15, 17-18 each recite(s) limitations that raise issues of indefiniteness as set forth above, and their dependent claims are rejected at least based on their direct and/or indirect dependency from the claims listed above. Appropriate explanation and/or amendment is required.
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 1-8 are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter. The claim(s) does/do not fall within at least one of the four categories of patent eligible subject matter because the claimed invention describes a computer program without further limiting the computer program with any structural limitations in its body. Claim 1 just recites a machine learning model comprising a multi-task learning model having multiple layers, which may be interpreted as a software. Thus, under the broadest reasonable interpretation, the machine learning model is a collection of instructions (software per se) and does not fall within at least one of the four statutory categories. A computer program is merely a set of instructions capable of being executed by a computer, the computer program itself is not a process and without the computer-readable medium, the computer program’s functionality is considered a non-statutory functional descriptive material. See MPEP § 2106(I) and MPEP § 2111.05. The examiner suggests using the terms “hardware processor” and “non-transitory memory” toward a system claim, or something else.
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.
Claim 18 is/are rejected under 35 U.S.C. 101 because the claimed invention is directed to non-statutory subject matter. The claims does/do not fall within at least one of the four categories of patent eligible subject matter (Step 1).
The claimed “non-transient storage medium” is a propagating signal when viewed in light of as-filed specification paragraph [0132]. The specification reads “It should be understood, however, that computer-readable storage media and data storage media do not include connections, carrier waves, signals, or other transient media, but are instead directed to non-transient, tangible storage media.” However, the specification does not clearly limit the term “non-transient storage medium” to non-transitory embodiments. Rather, under the broadest reasonable interpretation (BRI), “non-transient storage medium” can be transitory or non-transitory. Thus, the claimed “non-transient storage medium” may include transitory forms of signal transmission (often referred to as “signals per se”), such as a propagating electrical or electromagnetic signal or carrier wave. Signal per se does not fall within at least one of the four statutory categories. The examiner suggests using the term “non-transitory computer-readable storage media” or something else.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
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(s) 1, 9, 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure)
Regarding claim 1
Pavia teaches
A machine learning model for metadata classification and labelling comprising:
(Pavia [fig(s) 2] [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table. To gauge generality, we have evaluated it on four large-scale datasets - CORD-19 [34] and Web Data Commons (WDC) [29], Lehmberg [19]. WDC has more than 15 million tables in English respectively and hundred thousands of different sources, storing data and forming tables in different ways. We have designed an ensemble, combining either regular Long Short Term Memory (LSTM) or Bi-directional LSTM [25] Recurrent Neural Network (RNN) [26] with a keras-embedding layer and a Naive Bayes Classifier. Most of the previous approaches are limited to only on table cell-level analysis, whereas we take into account the cell context (i.e. the whole tuple or a column) along with the position of the cell and the surroundings of that cell in the table. The first-layer model is order-sensitive as order matters for the terms inside a cell (i.e. First Name is different from Name First). However, the second-layer model is order-insensitive as the order of cells does not matter in a tuple. For example, a tuple having artist and then album value is the same as first album then artist. Finally, we designed an algorithm that using our hybrid metadata classification ensemble can distinguish different kinds of metadata in a table - row/column-based or hierarchical.”;)
a multi-task learning model comprising:
(Paiva [fig(s) 2] [sec(s) Abs] “Metadata location and classification is an important problem for large-scale structured datasets. For example, Web tables [29] have hundreds of millions of tables, but often have missing or incorrect labels for rows (or columns) with attribute names. Such errors [24] significantly complicate all data management tasks such as query processing, data integration, indexing, etc.” [sec(s) 3] “The final output is binary, whether the input sequence of cells, i.e. a tuple or column is a Metadata row/column or not. The output of the Naive Bayes Classifier is then used with the decision-tree rules to produce the final output having the Metadata type.” [sec(s) 5] “Except many fundamental data management activities, naturally dependent on metadata such as query processing, data integration, warehousing, replication and many, another important application of Metadata location and classification is distinguishing relational and non-relational tables. In our datasets, we had both kinds of tables and we have evaluated performance of our ensemble on this task as well. … We finish by discussing Table 4, which illustrates the comparative evaluation results of our ensemble for the same classification task of three different kinds of Metadata - Metadata on the Top, Metadata on the Left and Hierarchical Metadata. This experiment is performed on 4 different datasets and the accuracy of locating these metadata types is presented.”;)
(Note: Hereinafter, if a limitation has bold brackets (i.e. [·]) around claim languages, the bracketed claim languages indicate that they have not been taught yet by the current prior art reference but they will be taught by another prior art reference afterwards.)
an input layer for receiving a first textual input characterizing one aspect of metadata for an input data element; and receiving a [second] textual input characterizing another aspect of metadata for the input data element;
(Pavia [fig(s) 2] [sec(s) 3] “The model architecture is depicted in Figure 2. We use an entire table tuple or column as the input of our model and the model predicts whether it contains Metadata or not. We do not check each tuple or column, because to the best of our knowledge metadata is never stored in the middle, at the very bottom or as a rightmost table column, but if the dataset specifics requires that, our architecture does not prevent that. Hence, we take the first table tuple, the first column and the second tuple as input for the model and the model processes them. The second tuple is taken as input, because for non-relational tables with hierarchical metadata, the second row contains the second layer of the Metadata as depicted in Figure 1(b). Finally, to post-filter false-positives, we have designed a custom decision-tree model, based on the number of spaces in the column or row” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table. To gauge generality, we have evaluated it on four large-scale datasets - CORD-19 [34] and Web Data Commons (WDC) [29], Lehmberg [19]. WDC has more than 15 million tables in English respectively and hundred thousands of different sources, storing data and forming tables in different ways.”;)
an embedding layer of the input layer, for embedding the first and [second] textual inputs [separately and independently] to a format suitable for long short term memory (LSTM) neural networks and each provided to a separate LSTM model in the input layer to generate a first and second output of the input layer respectively;
(Pavia [fig(s) 2] “Embedding” [sec(s) 3] “The model architecture is depicted in Figure 2. We use an entire table tuple or column as the input of our model and the model predicts whether it contains Metadata or not. We do not check each tuple or column, because to the best of our knowledge metadata is never stored in the middle, at the very bottom or as a rightmost table column, but if the dataset specifics requires that, our architecture does not prevent that. Hence, we take the first table tuple, the first column and the second tuple as input for the model and the model processes them. The second tuple is taken as input, because for non-relational tables with hierarchical metadata, the second row contains the second layer of the Metadata as depicted in Figure 1(b). Finally, to post-filter false-positives, we have designed a custom decision-tree model, based on the number of spaces in the column or row. … In this section, we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table. To gauge generality, we have evaluated it on four large-scale datasets - CORD-19 [34] and Web Data Commons (WDC) [29], Lehmberg [19]. WDC has more than 15 million tables in English respectively and hundred thousands of different sources, storing data and forming tables in different ways.”;)
a shared layer for receiving the first and second output of the input layer from each LSTM and concatenating the outputs to form a concatenated output, the shared layer subsequently applying [hard parameter sharing for sharing model parameters including hidden layers across all tasks]; and
(Pavia [fig(s) 2] “LSTM”, “Concatenate & Flatten” [sec(s) 3] “The model architecture is depicted in Figure 2. We use an entire table tuple or column as the input of our model and the model predicts whether it contains Metadata or not. We do not check each tuple or column, because to the best of our knowledge metadata is never stored in the middle, at the very bottom or as a rightmost table column, but if the dataset specifics requires that, our architecture does not prevent that. Hence, we take the first table tuple, the first column and the second tuple as input for the model and the model processes them. The second tuple is taken as input, because for non-relational tables with hierarchical metadata, the second row contains the second layer of the Metadata as depicted in Figure 1(b). Finally, to post-filter false-positives, we have designed a custom decision-tree model, based on the number of spaces in the column or row. … In this section, we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have taken the output of the intermediate dense layer as an encoding of the cell and for a whole row or column, we concatenated all such encodings to form the feature vector, input to the Naive Bayes Classifier. Figure 2 illustrates the ensemble architecture.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table. To gauge generality, we have evaluated it on four large-scale datasets - CORD-19 [34] and Web Data Commons (WDC) [29], Lehmberg [19]. WDC has more than 15 million tables in English respectively and hundred thousands of different sources, storing data and forming tables in different ways.” [sec(s) 4] “We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”;)
a task specific layer for receiving the concatenated output [including the hard parameter sharing] to learn parameters specific to each task and classifying the concatenated output into at least one of a possible set of tasks corresponding to separate metadata classifications using a set of simultaneously trained classifiers, the multi-task learning model being a single model trained to simultaneously learn, during a training phase, multiple classification tasks corresponding to different metadata classifications.
(Pavia [fig(s) 2] “LSTM”, “Concatenate & Flatten” [sec(s) 3] “The model architecture is depicted in Figure 2. We use an entire table tuple or column as the input of our model and the model predicts whether it contains Metadata or not. We do not check each tuple or column, because to the best of our knowledge metadata is never stored in the middle, at the very bottom or as a rightmost table column, but if the dataset specifics requires that, our architecture does not prevent that. Hence, we take the first table tuple, the first column and the second tuple as input for the model and the model processes them. The second tuple is taken as input, because for non-relational tables with hierarchical metadata, the second row contains the second layer of the Metadata as depicted in Figure 1(b). Finally, to post-filter false-positives, we have designed a custom decision-tree model, based on the number of spaces in the column or row. … In this section, we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table. To gauge generality, we have evaluated it on four large-scale datasets - CORD-19 [34] and Web Data Commons (WDC) [29], Lehmberg [19]. WDC has more than 15 million tables in English respectively and hundred thousands of different sources, storing data and forming tables in different ways.” [sec(s) 4] “After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. To alleviate that we have used a dropout layer after the activation layer with a value of 0.2, known to be a balance between dropping too many features and degree of overfitting. There is a node having sigmoid as the activation function. Before the activation and dropout layers we have used a Dense layer that’s output serves as input for the Multinomial Naive Bayes Classifier. First we have trained the LSTM model as a binary classifier i.e. to classify a cell being a Metadata cell or not. We have taken the output feature vector of the dense layer having 256 nodes as an encoding of the input cell, i.e. the dimensionality of an encoded vector is 256. For a tuple or column, there are several cells. We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”;)
However, Pavia does not appear to explicitly teach:
an input layer for receiving a first textual input characterizing one aspect of metadata for an input data element; and receiving a [second] textual input characterizing another aspect of metadata for the input data element;
an embedding layer of the input layer, for embedding the first and [second] textual inputs [separately and independently] to a format suitable for long short term memory (LSTM) neural networks and each provided to a separate LSTM model in the input layer to generate a first and second output of the input layer respectively;
the shared layer subsequently applying [hard parameter sharing for sharing model parameters including hidden layers across all tasks]; and
a task specific layer for receiving the concatenated output [including the hard parameter sharing] to learn parameters specific to each task.
(Note: Hereinafter, if a limitation has one or more bold underlines, the one or more underlined claim languages indicate that they are taught by the current prior art reference, while the one or more non-underlined claim languages indicate that they have been taught already by one or more previous art references.)
Bao teaches
an input layer for receiving a first textual input characterizing one aspect of metadata for an input data element; and receiving a second textual input characterizing another aspect of metadata for the input data element;
(Bao [fig(s) 1] “Input Layer” [sec(s) IV] “SemEval2014 evaluation corpus. Semantic Evaluation (SemEval) is an important evaluation for STS task. The Sentences Involving Com-positional Knowledge (SICK) data set[8] released in SemEval 2014 is to evaluate semantic relatedness between the sentence pair. It consists of 10,000 English sentence pairs, [training, 4500], [development, 500], [test, 5000]. Each sentence has an similarity label ∈[1, 5], with 1 indicating that two sentences in a given pair are completely unrelated while 5 indicating that two sentences are completely related.” [sec(s) Abs] “Semantic Textual Similarity (STS) is important for many applications such as Plagiarism Detection (PD), Text Paraphrasing and Information Retrieval (IR).” [sec(s) III] “Attentive Siamese LSTM model proposed in our paper is shown in Figure 1, which contains two networks, LSTMa and LSTMb, which share the same weight. LSTMa and LSTMb process one of the sentence in a given pair. Each network has five layers, 1) input layer: input the given sentence pair, 2) embedding layer: represent words in a low dimension, 3) hidden layer: learn highlevel features, 4) attention layer: produce weight vector, 5) output layer: output predicting similarity (or label). Instead of inputting numerous of handcraft features, sentence pair with similarity and word embedding is needed in our model.”;)
an embedding layer of the input layer, for embedding the first and second textual inputs separately and independently to a format suitable for long short term memory (LSTM) neural networks and each provided to a separate LSTM model in the input layer to generate a first and second output of the input layer respectively;
(Bao [fig(s) 1] “Input Layer” [sec(s) IV] “SemEval2014 evaluation corpus. Semantic Evaluation (SemEval) is an important evaluation for STS task. The Sentences Involving Com-positional Knowledge (SICK) data set[8] released in SemEval 2014 is to evaluate semantic relatedness between the sentence pair. It consists of 10,000 English sentence pairs, [training, 4500], [development, 500], [test, 5000]. Each sentence has an similarity label ∈[1, 5], with 1 indicating that two sentences in a given pair are completely unrelated while 5 indicating that two sentences are completely related.” [sec(s) Abs] “Semantic Textual Similarity (STS) is important for many applications such as Plagiarism Detection (PD), Text Paraphrasing and Information Retrieval (IR).” [sec(s) III] “Attentive Siamese LSTM model proposed in our paper is shown in Figure 1, which contains two networks, LSTMa and LSTMb, which share the same weight. LSTMa and LSTMb process one of the sentence in a given pair. Each network has five layers, 1) input layer: input the given sentence pair, 2) embedding layer: represent words in a low dimension, 3) hidden layer: learn highlevel features, 4) attention layer: produce weight vector, 5) output layer: output predicting similarity (or label). Instead of inputting numerous of handcraft features, sentence pair with similarity and word embedding is needed in our model.”;)
the shared layer subsequently applying hard parameter sharing for sharing model parameters including hidden layers across all tasks; and
(Bao [fig(s) 1] “Input Layer” [sec(s) Abs] “Semantic Textual Similarity (STS) is important for many applications such as Plagiarism Detection (PD), Text Paraphrasing and Information Retrieval (IR).” [sec(s) III] “Wi, Wf, Wc, Wo, Ui, Uf, Uc, Uo is the weight matrices and bi, bf, bc, bo is the bias vector. Siamese LSTM architecture in (Mueller and Thyagarajan, 2016) has three layers. There are two LSTM networks where each process one of the sentences of the input sentence pair. Two LSTM share the same weights in Siamese architecture. … Attentive Siamese LSTM model proposed in our paper is shown in Figure 1, which contains two networks, LSTMa and LSTMb, which share the same weight. LSTMa and LSTMb process one of the sentence in a given pair. Each network has five layers, 1) input layer: input the given sentence pair, 2) embedding layer: represent words in a low dimension, 3) hidden layer: learn highlevel features, 4) attention layer: produce weight vector, 5) output layer: output predicting similarity (or label). Instead of inputting numerous of handcraft features, sentence pair with similarity and word embedding is needed in our model.”;)
a task specific layer for receiving the concatenated output including the hard parameter sharing to learn parameters specific to each task.
(Bao [fig(s) 1] “Input Layer” [sec(s) Abs] “Semantic Textual Similarity (STS) is important for many applications such as Plagiarism Detection (PD), Text Paraphrasing and Information Retrieval (IR).” [sec(s) III] “Wi, Wf, Wc, Wo, Ui, Uf, Uc, Uo is the weight matrices and bi, bf, bc, bo is the bias vector. Siamese LSTM architecture in (Mueller and Thyagarajan, 2016) has three layers. There are two LSTM networks where each process one of the sentences of the input sentence pair. Two LSTM share the same weights in Siamese architecture. … Attentive Siamese LSTM model proposed in our paper is shown in Figure 1, which contains two networks, LSTMa and LSTMb, which share the same weight. LSTMa and LSTMb process one of the sentence in a given pair. Each network has five layers, 1) input layer: input the given sentence pair, 2) embedding layer: represent words in a low dimension, 3) hidden layer: learn highlevel features, 4) attention layer: produce weight vector, 5) output layer: output predicting similarity (or label). Instead of inputting numerous of handcraft features, sentence pair with similarity and word embedding is needed in our model.”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia with the separate inputs and parameter sharing of Bao.
One of ordinary skill in the art would have been motived to combine in order to gain prediction performance improvement than the corresponding baseline model by proving that attentive models can capture important semantic information in sentences.
(Bao [sec(s) IV] “The first and most important thing to note is that in both SemEval and MSRP corpus, all of the attentive models showed better performance than the corresponding baseline models without any handcraft features and external sources. Pearson’s correlation of English sentence pair can reach 0.7832, which indicates predicting similarity and human annotated similarity is strong correlated, near perfect correlated. In MSRP task, AttSiaLSTM and AttSiaBiLSTM model gain 5.27% and 1.51% improvement than the corresponding baseline model. It proves that attentive Siamese LSTM model can capture important semantic information in a sentence, which is effective for our STS task. We note that Siamese LSTM model obtain better performance than Siamese bidirectional LSTM model. This conclusion is in keeping with (Mueller and Thyagarajan, 2016).”)
Regarding claim 9
The claim is a method claim corresponding to the machine learning model claim 1, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Pavia further teaches
automatically tagging the input data element with the metadata classifications; and
(Pavia [fig(s) 2] “(0/1)” [sec(s) 3] “we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have used 64 LSTM/Bi-LSTM units in this case followed by a fully connected layer, containing 256 dense nodes. Then we have added a dropout layer, usually used to avoid overfitting and finally, the output layer has one dense unit with the sigmoid function as the activation function, which outputs the probability of the input cell is a Metadata cell, however it does not generate the final decision for an entire tuple or column. … The final output is binary, whether the input sequence of cells, i.e. a tuple or column is a Metadata row/column or not. The output of the Naive Bayes Classifier is then used with the decision-tree rules to produce the final output having the Metadata type.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table.” [sec(s) 4] “After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. … We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”;)
communicating the tagged input data element with the metadata classifications to a requesting computing device, via a communication device coupled to the metadata classifier to process the tagged input data element.
(Pavia [sec(s) 4] “In this section, we first describe the training and test sets construction followed by the experimental evaluation on several large-scale datasets. Hardware: We run all our experiments on Amazon AWS EC2 p3.8xlarge instances, each having 4 NVIDIA® V100 Tensor Core GPUs, 32 vCPUs, 244 GiB of memory, 10 Gbps network. Software: For implementing the RNN model, we have used Keras [3], a popular python library for deep learning, having Tensorflow [8] framework as the backend. The second step i.e. the Naive Bayes classifier is implemented using Scikit-learn [7], a popular machine learning library for python. Throughout the experimentation and implementation, we have used python as the programming language.”;)
Regarding claim 18
The claim is a computer program product claim corresponding to the method claim 9, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the method claim.
Claim(s) 2, 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Erb et al. (US 2015/0120915 A1)
Regarding claim 2
The combination of Pavia, Bao teaches claim 1.
Bao further teaches
wherein the first textual input comprises: a [business] name field and a description field for the input data element, and the second textual input comprises: a [malicious code] field and a technical name field for the input data element.
(Bao [fig(s) 1] “Input Layer” [fig(s) 2] [sec(s) IV] “1) SemEval2014 evaluation corpus. Semantic Evaluation (SemEval) is an important evaluation for STS task. The Sentences Involving Com-positional Knowledge (SICK) data set[8] released in SemEval 2014 is to evaluate semantic relatedness between the sentence pair. It consists of 10,000 English sentence pairs, [training, 4500], [development, 500], [test, 5000]. Each sentence has an similarity label ∈[1, 5], with 1 indicating that two sentences in a given pair are completely unrelated while 5 indicating that two sentences are completely related. 2) SemEval2014 translated corpus. In addition, English annotated sentence pairs were translated into Chinese and Tibetan to create additional tracks in our experiments, which will be used for Chinese Mandarin and Tibetan Plagiarism Detection. Chinese sentences are translated from English via Google Translator1 and Tibetan sentences are translated from Chinese via Niu Translator2. For each language we got 10,000 sentence pairs in total. Two samples in English, Chinese Mandarin and Tibetan are shown in Figure 2. Related score of two samples are 1.7 and 4.5. 3) MSRP. Another paraphrasing corpus is Microsoft Research Paraphrase corpus (MSRP), including 5801 English sentence pairs with human annotated label. Training set contains 4076 sentence pairs and test set contains 1725 sentence pairs. 67% sentence pairs are paraphrasing sample and 33% are not.” [sec(s) Abs] “Semantic Textual Similarity (STS) is important for many applications such as Plagiarism Detection (PD), Text Paraphrasing and Information Retrieval (IR).” [sec(s) III] “Attentive Siamese LSTM model proposed in our paper is shown in Figure 1, which contains two networks, LSTMa and LSTMb, which share the same weight. LSTMa and LSTMb process one of the sentence in a given pair. Each network has five layers, 1) input layer: input the given sentence pair, 2) embedding layer: represent words in a low dimension, 3) hidden layer: learn highlevel features, 4) attention layer: produce weight vector, 5) output layer: output predicting similarity (or label). Instead of inputting numerous of handcraft features, sentence pair with similarity and word embedding is needed in our model.”;)
The combination of Pavia, Bao is combinable with Bao for the same rationale as set forth above with respect to claim 1.
However, the combination of Pavia, Bao does not appear to explicitly teach:
wherein the first textual input comprises: a [business] name field and a description field for the input data element, and the second textual input comprises: a [malicious code] field and a technical name field for the input data element.
Erb teaches
wherein the first textual input comprises: a business name field and a description field for the input data element, and the second textual input comprises: a malicious code field and a technical name field for the input data element.
(Erb [fig(s) 4]
PNG
media_image1.png
317
1291
media_image1.png
Greyscale
[par(s) 3] “Network policies can be used to prevent undesirable material from being retrieved by a client computer. Such material can include malicious code that detrimentally modifies the behaviour of the retrieving computer or adult-oriented material that is unsuitable for viewing by a child that has access to the computer, to name just a few examples.”[par(s) 127-162] “An IP address hash node is configured divert log entries to different downstream neighbouring nodes depending on requesting computers’ IP addresses 83 (FIG. 4) of the log entries. In other respects, the IP address hash node is the same as the group hash node. A username hash node is configured divert log entries to different downstream neighbouring nodes depending on the usernames 84 (FIG. 4) associated with requests in the log entries. In other respects, the username hash node is the same as the group hash node. A denied hash node is configured divert log entries to different downstream neighbouring nodes depending on the policy decisions 88 (FIG. 4) made for the requests in the log entries. In this embodiment, the policy decisions are “denied” and “allowed”, and accordingly, the denied hash node requires two downstream neighbouring nodes. In other respects, the denied hash node is the same as the group hash node. … A URL hash node is configured divert log entries to different downstream neighbouring nodes depending on the requested URLs 82 (FIG. 4) in the log entries. In other respects, the URL hash node is the same as the group hash node. … The setting indicates one or more criterions for matching one or more fields 82-88 of the log entries 80. The one or more criterions can be specified in the same or similar manner as with the filter nodes described above. For example, one criterion may be that the category 86 matches a specified string (e.g., “malicious”). Another criterion may be that the group 85 matches a specified string (e.g., “student”). Criterions may be combined. The aggregate node then, for each received packet, updates a data file with the running count of log entries that meet the specified criteria. Running counts may be recorded over a specified period and with time or”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao with the log entries of Erb.
One of ordinary skill in the art would have been motived to combine in order to process large amounts of log data quickly and efficiently, and allow for relatively quick and simple deployment and maintenance of a geographically distributed and scalable logging service.
(Erb [par(s) 163-169] “One of the advantages of the configurable graph structures described herein is that large amounts of log data can be processed quickly and efficiently. Moreover, configuring or updating the configurations of the graph structures that define the logging service allows for relatively quick and simple deployment and maintenance of a geographically distributed and scalable logging service.”)
Regarding claim 10
The claim is a method claim corresponding to the machine learning model claim 2, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Claim(s) 3, 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Gao et al. (A Neural Model for Method Name Generation from Functional Description) in view of Zhou et al. (Translating Natural Language Instructions for Behavioral Robot Indoor Navigation with Attention-History Based Attention)
Regarding claim 3
The combination of Pavia, Bao teaches claim 1.
However, the combination of Pavia, Bao does not appear to explicitly teach:
a sequence to sequence deep learning model for converting metadata fields for data elements into a recognizable format for the multi-task learning model, the sequence to sequence model having an encoder layer, an attention layer and a decoder layer for receiving the data elements as a sequence containing textual input representing a first domain including a name with acronyms and translating to a second domain including an understandable text for each data element, the understandable text provided as input to the multi-task learning model for further processing as a further aspect of the metadata.
Gao further teaches
a sequence to sequence deep learning model for converting metadata fields for data elements into a recognizable format for the multi-task learning model, the sequence to sequence model having an encoder layer, an attention layer and a decoder layer for receiving the data elements as a sequence containing textual input representing a first domain including a name with acronyms and translating to a second domain including an understandable text for each data element, the understandable text [provided as input to the multi-task learning model for further processing as a further aspect of the metadata].
(Gao [table(s) III] “am”, “pm” [sec(s) I] “Therefore, we formulate the problem as a sequence to sequence learning problem which targets at automatic method name subtoken sequence generation from complete natural language sentences. The basic sequence to sequence model utilizes RNN encoder to pass the final state of encoder to the decoder for generation, but the final state may not be effective in capturing the whole information of a long sentence. Therefore, we adopt the attention mechanism [3] in our model, which weights the different parts in input sequence for each subtoken in the output sequence” [sec(s) III] “For example, suppose that the function description of a method is “return id of this item”. We easily know that the name should be “get ”, but we are not sure whether the name is “get id”, “get item” or “get item id”. The copying mechanism explicitly models this problem and learns the word choosing convention from massive training instances.” [sec(s) V.B] “Then we also analyze the failed cases in test set, and find that although some of the generated method names are different from the ground truth, they are still reasonable. We observe that the generated method names sometimes convey the same meaning with the ground truth (e.g., has error vs. contains error), even if there might be different word choices. It is also frequently seen that the generated method name is an abbreviation or acronym of the ground truth, e.g., setDocumentId and setDocId, where Doc is short for Document. In addition, since the quality of Github projects varies, some of the method descriptions are not clear enough to generate correct method names. These examples indicate that exact match does not suffice to evaluate the true efficacy of the method name generation techniques. … In summary, we could answer the RQ1 and conclude that our encoder-decoder model with attention and copying mechanisms could learn the naming conventions and suggest readable method names.”; e.g., “am” and/or “pm” and/or “id” read(s) on “acronyms”.)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao with the sequence to sequence deep learning model of Gao.
One of ordinary skill in the art would have been motived to combine in order to generate meaningful and accurate prediction and achieve significant improvement over the state-of-the-art baseline models.
(Gao [sec(s) Abs] “Our experiments show that our method can generate meaningful and accurate method names and achieve significant improvement over the state-of-the-art baseline models. We also address the cold-start problem using a training trick to utilize big data in Github for specific projects.”)
However, the combination of Pavia, Bao, Gao does not appear to explicitly teach:
the understandable text [provided as input to the multi-task learning model for further processing as a further aspect of the metadata].
Zhou teaches
the understandable text provided as input to the multi-task learning model for further processing as a further aspect of the metadata.
(Zhou [fig(s) 1-3] [sec(s) 7] “In this paper, based on the encoder-decoder model with attention mechanism, we propose an attention-history reader network for capturing the patterns in the attention history to improve translating natural translations to a high-level sequence of behaviors that is understandable and executable by robots. we also use a two-layer soft-attention for fusing multimodal inputs (instructions and graph representations). Experiments show that our method improved model performance.” [sec(s) 1] “In this work, we add an attention-history reader network using GRU network to the typical attentional encoder decoder model trying to capture the translating patterns from attention distributions in the past. Inspired by coverage mechanism [5, 6] and its variants [7, 8] which had performance improvement in machine translation and abstract summarization tasks, we add an attention-history reader network to the attentional encoder decoder model and a coverage vector is input to the additional network. The state of attention-history reader network is additionally given to the attention mechanism together with the current decoder state, which can capture the patterns in the history attention distributions during the decoder process and can generate much better future attention distributions. And in order to get better fusion information from parallel navigation instructions and environment graph presentation, we use a two-layer soft attention method to blend navigation instructions and environment graph information.”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao, Gao with the further processing of Zhou.
One of ordinary skill in the art would have been motived to combine in order to improve translating natural translations to a high-level sequence of behaviors that is understandable and executable by robots, and show improved model performance.
(Zhou [sec(s) 7] “In this paper, based on the encoder-decoder model with attention mechanism, we propose an attention-history reader network for capturing the patterns in the attention history to improve translating natural translations to a high-level sequence of behaviors that is understandable and executable by robots. we also use a two-layer soft-attention for fusing multimodal inputs (instructions and graph representations). Experiments show that our method improved model performance.”)
Regarding claim 11
The claim is a method claim corresponding to the machine learning model claim 3, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Claim(s) 4, 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Gao et al. (A Neural Model for Method Name Generation from Functional Description) in view of Zhou et al. (Translating Natural Language Instructions for Behavioral Robot Indoor Navigation with Attention-History Based Attention) in view of Masood et al. (A Multi-Step Time-Series Clustering-Based Seq2Seq LSTM Learning for a Single Household Electricity Load Forecasting)
Regarding claim 4
The combination of Pavia, Bao, Gao, Zhou teaches claim 3.
However, the combination of Pavia, Bao, Gao, Zhou does not appear to explicitly teach:
wherein the sequence to sequence deep learning model further applies long short term memory as the encoder and the decoder layers.
Masood teaches
wherein the sequence to sequence deep learning model further applies long short term memory as the encoder and the decoder layers.
(Masood [fig(s) 5] “LSTM encoder”, “LSTM decoder”, “A Seq2Seq LSTM network model for the time-series load forecasting. At each timestep, the encoder takes one series of data xt at time t, and its previous state ht−1 and produces an output vector ht state and cell state Ct. The next decoder generates an output sequence yt, at each step taking at time t, the previous state, and a weighted combination of all the encoder outputs (i.e., encoder state vector).” [sec(s) 3] “This paper adopts the Seq2Seq LSTM model for a more substantial analysis of our multi-step time-series load forecasting problem. The architecture of a Seq2Seq LSTM model is shown in Figure 5, where each rectangular block holds an LSTM cell. Furthermore, each LSTM cell contains a hidden state, ht, and cell state, ct, at timestep, t. The architecture is mainly divided into three parts: the encoder that is the input to the model, the decoder, which is the model’s output, and the encoder state vector. It can be seen that the encoder part is stacked with LSTM cells, and each of the cells allow a single element from the sequence as input.”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao, Gao, Zhou with the LSTM encoder and LSTM decoder layers of Masood.
One of ordinary skill in the art would have been motived to combine in order to significantly improve forecasting performance based on the reliable approach of multi-step time-series learning.
(Masood [sec(s) 5] “Furthermore, the simulation results showed that cluster-based multistep time-series Seq2Seq LSTM learning significantly improves single household load forecasting. This confirms that the cluster based multi-step time-series learning is a reliable approach for the future load forecasting of households. Furthermore, the limitation of the univariate analysis can be extended to the multivariate with multi-step load forecasting in future work. This research will open further challenges for applying DL techniques to the SG network in the future.”)
Regarding claim 12
The claim is a method claim corresponding to the machine learning model claim 4, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Claim(s) 5-7, 13-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Raj et al. (An Application to Detect Cyberbullying Using Machine Learning and Deep Learning Techniques)
Regarding claim 5
The combination of Pavia, Bao teaches claim 1.
However, the combination of Pavia, Bao does not appear to explicitly teach:
an optimization layer coupled to the task specific layer, for receiving outputs from each task sublayer of a plurality of task sublayers providing multi task classifiers comprised in the task specific layer for determining an indication of a likelihood of an input to the task specific layer corresponding to one of the task sublayers, the outputs provided to a root mean square propagation in the optimization layer for increasing a learning rate for the task specific layer.
Raj teaches
an optimization layer coupled to the task specific layer, for receiving outputs from each task sublayer of a plurality of task sublayers providing multi task classifiers comprised in the task specific layer for determining an indication of a likelihood of an input to the task specific layer corresponding to one of the task sublayers, the outputs provided to a root mean square propagation in the optimization layer for increasing a learning rate for the task specific layer.
(Raj [fig(s) 1] [fig(s) 2] “CNN”, “BiLSTM”, “Concatenate”, “optimizer(adam)” [sec(s) Results and Discussion] “Adam and RMSProp as optimizers are used and for activation layers ReLU and Sigmoid are used thus making it a total of 4 combinations. … The Adam optimizer is computationally more efficient, requires slight memory, is invariant to diagonal resizing of gradients, and it is well suited for problems with a lot of data/parameters, whereas the RMSProp optimization algorithm keeps the sections under control the entire time because of the decay rate, which makes RMSProp faster than Adam. Adam obtains his speed from momentum, while RMSProp gives him the capability to adjust gradients in various directions. It’s powerful because of the mix of the two. Whereas RMSProp just uses the second moment and speeds it up with a decay rate, Adam employs both first and second moments and is usually the best option (Figs. 4, 5).” [sec(s) CNN-BiLSTM Architecture] “The Adam optimizer is computationally more efficient, requires slight memory, is invariant to diagonal resizing of gradients, and it is well suited for problems with a lot of data/parameters. We will perform the best parameter using grid search and 10-fold cross validation. Now, Convolutional Neural Network (CNN) models are built to classify encoded documents as either cyberbullying or non-cyberbullying. Now, the CNN model can be defined as follows as shown in Fig. 2:”; e.g., “optimizer” read(s) on “optimization layer”. In addition, e.g., “CNN” and “BiLSTM” read(s) on “task sublayers”.)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao with the root mean square propagation of Raj.
One of ordinary skill in the art would have been motived to combine in order to provide a more accurate prediction accuracy compared to the conventional methods by learning global features and long-term dependencies.
(Raj [sec(s) Conclusion] “The model for automatically detecting cyberbullying text on multilingual data is addressed and proposed in this work. Solving this issue is critical for controlling social media material in multiple languages and protecting users from the negative impacts of toxic comments like verbal assaults and offensive language. The performance of our various models of neural networks is examined. The CNN-BiLSTM network has the best accuracy. While the CNN alone can only train local characteristics from word n-grams, with its LSTM layer, the CNN-BiLSTM can also learn global features and long-term dependencies.”)
Regarding claim 6
The combination of Pavia, Bao, Raj teaches claim 5.
Pavia further teaches
wherein the task specific layer applies deep multi-task learning and comprises an input layer, a hidden layer and an output layer, each node in the output layer associated with a particular task of a set of tasks and sharing common features therebetween for optimization of the multi-task learning.
(Pavia [fig(s) 2] “Dense”, “Activation”, “Dropout”, “Multinomial Naïve Bayes” [sec(s) 3] “we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have used 64 LSTM/Bi-LSTM units in this case followed by a fully connected layer, containing 256 dense nodes. Then we have added a dropout layer, usually used to avoid overfitting and finally, the output layer has one dense unit with the sigmoid function as the activation function, which outputs the probability of the input cell is a Metadata cell” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table.” [sec(s) 4] “After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. To alleviate that we have used a dropout layer after the activation layer with a value of 0.2, known to be a balance between dropping too many features and degree of overfitting. There is a node having sigmoid as the activation function. Before the activation and dropout layers we have used a Dense layer that’s output serves as input for the Multinomial Naive Bayes Classifier. First we have trained the LSTM model as a binary classifier i.e. to classify a cell being a Metadata cell or not. We have taken the output feature vector of the dense layer having 256 nodes as an encoding of the input cell, i.e. the dimensionality of an encoded vector is 256. For a tuple or column, there are several cells. We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”;)
Regarding claim 7
The combination of Pavia, Bao, Raj teaches claim 6.
Pavia further teaches
wherein the task specific layer applies a binary threshold to at least some output nodes to determine a likelihood of whether input to the task specific layer falls within the particular task for metadata the output layer to provide the classification.
(Pavia [fig(s) 2] “(0/1)” [sec(s) 3] “we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have used 64 LSTM/Bi-LSTM units in this case followed by a fully connected layer, containing 256 dense nodes. Then we have added a dropout layer, usually used to avoid overfitting and finally, the output layer has one dense unit with the sigmoid function as the activation function, which outputs the probability of the input cell is a Metadata cell, however it does not generate the final decision for an entire tuple or column. … The final output is binary, whether the input sequence of cells, i.e. a tuple or column is a Metadata row/column or not. The output of the Naive Bayes Classifier is then used with the decision-tree rules to produce the final output having the Metadata type.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table.” [sec(s) 4] “After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. … We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”; e.g., “(0/1)” read(s) on “binary threshold”.)
Regarding claim 13
The claim is a method claim corresponding to the machine learning model claim 5, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Regarding claim 14
The claim is a method claim corresponding to the machine learning model claim 6, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Regarding claim 15
The claim is a method claim corresponding to the machine learning model claim 7, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Claim(s) 8, 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Raj et al. (An Application to Detect Cyberbullying Using Machine Learning and Deep Learning Techniques) in view of Wynter et al. (Leveraging External Knowledge for Out-Of-Vocabulary Entity Labeling)
Regarding claim 8
The combination of Pavia, Bao, Raj teaches claim 6.
Pavia further teaches
wherein the task specific layer applies a [soft max] threshold to at least some output nodes to determine a likelihood of whether input to the task specific layer falls within the particular task associated with one of the output nodes.
(Pavia [fig(s) 2] “(0/1)” [sec(s) 3] “we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have used 64 LSTM/Bi-LSTM units in this case followed by a fully connected layer, containing 256 dense nodes. Then we have added a dropout layer, usually used to avoid overfitting and finally, the output layer has one dense unit with the sigmoid function as the activation function, which outputs the probability of the input cell is a Metadata cell, however it does not generate the final decision for an entire tuple or column. … The final output is binary, whether the input sequence of cells, i.e. a tuple or column is a Metadata row/column or not. The output of the Naive Bayes Classifier is then used with the decision-tree rules to produce the final output having the Metadata type.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table.” [sec(s) 4] “After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. … We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”; e.g., “(0/1)” read(s) on “threshold”.)
However, the combination of Pavia, Bao, Raj does not appear to explicitly teach:
wherein the task specific layer applies a [soft max] threshold to at least some output nodes to determine a likelihood of whether input to the task specific layer falls within the particular task associated with one of the output nodes.
Wynter teaches
wherein the task specific layer applies a soft max threshold to at least some output nodes to determine a likelihood of whether input to the task specific layer falls within the particular task associated with one of the output nodes.
(Wynter [fig(s) 1] “Softmax, threshold” [fig(s) 2] “SOFTMAX” [sec(s) 2] “Finally, we pass in the distances into a softmax function
PNG
media_image2.png
112
436
media_image2.png
Greyscale
to obtain their probablistic representation. Softmax was the preferred layer due to the fact that it tends to skew results in such a way that allow us to minimize the distance between the obtained and the expected probability distributions.
PNG
media_image3.png
76
1264
media_image3.png
Greyscale
(5) Due to the nature of our approach and the normalization from s(x)i , there are many low-confidence distances that will add noise, and could be detrimental to our training goal. To mitigate this, we implement a learnable threshold τ that allows us to prune out results:
PNG
media_image4.png
95
513
media_image4.png
Greyscale
(6)”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao, Raj with the soft max threshold of Wynter.
One of ordinary skill in the art would have been motived to combine in order to obtain a prediction accuracy increase compared to the state-of-the-art model.
(Wynter [sec(s) Abs] “In particular, we evaluate this approach by training a state-of-the-art model with candidates generated from our network, and obtained relative increases of 57.7% and 82.7% in F1 score and accuracy, respectively, for the aforementioned model, when compared to the current candidate generation strategy”)
Regarding claim 16
The claim is a method claim corresponding to the machine learning model claim 8, and is directed to largely the same subject matter. Thus, it is rejected for the same reasons as given in the rejections of the machine learning model claim.
Claim(s) 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Pavia et al. (HYBRID METADATA CLASSIFICATION IN LARGE-SCALE STRUCTURED DATASETS) in view of Bao et al. (Attentive Siamese LSTM Network for Semantic Textual Similarity Measure) in view of Raj et al. (An Application to Detect Cyberbullying Using Machine Learning and Deep Learning Techniques) in view of Gopalan et al. (US 2023/0091036 A1)
Regarding claim 17
The combination of Pavia, Bao, Raj teaches claim 14.
Pavia further teaches
detecting a trigger event prior to performing the metadata classifications and automatically tagging the input data element, the trigger event including [receiving an input, at the requesting computing device to initiate migration of] the input data element to a computing cloud.
(Pavia [fig(s) 2] “(0/1)” [sec(s) 3] “we describe the two-layer ensemble we have designed for tabular Metadata classification and the ideas behind its design. The first layer consists of a Recurrent Neural Network (RNN) [26] model containing either regular LSTM or Bidirectional LSTM [25] units for the analysis of a table cell. … We have used 64 LSTM/Bi-LSTM units in this case followed by a fully connected layer, containing 256 dense nodes. Then we have added a dropout layer, usually used to avoid overfitting and finally, the output layer has one dense unit with the sigmoid function as the activation function, which outputs the probability of the input cell is a Metadata cell, however it does not generate the final decision for an entire tuple or column. … The final output is binary, whether the input sequence of cells, i.e. a tuple or column is a Metadata row/column or not. The output of the Naive Bayes Classifier is then used with the decision-tree rules to produce the final output having the Metadata type.” [sec(s) 1] “Here, we describe and evaluate an ensemble of a Deep- and Machine-Learning model to classify metadata rows/columns in a table.” [sec(s) 4] “In this section, we first describe the training and test sets construction followed by the experimental evaluation on several large-scale datasets. Hardware: We run all our experiments on Amazon AWS EC2 p3.8xlarge instances, each having 4 NVIDIA® V100 Tensor Core GPUs, 32 vCPUs, 244 GiB of memory, 10 Gbps network. Software: For implementing the RNN model, we have used Keras [3], a popular python library for deep learning, having Tensorflow [8] framework as the backend. The second step i.e. the Naive Bayes classifier is implemented using Scikit-learn [7], a popular machine learning library for python. Throughout the experimentation and implementation, we have used python as the programming language. … After the LSTM/Bi-LSTM units, we have added a fully connected layer with 256 nodes. We have used rectified linear activation function for all nodes. We are aware that such models tend to overfit even on a large training set. … We have concatenated their encodings to form an encoding for a tuple/column. This feature vector is the input of the Multinomial Naive Bayes Classifier, which is also a binary classifier that predicts the input tuple or column being Metadata or not.”;)
However, the combination of Pavia, Bao, Raj does not appear to explicitly teach:
the trigger event including [receiving an input, at the requesting computing device to initiate migration of] the input data element to a computing cloud.
Gopalan teaches
the trigger event including receiving an input, at the requesting computing device to initiate migration of the input data element to a computing cloud.
(Gopalan [par(s) 10] “In some embodiments, to receive the request to forecast contact center data using the cloud system may include to receive a request to migrate the data from the on-premises system to the cloud system.” [par(s) 40] “The illustrative method 101 begins with flow 150 in which the user 106 requests that data be transferred from the on-premises system 102 to the cloud system 104 for forecasting and/or other processing. For example, in some embodiments, the user 106 requests a historical data transfer to preload the cloud system 104 with interaction data, which provides preliminary data for forecasting interaction volumes in the cloud system 104. In some embodiments, the request may occur as a one-time operation (or infrequent operation) in order to request that the on-premises data be transferred to the cloud system 104 for permanent cloud-based operation. For example, in some embodiments, the one-time operation may be performed in conjunction with migrating a customer from an on-premises workforce management system to a cloud-based system. In other embodiments, the user 106 may make the request on-demand via a user interface as many times as desired. For example, in some embodiments, the customer may maintain usage of the on-premises system 102 but leverage the forecasting algorithms of the cloud system 104 via a user interface accessible to the user 106 of the on-premises system 102.”;)
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the system of Pavia, Bao, Raj with the data migration of Gopalan.
One of ordinary skill in the art would have been motived to combine in order to improve overall contact center performance and the customer experience by allowing to tailor interactions based on predictions or to allocate resources in preparation for predicted characteristics of future interactions.
(Gopalan [par(s) 74] “The behavior models may be used to predict behaviors of, for example, customers or agents, in a variety of situations, thereby allowing embodiments of the present invention to tailor interactions based on such predictions or to allocate resources in preparation for predicted characteristics of future interactions, thereby improving overall contact center performance and the customer experience. It will be appreciated that, while the analytics module is described as being part of a contact center, such behavior models also may be implemented on customer systems ( or, as also used herein, on the "customer-side" of the interaction) and used for the benefit of customers.”)
Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Liang et al. (AC-BLSTM: Asymmetric Convolutional Bidirectional LSTM Networks for Text Classification) teaches generating the input sequence of BLSTM.
Raychev et al. (Predicting Program Properties from ‘Big Code) teaches predicting program properties from large codebases.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SEHWAN KIM whose telephone number is (571)270-7409. The examiner can normally be reached Mon - Fri 9:00 AM - 5:00 PM.
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, Michael J Huntley can be reached on (303) 297-4307. 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.
/SEHWAN KIM/Examiner, Art Unit 2129 12/19/2025