DETAILED ACTION
This action is responsive to the application filed on 01/30/2024. Claims 1-20 are pending and have been examined.
This action is Non-Final.
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 .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition
of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the
conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1,
Step 1: This claim is directed to a method, which is one of the four statutory categories. Therefore, claim 1 satisfies Step 1.
Step 2A Prong 1:
(a) “training a domain-invariant attribute extraction model based on the training data, wherein the domain-invariant attribute extraction model models the semantics of the predetermined attributes across the multiple domains” - This limitation is directed to a mathematical concept because training a machine learning model and modeling semantics using the model involves mathematical relationships, mathematical calculations, and optimization of model parameters.
(b) “extracting, based on the trained domain-invariant attribute extraction model, one or more of the predetermined attributes according to the semantics of the one or more of the predetermined attributes” - This limitation is directed to a mental process because identifying attributes from textual content according to the meaning or semantics of the text is an evaluation or judgment that can be performed in the human mind or with pen and paper. This limitation is also directed to a mathematical concept to the extent the extraction is performed by a trained model using calculated model parameters.
Step 2A Prong 2 and Step 2B:
(a) “obtaining training data... wherein the training data includes a plurality of training samples, each of which comprises textual content, one or more of the predetermined attributes extracted from the textual content, and a label corresponding to one of the multiple domains that produces the textual content” - The limitation is directed to obtaining and collecting information used as input to the abstract idea. This is mere data gathering and insignificant extra-solution activity that does not integrate the judicial exception into a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, obtaining training data is well-understood, routine, and conventional data gathering and does not provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
(b) “receiving textual content from any one of the multiple domains” - The limitation is directed to receiving input data for analysis. This is mere data gathering and insignificant extra-solution activity that does not integrate the judicial exception into a practical application under MPEP 2106.05(g)). Further under Step 2B, receiving textual content is well-understood, routine, and conventional data gathering and does not provide significantly more than the judicial exception under MPEP 2106.05(d)(II)).
Thus, claim 1 is non-patent eligible. Claims 8 and 15 are analogous to claim 1, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 2,
Step 1: This claim depends from claim 1 and is directed to a method, which is one of the four statutory categories. Therefore, claim 2 satisfies Step 1.
Step 2A Prong 1:
(a) “the predetermined attributes to be extracted from the textual content include features associated with different aspect of the transaction” - This limitation is directed to a mental process because identifying transaction features or attributes from textual content is an observation, evaluation, or judgment that can be performed by a person reading the text.
Step 2A Prong 2 and Step 2B:
(a) “each of the multiple domains corresponds to a web platform through which a user conducts online activities” - The limitation merely links the judicial exception to a particular technological environment, namely web platforms and online activities, and does not integrate the judicial exception into a practical application, does not provide significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 2 is non-patent eligible. Claims 9 and 16 are analogous to claim 2, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 3,
Step 1: This claim depends from claim 1 and is directed to a method, which is one of the four statutory categories. Therefore, claim 3 satisfies Step 1.
Step 2A Prong 1:
(a) “the domain-invariant attribute extraction model is trained using domain adversarial learning” - This limitation is directed to a mathematical concept because domain adversarial learning involves mathematical model training, optimization, calculated losses, adversarial objectives, and adjustment of model parameters.
There are no new elements to be evaluated under Step 2A Prong 2 and Step 2B.
Thus, claim 3 is non-patent eligible. Claim 10 and part of claim 17 are analogous to claim 3, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 4,
Step 1: This claim depends from claim 3 and is directed to a method, which is one of the four statutory categories. Therefore, claim 4 satisfies Step 1.
Step 2A Prong 1:
“the first part is for modeling domain-specific characteristics with respect to the multiple domains; the second part is for modeling domain-invariant semantics of the predetermined attributes;” -- This limitation is directed to a mathematical concept because modeling domain-invariant/specific semantics involves mathematical representation and processing of semantic information. This limitation is also directed to a mental process because determining the meaning or semantics of attributes in textual content is an observation, evaluation, or judgment that can be performed in the human mind.
Step 2A Prong 2 and Step 2B:
“the first part and the second part interact.” -- The limitation recites that the first and second part interact. The limitation amounts to mere transmitting of data which is seen as an insignificant, extra-solution activity, and thus it cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, transmitting data over a network is a well-understood, routine, and conventional activity (WURC) that cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
Thus, claim 4 is non-patent eligible. Claim 11 and part of claim 17 are analogous to claim 4, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 5,
Step 1: This claim depends from claim 4 and is directed to a method, which is one of the four statutory categories. Therefore, claim 5 satisfies Step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
“the first part influences the modeling of the second part by sending information representing domain-specific characteristics learned by the first part to the second part;” -- The limitation recites sending information representing domain-specific characteristics that are learned from the first part to be sent to the second part. The limitation amounts to mere transmitting of data which is seen as an insignificant, extra-solution activity, and thus it cannot be integrated to a practical application (see MPEP 2106.05(g)). Furthermore, under Step 2B, transmitting data over a network is a well-understood, routine, and conventional activity (WURC) that cannot provide significantly more than the judicial exception (see MPEP 2106.05(d)(II)).
“the second part excludes, based on the information from the first part, influence of the domain-specific characteristics in learning domain-invariant semantics of the predetermined attributes by negating, impact of the domain-specific characteristics, in optimization in training the second part” -- The limitation recites excluding the second part to influence the domain-specific characteristics in learning domain-invariant semantics of the predetermined attributes by negating impact on training of the second part. The limitation recites no more than mere limiting to a field of use/environment, and thus the limitation cannot be integrated to a practical application, nor provide significantly more than the judicial exception (see MPEP 2106.05(h)).
Thus, claim 5 is non-patent eligible. Claims 12 and 18 are analogous to claim 5, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 6,
Step 1: This claim depends from claim 4 and is directed to a method, which is one of the four statutory categories. Therefore, claim 6 satisfies Step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
(a) “The method of claim 4,the second part is constructed using an encoder-decoder architecture with an attention layer between an encoder layer and a decoder layer” - This limitation recites a encoder-decoder architectures and attention layers merely being constructed to be used in the second part. The limitation amounts to no more than mere instructions to apply onto a computer, and it cannot be integrated to a practical application, nor provide significantly more than the judicial exception (see MPEP 2106.05(f)).
Thus, claim 6 is non-patent eligible. Claims 13 and 19 are analogous to claim 6, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 7,
Step 1: This claim depends from claim 6 and is directed to a method, which is one of the four statutory categories. Therefore, claim 7 satisfies Step 1.
There are no elements to be evaluated under Step 2A Prong 1.
Step 2A Prong 2 and Step 2B:
(a) “the encoder layer comprises a first multiple serially connected long short-term memory (LSTM) units, each of which corresponds to a token from the textual content and the decoder layer comprises a second multiple serially connected long short-term memory (LSTM) units, each of which corresponds to one of the predetermined attributes recognized from the textual content” - This limitation recites that the encoder layer will comprise LSTM units process tokens that will correspond to a textual context, and a decoder layer that comprises multiple connected LSTM units for which correspond to one of the predetermined attributes of the textual context. The limitation amount to no more than mere instructions to apply onto a computer, and thus it does not integrate to a practical application, nor provide significantly more than the judicial exception (see MPEP 2106.05(f)).
Thus, claim 7 is non-patent eligible. Claims 14 and 20 are analogous to claim 7, aside from claim type and minute differences, hence the same rejection can apply.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are
summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-5. 8-12, 15-18 are rejected under 35 U.S.C. 103 as being unpatentable over NPL reference “Domain-adversarial training of neural networks”, by Ganin et. al. (referred herein as Ganin) in view of NPL reference “Open attribute value extraction from product profiles”, by Zheng et. al. (referred herein as Zheng) further in view of NPL reference “Multinomial adversarial networks for multi-domain text classification.”, by Chen et. al. (referred herein as Chen).
Regarding claim 1, Ganin teaches:
A method, comprising: obtaining training data for training a domain-invariant attribute extraction model ([Ganin, page 5], “ An unsupervised domain adaptation learning algorithm is then provided with a labeled source sample S drawn i.i.d. from DS, and an unlabeled target sample T drawn i.i.d. from DX_T”, AND [Ganin, page 2], “We thus focus on learning features that combine (i) discriminativeness and (ii) domain-invariance”, wherein the examiner interprets “labeled source sample S...and an unlabeled target sample T” to be the same as obtaining training data having textual content with labels indicating domains, because both are directed to gathering and using training information that includes samples from multiple source and target data with domain labels for training purposes. The examiner further interprets “learning features that combine (i) discriminativeness and (ii) domain-invariance” to be the same as training a domain-invariant model because they are both directed to learning representations that remain useful for a task while reducing dependence on domain differences.)
training a domain-invariant attribute extraction model based on the training data, wherein the domain-invariant attribute extraction model models the semantics of the predetermined attributes across the multiple domains; ([Ganin, Abstract] “the approach promotes the emergence of features that are (i) discriminative for the main learning task on the source domain and (ii) indiscriminate with respect to the shift between the domains”, wherein the examiner interprets “features that are discriminative for the main learning task on the source domain and indiscriminate with respect to the shift between the domains” to be the same as a model that learns the semantics of predetermined attributes across multiple domains, because both are directed to training a model to learn task-relevant representations that remain invariant across domain variations.)
receiving textual content from any one of the multiple domains; ([Ganin, page 5] “The goal of the learning algorithm is to build a classifier η : X → Y with a low target risk RDT(η) = Pr(x,y)∼DT [η(x) ≠ y], while having no information about the labels of DT (target domain)”, wherein the examiner interprets “to build a classifier η with input from DT (target domain)” to be the same as receiving textual content from any one of the multiple domains, because both are directed to accepting input data that originates from any domain in a set of multiple domains.)
extracting, based on the trained domain-invariant attribute extraction model, one or more of the predetermined attributes according to the semantics of the one or more of the predetermined attributes. ([Ganin, page 7] “That is, to learn a model that can generalize well from one domain to another, we ensure that the internal representation of the neural network contains no discriminative information about the origin of the input (source or target), while preserving a low risk on the source (labeled) examples”, wherein the examiner interprets “the neural network preserving a low risk on labeled examples” to be the same as extracting predetermined attributes according to their semantics using a trained model, because both are directed to using a trained neural network representation to perform prediction or extraction of target information based on learned semantic representations.)
Ganin does not teach to be used to extract predetermined attributes from textual content from multiple domains, wherein the training data includes a plurality of training samples, each of which comprises textual content, one or more of the predetermined attributes extracted from the textual content, and a label corresponding to one of the multiple domains that produces the textual content.
Zheng teaches:
to be used to extract predetermined attributes from textual content … ([Zheng, page 1049], “a set of pre-defined target attributes (e.g., brand, flavor, size), our objective is to extract corresponding attribute values from unstructured text.” wherein the examiner interprets “a set of pre-defined target attributes” to be the same as predetermined attributes because they are both directed to attribute categories specified before extraction, and “extract corresponding attribute values from unstructured text” to be the same as to be used to extract predetermined attributes from textual content because they are both directed to extracting attribute information from text-based content.)
wherein the training data includes a plurality of training samples, ([Zheng, page 1053] “Starting with a small set of labeled instances as an initial training set L”, wherein the examiner interprets “a small set of labeled instances as an initial training set L” to be the same as training data includes a plurality of training samples because they are both directed to multiple labeled examples used for training.)
each of which comprises textual content, one or more of the predetermined attributes extracted from the textual content, ([Zheng, page 1050] “Given a set of products I, corresponding profiles X = {xi : i 2 I}, and a set of attributes A = {a1,...,am}, extract all attribute-values Vi = h{i,j,1,...,i,j, `i,j }, aji for i 2 I and j 2 [1,m] with an open world assumption (OWA); we use i,j to denote the set of values (of size `i,j) for attribute aj for the i th product, and the product profile (title, description, bullets) consists of a sequence of words/tokens xi = {wi,1,wi,2, ···wi,ni }. Note that we want to discover multiple values for a given set of attributes.”, wherein the examiner interprets “extract all attribute-values Vi” and “multiple values for a given set of attributes” to be the same as one or more of the predetermined attributes extracted from the textual content because they are both directed to extracting one or more values associated with specified attributes from the input text.).
Ganin and Zheng does not teach … from multiple domains … and a label corresponding to one of the multiple domains that produces the textual content;
Chen teaches … from multiple domains … and a label corresponding to one of the multiple domains that produces the textual content; ([Chen, page 1227-1228] “texts come from a variety of domains … Mini-batch of documents from domain di ∈ Δ … Domain Label”; [Chen, page 1229] “where d is the domain index of some sample”, wherein the examiner interprets “a variety of domains” to be the same as multiple domains because they are both directed to more than one domain/source of text. The examiner further interprets “Domain Label” and “where d is the domain index of some sample” to be the same as a label corresponding to one of the multiple domains because they are both directed to identifying the particular domain associated with the sample, and “documents from domain di ∈ Δ” to be the same as that produces the textual content because they are both directed to documents/text samples originating from a domain within the set of domains).
Ganin, Zheng, Chen, and the instant application are analogous art because they are all directed to machine-learning models for processing textual or domain-based data and learning task-relevant representations for prediction or extraction across domains.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method of domain-invariant features disclosed by Ganin to include the attribute extraction technique disclosed by Zheng. One would be motivated to do so to effectively adapt Ganin’s domain-invariant feature-learning method to a textual attribute-extraction task in which specified attribute values are extracted from unstructured text, as suggested by Zheng ([Zheng, page 1050] “a set of pre-defined target attributes (e.g., brand, flavor, size), our objective is to extract corresponding attribute values from unstructured text … OpenTag does not use any dictionary or hand-crafted features.”).
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify to include the domain labeling technique disclosed by Chen. One would be motivated to do so to effectively identify the domain associated with each textual sample and use such domain information when learning features across multiple domains, as suggested by Chen ([Chen, page 1226] “The objective of MDTC is to leverage all the available resources in order to improve the system performance over all domains simultaneously.”). Claims 8 and 15 are analogous to claim 1, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 2, Ganin, Zheng, Chen teach The method of claim 1, (see rejection of claim 1)
Zheng further teaches wherein each of the multiple domains corresponds to a web platform through which a user conducts online activities; [Zheng, page 1049] “Product catalogs are a valuable resource for eCommerce retailers that allow them to organize, standardize, and publish information to customers…For a concrete example, refer to Figure 1 showing a snapshot of the product profile of a 'dog food' in Amazon.com with unstructured data such as title, description, and bullets.”, wherein the examiner interprets “Product catalogs are a valuable resource for eCommerce retailers that allow them to organize, standardize, and publish information to customers” to be the same as each of the multiple domains corresponding to a web platform through which a user conducts online activities, because they are both directed to recognizing that web-based e-commerce platforms serve as domains where users engage in online purchasing activities.)
the textual content from one of the multiple domains describes a transaction a user carries out via the domain;… the predetermined attributes to be extracted from the textual content include features associated with different aspect of the transaction. ([Zheng, page 1049] “The product title 'Variety Pack Fillet Mignon and Porterhouse Steak Dog Food (12 Count)' contains two attributes of interest namely size and flavor. We want to discover corresponding values for the attributes like '12 count' (size), 'Fillet Mignon' (flavor) and 'Porterhouse Steak' (flavor).” AND [Zheng, page 1050] “OpenTag does not make any assumptions about the structure of input text and could be applied to any kind of textual data like profile pages of a given product.” AND [Zheng, page 1055], “We perform experiments in 3 domains, namely, (i) dog food, (ii) detergents, and (iii) camera. For each domain, we use the product profiles (like titles, descriptions, and bullets) from Amazon.com public pages.”, wherein the examiner interprets “the textual content from different web-domains may be processed...to identify relevant attributes” to be the same as the textual content from one of the multiple domains describing a transaction a user carries out via the domain, because they are both directed to processing text content originating from a specific web platform that contains information about user activities or interactions on that platform. Furthermore, the examiner interprets “We perform experiments in 3 domains, namely, (i) dog food, (ii) detergents, and (iii) camera. For each domain, we use the product profiles (like titles, descriptions, and bullets) from Amazon.com public pages” to be the same as “the textual content from one of the multiple domains describes a transaction a user carries out via the domain, because they are both describing textual content originating from multiple distinct domains (dog food, detergents, camera representing different eCommerce product categories/domains) where each domain's textual content (product titles, descriptions, and bullets) describes transactions that users carry out through the eCommerce platform (Amazon.com) when purchasing products within that domain.)
Ganin, Zheng, Chen, and the instant application are analogous art because they are all directed to machine-learning techniques for processing domain-dependent data using learned representations.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method of claim 1 disclosed by Ganin, Zheng, and Chen to include the attribute extraction technique disclosed by Zheng. One would be motivated to do so to effectively extract product-related attribute values from e-commerce textual content and supplement product catalog information, as suggested by Zheng (Zheng, [Zheng, page 1049] “The product title 'Variety Pack Fillet Mignon and Porterhouse Steak Dog Food (12 Count)' contains two attributes of interest namely size and flavor. We want to discover corresponding values for the attributes like '12 count' (size), 'Fillet Mignon' (flavor) and 'Porterhouse Steak' (flavor).” AND [Zheng, page 1050] “OpenTag does not make any assumptions about the structure of input text and could be applied to any kind of textual data like profile pages of a given product.” AND [Zheng, page 1055], “We perform experiments in 3 domains, namely, (i) dog food, (ii) detergents, and (iii) camera. For each domain, we use the product profiles (like titles, descriptions, and bullets) from Amazon.com public pages.”). Claims 9 and 16 are analogous to claim 2, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 3, Ganin, Zheng, Chen teach The method of claim 1, (see rejection of claim 1)
Ganin further teaches: wherein the domain-invariant attribute extraction model ([Ganin , page 2] “We thus focus on learning features that combine (i) discriminativeness and (ii) domain-invariance”, wherein the examiner interprets “features that combine discriminativeness and domain-invariance” to be the same as a domain-invariant attribute extraction model, because they are both directed to a neural network model that learns to extract features or attributes that remain invariant across domain shifts while maintaining discriminative capability.)
is trained using domain adversarial learning. ([Ganin, page 10] “In words, during training, the neural network (parameterized by W, b, V, c) and the domain regressor (parameterized by u, z) are competing against each other, in an adversarial way, over the objective of Equation (9). For this reason, we refer to networks trained according to this objective as Domain-Adversarial Neural Networks (DANN)”, wherein the examiner interprets “the neural network and domain regressor competing against each other in an adversarial way, referred to as Domain-Adversarial Neural Networks trained according to this objective” to be the same as training using domain adversarial learning, because they are both directed to a training methodology where a feature extraction network and a domain classifier are optimized adversarially against each other to produce domain-invariant representations.) Claim 10 and part of claim 17 are analogous to claim 3, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 4, Ganin, Zheng, Chen teach The method of claim 3, (see rejection of claim 3).
Chen further teaches:
wherein the domain-invariant ([Chen, page 1226-1227] “MAN learns features that are invariant across multiple domains … The main idea of MAN is to explicitly model the domain-invariant features that are beneficial to the main classification task across all domains”, wherein the examiner interprets “The main idea of MAN is to explicitly model the domain-invariant features that are beneficial to the main classification task across all domains” and “features that are invariant across multiple domains” to be the same as the domain-invariant method of claim 3 because they are both directed to an adversarial learning model that learns features that are invariant across multiple domains).
comprises a first part and a second part, wherein the first part is for modeling domain-specific characteristics with respect to the multiple domains;… the second part is for modeling domain-invariant ([Chen, page 1227] “consists of four components: a shared feature extractor Fs, a domain feature extractor Fdi for each labeled domain di 2 ∆ L,atext classifier C, and a domain discriminator D. The main idea of MAN is to explicitly model the domain-invariant features that are beneficial to the main classification task across all domains (i.e. the shared features, extracted by Fs), as well as the domain-specific features that mainly contribute to the classification in its own domain (the domain features, extracted by Fd).”, wherein the examiner interprets “a shared feature extractor Fs” and “a domain feature extractor Fdi” to be the same as a first part and a second part because they are both directed to distinct model components within the adversarial network) and also interprets “a domain feature extractor Fdi for each labeled domain” and “the domain-specific features” to be the same as the first part is for modeling domain-specific characteristics with respect to the multiple domains because they are both directed to a model component that learns features specific to particular domains).
… and the first part and the second part interact. ([Chen, page 1228] “The parameters of Fs, Fd, C are updated together”, wherein the examiner interprets “Fs, Fd, C are updated together” to be the same as the first part and the second part interact because they are both directed to the shared-feature component and domain-feature component being combined and jointly used during model training.).
Zheng further teaches …attribute extraction model… semantics of the predetermined attributes; ([Zheng, page 1049] “We develop a novel deep tagging model OpenTag for this extraction problem with the following contributions: (1) we formalize the problem as a sequence tagging task, and propose a joint model exploiting recurrent neural networks (specifically, bidirectional LSTM) to capture context and semantics,”, wherein the examiner interprets “a novel deep tagging model OpenTag for this extraction problem…to capture context and semantics” to be the same as attribute extraction model because they are both directed to a trained model for extracting attribute information from text.).
Ganin, Chen, Zheng, and the instant application are analogous art because they are all directed to machine-learning models for processing textual or domain-dependent data using learned representations.
It would have also been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method of claim 3 disclosed by Ganin, Chen, and Zheng to include the attribute extraction approach disclosed by Zheng. One would be motivated to do so to effectively adapt the domain-invariant multi-domain learning model to extract semantic attribute information from textual content, as suggested by Zheng ([Zheng, page 1049] “We develop a novel deep tagging model OpenTag for this extraction problem with the following contributions: (1) we formalize the problem as a sequence tagging task, and propose a joint model exploiting recurrent neural networks (specifically, bidirectional LSTM) to capture context and semantics,”). Claim 11 and part of claim 17 are analogous to claim 4, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 5, Ganin, Zheng, Chen teach The method of claim 4, (see rejection of claim 4)
Ganin further teaches, wherein the first part influences the modeling of the second part by sending information representing domain-specific characteristics learned by the first part to the second part; ([Ganin, page 2] “(ii) the domain classifier that discriminates between the source and the target domains during training” AND [Ganin, page 12] “domain classifier (red) connected to the feature extractor via a gradient reversal layer … the GRL takes the gradient from the subsequent level and changes its sign, i.e., multiplies it by -1, before passing it to the preceding layer.”, wherein the examiner interprets “the domain classifier that discriminates between the source and the target domains during training” to be the same as the first part learning information representing domain-specific characteristics because they are both directed to learning information that distinguishes the domain of the input, and “domain classifier (red) connected to the feature extractor via a gradient reversal layer” and “before passing it to the preceding layer” to be the same as the first part influences the modeling of the second part by sending information to the second part because they are both directed to passing domain-classification training information from the domain classifier side back to the feature extractor side during training.)
the second part excludes, based on the information from the first part, influence of the domain-specific characteristics in learning domain-invariant semantics of the predetermined attributes by negating, impact of the domain-specific characteristics, in optimization in training the second part. ([Ganin , page 2-3] “the parameters of the underlying deep feature mapping are optimized in order to minimize the loss of the label classifier and to maximize the loss of the domain classifier. The latter update thus works adversarially to the domain classifier, and it encourages domain-invariant features to emerge in the course of the optimization.”, AND [Ganin , page 12] “During the backpropagation however, the GRL takes the gradient from the subsequent level and changes its sign, i.e., multiplies it by -1, before passing it to the preceding layer.”, wherein the examiner interprets “the parameters of the deep feature mapping are optimized to maximize the loss of the domain classifier” combined with “the GRL changes its sign, multiplies it by -1, during backpropagation” to be the same as the second part excluding influence of domain-specific characteristics by negating impact of domain-specific characteristics in optimization, because they are both directed to employing a negation mechanism during training that causes the feature extractor to learn representations that prevent the domain classifier from successfully discriminating between domains, thereby excluding domain-specific information and learning domain-invariant semantics by inverting the optimization signal so that domain classification failure becomes the training objective.) Claims 12 and 18 are analogous to claim 5, aside from claim type and minute differences, hence the same rejection can apply.
Claim(s) 6-7, 13-14, and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Ganin in view of Zheng in view of Chen further in view of NPL reference “Attention-Based Recurrent Neural Network Models for Joint Intent Detection and Slot Filling”, by NPL reference Liu et. al. (referred herein as Liu).
Regarding claim 6, Ganin, Zheng, Chen teach The method of claim 4, (see rejection of claim 4).
Ganin, Zheng, Chen do not teach wherein the second part is constructed using an encoder-decoder architecture with an attention layer between an encoder layer and a decoder layer.
Liu teaches wherein the second part is constructed using an encoder-decoder architecture with an attention layer between an encoder layer and a decoder layer. ([Liu, page 1] “Attention-based encoder-decoder neural network models have recently shown promising results in machine translation and speech recognition. In this work, we propose an attention-based neural network model for joint intent detection and slot filling ... The main idea behind the encoder-decoder model is to encode input sequence into a dense vector, and then use this vector to generate corresponding output sequence. The attention mechanism introduced in [12] enables the encoder-decoder architecture to learn to align and decode simultaneously” AND [Liu, page 2] “The encoder and decoder are two separate RNNs. The encoder reads a sequence of input (x1, ..., xT) to a vector c. This vector encodes information of the whole source sequence, and is used in decoder to generate the target output sequence”, wherein the examiner interprets Liu’s attention-based neural network model for joint intent detection and slot filling to be the same as the second part because they are both directed to a neural-network component used to identify semantic constituents or extracted information from input language. The examiner further interprets Liu’s encoder-decoder model and encoder-decoder architecture to be the same as constructed using an encoder-decoder architecture because they are both directed to a model structure in which an encoder processes an input sequence and a decoder generates an output sequence. Also, Liu’s attention mechanism to be the same as an attention layer because they are both directed to an attention component used with the encoder-decoder architecture to align input information during decoding.)
Ganin, Zheng, Chen, Liu, and the instant application are analogous art because they are all directed to neural-network models using learned representations to process input data and generate task-specific outputs.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method of claim 4 disclosed by Ganin, Zheng, and Chen to include the “attention-based encoder-decoder neural network” disclosed by Liu. One would be motivated to do so to effectively encode input language sequences, generate corresponding output sequences, and improve alignment during decoding, as suggested by Liu ([Liu, page 1] “The attention mechanism introduced in [12] enables the encoder-decoder architecture to learn to align and decode simultaneously.”). Claims 13 and 19 are analogous to claim 6, aside from claim type and minute differences, hence the same rejection can apply.
Regarding claim 7, Ganin, Zheng, Chen, and Liu teach The method of claim 6, (see rejection of claim 6)
Liu further teaches wherein the encoder layer comprises a first multiple serially connected long short-term memory (LSTM) units, each of which corresponds to a token from the textual content; ([Liu, page 1] “Recently, encoder-decoder neural network models have been successfully applied in many sequence learning problems” AND [Liu, page 2] “The encoder and decoder are two separate RNNs … On encoder side, we use a bidirectional RNN. Bidirectional RNN has been successfully ap plied in speech recognition [15] and spoken language understanding [6]. We use LSTM [16] as the basic recurrent network unit for its ability to better model long-term dependencies com paring to simple RNN… The bidirectional RNN encoder reads the source word sequence forward and backward. The forward RNN reads the word sequence in its original order and generates a hidden state f_hi at each time step”, wherein the examiner interprets “On encoder side, we use a bidirectional RNN” and “We use LSTM [16] as the basic recurrent network unit” to be the same as the encoder layer comprises long short-term memory (LSTM) units because they are both directed to implementing the encoder with LSTM recurrent units, “reads the word sequence in its original order” and “at each time step” to be the same as a first multiple serially connected because they are both directed to recurrent sequence processing over ordered positions, and “word sequence” to be the same as a token from the textual content because they are both directed to text input units processed by the model.)
and the decoder layer comprises a second multiple serially connected long short-term memory (LSTM) units, each of which corresponds to one of the predetermined attributes recognized from the textual content. ([Liu, page 1] “extracting semantic constituents from the natural language query … Slot filling can be treated as a sequence labeling task” AND [Liu, page 2] “we want to map a word sequence x = (x1, ..., xT) to its corresponding slot label sequence y = (y1, ..., yT)” AND [Liu, page 2] “The decoder is a unidirectional RNN. Again, we use an LSTM cell as the basic RNN unit. At each decoding step i, the decoder state si is calculated as a function of the previous decoder state si-1, the previous emitted label yi-1, the aligned encoder hidden state hi, and the context vector c_i”, wherein the examiner interprets “The decoder is a unidirectional RNN” and “Again, we use an LSTM cell as the basic RNN unit” to be the same as the decoder layer comprises long short-term memory (LSTM) units because they are both directed to implementing the decoder with LSTM recurrent units, “At each decoding step i” and “the previous decoder state si-1” to be the same as a second multiple serially connected because they are both directed to sequential decoder processing in which a decoder state depends on a previous decoder state, and “extracting semantic constituents from the natural language query,” “Slot filling,” and “corresponding slot label sequence y = (y1, ..., yT)” to be the same as each of which corresponds to one of the predetermined attributes recognized from the textual content because they are both directed to recognizing semantic slot or attribute information from the input text and producing corresponding output labels.)
Ganin, Zheng, Chen, Liu, and the instant application are analogous art because they are all directed to neural-network model training and sequence-processing architectures.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to modify the method of claim 6 disclosed by Ganin, Zheng, Chen, and Liu to include the “encoder-decoder neural network models” using “LSTM” recurrent units disclosed by Liu. One would be motivated to do so to effectively process ordered text sequences and generate corresponding output labels using a recurrent encoder-decoder architecture, as suggested by Liu ([Liu, page 1] “strong modeling capacity of the sequence models … provide additional information to the intent classification and slot label prediction.”). Claims 14 and 20 are analogous to claim 7, aside from claim type and minute differences, hence the same rejection can apply.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DEVAN KAPOOR whose telephone number is (703)756-1434. The examiner can normally be reached Monday - Friday: 9:00AM - 5:00 PM EST (times may vary).
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, David Yi can be reached at (571) 270-7519. 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.
/DEVAN KAPOOR/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights