DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 10/27/2025 has been entered.
Response to Arguments
1. Regarding the rejection of claims 1, 2, 4-13, and 15-21 under 35 U.S.C. § 101, Applicant's arguments filed 10/27/2025 have been fully considered but they are not persuasive.
Applicant argues on pgs. 9-10 that that the claims recite significantly more than an abstract idea and is directed to a specific technological improvement in speech recognition (see pg. 9, 2nd para.). Specifically, Applicant argues regarding claim 1 that the claimed storing operation implemented by the recited data-storage module (pg. 9, 2nd para.), the model-building module configured to construct and update classified language models (pg. 9, 3rd para.), and the operation of updating probability scores only between words that appear in the piece of new corpus data (pg. 10, 1st para.) provide a technical improvement. The Examiner respectfully disagrees.
The claims as currently written are directed to abstract ideas without significantly more. Under Step 2A Prong 1 analysis, the claims recite both mental processes which can be performed by a person with the aid of pen and paper, as well as mathematical concepts. Under Step 2A Prong 2 analysis, the additional elements present in claim 1 recite mere instructions to implement the judicial exception using generic computer components, which do not integrate the judicial exception into a practical application as they do not impose any meaningful limits on practicing the recited abstract ideas.
(1) The data storage module recited in claim 1 implements a step which can be performed mentally by a person using pen and paper. Specifically, a person can write down multiple pieces of corpus data on a piece of paper, and can then add or strikethrough data as the data is updated. The data-storage module itself is described at a high level of generality and amounts to mere instructions to implement the judicial exception using a generic computer component under Step 2A Prong 2 analysis.
(2) Furthermore, the step performed by the model-building module can be performed mentally with the aid of pen and paper. Specifically, a person can update a set of rules for a particular model and its domain using pen and paper (e.g. can decide that a new utterance refers to a particular model related to ‘history’, and update that particular model with the new data). The model-building model as well as the classified language models themselves are described at a high level of generality, and additionally amount to mere instructions to implement the judicial exception using generic computer components.
(3) Finally, the step of updating the probabilities by only updating probability scores between words that appear in the piece of new corpus data and does not update probability scores between words that do not appear in the piece of new corpus data does not integrate the judicial exception into a practical application. Specifically, a person can choose which probability scores need updated (e.g. determine that a particular n-gram needs updated probability score if that n-gram is present in a particular new corpus data). The updating of the probability score itself further recites a mathematical calculation, which falls under the abstract idea grouping of mathematical concept.
Hence, Applicant’s arguments are not persuasive.
2. Regarding the rejection of claims 1, 2, 4, 5, 7-10, 12, 13, 15, 16, and 18-21 under 35 U.S.C. § 103 as being unpatentable over Rao in view of Biadsy, the rejection of claims 6 and 17 under 35 U.S.C. § 103 as being unpatentable over Rao in view of Biadsy and further in view of Conneau, and the rejection under 35 U.S.C> § 103 as being unpatentable over Rao in view of Biadsy and further in view of Strimel, Applicant’s arguments have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Claim Rejections - 35 USC § 101
1. Claims 1-2, 4-13, and 15-21 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1, “A system” is recited, which is directed to one of the four statutory categories of invention (machine) (Step 1: YES). However, the claims limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts which fall under the category of abstract idea (Step 2A Prong 1: YES).
The following limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts:
“… store multiple pieces of corpus data corresponding to multiple categories”: a person keeps a list of corpus data on pen and paper, which is sorted by category.
“… store a piece of new corpus data …, wherein the piece of new corpus data corresponds to one of the categories”: a person takes a new piece of corpus data, and stores it in the list of corpus data kept on pen and paper, and writes it under the category it corresponds to.
“… construct a plurality of classified language models, and update one of the classified language models based on the piece of new corpus data stored …, wherein the classified language model updated corresponds to the category that corresponds to the piece of new corpus data.”: a person constructs a model (set of rules) for each category of corpus data using pen and paper, and then updates a particular model depending on the category of the new corpus data.
…record sentences that are unrecognizable…and convert the sentences into the multiple pieces of new corpus data: a person listens to words, and writes them down as sentences using pen and paper, before splitting the sentences into pieces of corpus data
“wherein the model-building module updates the classified language model by only updating probability scores between words in the piece of new corpus data, and not updating probability scores of words that are not in the piece of new corpus data, wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data”: a person updates probability scores between only words in a particular piece of text data (e.g. looks to update only n-gram probabilities which have n-grams present in new data), using pen and paper, the updating of the probability scores itself using the occurrence frequency amounts to a mathematical calculation.
Claim 1 does not contain any additional elements which integrate the judicial exception into a practical application (Step 2A Prong 2: NO). The only additional limitations are “a data-storage module, implemented using a memory, configured to store”, “a data-update module, implemented using a processing device, configured to store”, and “a model-building module, implemented using the processing device, configured to construct”, “a data collection module, implemented by a client device, configured to…by the client device through speech recognition technologies”. All of these limitations are recited at a high level of generality, and amount to mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer is not enough to integrate the judicial exception into a practical application as it does not impose any meaningful limits on practicing the abstract idea. Accordingly, claim 1 is directed to an abstract idea.
Claim 1 does not contain any additional elements that amount to significantly more than the judicial exception (Step 2B: NO). As discussed above with respect to integrating the judicial exception into a practical application, the additional limitations are mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer cannot amount to significantly more than the judicial exception as they do not provide an inventive concept. Therefore, claim 1 is not patent eligible.
Regarding dependent claims 2 and 4-11, “The system” is recited, which is directed to one of the four statutory categories of invention (machine) (Step 1: YES). However, the claims limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts which fall into the category of abstract idea (Step 2A Prong 1: YES).
Claims 2 and 4-11 recite the mental processes and mathematical concepts of claim 1 due to their dependence on claim 1. Additionally, the following limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts, both of which fall under the category of abstract idea (Step 2A Prong 1: YES):
Claim 2: “wherein the classified language model uses n-grams to calculate probability scores between words”: calculating probabilities using n-grams amounts to mathematical calculation.
Claim 4: “wherein the model-building module further updates a generic language model based on the piece of new corpus data stored in the data-storage module”: a person uses a new piece of data to update a generic language model, representing a generic set of rules for interpreting the corpus data
Claim 5: “using a classification model to determine the category that corresponds to the piece of new corpus data”: a person reads the new corpus data, then uses a classification model as a set of rules to classify the new data into a category
Claim 7: “extracts a feature vector of the piece of new corpus data, inputs the feature vector into the classification model, and determines the category that corresponds to the piece of new corpus data according to a result output by the classification model”: a person reads the new corpus data, writes down a vector of numbers reflecting the words in the corpus data, and uses the classification model as a set of rules to determine a category based on the feature vector
Claim 8: “wherein the corpus-classification module uses a term frequency-inverse document frequency (tf-idf) approach to extract the feature vector from the piece of new corpus data”: using tf-idf to extract the feature vector is a mathematical concept
Claim 9: “stores the corpus data that correspond to the category as a classified corpus”: a person writes down corpus data on a piece of paper in groups based on their classification
Claim 10: “stores a category label of the category that corresponds to each piece of corpus data”: a person writes down a category label next to each piece of corpus data using pen and paper.
Claim 11: “wherein in response to the amount of new corpus data accumulated exceeding a threshold…updating the classified model”: a person writes down corpus data using pen and paper, and decides to update the model after a certain number of corpus data have been written down.
Claims 2 and 4-11 do not contain any additional elements which integrate the judicial exception into a practical application (Step 2A Prong 2: NO). The only additional limitations are: those discussed above for claim 1, “a corpus-classification module, implemented using the processing device, configured to use” (claim 5), “wherein the classification model is a Fully-Connected Neural Network” (claim 6), “uploads the accumulated new corpus data to a backend server” (claim 11), and “wherein the data-collection module is executed by the client device, and the data-update module, the data-storage module and the model-building module are executed by the backend server” (claim 11). All of these limitations are recited at a high level of generality, and amount to mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer is not enough to integrate the judicial exception into a practical application as they do not impose any meaningful limits on practicing the abstract idea. Accordingly, claims 2 and 4-11 are directed to an abstract idea.
Claims 2 and 4-11 do not contain any additional elements that amount to significantly more than the judicial exception (Step 2B: NO). As discussed above with respect to integrating the judicial exception into a practical application, the additional limitations are mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer cannot amount to significantly more than the judicial exception as they do not provide an inventive concept. Therefore, claims 2 and 4-11 are not patent eligible.
Regarding claim 12, “A method” is recited, which is directed to one of the four statutory categories of invention (process) (Step 1: YES). However, the claims limitations, under their broadest reasonable interpretation, recite mental processes and mathematical concepts which fall into the category of abstract idea.
The following limitations, under their broadest reasonable interpretation, recite mental processes (Step 2A Prong 1: YES):
“storing a piece of new corpus data …storing multiple pieces of corpus data corresponding to multiple categories, and the piece of new corpus data corresponds to one of the categories”: a person keeps a list of corpus data on pen and paper, which are each sorted into a particular category.
“updating one of a plurality of classified language models based on the piece of new corpus data stored …, wherein the classified language model updated corresponds to the category that corresponds to the piece of new corpus data.”: a person constructs a model (set of rules) for each category of corpus data using pen and paper, and then updates a particular model depending on the category of the new corpus data.
“wherein the model-building module updates the classified language model by only updating probability scores between words in the piece of new corpus data, and not updating probability scores of words that are not in the piece of new corpus data, wherein the probabilities scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data”: a person updates probability scores between only words in a particular piece of text data (e.g. looks to update only n-gram probabilities which have n-grams present in new data), using pen and paper, the updating of the probability scores itself using the occurrence frequency amounts to a mathematical calculation.
wherein sentences that are unrecognizable…is recorded and converted into the multiple pieces of new corpus data: a person listens to words, and writes them down as sentences using pen and paper, before splitting the sentences into pieces of corpus data
Claim 12 does not contain any additional elements which integrate the judicial exception into a practical application (Step 2A Prong 2: NO). The only additional limitations are “a computer system”, “a data-storage module of the computer system”, and “by a client device through speech recognition technologies…”. These limitations are recited at a high level of generality, and amount to mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer is not enough to integrate the judicial exception into a practical application as they do not impose any meaningful limits on practicing the abstract idea. Accordingly, claim 12 is directed to an abstract idea.
Claim 12 does not contain any additional elements that amount to significantly more than the judicial exception (Step 2B: NO). As discussed above with respect to integrating the judicial exception into a practical application, the additional limitation is mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer cannot amount to significantly more than the judicial exception as they do not provide an inventive concept. Therefore, claim 12 is not patent eligible.
Regarding dependent claims 13 and 15-21, “The method” is recited, which is directed to one of the four statutory categories of invention (process) (Step 1: YES). However, the claims limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts which fall into the category of abstract idea (Step 2A Prong 1: YES).
Claims 13 and 15-21 recite the mental processes of claim 12 due to their dependence on claim 12. Additionally, the following limitations, under their broadest reasonable interpretation, recite mental processes or mathematical concepts, both of which fall under the category of abstract idea:
Claim 13: “wherein the classified language model uses n-grams to calculate probability scores between words”: a person uses n-grams to calculate probabilities between words using pen and paper.
Claim 15: “updating a generic language model based on the piece of new corpus data stored in the data-storage module”: a person uses a new piece of data to update a generic language model, representing a generic set of rules for interpreting the corpus data
Claim 16: “using a classification model to determine the category that corresponds to the piece of new corpus data”: a person reads the new corpus data, then uses a classification model as a set of rules to classify the new data into a category
Claim 18: “extracting a feature vector of the piece of new corpus data, inputting the feature vector into the classification model, and determining the category that corresponds to the piece of new corpus data according to a result output by the classification model”: a person reads the new corpus data, writes down a vector of numbers reflecting the words in the corpus data, and uses the classification model as a set of rules to determine a category based on the feature vector
Claim 19: “using a term frequency-inverse document frequency (tf-idf) approach to extract the feature vector from the piece of new corpus data”: using tf-idf to extract the feature vector is a mathematical concept
Claim 20: “storing the corpus data that correspond to the category as a classified corpus”: a person writes down corpus data on a piece of paper in groups based on their classification
Claim 21: “storing a category label of the category that corresponds to the piece of new corpus data into the data-storage module”: a person writes down a category label next to each piece of corpus data using pen and paper.
Claims 13 and 15-21 do not contain any additional elements which integrate the judicial exception into a practical application (Step 2A Prong 2: YES). The only additional limitations are: those discussed above for claim 12 and “wherein the classification model is a Fully-Connected Neural Network” (claim 17). All of these limitations are recited at a high level of generality, and amount to mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer is not enough to integrate the judicial exception into a practical application as they do not impose any meaningful limits on practicing the abstract idea. Accordingly, claims 13 and 15-21 are directed to an abstract idea.
Claims 13 and 15-21 do not contain any additional elements that amount to significantly more than the judicial exception (Step 2B: NO). As discussed above with respect to integrating the judicial exception into a practical application, the additional limitations are mere instructions to implement the judicial exception using a generic computer. Mere instructions to implement the judicial exception using a generic computer cannot amount to significantly more than the judicial exception as they do not provide an inventive concept. Therefore, claims 13 and 15-21 are not patent eligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
4. Claims 1-2, 4-5, 7-9, 11-13, 15-16, and 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Rao et al. (US PGPUB No. 2014/0214419, hereinafter Rao) in view of Strimel et al. (US 11,887,583 B1, hereinafter Strimel) and further inv iew of Itoh et al. (US 2015/0294665 A1, hereinafter Itoh).
Regarding claim 1, Rao discloses a data-storage module, implemented using a memory (Fig. 12, “Classifying process module 1218” in “Memory 1206”), configured to store multiple pieces of corpus data corresponding to multiple categories (Fig. 6, Output of “Classifying processing module 602” is classified corpuses corresponding to category 1-m; para. 0073 “Classifying processing module 602, configured to carry out the corpus classification calculation for the raw corpus so as to obtain a different categories of more than one classifying corpus.”; Corpus data is stored as “Classifying process module” resides in memory (Fig. 12, 1216)); a data-update module, implemented using a processing device (Fig. 10, “Classifying process module 1216” in “Memory 1206” run on CPU 1202 ), configured to store a piece of new corpus data into the data-storage module (Fig. 6, “Classifying processing module 602” takes as input new corpus data “Raw corpus” and classifies the data first so that it can then be stored in one of the classifying corpuses; Fig. 2, Step 201; para. 0037 “Step 201, carry out the corpus classification calculation for the raw corpus so as to obtain different categories of more than one classifying corpus.”), wherein the piece of new corpus data corresponds to one of the categories (New corpus data “Raw corpus” is corresponded into one of m categories; para. 0037 “…For example, the mentioned classifying corpus can be divided into many types, such as person name, place name, computer term, medical terminology, etc. For example, "isatis root" belongs to the classification of medical terminology. A term may belong to multi-classification.”); a model-building module, implemented using the processing device (Fig. 12, system contains a model-building module comprised of two modules: “Classifying language model training module 1250” and “Primary language model training module 1216” run on CPU 1202), configured to construct a plurality of classified language models (Fig. 6 “Classifying language model training module 603” takes as input the m different classified corpuses, and builds m different “Classifying language model(s)”; Fig. 2, step 202; para. 0038 “Step 202, carry out a language model training calculation for every mentioned classifying corpus to obtain more than one corresponding classifying language models.”), and update one of the classified language models based on the piece of new corpus data stored in the data-storage module (Fig. 6, “Classifying language model” 1-m are updated using new corpus data that has been stored (using “Raw corpus” data that has been stored in a “Classifying corpus” 1-m))… wherein the classified language model updated corresponds to the category that corresponds to the piece of new corpus data (Each “Classifying language model” 1-m is trained using the corresponding “Classifying corpus” 1-m; Fig. 2, step 202; para. 0038 “Step 202, carry out a language model training calculation for every mentioned classifying corpus to obtain more than one corresponding classifying language models.”).
Rao does not specifically disclose a data collection module, implemented by a client device, configured to record sentences that are unrecognizable by the client device through speech recognition technologies, and convert the sentences into the multiple pieces of new corpus data.
Strimel teaches a data-collection module, implemented by the client device (Fig. 1, 160, in device 110), configured to record sentences that are unrecognizable by a client device through speech recognition technologies (Col. 6 Lines 37-40 ”At a stage 160, the device 110 may process input data 111 from the user 5 using the updated model 131b. The input data 111 may include spoken and/or written natural language input,”), and converting the sentences into multiple pieces of new corpus data (Model 131 is taught to include automatic speech recognition models, configured to output recognized speech as text data; Col. 5 Lines 23-31 “The trained model 131 can include, for example and without limitation…an ASR acoustic and/or language model…”; Col. 20 Lines 5-10 “The ASR component 1250 may transcribe the audio data 1211 into text data. The text data output by the ASR component 1250 represents one or more than one (e.g., in the form of an N-best list) ASR hypotheses representing speech represented in the audio data 1211”)
Rao and Strimel are considered to be analogous to the claimed invention as
they both are in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rao to incorporate the teachings of Strimel in order to implement a client device has a data collection module configured to record sentences unrecognizable by the client device through speech recognition technologies, and convert the sentences into multiple pieces of new corpus data. Doing so would beneficial, as this would collect new data which can be used to further update the model to improve accuracy with respect to the new data (Strimel, Col. 3 Lines 1-36).
Rao in view of Strimel does not specifically disclose wherein the model-building model updates the classified language model by updating probability scores only between words that appear in the piece of new corpus data, and does not update probability scores between words that do not appear in the piece of new corpus data, wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data.
Itoh teaches wherein the model-building model (unsupervised training system 202) updates the classified language model (baseline model: para. 0047 “The corpus A302 is a baseline corpus used to build a part that forms the basis for the N-gram language model. As an example, the corpus A302 may be a corpus having a domain and a style consistent with a target application.”) by updating probability scores only between words that appear in the piece of new corpus data, and does not update probability scores between words that do not appear in the piece of new corpus data (model updated with recognition results of speech data: para. 0048 “The corpus B304 is a corpus composed of recognition results of automatic speech recognition of speech data without manual intervention.”; para. 0054 “The language model training section 308 uses both the corpus A302 and the corpus B304 as training text to build an N-gram language model.”; may use only new corpus data B304 to add on to existing n-gram language model by only adding probabilities of new corpus data n-grams to the original model, instead of building from scratch (updating all probabilities): para. 0057 “Note that the use of the corpus A302 is optional, and N-gram entries may be selected without using the corpus A302.”; para. 0062 “The language model training section 308 may add, to the base N-gram language model, the selected one or more N-gram entries and their probabilities obtained as a result of the training, or build the N-gram language model from scratch.”), wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data (probabilities for N-grams based on occurrence frequencies of word sequences: para. 0061 “The probability calculation section 312 uses all the recognition results included in the corpus A302 and the corpus B304 to train the N-gram language model about one or more N-gram entries selected by the selection section 310, or when only the corpus B is used as training data, the probability calculation section 312 uses all the recognition results included in the corpus B to train the N-gram language model about the selected one or more N-gram entries.”; para. 0040 “The conditional probabiliites of the above mathematical expression (2) can be determined by using maximum likelihood estimation from the number of appearances of a string of N words and a string of (N-1) words appears in a corpus…”, see Eqs. (1)-(3)).
Rao, Strimel, and Itoh are considered to be analogous to the claimed invention as they are all in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rao in view of Strimel to incorporate the teachings of Itoh in order to have the model building model update the classified language model by updating probability scores only between words that appear in the piece of new corpus data, and does not update probability scores between words that do not appear in the piece of new corpus data, wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data. Doing so would beneficial, as this would allow for necessary updates to language models to be obtained without the need to necessarily train from scratch (Itoh, para. 0004, para. 0062).
Regarding claim 2, Rao in view of Strimel in further view of Itoh discloses wherein the classified language model uses n-grams to calculate probabilities between words (Itoh, Equations (1)-(3); para. 0061 “The probability calculation method is as described in the outline of the N-gram language model.”).
Rao, Strimel, and Itoh are considered to be analogous to the claimed invention as they are all in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rao to incorporate the teachings of Itoh in order to use n-grams to calculate probabilities between words. Doing so would beneficial, as n-gram language models are simple yet effectively, commonly used in large vocabulary speech recognition tasks (Itoh, para. 0037).
Regarding claim 4, Rao in view of Strimel in further view of Itoh discloses wherein the model-building module further updates a generic language model based on the piece of new corpus data stored in the data-storage module (Rao, Fig. 10, “Primary Language Model Training Module 1216”; Fig. 6, a generic language model is trained via “Primary language model training module 601” to obtain “Primary language model” using pieces of new corpus data “Raw corpus”; para. 0043 “Step 301, carry out a calculation of the language model training according to the raw corpus to obtain the primary language model. Here, the language model training is the conventional regular language model training.”).
Regarding claim 5, Rao in view of Strimel in further view of Itoh discloses further comprising a corpus-classification module, implemented using the processing device (Rao, Fig. 12, “Classifying process module 1218” in “Memory 1206” run on CPU 1202), configured to use a classification model to determine the category that corresponds to the piece of new corpus data (Rao, Fig. 7, A category (which “Classifying corpus 1-m” the piece of corpus corresponds to) is determined for each piece of “Raw corpus”, using a classification model “Classifier 704”; para. 0083 “Classifier 704, configured to put the term characteristic after the processing of dimensionality reduction into the classifier for training, output different categories of more than one classifying corpus. In a preferred embodiment, the mentioned classifier is a SVM classifier.”).
Regarding claim 7, Rao in view of Strimel in further view of Itoh discloses wherein the corpus-classification module extracts a feature vector of the piece of new corpus data (Rao, Fig. 7 discloses structure of the corpus-classification module; “Characteristic extracting module 702” extracts a feature vector of the new corpus data “Raw corpus”, which is then reduced by dimensionality by “Dimensionality reduction module 703”; para. 0081 “Characteristic extracting module 702, configured to use TF-IDF method to extract the term characteristic from the raw corpus”; para. 0082 “Dimensionality reduction module 703, configured that according to the mentioned affinity matrix, use the dimensionality reduction method to process dimension reduction for the extracted term characteristic. In a preferred embodiment, the mentioned dimensionality reduction module is PCA dimensionality reduction module.”), inputs the feature vector into the classification model (Rao, Feature vectors (output of “Dimensionality reduction module 703”) is fed into “Classifier 704”), and determines the category that corresponds to the piece of new corpus data according to a result output by the classification model (Rao, “Classifier 704“ determines a category corresponds to the piece of new corpus data (a category 1-m) according to a result output by classifier (output of SVM); para. 0083 “Classifier 704, configured to put the term characteristic after the processing of dimensionality reduction into the classifier for training, output different categories of more than one classifying corpus. In a preferred embodiment, the mentioned classifier is a SVM classifier.”).
Regarding claim 8, Rao in view of Strimel in further view of Itoh discloses wherein the corpus-classification module uses a term frequency-inverse document frequency (tf-idf) approach to extract the feature vector from the piece of new corpus data (Rao, “Characteristic extracting module 702” extracts features using tf-idf; para. 0081 “Characteristic extracting module 702, configured to use TF-IDF method to extract the term characteristic from the raw corpus.”).
Regarding claim 9, Rao in view of Strimel in further view of Itoh discloses wherein the data-storage module further stores the corpus data that correspond to the category as a classified corpus (Rao, Data storage module (“Fig. 12, “Classifying process module 1218” in “Memory 1206”) classifies the data into a plurality of classified corpuses 1-m, which are stored as they reside in memory 1206).
Regarding claim 11, Rao in view of Strimel and Itoh discloses wherein in response to the amount of new corpus data accumulated exceeding a threshold (Strimel, Col. 7 Lines 23-27 “At some point, perhaps based on a predetermined schedule, a change in accuracy of the first-generation model M.sub.0, or a threshold amount of usage data d.sub.t collected, the system 100 may commence operations to update the on-device model with the next-generation model M.sub.1.”), the data-collection module uploads the accumulated new corpus data to a backend server for updating the classified language model (Strimel, Fig. 1, in step 170, use data is sent to a backend server (System 120, Col. 54 Lines 66-67 “A system (120/125) may include one or more servers.”) at step 130, the “use data 171” including output of the model; Col. 5 Lines 32-35 “The operations of system 100 may begin by receiving, at a stage 130, usage data representing interactions between the user 5 and the system 100. The interactions may include user input data 111 and/or system output data 112.”; received usage data at step 130 is used to update the language model at step 140); and wherein the data-collection module is executed by the client device (Strimel, Fig. 1, Data collection module “160” exists on a client device (user device 110 operated by user 5)), and the data-update module, the data-storage module and the model-building module are executed by the backend server (Strimel, data is updated and stored (step 130), and a model is built (step 140) using the server (system 120)).
Rao, Strimel, and Itoh are considered to be analogous to the claimed invention as they are all in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of Strimel in order to implement a client and server, and uploading data to the server to update a language model. Doing so would beneficial, as utilizing a client-server architecture for implementing the system of Rao would allow for benefits such as scalability and centralized management.
Regarding claim 12, Rao discloses storing a piece of new corpus data into a data-storage module of the computer system (Rao, Fig. 12, “Classifying process module 1218” in “Memory 1206”), wherein the data storage module is used for storing multiple pieces of corpus data corresponding to multiple categories (Rao, Fig. 6, Output of “Classifying processing module 602” is classified corpuses corresponding to category 1-m; para. 0073 “Classifying processing module 602, configured to carry out the corpus classification calculation for the raw corpus so as to obtain a different categories of more than one classifying corpus.”; Corpus data is stored as “Classifying process module” resides in memory (Fig. 12, 1216)), and the piece of new corpus data corresponds to the categories (New corpus data “Raw corpus” is corresponded into one of m categories; para. 0037 “…For example, the mentioned classifying corpus can be divided into many types, such as person name, place name, computer term, medical terminology, etc. For example, "isatis root" belongs to the classification of medical terminology. A term may belong to multi-classification.”); and updating one of a plurality of classified language models based on the piece of new corpus data stored in the data-storage module (Fig. 6, “Classifying language model” 1-m are updated using new corpus data that has been stored (using “Raw corpus” data that has been stored in a “Classifying corpus” 1-m)), wherein the classified language model updated corresponds to the category that corresponds to the piece of new corpus data (Each “Classifying language model” 1-m is trained using the corresponding “Classifying corpus” 1-m; Fig. 2, step 202; para. 0038 “Step 202, carry out a language model training calculation for every mentioned classifying corpus to obtain more than one corresponding classifying language models.”)….
Rao does not specifically disclose wherein sentences that are unrecognizable by a client device through speech recognition technologies is recorded and converted into the multiple pieces of new corpus data.
Strimel teaches wherein sentences that are unrecognizable by a client device (Fig. 1, 160, in device 110), through speech recognition technologies is recorded (Col. 6 Lines 37-40 ”At a stage 160, the device 110 may process input data 111 from the user 5 using the updated model 131b. The input data 111 may include spoken and/or written natural language input,”), and converted into the multiple pieces of new corpus data (Model 131 is taught to include automatic speech recognition models, configured to output recognized speech as text data; Col. 5 Lines 23-31 “The trained model 131 can include, for example and without limitation…an ASR acoustic and/or language model…”; Col. 20 Lines 5-10 “The ASR component 1250 may transcribe the audio data 1211 into text data. The text data output by the ASR component 1250 represents one or more than one (e.g., in the form of an N-best list) ASR hypotheses representing speech represented in the audio data 1211”)
Rao and Strimel are considered to be analogous to the claimed invention as
they both are in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rao to incorporate the teachings of Strimel in order to implement a client device configured to record sentences unrecognizable by the client device through speech recognition technologies, and convert the sentences into multiple pieces of new corpus data. Doing so would beneficial, as this would collect new data which can be used to further update the model to improve accuracy with respect to the new data (Strimel, Col. 3 Lines 1-36).
Rao in view of Strimel does not specifically disclose wherein the model-building module updates the classified language model by only updating probability scores between words in the piece of new corpus data, and not updating probability scores of words that are not in the piece of new corpus data, wherein the probability scores are calculated based on occurrence frequency of word sequences in the piece of new corpus data.
Itoh teaches wherein the model-building model (unsupervised training system 202) updates the classified language model (baseline model: para. 0047 “The corpus A302 is a baseline corpus used to build a part that forms the basis for the N-gram language model. As an example, the corpus A302 may be a corpus having a domain and a style consistent with a target application.”) by only updating probability scores between words in the piece of new corpus data, and not updating probability scores of words that are not in the piece of new corpus data (model updated with recognition results of speech data: para. 0048 “The corpus B304 is a corpus composed of recognition results of automatic speech recognition of speech data without manual intervention.”; para. 0054 “The language model training section 308 uses both the corpus A302 and the corpus B304 as training text to build an N-gram language model.”; may use only new corpus data B304 to add on to existing n-gram language model by only adding probabilities of new corpus data n-grams to the original model, instead of building from scratch (updating all probabilities): para. 0057 “Note that the use of the corpus A302 is optional, and N-gram entries may be selected without using the corpus A302.”; para. 0062 “The language model training section 308 may add, to the base N-gram language model, the selected one or more N-gram entries and their probabilities obtained as a result of the training, or build the N-gram language model from scratch.”), wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data (probabilities for N-grams based on occurrence frequencies of word sequences: para. 0061 “The probability calculation section 312 uses all the recognition results included in the corpus A302 and the corpus B304 to train the N-gram language model about one or more N-gram entries selected by the selection section 310, or when only the corpus B is used as training data, the probability calculation section 312 uses all the recognition results included in the corpus B to train the N-gram language model about the selected one or more N-gram entries.”; para. 0040 “The conditional probabiliites of the above mathematical expression (2) can be determined by using maximum likelihood estimation from the number of appearances of a string of N words and a string of (N-1) words appears in a corpus…”, see Eqs. (1)-(3)).
Rao, Strimel, and Itoh are considered to be analogous to the claimed invention as they are all in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Rao in view of Strimel to incorporate the teachings of Itoh in order to have the model building model update the classified language model by updating probability scores only between words that appear in the piece of new corpus data, and does not update probability scores between words that do not appear in the piece of new corpus data, wherein the probability scores are calculated based on occurrence frequencies of word sequences in the piece of new corpus data. Doing so would beneficial, as this would allow for necessary updates to language models to be obtained without the need to necessarily train from scratch (para. 0004, para. 0062).
Regarding claim 13, claim 13 is rejected for analogous reasons to claim 2.
Regarding claim 15, claim 15 is rejected for analogous reasons to claim 4.
Regarding claim 16, claim 16 is rejected for analogous reasons to claim 5.
Regarding claim 18, claim 18 is rejected for analogous reasons to claim 7.
Regarding claim 19, claim 19 is rejected for analogous reasons to claim 8.
Regarding claim 20, claim 20 is rejected for analogous reasons to claim 9.
5. Claims 6 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Rao in view of Strimel and Itoh, and further in view of Conneau et al. (NPL Very Deep Convolutional Networks for Text Classification, hereinafter Conneau).
Regarding claim 6, Rao in view of Strimel and Itoh does not specifically disclose wherein the classification model is a Fully-Connected Neural Network.
Conneau teaches wherein the classification model is a Fully-Connected Neural Network (a text classification system is taught (pg. 4, Fig. 1), which is used to classify text in one or more categories (see pg. 6, Table 3, e.g. classifying DBPedia text dataset into one of 14 classes). The system uses a classification model which is a fully-connected neural network (pg. 4, Fig. 1, top 3 layers denoted “fc”; pg. 4, para 3 “The 512 x k resulting features are transformed into a single vector which is the input to a three layer fully connected classifier with ReLU hidden units and softmax outputs. The number of output neurons depends on the classification task…”))
Rao, Strimel, Itoh, and Conneau are considered to be analogous to the claimed invention as they are in the same field of classifying text into corpus categories. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of Conneau in order to use a fully-connected neural network as the classification model. Doing so would beneficial, as fully-connected neural networks are structure agnostic, allowing for greater model input flexibility (NPL Ramsundar et al., TensorFlow for Deep Learning, Chapter 4, first paragraph).
Regarding claim 17, claim 17 is rejected for analogous reasons to claim 6.
6. Claims 10 and 21 are rejected under 35 U.S.C. 103 as being unpatentable over Rao in view of Strimel and Itoh, and further in view of Biadsy et al. (US 2018/0053502 A1, hereinafter Biadsy).
Regarding claim 10, Rao in view of Strimel and Itoh does not specifically disclose wherein the data-storage module further stores a category label of the category that corresponds to each piece of corpus data.
Biadsy teaches wherein the data-storage module further stores a category label of the category that corresponds to each piece of corpus data (Biadsy, Fig. 8, for “Training Example”, there are accompanying labels “Application ID: Maps”, and “Location: New York City, NY”, both of which are labels indicating respective domains (see Fig. 7, domains 710c and 710a respectively)).
Rao, Strimel, Itoh, and Biadsy are considered to be analogous to the claimed invention as they are all in the same field of updating language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to incorporate the teachings of Biadsy in order to store the category label that corresponds to each piece of corpus data. Doing so would beneficial, as the category label could be later used as a feature for indicating which classified language models to update for a particular training example.
Regarding claim 21, claim 21 is rejected for analogous reasons to claim 10.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Houghton et al. (US 10,559,301 B2): generating topic specific language models (Fig. 2)
Nakazawa et al. (US 2009/0313017 A1): updating language models by updating the update target word(s) (Fig. 1, Fig. 7)
Any inquiry concerning this communication or earlier communications from the examiner should be directed to CODY DOUGLAS HUTCHESON whose telephone number is (703)756-1601. The examiner can normally be reached M-F 8:00AM-5:00PM EST.
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, Pierre-Louis Desir can be reached at (571)-272-7799. 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.
/CODY DOUGLAS HUTCHESON/Examiner, Art Unit 2659
/PIERRE LOUIS DESIR/Supervisory Patent Examiner, Art Unit 2659