DETAILED ACTION
Notice of AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
Applicant’s Amendment and remarks dated 8/5/2025 have been considered. Claims 6-7 are cancelled.
Drawing Objections. The objections to the drawings are withdrawn in view of the replacement drawings submitted.
Claim Objections. The objections to claims 1 and 3 are withdrawn in view of Applicant’s amendments to such claims. However, see new claim objections below raised by Applicant’s claim amendments.
35 U.S.C. 112(b) Rejections. The previous rejections under 35 U.S.C. 112(b) are withdrawn in view of Applicant’s amendments to the claims. However, new rejections under 35 U.S.C. 112(b) are made below.
Response to Arguments
On page 7 of Applicant’s 8/5/2025 Amendment and remarks, Applicant asserts that no “new matter” has been added.
The examiner respectfully disagrees. While no “new matter” has been added to the specification (as there have been no specification amendments), as set forth in the rejections below with respect to 35 U.S.C. 112(a), Applicant has added new matter into the claims that is not supported by the specification. See MPEP 2163.06.
On page 7 of Applicant’s 8/5/2025 Amendment and remarks, with respect to the 35 U.S.C. 112(b) rejections, Applicant asserts that “In response, claims are amended according to the Office Action, and no new matter has been entered.”
The examiner acknowledges that Applicant’s amendments resolve many of the indefiniteness issues identified in the previous rejection. However, the examiner respectfully submits that some of the issues raised by 112(b) rejections were not addressed, and that the amendments raise new 112(b) issues. A new set of rejections under 35 U.S.C. 112(b) is respectfully provided below.
On page 8 of Applicant’s 8/5/2025 Amendment and remarks, with respect to the rejections to claim 1 under 35 U.S.C. 103, Applicant argues:
PNG
media_image1.png
636
646
media_image1.png
Greyscale
The examiner respectfully disagrees. In response to applicant's argument that the references fail to show certain features of the invention, it is noted that the features upon which applicant relies (i.e., “group-level similarity penalties”) are not recited in the rejected claim(s). Although the claims are interpreted in light of the specification, limitations from the specification are not read into the claims. See In re Van Geuns, 988 F.2d 1181, 26 USPQ2d 1057 (Fed. Cir. 1993).
Further, in response to Applicant's arguments against the references individually, one cannot show nonobviousness by attacking references individually where the rejections are based on combinations of references. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981); In re Merck & Co., 800 F.2d 1091, 231 USPQ 375 (Fed. Cir. 1986). The rejection is based on the combination of LE with VASWANI, MADOTTO, and EDUNOV, where LE specifically teaches a length penalty used in the context of a natural language machine learning model, and specifically teaches a “beam search with beam size 5 and a length penalty 1.0”, where groups are in the size of 5 (corresponding to the beam size, or the 5 most likely hypotheses), and groups are separated using length penalty 1.0 to increase differences between groups, and where VASWANI teaches “top B words by probability at every decoding time are selected as an output at the current time” limitation, so LE takes the “top B words” of VASWANI, groups them according to most likely hypotheses in a beam search, and applies a length penalty, corresponding to the recited claim limitations. In other words, LE is not being used for the “label smoothing” teachings as Applicant alleges. LE teaches a beam form search, in a natural language context, which uses a length penalty, and when applied to the “top B words by probability” as taught by VASWANI, one of ordinary skill would understand that such top B words can be grouped using a beam search (e.g., beam size of 5, so best 5 hypotheses at a time), and a length penalty of 1.0.
The examiner further respectfully submits that dependent claims 2-5 are not allowable as argued by Applicant, for the same reasons argued with respect to claim 1.
Claim Objections
Claims 3 and 4 are objected to because of the following informalities:
In claim 3, line 2, “sequence of model” should read “sequence of the model”
In claim 3, line 2, “in the step 2” should read “in step 2”
In claim 3, line 12, “the loss function” lacks antecedent basis. The examiner suggests amending this to read “a loss function”.
In claim 4, last line “the output projection matrix” lacks antecedent basis. The examiner suggests amending this to read “an output projection matrix”, however, see the rejections under 35 U.S.C. 112(a) below with respect to this limitation.
Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
Claims 1-5 are rejected under 35 U.S.C. 112(a) as failing to comply with the written description requirement. The claims contain subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, at the time the application was filed, had possession of the claimed invention.
In claim 1, with respect to recited step 2, Applicant added the following claim language: “wl and b1 represent the weight matrix and bias vector of the first linear transformation in the feedforward neural network layer, and w2 and b2 represent the weight matrix and bias vector of the second linear transformation in the feedforward neural network layer”. This limitation has no written description support in the original disclosure as required by 35 U.S.C. 112(a), and one of ordinary skill would not have understood the inventors to have possession of the particular explanations of W1, b1, W2, and b2 recited in the newly-amended claim, as these are merely variable names that could refer to anything. Moreover, one of ordinary skill would not understand that the inventors possessed a feedforward neural network layer with both first and second linear transformations (and corresponding weight matrices), as one of ordinary skill would understand that a neural network layer typically uses a single activation function, and not multiple activation functions. Therefore, claim 1 is rejected under 35 U.S.C. 102(a) for lack of written description for this limitation.
In claim 1, on line 15 of page 3, Applicant added the following claim language: “DKL denotes the Kullback-Leibler Divergence”. This limitation has no written description support in the original disclosure as required by 35 U.S.C. 112(a), and one of ordinary skill would not have understood the inventors to have possession of the particular explanation of DKL recited in the newly-amended claim. For example, DKL could also stand for Karhunen Loive Decomposition. Therefore, claim 1 is rejected under 35 U.S.C. 102(a) for lack of written description for this limitation.
Claims 2-5 depend from claim 1, do not remedy the deficiencies of claim 1, and are therefore rejected for the same reasons explained with respect to claim 1.
In claim 2, Applicant added the following claim language: “dmodel indicates embedding dimension of the model.” This limitation has no written description support in the original disclosure as required by 35 U.S.C. 112(a), and one of ordinary skill would not have understood the inventors to have possession of the particular explanation of dmodel recited in the newly-amended claim. Therefore, claim 2 is further rejected under 35 U.S.C. 102(a) for lack of written description for this limitation.
In claim 4, Applicant added the following claim language: “Q indicates query vectors, K indicates key vectors, V indicates value vectors... Wi indicates head-specific weight matrices, dk indicates the dimension of query vectors and key vectors, Wo indicates the output projection matrix.” With respect to at least the Wi, dk, and Wo variables, this limitation has no written description support in the original disclosure as required by 35 U.S.C. 112(a), and one of ordinary skill would not have understood the inventors to have possession of the particular explanation of the Wi, dk, and Wo variables recited in the newly-amended claim. Therefore, claim 4 is further rejected under 35 U.S.C. 102(a) for lack of written description for this limitation.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
Claims 1-5 are rejected under 35 U.S.C. 112(b) as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor regards as the invention.
In claim 1, with respect to recited step 2, Applicant added the following claim language: “wl and b1 represent the weight matrix and bias vector of the first linear transformation in the feedforward neural network layer, and w2 and b2 represent the weight matrix and bias vector of the second linear transformation in the feedforward neural network layer”.
The terms “the first linear transformation” and “the second linear transformation” lack antecedent basis. It is not clear which first and second “linear transformations” are being referred to, as such terms were not previously introduced, and because one of ordinary skill would understand that neural network layers typically use a single non-linear activation function, so it’s unclear why a single neural network layer would have 2 separate linear transformations, and why such single layer would use linear transformations, as opposed to non-linear transformations. Therefore, claim 1 is rejected under 35 U.S.C. 112(b) because the metes and bounds of “the first linear transformation” and “the second linear transformation” are not clear. See MPEP 2173.02.
In claim 1, with respect to step 4, “an encoding-decoding attention mechanism having a same multi-head attention structure” is indefinite, because it is unclear what claim limitation is the basis for the “multi-head attention structure” that is the “same.” For example, is it the same multi-head attention structure of the multi-head attention layer of the encoder of step 3? Or is it a multi-head attention structure of the recited feedforward neural network layer of step 3? Or is it some other multi-head attention structure that the recited “an encoding-decoding attention mechanism” is now being configured the same as?
In claim 1, with respect to step 5, “the generated sentence sequence” lacks antecedent basis. For purposes of compact prosecution, this phrase is being interpreted as “the output of the decoding stage”.
Claims 2-5 depend from claim 1, do not remedy the deficiencies of claim 1, and are therefore rejected for the same reasons explained with respect to claim 1.
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.
Claims 1-5 are rejected under 35 U.S.C. 103 as being unpatentable over Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017), hereinafter referenced as VASWANI, in view of Madotto, Andrea, et al. "Personalizing dialogue agents via meta-learning." Proceedings of the 57th annual meeting of the association for computational linguistics (July 2019), hereinafter referenced as MADOTTO, and further in view of Edunov, Sergey, et al. "Classical structured prediction losses for sequence to sequence learning." arXiv preprint arXiv:1711.04956 (2018), hereinafter referenced as EDUNOV, and further in view of Le, Hung, et al. "Multimodal transformer networks for end-to-end video-grounded dialogue systems." arXiv preprint arXiv:1907.01166 (July 2, 2019), hereinafter referenced as LE.
Regarding Claim 1
VASWANI teaches:
step 2: defining an input X={x1, x2, ..., xn} sequence of the model, which includes n words in an input sentence sequence; (VASWANI, p. 2, section 3: “the encoder maps an input sequence of symbol representations (x1, ..., xn)”;
VASWANI, p. 7, section 5.1: “We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs. Sentences were encoded using byte-pair encoding”;
Examiner’s Note (EN): VASWANI teaches defining an input sequence, where the training data comprises sentence sequences)
word embedding all of the words in the input sequence to obtain corresponding word embeded vectors, (VASWANI, p. 5, section 3.4: “we use learned embeddings to convert the input tokens and output tokens to vectors”) then performing a position encoding to obtain corresponding position encoded vectors, and correspondingly adding the word embeded vectors and position encoded vectors to obtain an input vector representation of the model; (VASWANI, p. 5, section 3.5: “we add "positional encodings" to the input embeddings ... The positional encodings have the same dimension dmodel as the embeddings, so that the two can be summed”)
step 3: entering an encoding stage, in which the word vectors in the sentence sequence is updated according to a context with a multi-head attention layer, so as to obtain an output of the encoding stage via a feedforward neural network layer with the following formula:
PNG
media_image2.png
44
320
media_image2.png
Greyscale
in which Z indicates output content of a multi-head attention layer, wl and b1 represent the weight matrix and bias vector of the first linear transformation in the feedforward neural network layer, and w2 and b2 represent the weight matrix and bias vector of the second linear transformation in the feedforward neural network layer; (VASWANI, p. 5, section 3.3: In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.
PNG
media_image3.png
30
442
media_image3.png
Greyscale
Examiner’s Note: p. 5, section 3.2.2 explains that W1 and W2 are parameter matrices with respect to multi-head attention, where the different token embeddings (corresponding to recited “word vectors in the sentence sequence”), provide context for one another because they are part of a sentence sequence, and where VASWANI, p. 2, section 3, describes the relationship between vectors x and z such that z = x(W1 + b1) because z is an encoded version of x)
step 4: entering a decoding stage, in which an input of the decoding stage is also subjected to word embedding and position encoding to obtain a vector representation of the input of the decoding stage; (VASWANI, p. 5, section 3.3: In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between
VASWANI, p. 5, section 3.4: “we use learned embeddings to convert the input tokens and output tokens to vectors”
VASWANI, p. 5, section 3.5: “we add "positional encodings" to the input embeddings ... The positional encodings have the same dimension dmodel as the embeddings, so that the two can be summed”;
Examiner’s Note: VASWANI teaches a decoder, which uses summed positional encodings and input embeddings, to obtain an output of the decoder)
the vector representation of the input of the decoding stage is updated with the multi-head attention mechanism, then influences of input content at different times, historical dialogue content and different personalized characteristics on an output at current time are determined by an encoding-decoding attention mechanism having a same multi-head attention structure, and (VASWANI, p. 3, section 3.1: “The decoder is also composed of a stack of N = 6 identical layers. In addition to the two sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head attention over the output of the encoder stack. Similar to the encoder, we employ residual connections around each of the sub-layers, followed by layer normalization.”; (EN): VASWANI explains that there is an encoder and decoder stack, so the output of the encoder is input into the decoder, and that the structure of the encoder and decoder is the same)
finally the output of the decoding stage is obtained via the feedforward neural network layer; and (VASWANI, p. 5, section 3.3: In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.”; (EN): see also Fig. 1, showing that the feed forward network of the decoder (on the right side) has an output)
top B words by probability at every decoding time are selected as an output at the current time, and specifically, conditional probabilities of all words on the top B words are respectively calculated at the current time according to a probability distribution of the top B words selected at a previous time in a predicting process, and top B word sequences by probability are selected as the output at the current time; (VASWANI, Fig. 1:
PNG
media_image4.png
556
414
media_image4.png
Greyscale
Examiner’s Note: As shown in Fig. 1, the output probabilities are output from a softmax layer, corresponding to the recited “top B words by probability at every decoding time are selected as an output at the current time” limitation, where the multi-head attention of VASWANI calculates conditional probabilities according to a probability distribution when determining the outputs)
However, VASWANI fails to explicitly teach:
A method for generating personalized dialogue content, comprising the following steps:
step 1: collecting a set of personalized dialogue data and preprocessing the data, dividing the set of personalized dialogue data into a training set, a verification set and a test set to provide a support for subsequent training of a model;
step 5: learning parameters of the model by minimizing a negative logarithm likelihood loss of the output of the decoding stage so as to obtain a personalized multi-turn dialogue content generation model, a formula for the negative logarithm likelihood loss being as follows:
PNG
media_image5.png
70
324
media_image5.png
Greyscale
where, t1, ..., ti indicates the i-th word in the generated sentence sequence.
wherein the method further comprising adding an optimization algorithm to improve the diversity of the generated content, in which firstly, a label smoothing term is added to the negative logarithm likelihood loss to prevent the model from excessively concentrating predicted values on a category with a higher probability, thus reducing a possibility of generating universal reply content, the loss function with the label smoothing term added being:
PNG
media_image6.png
76
496
media_image6.png
Greyscale
where, f indicates a uniform prior distribution independent of the input, f = 1/v, V is a size of a wordlist, DKL denotes the Kullback-Leibler Divergence; and then the diversified bundle search algorithm with length penalty is added in a test stage, so that with a punishing of a sequence length, a probability of generating a short sequence is reduced and a possibility of generating a long sequence by the model is improved;
and B sentence sequences are grouped with similarity penalty added between groups to reduce the probability of generating similar content and improve the diversity of the content generated by the model.
However, in a related field of endeavor (implementing a transformer-based model using the teachings of VASWANI, see p. 5456, section 3), MADOTTO teaches:
A method for generating personalized dialogue content, comprising the following steps: (MADOTTO, p. 5454, section 1: “The main contribution of this paper is to cast the personalized dialogue learning as a meta-learning problem, which allows our model to generate personalized responses by efficiently leveraging only a few dialogue samples”)
step 1: collecting a set of personalized dialogue data and preprocessing the data, dividing the set of personalized dialogue data into a training set, a verification set and a test set to provide a support for subsequent training of a model; (MADOTTO, p. 5456, section: “The experiments are conducted using Persona-chat.... To create the meta-sets D, we match the dialogues by their persona description separately for train, validation and test”;
MADOTTO, p. 5456, footnote 1: “The model and pre-processing scripts are available at ...”
Examiner’s Note: MADOTTO describes using a dataset, Persona-chat (see also p. 5454, section 1, describing Persona-Chat as a “multi-turn conversational dataset”), and splitting that into 3 test sets (training, validation, and test) (corresponding to recited (training, verification, and test), where pre-processing scripts are used to pre-process the data for the model and made available by the authors at Github; in combination with VASWANI, the transformer-based attention model of VASWANI now trains using the train, validation, and test datasets as taught by MADOTTO in order to train a personalized dialogue generator as in MADOTTO).
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the transformer architecture of VASWANI with the personalized dialog generator and datasets of MADOTTO. As disclosed by MADOTTO, one of ordinary skill would have been motivated to do so because MADOTTO teaches “a novel meta-learning setting for personalizing dialogue agents without conditioning the model response to the persona description. This is especially useful since obtaining such persona description requires human effort. Moreover, we show that a dialogue agent trained with meta-learning achieves a more consistent dialogue by both of automatic measures and human evaluation.” (p. 5457, section 5). One of ordinary skill would further be motivated to combine VASWANI with MADOTTO because MADOTTO explicitly cites to VASWANI (see p. 5456, section 3) for the transformer-architecture).
However, VASWANI and MADOTTO fail to explicitly teach:
step 5: learning parameters of the model by minimizing a negative logarithm likelihood loss of the output of the decoding stage so as to obtain a personalized multi-turn dialogue content generation model, a formula for the negative logarithm likelihood loss being as follows:
PNG
media_image5.png
70
324
media_image5.png
Greyscale
where, t1, ..., ti indicates the i-th word in the generated sentence sequence.
wherein the method further comprising adding an optimization algorithm to improve the diversity of the generated content, in which firstly, a label smoothing term is added to the negative logarithm likelihood loss to prevent the model from excessively concentrating predicted values on a category with a higher probability, thus reducing a possibility of generating universal reply content, the loss function with the label smoothing term added being:
PNG
media_image6.png
76
496
media_image6.png
Greyscale
where, f indicates a uniform prior distribution independent of the input, f = 1/v, V is a size of a wordlist, DKL denotes the Kullback-Leibler Divergence; and then the diversified bundle search algorithm with length penalty is added in a test stage, so that with a punishing of a sequence length, a probability of generating a short sequence is reduced and a possibility of generating a long sequence by the model is improved;
and B sentence sequences are grouped with similarity penalty added between groups to reduce the probability of generating similar content and improve the diversity of the content generated by the model.
However, in a related field of endeavor (sequence to sequence learning, which one of ordinary skill would understand is pertinent to natural language processing, see p. 1, section 1), EDUNOV teaches:
step 5: learning parameters of the model by minimizing a negative logarithm likelihood loss of the output of the decoding stage so as to obtain a personalized multi-turn dialogue content generation model, a formula for the negative logarithm likelihood loss being as follows:
PNG
media_image5.png
70
324
media_image5.png
Greyscale
where, t1, ..., ti indicates the i-th word in the generated sentence sequence.
(EDUNOV, pp. 2-3, section 3.1 and Equation 1: “Token-level likelihood (TokNLL, Equation 1) minimizes the negative log likelihood of individual reference tokens t = (t1, ... tn). It is the most common loss function optimized in related work and serves as a baseline for our comparison.
PNG
media_image7.png
48
444
media_image7.png
Greyscale
Examiner’s Note: in combination with VASWANI and MADOTTO, the personalized dialog generator of MADOTTO, implemented using the VASWANI transformer architecture, now utilizes the negative logarithm likelihood loss function of EDUNOV to train the model)
wherein the method further comprising adding an optimization algorithm to improve the diversity of the generated content, in which firstly, a label smoothing term is added to the negative logarithm likelihood loss to prevent the model from excessively concentrating predicted values on a category with a higher probability, thus reducing a possibility of generating universal reply content, the loss function with the label smoothing term added being:
PNG
media_image6.png
76
496
media_image6.png
Greyscale
where, f indicates a uniform prior distribution independent of the input, f = 1/v, V is a size of a wordlist, DKL denotes the Kullback-Leibler Divergence (EDUNOV, pp. 2-3, section 1 and Equation 2: “Likelihood training forces the model to make extreme zero or one predictions to distinguish between the ground truth and alternatives. This may result in a model that is too confident in its training predictions, which may hurt its generalization performance. Label smoothing addresses this by acting as a regularizer that makes the model less confident in its predictions. Specifically, we smooth the target distribution with a prior distribution f that is independent of the current input x .... We use a uniform prior distribution over all words in the vocabulary, f = 1/V . One may also use a unigram distribution which has been shown to work better on some tasks”
PNG
media_image8.png
56
590
media_image8.png
Greyscale
Examiner’s Note: The equation in this claim 7 is identical to equation 2 of EDUNOV).
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the transformer architecture of VASWANI with the personalized dialog generator and datasets of MADOTTO, and further with the negative logarithm likelihood loss function of EDUNOV. As disclosed by EDUNOV, one of ordinary skill would have been motivated to do so because the negative logarithm likelihood loss function is “the most common loss function optimized in related work”, and therefore one of ordinary skill would be motivated to utilize a well-known and peer-reviewed technique. (p. 2, section 3.1).
However, VASWANI, MADOTTO, and EDUNOV fail to explicitly teach:
and then the diversified bundle search algorithm with length penalty is added in a test stage, so that with a punishing of a sequence length, a probability of generating a short sequence is reduced and a possibility of generating a long sequence by the model is improved;
and B sentence sequences are grouped with similarity penalty added between groups to reduce the probability of generating similar content and improve the diversity of the content generated by the model.
However, in a related field of endeavor (dialogue systems that “generates appropriate conversational response to queries of humans”, see p. 1, section 1), LE teaches:
and then the diversified bundle search algorithm with length penalty is added in a test stage, so that with a punishing of a sequence length, a probability of generating a short sequence is reduced and a possibility of generating a long sequence by the model is improved; (LE, p. 6, section 4.2: “Label Smoothing (Szegedy et al., 2016) is also applied during training. For all models, we select the latest checkpoints that achieve the lowest perplexity on the validation set. We used beam search with beam size 5 an a length penalty 1.0.”; (EN): the beam search corresponds to the recited “diversified bundle search algorithm with length penalty”; in combination with VASWANI, MADOTTO, and EDUNOV, the personalized dialog agent of MADOTTO, using the architecture of VASWANI and loss function of EDUNOV, now utilizes the beam search algorithm, with label smoothing and length penalty as in LE)
and B sentence sequences are grouped with similarity penalty added between groups to reduce the probability of generating similar content and improve the diversity of the content generated by the model. (LE, p. 6, section 4.2: “Label Smoothing (Szegedy et al., 2016) is also applied during training. For all models, we select the latest checkpoints that achieve the lowest perplexity on the validation set. We used beam search with beam size 5 an a length penalty 1.0.”; (EN): the length penalty of LE is used during the beam search, which groups sentence sequences, corresponding to recited “B sentence sequences are grouped with similarity penalty added between groups” limitation; in combination with VASWANI, MADOTTO, and EDUNOV, the personalized dialog agent of MADOTTO, using the architecture of VASWANI and loss function of EDUNOV, now utilizes the beam search algorithm, with label smoothing and length penalty as in LE)
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the transformer architecture of VASWANI with the personalized dialog responses of MADOTTO, and further with the negative logarithm likelihood loss function of EDUNOV, and further with the beam search algorithm, with label smoothing and length penalty as in LE. As disclosed by LE, one of ordinary skill would have been motivated to do so because LE was “inspired” by VASWANI and utilizes parts of its architecture. (p. 2, section 1). One of ordinary skill would further be motivated to do so because using optimizers to train neural networks would improve the efficiency of the personalized dialog agent of MADOTTO, using the architecture of VASWANI and loss function of EDUNOV.
Regarding Claim 2
VASWANI, MADOTTO, and EDUNOV teach the method of claim 1. VASWANI further teaches:
wherein a position encoding formula in step 2 is:
PNG
media_image9.png
130
302
media_image9.png
Greyscale
Where PE(pos, 2i) indicates a value in a 2i-th dimension of the pos-th word in the sentence sequence, and PE(pos, 2i+1) indicates a value in a 2i+1-th dimension of the pos-th word in the sentence sequence, dmodel indicates embedding dimension of the model. (VASWANI, p. 6, section 3:
PNG
media_image10.png
262
626
media_image10.png
Greyscale
Examiner’s Note: the formulas in VASWANI appear to be identical to the formulas claimed in this claim 2)
Regarding Claim 3
VASWANI, MADOTTO, and EDUNOV teach the method of claim 1. However, VASWANI fails to explicitly teach:
wherein the input of model in the step 2 comprises not only a current dialogue content, but also all of the historical dialogue content that has occurred as well as specific personalized characteristics.
However, in a related field of endeavor (implementing a transformer-based model using the teachings of VASWANI, see p. 5456, section 3), MADOTTO teaches:
wherein the input of model in the step 2 comprises not only a current dialogue content, but also all of the historical dialogue content that has occurred as well as specific personalized characteristics. (MADOTTO, p. 5455, section 2.2: “we first adapt θ to the set of dialogue made by a persona P and then we only use the dialogue history to condition our response. ... The main idea of our work is to use Model-Agnostic Meta-Learning (MAML) (Finn et al., 2017) to learn an initial set of parameters that can quickly learn a persona from few dialogues sample.”; (EN): in combination with VASWANI and EDUNOV, the personalized dialog agent of MADOTTO, as modified to use the transformer architecture of VASWANI and the loss function of EDUNOV, utilizes the set of dialogue made by a persona P (corresponding to recited “current dialogue content”), the dialogue history, and the learned attributes of the persona (corresponding to recited “specific personalized characteristics”) of MADOTTO when generating content).
Before the effective filing date of the present application, it would have been obvious to one of ordinary skill in the art to combine the transformer architecture of VASWANI with the personalized dialog responses of MADOTTO, and further with the negative logarithm likelihood loss function of EDUNOV. As disclosed by MADOTTO, one of ordinary skill would have been motivated to do so because MADOTTO teaches “a novel meta-learning setting for personalizing dialogue agents without conditioning the model response to the persona description. This is especially useful since obtaining such persona description requires human effort. Moreover, we show that a dialogue agent trained with meta-learning achieves a more consistent dialogue by both of automatic measures and human evaluation.” (p. 5457, section 5). One of ordinary skill would further be motivated to combine VASWANI with MADOTTO because MADOTTO explicitly cites to VASWANI (see p. 5456, section 3) for the transformer-architecture). One of ordinary skill would further be motivated to make this combination because one of ordinary skill would understand that historical dialogue impacts future dialogue (e.g., a person has similar tendencies reflected in the historical dialogue), and one of ordinary skill would therefore utilize such historical dialogue information if available.
Regarding Claim 4
VASWANI, MADOTTO, and EDUNOV teach the method of claim 1. VASWANI further teaches:
wherein a formula for updating of the word vector in the step 3 is as follows:
PNG
media_image11.png
208
444
media_image11.png
Greyscale
where, Q indicates query vectors, K indicates key vectors, V indicates value vectors, Q, K, V are respectively obtained by multiplying three different weight matrices by the input vector of the model, and headi indicates an attention head in the multi-head attention mechanism, Wi indicates head-specific weight matrices, dk indicates the dimension of query vectors and key vectors, Wo indicates the output projection matrix. (VASWANI, p. 5, section 3.2.2:
PNG
media_image12.png
196
640
media_image12.png
Greyscale
VASWANI, p. 4, section 3.2.1:
PNG
media_image13.png
328
628
media_image13.png
Greyscale
Examiner’s Note: the formulas in VASWANI appear to be identical to the formulas claimed in this claim 4)
Regarding Claim 5
VASWANI, MADOTTO, and EDUNOV teach the method of claim 1. VASWANI further teaches:
wherein a residual connection and layer normalization process is added to the multi-head attention layer and feedforward neural network layer in the encoding stage in step 3, and the residual connection and layer normalization processes is also added to each sublayer in the decoding stage in step 4, a formula for the residual connection and layer normalization process is as follows:
PNG
media_image14.png
42
400
media_image14.png
Greyscale
where, SubLayer indicates the multi-head attention layer or feedforward neural network layer. (VASWANI, p. 3, section 3.1:
PNG
media_image15.png
234
632
media_image15.png
Greyscale
Examiner’s Note: output of each sub-layer has the same formula as in VASWANI, where it further explains that the decoder has “residual connections around each of the sub-layers, followed by layer normalization”)
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MICHAEL C LEE whose telephone number is (571)272-4933. The examiner can normally be reached M-F 12:00 pm - 8:00 pm ET.
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, Omar Fernandez Rivas can be reached at 571-272-2589. 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.
/MICHAEL C. LEE/Examiner, Art Unit 2128
/OMAR F FERNANDEZ RIVAS/Supervisory Patent Examiner, Art Unit 2128