DETAILED ACTION
This examination is in response to the communication filed on 12/16/2025. Claims 1-20 are currently pending, where claims 1, 4-6, 14, 17 and 18 have been amended.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendments/Arguments
Applicant’s amendments, filed 12/16/2025, with respect to the claim objection and rejections under §112 and §101 have been fully considered and are persuasive. The objection to claim 6, rejection of claim 4 under §112, and rejection of claims 1-3, 14-16, and 18-20 under §101 has been withdrawn.
Applicant's argument that claims 1, 14, and 18 have amended “to incorporate novel features corresponding to the allowable subject matter of claim 6” has been fully considered but they are not persuasive. Specifically, the Examiner notes that all of the limitations of claim 6 and intervening claims 5 and 4 have not been incorporated into independent claim 1. Furthermore, the broader concepts incorporated into amended claims 1, 14, and 18 are taught by the prior art of record as noted below in the updated rejection.
Claim Rejections - 35 USC § 102
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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 1-3, 14-16 and 18-20 are rejected under 35 U.S.C. 102(a)(2) as being anticipated by Romeo et al. (US 11568145 B1; herein “Romeo”).
Regarding claim 1, Romeo teaches a system for human-machine collaborative conversation interaction, comprising: one or more processors (Fig. 9, Processors (910A, 910B…910N) configured to execute instructions (Col. 8, lines 20-24 teaches “Some or all of the operations 500 (or other processed described herein…) are performed under the control of one or more computer systems configured with executable instructions…) to cause the system to perform operations comprising:
outputting, according to conversation data to be processed, structural information of the conversation data, wherein the conversation data comprises multiple turns of conversation (Fig. 5, step 502 and col. 8, lines 35-37 teaches “At 502 a user utterance provided by a user within a multi-turn dialog between the user and a conversational agent is received”);
obtaining, according to the structural information, a semantic representation vector carrying phrase-dimensional semantic information, sentence-dimensional semantic information and topic-dimensional semantic information corresponding to the conversation data (Col. 5, lines 50-56 teaches “a pre-trained language model (ML) produces a contextualized encoding of the input utterance tokens which are fused with encoded contextual signals. Specifically, in some embodiments, a pre-trained BERT model encodes the input utterance…”, Fig. 5, step 506 and col. 8, lines 47-49 teaches “At 506, an intent classification and one or more slot labels for the user utterance are obtained from the cNLU framework” and col. 7, lines 25-30 teaches “the sequence fed to the self-attention layers is composed of the contextual signal encodings, followed by the normalized output embedding, which are followed by the (down-projected) token encodings, i.e., <CSIC, CSSL, SPI, SDA,SES, [CLS], T1,…,TN>…” the token encodings are interpreted at phrase-dimensional and sentence-dimensional information and the contextual signal encodings is interpreted as topic-dimensional information);
obtaining semantic transfer relationships between each turn of conversation according to the semantic representation vector (col. 7, lines 37-39 teaches “the IC classifier 333 is fed with the concatenation of the encodings of CSIC and [CLS] output from the stack of self-attention layers 331…” and col. 14, lines 53-54 teaches “generating an intent contextual signal and a slot label contextual signal from the previous intent, the previous dialog act, and the elicited slot” Accordingly, the system applies cross-turn semantics to interpret the current turn intent, the cross-turn semantics are interpreted as semantic transfer); and
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data matching service requirements so as to perform preset service processing through the determined conversation data (Fig. 5, step 508 and col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508” ),
wherein the structural information of the conversation data is outputted by a pre-trained language model trained with respect to a conversation structure based on conversation data samples of a ser service, the conversation data samples comprise data samples of multiple turns of conversation ( col. 5, lines 50-56 teaches “…a pre-trained language model (LM) produces a contextualized encoding of the input utterances tokens which are fused with encoded contextual signals. Specifically, in some embodiments, a pre-trained BERT model encodes the input utterance…”), and wherein the pre-trained language model is trained by:
token feature extraction is performed using a first sub-machine learning model to obtain a first sub-representation vector (Fig. 6, step 602 and col. 9, lines 9-11 teaches “At 602 the user utterance is encoded as a token sequence…”);
syntactic structure feature extraction is performed using a second sub-machine learning model to obtain a second sub-representation vector (Fig. 6, step 608 and col. 9, lines 14-15 teaches “At 608, embeddings for the previous intent, previous dialog act, and elicited slot are generated” ); and
a token structure representation vector is obtained carrying information of an intra-sentence token structure of the conversation data samples according to the first sub-representation vector and the second sub-representation vector (Fig. 6, step 614 Col. 9, lines 29-30 “The contextual signals and (normalized and/or down projected) token sequence are fused at 614” ).
Regarding claim 2, Romeo teaches all the elements of claim 1 (see detailed element listing above). In addition, Romeo further teaches the operation of outputting, according to the conversation data to be processed, further comprises
outputting, according to conversion data to be processed, structural information of an intra-sentence token structure of each turn of conversation and structural information of conversation dependencies of the multiple turns of conversation in the conversation data (col. 7, lines 25-30 teaches “the sequence fed to the self-attention layers is composed of the contextual signal encodings, followed by the normalized output embedding, which are followed by the (down-projected) token encodings, i.e., <CSIC, CSSL, SPI, SDA,SES, [CLS], T1,…,TN>…” the token encodings are interpreted as intra-sentence token structure and the contextual signal encodings are interpreted as structural information of conversation dependencies of multiple turns).
Regarding claim 3, Romeo teaches all the elements of claim 1 (see detailed element listing above). In addition, Romeo further teaches the operations further comprising:
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for performing service training, and generate simulated conversation test questions according to the determined conversation data and user profile data (the “or” term makes this limitation optional); or
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for assisting services, and perform, based on the determined conversation data, at least one of streaming intention recognition processing, conversation context- based service response screening processing, or preset goal-based guided conversation screening processing (In view of ¶[0102] of the specification “preset goal-based guided conversation screening” is interpreted as guiding the direction of the conversation. Fig. 5, step 508 and col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508”); or
acquiring conversation data during service processing and perform at least one of conversation flow mining processing, key conversation mining processing, or dialogue summary generation processing on the acquired conversation data (the “or” term makes this limitation optional).
Regarding claims 14 and 18, Romeo teaches a method for human-machine collaborative conversation interaction applied to a human-machine collaborative conversation interaction system and a non-transitory computer-readable storage medium storing a set of instructions Col. 8, lines 20-24 teaches “Some or all of the operations 500 (or other processed described herein…) are performed under the control of one or more computer systems configured with executable instructions…) that are executable by one or more processors of a device to cause the device to perform the method comprising:
receiving conversation data to be processed, wherein the conversation data comprises multiple turns of conversation (Fig. 5, step 502 and col. 8, lines 35-37 teaches “At 502 a user utterance provided by a user within a multi-turn dialog between the user and a conversational agent is received”);
obtaining structural information of the conversation data (Col. 5, lines 50-56 teaches “a pre-trained language model (ML) produces a contextualized encoding of the input utterance tokens which are fused with encoded contextual signals. Specifically, in some embodiments, a pre-trained BERT model encodes the input utterance…”);
according to the structural information, obtaining a semantic representation vector carrying phrase-dimensional semantic information, sentence-dimensional semantic information and topic-dimensional semantic information corresponding to the conversation data (Col. 5, lines 50-56 teaches “a pre-trained language model (ML) produces a contextualized encoding of the input utterance tokens which are fused with encoded contextual signals. Specifically, in some embodiments, a pre-trained BERT model encodes the input utterance…”, Fig. 5, step 506 and col. 8, lines 47-49 teaches “At 506, an intent classification and one or more slot labels for the user utterance are obtained from the cNLU framework” and col. 7, lines 25-30 teaches “the sequence fed to the self-attention layers is composed of the contextual signal encodings, followed by the normalized output embedding, which are followed by the (down-projected) token encodings, i.e., <CSIC, CSSL, SPI, SDA,SES, [CLS], T1,…,TN>…” the token encodings are interpreted at phrase-dimensional and sentence-dimensional information and the contextual signal encodings is interpreted as topic-dimensional information);
according to the semantic representation vector, obtaining semantic transfer relationships between each turn of conversation (col. 7, lines 37-39 teaches “the IC classifier 333 is fed with the concatenation of the encodings of CSIC and [CLS] output from the stack of self-attention layers 331…” and col. 14, lines 53-54 teaches “generating an intent contextual signal and a slot label contextual signal from the previous intent, the previous dialog act, and the elicited slot” Accordingly, the system applies cross-turn semantics to interpret the current turn intent, the cross-turn semantics are interpreted as semantic transfer); and
according to the semantic representation vector and the semantic transfer relationships, determining conversation data matching service requirements so as to perform preset service processing through the determined conversation data (Fig. 5, step 508 and col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508”),
wherein the structural information of the conversation data is outputted by a pre-trained language model trained with respect to a conversation structure based on conversation data samples of a ser service, the conversation data samples comprise data samples of multiple turns of conversation ( col. 5, lines 50-56 teaches “…a pre-trained language model (LM) produces a contextualized encoding of the input utterances tokens which are fused with encoded contextual signals. Specifically, in some embodiments, a pre-trained BERT model encodes the input utterance…”), and wherein the pre-trained language model is trained by:
token feature extraction is performed using a first sub-machine learning model to obtain a first sub-representation vector (Fig. 6, step 602 and col. 9, lines 9-11 teaches “At 602 the user utterance is encoded as a token sequence…”);
syntactic structure feature extraction is performed using a second sub-machine learning model to obtain a second sub-representation vector (Fig. 6, step 608 and col. 9, lines 14-15 teaches “At 608, embeddings for the previous intent, previous dialog act, and elicited slot are generated” ); and
a token structure representation vector is obtained carrying information of an intra-sentence token structure of the conversation data samples according to the first sub-representation vector and the second sub-representation vector (Fig. 6, step 614 Col. 9, lines 29-30 “The contextual signals and (normalized and/or down projected) token sequence are fused at 614” )..
Regarding claims 15 and 19, Romeo teaches all the elements of claims 14 and 18 (see detailed element listing above). In addition, Romeo further teaches outputting, according to the conversation data to be processed, structural information of an intra-sentence token structure of each turn of conversation and structural information of conversation dependencies of the multiple turns of conversation in the conversation data (col. 7, lines 25-30 teaches “the sequence fed to the self-attention layers is composed of the contextual signal encodings, followed by the normalized output embedding, which are followed by the (down-projected) token encodings, i.e., <CSIC, CSSL, SPI, SDA,SES, [CLS], T1,…,TN>…” the token encodings are interpreted as intra-sentence token structure and the contextual signal encodings are interpreted as structural information of conversation dependencies of multiple turns).
Regarding claims 16 and 20, Romeo teaches all the elements of claims 14 and 18 (see detailed element listing above). In addition, Romeo further teaches the operations further comprising:
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for performing service training, and generate simulated conversation test questions according to the determined conversation data and user profile data (the “or” term makes this limitation optional); or
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for assisting services, and perform, based on the determined conversation data, at least one of streaming intention recognition processing, conversation context- based service response screening processing, or preset goal-based guided conversation screening processing (In view of ¶[0102] of the specification “preset goal-based guided conversation screening” is interpreted as guiding the direction of the conversation. Fig. 5, step 508 and col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508”); or
acquiring conversation data during service processing and perform at least one of conversation flow mining processing, key conversation mining processing, or dialogue summary generation processing on the acquired conversation data (the “or” term makes this limitation optional).
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.
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 non-obviousness.
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.
Claims 4, 5, 7, 13 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Romeo as applied to claims 1 and 14 above, and further in view of Chen et al. (US 2022/0129556 A1; herein “Chen”).
Regarding claims 4 and 17, Romeo teaches all the elements of claims 1 and 14 (see detailed element listings above). In addition, Romeo further teaches the system for human-machine collaborative conversation interaction is obtained through training using operations comprising:
training the pre-trained language model with respect to the conversation structure based on the conversation data samples of the set service to obtain the pre-trained language model capable of outputting the structural information carrying the conversation data samples (Col. 5, lines 50-53 teaches “a pre-trained language model (LM) produces a contextualized encoding of the input utterance tokens which are fused with encoded contextual signals. Specifically…a pre-trained BERT model encodes the input utterance” );
according to a representation vector output by the trained pre-trained language model, performing phrase representation training, sentence vector representation training, and topic representation training for the conversation data samples by a machine learning model to obtain a machine learning model capable of outputting a representation vector carrying corresponding phrase-dimensional semantic information, sentence-dimensional semantic information, and topic- dimensional semantic information (Col. 6, lines 12-37 teaches “Each of the contextual signals (PI, DA, and ES) is encoded using embedding layers 307, 311, or 315…As such, the normalized output embeddings (s) and contextual signal (CS) encodings are formally defined in some embodiments as…
C
S
I
C
=
S
P
I
,
S
D
A
,
S
E
S
W
C
S
I
C
T
+
b
C
S
I
C
C
S
S
L
=
S
P
I
,
S
D
A
,
S
E
S
W
C
S
S
L
T
+
b
C
S
C
S
I
C
where [▪,▪,▪] is a concatenation operator…” encoding layers 307, 311, and 315 inherently requiring training to performed the disclosed encoding);
according to the representation vector output by the trained machine learning model, performing training of performing semantic analysis for the multiple turns of conversation to obtain the semantic transfer relationships between the turns of conversation (Col. 4, lines 39-43 teaches “cNLU framework explicitly models more comprehensive contextual information training IC and SL tasks by leveraging previous utterances, dialogue acts, and previous intent classes and slot labels among other possible signals” ).
Although Romeo teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508” Romeo fails to explicitly disclose that the dialog manager is a trained model. Therefore, Romeo fails to explicitly disclose performing training of determining the conversation data matching service requirements of the set service according to the semantic representation vector and the semantic transfer relationships, so as to perform the preset service processing through the determined conversation data as recited in claims 4 and 17.
Chen teaches systems and method for implementing smart assistant system which provided guided conversation. The system includes, inter alia, a natural-language understanding (NLU) module and a dialog manager. More specifically, Chen teaches performing training of determining the conversation data matching service requirements of the set service according to the semantic representation vector and the semantic transfer relationships, so as to perform the preset service processing through the determined conversation data (¶[0119] teaches “In particular embodiments, the output of the entity resolution module 212 may be sent to the dialog manager 216 to advance the flow of the conversation with the user. The dialog manager 216 may be an asynchronous state machine that repeatedly updates the state and selects actions based on the new state. The dialog manager 216 may additionally store previous conversations between the user and the assistant system 140…In particular embodiments, the dialog manager 216 may implement reinforcement learning frameworks to improve the dialog optimization. The dialog manager 216 may comprise dialog intent resolution 356, the dialog state tracker 218, and the action selector 222 ... Each action selected may depend on the execution result from previous actions.” )
Romeo differs from the claimed invention as defined in claims 4 and 17 in that Romeo fails to disclose utilizing a trained model for performing the guided conversation. Utilizing a trained models, e.g., a reinforcement learning (RL) based dialogue manager, for performing guided conversation are known in the art as evidenced by Chen. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have implemented the dialog manager using a reinforcement learning framework as taught by Chen “to improve the dialog optimization” (Chen, ¶[0119]).
Regarding claim 5, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). In addition, Romeo further teaches the operation of training of the pre-trained language model with respect to the conversation structure comprises:
training for the intra-sentence token structure in the conversation data samples, and training for conversation dependencies between the multiple turns of conversation in the conversation data samples (Col. 5, lines 50-53 teaches “a pre-trained language model (LM) produces a contextualized encoding of the input utterance tokens which are fused with encoded contextual signals. Specifically…a pre-trained BERT model encodes the input utterance” and Col. 8, lines 7-10 teaches “Encoder 401 determines the CLS token sequence of FIG. 3 and a second sequence (kicked off with a SEP token) for a previous utterance determine a token sequency. As shown, a separator (SEP) and other tokens per word are encoded”).
Regarding claim 7, the combination of Romeo and Chen teaches all the elements of claim 5 (see detailed element listing above). In addition, Romeo further teaches training for the conversation dependencies between the multiple turns of conversation in the conversation data samples comprises:
training, by a third sub-machine learning model based on semantic similarities between sample features of the data samples of the turns of conversation, to obtain a conversation structure representation vector characterizing dependencies between the turns of conversation (Col. 4, lines 39-43 teaches “cNLU framework explicitly models more comprehensive contextual information training IC and SL tasks by leveraging previous utterances, dialogue acts, and previous intent classes and slot labels among other possible signals” As shown in Figs. 3 and 4, the intent classifier 333 and slot label classifier 335 are different from the embedding encoding layers 337, 307, 311 and 315. Accordingly, Romeo teaches three or more sub-machine learning models.).
Regarding claim 13, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). In addition, Romeo further teaches
the operation of performing training comprises: determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for performing service training; and generating simulated conversation test questions according to the determined conversation data and user profile data, and performing the service training through the simulated conversation test questions (the “or” term makes this limitation optional); or
the operation of performing training comprises:
determining, according to the semantic representation vector and the semantic transfer relationships, conversation data for assisting services (Figs 3 and 4; col. 7, lines 24-30 teaches “…the sequence fed to the self-attention layers is composed of the contextual signal encodings, followed by the normalized output embeddings, which followed by the (down-projected) token encodings”);
performing, based on the determined conversation data, at least one of streaming intention recognition processing, conversation context-based service response screening processing, or preset goal-based guided conversation screening processing (In view of ¶[0102] of the specification “preset goal-based guided conversation screening” is interpreted as guiding the direction of the conversation. Fig. 5, step 508 and col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508”); and
performing service processing according to the result of the processing (col. 8, lines 50-52 teaches “the intent classification and slot label are provided to a dialog manager to determine a next dialog (if any) to present in the multi-turn chat at 508”); or
the operation of performing training comprises: acquiring conversation data during service processing and performing at least one of conversation flow mining processing, key conversation mining processing, or dialogue summary generation processing on the acquired conversation data (the “or” term makes this limitation optional).
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Romeo and Chen as applied to claim 7 above, and further in view of Devlin et al. (“BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding” Proceedings of NAACL-HLT 2019, pp. 4171-4186; herein “Devlin”).
Regarding claim 8, the combination of Romeo and Chen teaches all the elements of claim 7 (see detailed element listing above). In addition, Romeo further teaches the using a pre-trained BERT encoder and determining, by the third sub-machine learning model, semantic similarities between sample features of the data samples of the turns of conversation based on the data samples of the multiple turns of conversation containing the masked data, and training to obtain the conversation structure representation vector characterizing the dependencies between the turns of conversation (Col. 4, lines 39-43 teaches “cNLU framework explicitly models more comprehensive contextual information training IC and SL tasks by leveraging previous utterances, dialogue acts, and previous intent classes and slot labels among other possible signals” As shown in Figs. 3 and 4, the intent classifier 333 and slot label classifier 335 are different from the embedding encoding layers 337, 307, 311 and 315. Accordingly, Romeo teaches three or more sub-machine learning models.).
However, Romeo is silent regarding whether the pre-training o of the BERT encoder includes masked language modeling. Therefore, the combination of Romeo and Chen fails to explicitly disclose that masking tokens in data samples of a part of the turns of conversation among the data samples of the multiple turns of conversation to obtain data samples of the multiple turns of conversation containing masked data;.
Pre-training BERT models using mask language modeling is known in the art as evidenced by Devlin. Specifically, Devlin teaches “we improve the fine-tuning based approaches by proposing BERT: Bidirectional Encoder Representations from Transformers. BERT alleviates the previously mentioned unidirectionality constraint by using a “masked language model” (MLM) pre-training objective…The masked language model randomly masks some of the tokens from the input, and the objective is to predict the original vocabulary id of the masked words based only on its context” (Devlin, p. 4171, 2nd col., last ¶ to p. 4172 1st col. 1st ¶.) Therefore, Delvin teaches masking tokens in data samples of a part of the turns of conversation among the data samples of the multiple turns of conversation to obtain data samples of the multiple turns of conversation containing masked data.
The combination of Romeo and Chen differs from the claimed invention, defined by claim 8, in that the combination fails to disclose using a masked language modeling to pre-train the BERT encoder. BERT encoders having been pre-trained using masked language modeling are known in the art as taught by Devlin. Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date of the invention, to have utilizes a BERT encoder that was pre-trained using masked language modeling as taught by Devlin as it merely constitutes the combination of known elements to achieve the predictable result of utilizing a pre-trained BERT model that can be fine-tuned to create models for a wide range of tasks (Devlin, Abstract).
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over the combination of Romeo and Chen as applied to claim 4 above, and further in view of Wu et al. (US 2022/0139384 A1; herein “Wu”).
Regarding claim 9, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). In addition, Romeo further teaches performing the phrase representation training for the conversation data samples by the machine learning model comprises:
performing word segmentation processing on the conversation data samples to obtain multiple segmented words (Col. 8, lines 7-10 teaches “Encoder 401 determines the CLS token sequence of FIG. 3 and a second sequence (kicked off with a SEP token) for a previous utterance…a separator (SEP) and other tokens per word are encoded” the word tokens are interpreted as segmented words).
The combination of Romeo and Chen fails to explicitly disclose performing the phrase representation training of the machine learning model with respect to the conversation data samples, according to the cohesion and degree of freedom of the multiple segmented words, the representation vector obtained through the pre-trained language model, and a preset contrastive learning loss function.
Wu teaches system and method for training task-oriented dialogue language models. Specifically, Wu teaches performing the phrase representation training of the machine learning model with respect to the conversation data samples, according to the cohesion and degree of freedom of the multiple segmented words, the representation vector obtained through the pre-trained language model, and a preset contrastive learning loss function (¶[0016] teaches “the user utterance and system response of the dialogue of the task oriented training data sets may be prepared to form an input training sequence by prefixing a start token to each and concatenating the pair of user utterance and system response. The input sequence may be used to pre-train the TOD language model via masked language loss. In some embodiments, different sets of dialogues may be selected for contrastive learning.” and ¶[0022] teaches “an example of the one or more loss functions can be the response contrastive loss (RCL) objective function. In some cases, pre-training TOD language models with RCL may be advantageous because RCL may not require any additional human annotation and allow for an improved representation for the [CLS] token.”)
Therefore, it would have been obvious to one having ordinary skill in the art, before the effective filing date, to have pre-trained the BERT module taught by the combination of Romeo and Chen with contrastive learning as taught by Wu to provide the advantage of not requiring any additional human annotation and allow for an improved representation for the [CLS] token. (Wu, ¶[0022]).
Allowable Subject Matter
Claims 6 and 10-12 objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
The following is a statement of reasons for the indication of allowable subject matter:
Regarding claim 6, the combination of Romeo and Chen teaches all the elements of claim 5 (see detailed element listing above). However, neither Romeo nor Chen, teaches or suggests training for the intra-sentence token structure in the conversation data samples comprises:
performing the token feature extraction using the first sub-machine learning model using tokens corresponding to the conversation data samples as an input to the first sub-machine learning model to obtain the first sub-representation vector;
performing syntactic structure feature extraction using the second sub-machine learning model using part-of-speech information of tokens corresponding to the conversation data samples and a syntactic dependency tree obtained by syntactic analysis of the conversation data samples based on the tokens as an input to the second sub-machine learning model, so as to obtain the second sub-representation vector; and
concatenating the first sub-representation vector and the second sub-representation vector to obtain a token structure representation vector carrying information of the intra-sentence token structure of the conversation data samples in combination with other elements of claims 5, 4 and 1 from which it depends.
Regarding claim 10, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). However, neither Romeo nor Chen, teaches or suggests performing the sentence vector representation training for the conversation data samples by the machine learning model comprises:
determining, from the conversation data samples, conversation sample data to be processed, and forming conversation sample data pairs;
obtaining, based on the pre-trained language model, representation vector pairs corresponding to the conversation sample data pairs; and
performing mutual representation processing of the representation vector pairs by the machine learning model, and performing the sentence vector representation training for the conversation data samples based on the result of the mutual representation processing in combination with other elements of claims 4 and 1 from which it depends.
Regarding claim 11, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). However, neither Romeo nor Chen, teaches or suggests the pre-trained language model comprises a first sub pre-trained model and a second sub pre-trained model; wherein the first sub pre-trained model is configured to perform training for the conversation structure according to token information corresponding to data samples of each turn of conversation; and the second sub pre-trained model is configured to perform training for the conversation structure according to turn information, role information, token information, and token position information corresponding to data samples of each turn of conversation; and
wherein the performing, according to the representation vector output by the trained pre- trained language model, topic representation training for the conversation data samples by the machine learning model comprises:
performing the topic representation training for the conversation data samples by the machine learning model according to a representation vector output by the trained second sub pre-trained model in combination with other elements of claims 4 and 1 from which it depends.
Regarding claim 12, the combination of Romeo and Chen teaches all the elements of claim 4 (see detailed element listing above). However, neither Romeo nor Chen, teaches or suggests the operation of performing training comprises:
performing discretization processing of the representation vector output by the trained machine learning model; and performing semantic analysis for the multiple turns of conversation according to the result of the discretization processing to obtain the semantic transfer relationships between the turns of conversation; or
the operation of performing training comprises:
performing auto-encoding processing based on the representation vector output by the trained machine learning model; and performing conversation task modeling according to the result of the auto-encoding processing and obtaining the semantic transfer relationships between the turns of conversation according to the result of the modeling in combination with other elements of claims 4 and 1 from which it depends.
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 PENNY L CAUDLE whose telephone number is (703)756-1432. The examiner can normally be reached M-Th 8:00 am to 5:00 pm eastern.
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, Daniel Washburn can be reached at 571-272-5551. 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.
/PENNY L CAUDLE/Examiner, Art Unit 2657
/DANIEL C WASHBURN/Supervisory Patent Examiner, Art Unit 2657