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 .
Claims 1, 2, 4-8, 10-16, and 21-26 are pending. This Office Action is responsive to the amendment filed on 06/11/2026.
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.
Claims 1-2, 5-8, 10-12, 14-16, and 21-25 are rejected under 35 U.S.C. 103 as being unpatentable over Fabbri et al. (“Improving Zero and Few-Shot Abstractive Summarization with Intermediate Fine-tuning and Data Augmentation”, published April 11, 2021), hereinafter Fabbri; in view of Lin et al. (“CSDS: A Fine-Grained Chinese Dataset for Customer Service Dialogue Summarization”, published 09/06/2021), hereinafter Lin. Fabbri was cited in the IDS filed 09/12/2024. Lin was cited as non-relied upon prior art in a previous Office Action dated 12/19/2025.
Regarding Claim 1, Fabbri teaches A system comprising a processor (Fabbri: “The experiments make use of V100 GPUs. We used up to 8 GPUs per experiment (depending on the experiment; sometimes a single GPU was used to run the maximum number of experiments in parallel).” [Section 8. Ethical Considerations]) to:
define a respective associated heuristic for unlabeled data, wherein the respective associated heuristic is configured to select at least one sentence based on a structural characteristic of the utterance to generate a weak summary (Fabbri: “Effect of summary sentence choice: The first M sentences of a given Wikipedia article were chosen as this introduction intuitively form a coherent summary of the article. We examine the effect of choosing the first sentences compared to choosing based on other criteria. As an alternative, we pick the sentences with the highest self-ROUGE (ROUGE score of a sentence when using all other sentences as the reference summary) in a greedy fashion (the equivalent of the IND-ORIG settings in Zhang et al. (2019)). As in Zhang et al. (2019), we use ROUGE-1 F1. The sentences chosen under this heuristic consistently corresponded to those which were longest, and the resulting summaries were hence longer. Thus, we also experimented with choosing important sentences by using ROUGE-1 Precision, IND-ORIG-P.” [Section 5.2. Effect of WikiTransfer Hyperparameters]);
generate weakly labeled data using the respective associated heuristic (Fabbri: “WikiTransfer Intermediate Fine-tuning: We propose a method for fine-tuning pretrained models using unsupervised Wikipedia data. We create dataset-specific unsupervised data for this intermediate fine-tuning, by making use of characteristics of the target dataset such as the average length of input documents, the average summary length, and the general bin of whether the summaries desired are very abstractive or very extractive, as discussed above. Assume that we want a summary of M sentences from source documents of N sentences on average, and that we know approximately how extractive the summaries are in the target dataset, as defined as the upper bound ROUGE (Lin, 2004) performance of an extractive model, the extractive oracle, on that dataset. We bin the level of extraction of the target summaries into extremely abstractive (ROUGE oracle 10-30), more abstractive (ROUGE oracle 20-30), more extractive (ROUGE oracle 30-50), and extremely extractive (ROUGE oracle 40-60). We then iterate the following procedure on all Wikipedia articles available in a Wikipedia dump: We remove the first M sentences from the Wikipedia article for use as a summary and the following N sentences for use as a source document.” [Section 3. Methods]);
inter-train a pre-trained model based on respective weakly labeled data to generate an inter-trained weak label-based model (Fabbri: “1) We introduce a novel method, called WikiTransfer, which creates pseudo-summaries with subaspects of the target dataset which can be used as unlabeled data for intermediate fine-tuning.” [Section 1. Introduction]; “We aim to build on recent work in pretrained models and improve zero-shot and few-shot summarization by encoding characteristics of the target summarization dataset in unsupervised, intermediate fine-tuning data.” [Section 1. Introduction]); and
fine-tune each inter-trained weak label-based model based on few-shot training data to generate a final model while maintaining accuracy of the final model (Fabbri: “2) We demonstrate the benefits of WikiTransfer in few-shot settings, and show additional improvements when applying WikiTransfer with data augmentation and a regularization term for training with potentially noisy augmented data.” [Section 1. Introduction]; “We introduced WikiTransfer, a novel and generalizable method for fine-tuning pretrained models on dataset-specific unsupervised data obtained from generic Wikipedia data. WikiTransfer models achieve state-of-the-art zero-shot abstractive summarization performance on the CNN-DailyMail dataset and generalize across three additional datasets. In few-shot settings, WikiTransfer models are robust to noise introduced through data augmentation and benefit from consistency loss on more abstractive datasets.” [Section 7. Conclusion]).
However, Fabbri fails to explicitly disclose steps to split unlabeled data into a plurality of groups corresponding to different perspectives; define a heuristic, generate weakly-labeled data, and train and fine-tune a model for each different perspective.
In the same field of endeavor, Lin teaches steps to split unlabeled data into a plurality of groups corresponding to different perspectives (Lin: “The input of the task is a dialogue with multiple turns. Each utterance is labeled with its speaker role (e.g., user or agent) and the specific user identity (e.g., customer or seller). The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary. Each summary should consist of several segments, and each segment represents the summary for a topic in the dialogue. Besides, models could use the annotated key utterance indexes as additional supervised signals during the training process.” [Section 4.1 Task Definition]);
define a heuristic, generate weakly-labeled data, and train and fine-tune a model for each different perspective (Lin: “Traditional summarization models treat the dialogue as a whole and do not specifically consider the structure of generated summaries. Thus we wonder whether these methods could generate summaries with correct topic structure as references. We analyze the overall summary results and consider each QA pair in the summary as a whole, judging whether each QA pair is contained by the results of different models.” [Section 6. Dataset Difficulties]; See [Table 3], which presents metric results for each summarization method/heuristic and evaluation metric applied to represent each of the overall summary, user summary, and agent summary independently from each other).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated steps to split unlabeled data into a plurality of groups corresponding to different perspectives; define a heuristic, generate weakly-labeled data, and train and fine-tune a model for each different perspective, as taught by Lin to the system of Fabbri because both of these systems are directed towards extractive and abstractive textual summarization based on defined scoring methods and heuristics. In making this combination and applying the training techniques of Fabbri to the role-based dialogue datasets and tasks of Lin, it would provide Fabbri access to “fine-grained annotations” in the form of role-oriented summaries, allowing for Fabbri’s trained models to be applied to reflect the differing viewpoints of the multiple speakers (Lin: [Section 1. Introduction]).
Regarding Claim 2, Fabbri and Lin teach the system of Claim 1, wherein the processor is to further: receive a conversation to summarize (Lin: “The input of the task is a dialogue with multiple turns. Each utterance is labeled with its speaker role (e.g., user or agent) and the specific user identity (e.g., customer or seller). The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary.” [Section 4.1 Task Definition]);
split the conversation into a second plurality of groups corresponding to the different perspectives (Lin: “The input of the task is a dialogue with multiple turns. Each utterance is labeled with its speaker role (e.g., user or agent) and the specific user identity (e.g., customer or seller). The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary.” [Section 4.1 Task Definition]); and
input each of the second plurality of groups into a respective final model for each different perspective and receive a generated conversation summary for each of the second plurality of groups from each respective final model (Lin: “The input of the task is a dialogue with multiple turns. Each utterance is labeled with its speaker role (e.g., user or agent) and the specific user identity (e.g., customer or seller). The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary.” [Section 4.1 Task Definition]).
Regarding Claims 11-12 and 16, they are method and computer program product claims corresponding to the system of Claims 1-2. Therefore they are rejected for the same reason as Claims 1-2 above.
Regarding Claim 5, Fabbri and Lin teach the system of Claim 1, wherein the weakly labeled data comprises a summary automatically generated using the respective associated heuristic (Fabbri: “Effect of summary sentence choice: The first M sentences of a given Wikipedia article were chosen as this introduction intuitively form a coherent summary of the article. We examine the effect of choosing the first sentences compared to choosing based on other criteria. As an alternative, we pick the sentences with the highest self-ROUGE (ROUGE score of a sentence when using all other sentences as the reference summary) in a greedy fashion (the equivalent of the IND-ORIG settings in Zhang et al. (2019)). As in Zhang et al. (2019), we use ROUGE-1 F1. The sentences chosen under this heuristic consistently corresponded to those which were longest, and the resulting summaries were hence longer. Thus, we also experimented with choosing important sentences by using ROUGE-1 Precision, IND-ORIG-P.” [Section 5.2. Effect of WikiTransfer Hyperparameters]).
Regarding Claim 6, Fabbri and Lin teach the system of Claim 1, wherein the pre-trained model comprises a pre-trained generative model (Fabbri: “In this work, we introduce a novel and generalizable method, called WikiTransfer, for fine-tuning pretrained models for summarization in an unsupervised, dataset-specific manner.” [Abstract]; “We also experiment with this setup using unlabeled examples from the target dataset with pseudo labels (for teacher forcing) generated by a model trained on the associated few-shot subset” [Section 3. Methods]).
Regarding Claim 7, Fabbri and Lin teach the system of Claim 1, wherein a respective associated heuristic for one of the plurality of groups is different from a respective associated heuristic for another of the plurality of groups (Lin: “In this section, we will introduce some widely-used extractive and abstractive summarization models on dialogue summarization. We also enhance some of the models using our special annotations. The extractive methods include: LONGEST: As the longer utterances in the dialogue may contain more useful information, we sort the utterances by their lengths and extract the top k longest utterances as the summary. Number k is decided by the maximum summary length limit. LexPageRank (Erkan and Radev, 2004): This method ranks dialogue utterances by PageRank algorithm and extracts utterances in order until the length of the summary reaches the limit...” [Section 4.2 Summarization Methods]).
Regarding Claim 8, Fabbri and Lin teach the system of Claim 1, wherein the plurality of groups each comprise a list of sentences associated with a particular perspective of the plurality of different perspectives (Lin: “The input of the task is a dialogue with multiple turns. Each utterance is labeled with its speaker role (e.g., user or agent) and the specific user identity (e.g., customer or seller). The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary.” [Section 4.1 Task Definition]).
Regarding Claim 10, Fabbri and Lin teach the system of Claim 1, wherein the weakly labeled data comprises dialog-summary pairs (Lin: “we employ Question Answer (QA) pairs as the annotation format since it is the basic granularity in a customer service dialogue. We ask annotators to divide the dialogue into segments based on the topic and summarize each segment into a QA pair.” [Section 1. Introduction]).
Regarding Claim 14, Fabbri and Lin teach the method of Claim 12, further comprising: concatenating, via the processor, the generated conversation summaries to generate a final multi-perspective summary (Lin: “Overall summary: The overall summary condenses the main information of the whole dialogue. We concatenated annotated QA pairs in sequence to obtain the overall summary. Each QA pair represents the summary for a single topic in the dialogue, and the order reflects the topic flow.” [Section 3.3. Summary Format]; See [Figure 1], in which the overall summary is a concatenation of the user and agent summaries); and
outputting, via the processor, the final multi-perspective summary (Lin: “The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary.” [Section 4.1 Task Definition]).
Regarding Claim 15, Fabbri and Lin teach the method of Claim 11, wherein inter-training the pre-trained model comprises masking a target utterance (Fabbri: “We use knowledge of how abstractive a dataset is as a type of summary style which an end-user would know ahead of time. We filter the Wikipedia data points so that only those which fall into the bin for a given dataset are used for fine-tuning. For datasets that are extremely abstractive, such examples may be hard to find, so we remove high-ROUGE sentences from the input until the desired ROUGE oracle score is reached.” [Section 3. Methods]).
Regarding Claim 21, Fabbri and Lin teach the system of Claim 1, wherein the processor is to further: identify an important part of an input with the fine-tuned inter-trained model for each perspective; and output the important part of the input as a summary (Fabbri: “Effect of summary sentence choice: The first M sentences of a given Wikipedia article were chosen as this introduction intuitively form a coherent summary of the article. We examine the effect of choosing the first sentences compared to choosing based on other criteria. As an alternative, we pick the sentences with the highest self-ROUGE (ROUGE score of a sentence when using all other sentences as the reference summary) in a greedy fashion (the equivalent of the IND-ORIG settings in Zhang et al. (2019)). As in Zhang et al. (2019), we use ROUGE-1 F1. The sentences chosen under this heuristic consistently corresponded to those which were longest, and the resulting summaries were hence longer. Thus, we also experimented with choosing important sentences by using ROUGE-1 Precision, IND-ORIG-P. The comparison of these methods is shown in Table 5. The choice of the summary sentence has a noticeable impact on performance. We hypothesize that the coherence lost in the summaries is especially important for the longer CNNDM summaries. Using important sentences other than the first sentence likely adds more diversity in the data, and finding a balance between coherence and output style is an interesting direction for additional work.” [Section 5.2. Effect of WikiTransfer Hyperparameters]).
Regarding Claim 22, Fabbri and Lin teach the system of Claim 1, wherein the processor is to further: generate an abstractive third-person summary with the fine-tuned inter-trained model for each perspective (Fabbri: “WikiTransfer fine-tunes pretrained models on pseudo-summaries, produced from generic Wikipedia data, which contain characteristics of the target dataset, such as the length and level of abstraction of the desired summaries. WikiTransfer models achieve state-of-the-art, zero-shot abstractive summarization performance on the CNN-DailyMail dataset and demonstrate the effectiveness of our approach on three additional diverse datasets.” [Abstract]; Lin: “we introduce a novel Chinese dataset for Customer Service Dialogue Summarization (CSDS). CSDS improves the abstractive summaries in two aspects: (1) In addition to the overall summary for the whole dialogue, role-oriented summaries are also provided to acquire different speakers’ viewpoints. (2) All the summaries sum up each topic separately, thus containing the topic-level structure of the dialogue.” [Abstract]).
Regarding Claim 23, Fabbri and Lin teach the system of Claim 1, wherein the few-shot training data includes conversation summaries from a plurality of perspectives wherein the few-shot training data includes conversation summaries from a plurality of perspectives (Fabbri: “We also experiment with this setup using unlabeled examples from the target dataset with pseudo labels (for teacher forcing) generated by a model trained on the associated few-shot subset, although this approach is very sensitive to the quality of the pseudo labels (see Appendix). We refer to the training setting of N supervised data points with consistency training as N-c.” [Section 3. Methods]; Lin: “Benefiting from the annotation, our labeled summaries are role-oriented (containing different speakers’ main viewpoints) and topic-based structural (summarizing different topics separately).” [Section 1. Introduction]).
Regarding Claim 24, Fabbri and Lin teach the system of Claim 1, generating a conversation perspective summary for each perspective with the final model for each perspective (Lin: “The task is to generate three different kinds of summaries as explained in Section 3.3, including the overall summary, user summary, and agent summary. Each summary should consist of several segments, and each segment represents the summary for a topic in the dialogue.” [Section 4.1 Task Definition]); and
using the conversation perspective summary for each perspective to output a final summary (Lin: “Overall summary: The overall summary condenses the main information of the whole dialogue. We concatenated annotated QA pairs in sequence to obtain the overall summary. Each QA pair represents the summary for a single topic in the dialogue, and the order reflects the topic flow.” [Section 3.3. Summary Format]).
Regarding Claim 25, Fabbri and Lin teach the system of Claim 24, wherein a concatenator receives the conversation perspective summaries and outputs a final summary (Lin: “Overall summary: The overall summary condenses the main information of the whole dialogue. We concatenated annotated QA pairs in sequence to obtain the overall summary. Each QA pair represents the summary for a single topic in the dialogue, and the order reflects the topic flow.” [Section 3.3. Summary Format]).
Claims 4, 13, and 26 are rejected under 35 U.S.C. 103 as being unpatentable over Fabbri in view of Lin, as applied to Claims 1, 2, 12, and 24 above, in further view of Zhu et al. (“A Hierarchical Network for Abstractive Meeting Summarization with Cross-Domain Pretraining”, published 09/20/2020). Zhu was cited in a previous Office Action dated 02/03/2025.
Regarding Claim 4, Fabbri and Lin teach the system of Claim 2. However, they fail to expressly disclose wherein the processor is to add a prefix of indirect speech clause to a generated conversation summary in response to detecting that a generated conversation summary does not begin with any prefix of indirect speech clause.
In the same field of endeavor, Zhu teaches wherein the processor is to add a prefix of indirect speech clause to a generated conversation summary in response to detecting that a generated conversation summary does not begin with any prefix of indirect speech clause (Zhu: “To incorporate the participants’ information, we integrate the speaker role component. In the experiments, each meeting participant has a distinct role, e.g., program manager, industrial designer. For each role, we train a vector to represent it as a fixed-length vector rp, 1 ≤ p ≤ P, where P is the number of roles... This vector is appended to the embedding of the speaker's turn” [Section 3.1.1 Role Vector], See [Table 1 - Summary from our model (23 sentences)] where each line of summary includes the role with a term of indirect speech).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated wherein the processor is to add a prefix of indirect speech clause to a generated conversation summary in response to detecting that a generated conversation summary does not begin with any prefix of indirect speech clause, as taught by Zhu to the system of Fabbri and Lin because both of these systems are directed towards training a model for speaker-based dialogue summarization. In making this combination and including a prefix before the summary of each segment of dialogue, it would allow the person reviewing the summary to know who is speaking by “incorporat[ing] the role of each speaker to encode different semantic styles and standpoints among participants” (Zhu: [Section 1. Introduction]).
Regarding Claim 13, it is a method that corresponds with the system of Claim 4. Therefore, it is rejected for the same reason as Claim 4 above.
Regarding Claim 26, Fabbri and Lin teach the system of Claim 24. However, they fail to expressly disclose post-processing the conversation perspective summary for each perspective.
In the same field of endeavor, Zhu teaches post-processing the conversation perspective summary for each perspective (Zhu: “To incorporate the participants’ information, we integrate the speaker role component. In the experiments, each meeting participant has a distinct role, e.g., program manager, industrial designer. For each role, we train a vector to represent it as a fixed-length vector rp, 1 ≤ p ≤ P, where P is the number of roles... This vector is appended to the embedding of the speaker's turn” [Section 3.1.1 Role Vector], See [Table 1 - Summary from our model (23 sentences)] where each line of summary includes the role with a term of indirect speech; In light of paragraph [0026] of the specification, which states “the generated conversation perspective summaries 206A, 206B, and 206C may be post-processed before concatenation by the concatenator 208. For example, in response to detecting that a conversation perspective summary does not begin with a prefix of indirect speech clause, a post-processing unit (not shown) may add a prefix of indirect speech clause to the conversation perspective summary”, BRI would support that “post-processing” encompasses appending a prefix of indirect speech clause to a summary).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the invention to have incorporated post-processing the conversation perspective summary for each perspective, as taught by Zhu to the system of Fabbri and Lin because both of these systems are directed towards training a model for speaker-based dialogue summarization. In making this combination and post-processing the conversation summaries for each perspective, it would allow the person reviewing the summary to know who is speaking by “incorporat[ing] the role of each speaker to encode different semantic styles and standpoints among participants” [Section 1. Introduction].
Response to Arguments
Examiner acknowledges the amendments made to Claims 1, 11, and 16.
Applicant’s arguments, filed 06/11/2026, with respect to the rejection of Claims 1, 2, 4-8, 10-16, and 21-26 under 35 U.S.C. § 101 have been fully considered and are persuasive. The rejection has been withdrawn.
Applicant’s arguments, filed 06/11/2026, with respect to the rejection of Claims 1, 2, 4-8, 10-16, and 21-26 under 35 U.S.C. § 103 have been fully considered and are found moot in light of the new grounds of rejection (see rejection above).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Feigenblat et al. (“TWEETSUMM – A Dialog Summarization Dataset for Customer Service”) discusses dialog summarization automation in customer service settings in which heuristics are applied to identify and discard erroneous or uninformative extractive summaries.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MEGAN E HWANG whose telephone number is (703)756-1377. The examiner can normally be reached Monday-Thursday 10:00AM-7:30PM 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, Jennifer Welch can be reached at (571) 272-7212. 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.
/M.E.H./Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights