Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) significantly more. The subject matter eligibility test for products and process is describe below for claim 1 in view of dependent claims.
Regarding claim 1:
Step 1: Is the claim to a process machine manufacture or composition of matter?
Yes – Claim 1 recites a method, which is a method that falls under the statutory categories.
Step 2A Prong 1: Does the claim recite an abstract idea, law of nature, or natural phenomenon?
Yes – The claim recites the following:
“[using the trained dialogue classifier to] determine a respective dialogue score for each dialogue of a set of dialogues;” - The limitations recites a mental process of determine a respective dialogue score for each dialogue of a set of dialogues (see MPEP 2106.04(a)(2)III).
“selecting a respective engagement content item of a particular dialogue of the set of dialogues based on the respective dialogue score for the particular dialogue;” - The limitations recites a mental process of selecting a respective engagement content item based on the dialogue score (see MPEP 2106.04(a)(2)III).
Step 2 Prong 2: Does the claim recite additional elements that integrate the judicial exception into a particular application? No –
The claim includes the additional element(s):
“A method comprising: using a set of large language model prompts to prompt a large language model to generate a set of completions; wherein each completion of the set of completions is generated by the large language model in response to a respective large language model prompt of the set of large language model prompts;”
The additional elements fall under “apply it” as using a generic computer to use a large language model to generate a set of completions when prompted. See Mere Instructions to Apply an Exemption (see MPEP 2106.05(f)).
“obtaining a set of engagement content items for a set of anchor content items from the set of completions;”
The additional elements fall under Insignificant Extra-Solution Activity as mere data gathering by obtaining a set of engagement content items. See MPEP 2106.5(g).
“training a dialogue classifier based on a set of dialogue examples to yield a trained dialogue classifier,”
The additional elements fall under “apply it” as using a generic computer to train a dialogue classifier. See MPEP 2106.05(f)).
“the set of dialogue examples comprising the set of anchor content items and the set of engagement content items;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g).
“using the trained dialogue classifier to [determine a respective dialogue score for each dialogue of a set of dialogues;]”
The additional elements fall under “apply it” as using a generic computer to use the trained dialogue classifier to determine a respective dialog score (see MPEP 2106.05(f)).
“wherein each dialogue of the set of dialogues comprises a respective anchor content item and a respective engagement content item associated with the respective anchor content item;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g).
“causing at least a portion of the respective engagement content item of the particular dialogue to be presented in a graphical user interface as a highlighted engagement content item.”
The additional elements fall under “apply it” as using a generic computer to preset a highlighted engagement content item (see MPEP 2106.05(f)).
Step 2B: Does the claim recite additional elements that amount to significantly more than the judicial exception?
No - The claim does not include additional elements that are sufficient to amount to a significantly more than the judicial exemption. As an order whole, the claim is directed to towards the mental process of determining a score for relevant content items for presentation. As discussed above with respect to integration of the abstract idea into a practical application, the additional elements of using an large language model, obtaining, training, using the trained dialogue classifier and causing the presentation of the content item fall under using generic computer to apply an exemption and mere data gathering. The method does not improve on the function of a computer, transforms an article into another article, nor is it applied by a particular machine, making the claim not patent eligible.
Regarding claim 2:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein: each large language model prompt of the set of large language model prompts instructs the large language model to generate a number of engagement content items for a respective anchor content item of the set of anchor content items;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“each large language model prompt of the set of large language model prompts comprises a respective set of engagement content item specifications;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“each engagement content item specification of the respective set of engagement content item specifications of each large language model prompt of the set of large language model prompts comprises a respective set of features of a respective engagement content item to be generated by the large language model.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 3:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 2, wherein, for each engagement content item specification of the respective set of engagement content item specifications of each large language model prompt of the set of large language model prompts, the respective set of features of the respective engagement content item to be generated by the large language model comprises one or more of: a length of the respective engagement content item to be generated by the large language model, a specification that the respective engagement content item to be generated by the large language model is to comment on a particular point made in the respective anchor content item, a specification that the respective engagement content item to be generated by the large language model is to comment on an overall topic of the respective anchor content item, a specification that the respective engagement content item to be generated by the large language model is to have a social interaction with an author of the respective anchor content item, an engagement content item depth type of the respective engagement content item to be generated by the large language model, or an engagement content item tone of the respective engagement content item to be generated by the large language model.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 4:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein: the trained dialogue classifier comprises a first trained bidirectional encoder representations from transformers model, a second trained bidirectional encoder representations from transformers model, and a trained fully connected layer;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises inputting the engagement content item of the particular dialogue into the first trained bidirectional encoder representations from transformers model, and inputting the anchor content item of the particular dialogue into the second trained bidirectional encoder representations from transformers model.”
The additional elements fall under “apply it” as using a generic computer to input the engagement content item of the particular dialogue into the first trained bidirectional encoder representations from transformers model, and input the anchor content item of the particular dialogue into the second trained bidirectional encoder representations from transformers model. (see MPEP 2106.05(f)).
Regarding claim 5:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein: the trained dialogue classifier comprises a trained bidirectional encoder representations from transformers model and a trained fully connected layer;”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
“wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained bidirectional encoder representations from transformers model.”
The additional elements fall under “apply it” as using a generic computer to inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained bidirectional encoder representations from transformers model. (see MPEP 2106.05(f)).
Regarding claim 6:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein the graphical user interface comprises a feed item, a notifications item, or an electronic mail message item; and wherein the engagement content item of the particular dialogue is presented as the highlighted engagement content item in the feed item, the notifications item, or the electronic mail message item.”
The additional elements fall under “apply it” as using a generic computer to include presenting the engagement content item in the feed item, the notifications item, or the electronic mail message item. (see MPEP 2106.05(f)). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Regarding claim 7:
Step 2A Prong 2, Step 2B: The additional element(s):
“The method of claim 1, wherein: each dialog example, of the set of dialogue examples, corresponds to an anchor content item of the set of anchor content items; each dialog example, of the set of dialogue examples, comprises an engagement content item, of the set of engagement content items, generated by the large language model for the anchor content item to which the dialog example corresponds; the set of dialog examples are associated with a set of labels; each label, of the set of labels, labels a respective dialog example of the set of dialog examples; and each label, of the set of labels, indicates whether the engagement content item of the respective dialog example is an insightful comment on the anchor content item to which the respective dialog example corresponds.”
The additional elements fall under Insignificant Extra-Solution Activity. See MPEP 2106.5(g). The judicial exemptions do not integrate into a practical application nor provide an improvement. The process does not provide an inventive concept nor provides a practical application.
Claims 8-14 recite a system and are analogous to the method of claims 1-7. Therefore, the rejections of claim 1-7 above applies to claims 8-14.
Claims 15-20 recite a computer readable medium product and are analogous to the method of claims 1-7. Therefore, the rejections of claim 1-7 above applies to claims 15-20.
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.
Claim(s) 1-3, 6, 7, 8-10, 13, 14-17, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Liu, Junling, et al. "LLMRec: Benchmarking Large Language Models on Recommendation Task.(2023)." arXiv preprint arXiv:2308.12241 (2023) (“Liu”) in view of Friedman, Luke, et al. "Leveraging large language models in conversational recommender systems." arXiv preprint arXiv:2305.07961 (2023). (“Friedman”) in view of Evans et al. (US11341748B2) (“Evans”).
Regarding claim 1 and analogous claims 8 and 15, Liu teaches a method comprising:
using a set of large language model prompts to prompt a large language model to generate a set of completions; wherein each completion of the set of completions is generated by the large language model in response to a respective large language model prompt of the set of large language model prompts (Liu page 3 Figure 1,
PNG
media_image1.png
778
1325
media_image1.png
Greyscale
[wherein each completion of the set of completions is generated by the large language model in response to a respective large language model prompt of the set of large language model prompts]
Page 3, Overall Architecture para 1 line 1-10, In this paper, we have designed an LLM-based recommender system called LLMRec in order to benchmark the performance of various LLM models on the aforementioned five tasks. The workflow of the proposed recommendation system is illustrated in Fig.1, which consists of three steps. Firstly, we generate task-specific prompts using task description, behavior injection, and format indicator modules. The task description module is utilized to adapt recommendation tasks to natural language processing tasks [using a set of large language model prompts to prompt a large language model to generate a set of completions;].));
obtaining a set of engagement content items for a set of anchor content items from the set of completions (Liu page 4,
PNG
media_image2.png
352
775
media_image2.png
Greyscale
[a set of anchor content items from the set of completions]
Page 4, Fine-Tuned LLM as Recommender para 1 line 12-21 In the inference process, sequential and direct recommendation tasks typically necessitate an item list as the target output. In that case, for sequential recommendation, beam search algorithms are employed to generate a list of potential next items, which is then evaluated under the all-item setting. In direct recommendation, recommended items are predicted from a candidate set. In this context, beam search is also utilized to decode a list of potential target items with the highest scores, after which evaluations are conducted [obtaining a set of engagement content items]);
However Liu does explicitly teach training a dialogue classifier based on a set of dialogue examples to yield a trained dialogue classifier, the set of dialogue examples comprising the set of anchor content items and the set of engagement content items;
using the trained dialogue classifier to determine a respective dialogue score for each dialogue of a set of dialogues;
wherein each dialogue of the set of dialogues comprises a respective anchor content item and a respective engagement content item associated with the respective anchor content item;
selecting a respective engagement content item of a particular dialogue of the set of dialogues based on the respective dialogue score for the particular dialogue;
and causing at least a portion of the respective engagement content item of the particular dialogue to be presented in a graphical user interface as a highlighted engagement content item.
However Friedman training a dialogue classifier based on a set of dialogue examples to yield a trained dialogue classifier, the set of dialogue examples comprising the set of anchor content items and the set of engagement content items (Friedman Page 2,
PNG
media_image3.png
653
1230
media_image3.png
Greyscale
Page 7 4 SIMULATION AND LARGE-SCALE TUNING
A major impediment to building a high-quality industrial CRS is a lack of data available for training and evaluation. Typically, largescale recommender systems are trained on user interaction data mined from the logs of existing products; … RecLLM deals with the data sparsity problem by exploiting the transfer learning ability of large language models using in-context few-shot learning or fine-tuning on a small number of manually generated examples. However, we hypothesize that ultimately there is a ceiling to the quality that can be achieved through these approaches, given the long-tail of different scenarios that can arise within a mixed-initiative CRS. In this section we discuss the use of LLM-powered user simulators to generate realistic data at scale and techniques for tuning system components using larger amounts of data.
Page 7. Figure. 8,
PNG
media_image4.png
492
584
media_image4.png
Greyscale
[the set of dialogue examples comprising the set of anchor content items and the set of engagement content items;]
Page 9 4.2 Tuning System Modules para 2 line 7-13, Given this data, we can tune a Generalized Dual Encoder Model (see Section 3.2.1), in which the initial context representation and item representations are each encoded by an LLM. Regardless of whether we choose to tune only the adapter layers of the two tower model or the LLM params as well, the loss is fully differentiable and normal supervised learning with gradient descent suffices [training a dialogue classifier based on a set of dialogue examples to yield a trained dialogue classifier,].
Page 9 4.2 Tuning System Modules para 4 line 4-11, In Section 3.2.2 we present an LLM based ranking module that jointly generates a score for each item and an explanation for that score. Using this data, we can tune the ranking LLM to predict the ground truth labels as a regression problem. Using only this relevancy data we cannot directly tune the LLM to generate better explanations, although this is still possible using bootstrapping methods that depend only on labels for the end task (in this case the scoring task) [35, 107].));
using the trained dialogue classifier to determine a respective dialogue score for each dialogue of a set of dialogues (Friedman page 2, A joint ranking / explanation module that uses an LLM to extract user preferences from an ongoing conversation and match them to textual artifacts synthesized from item metadata. As a byproduct of intermediate chain-of-thought reasoning [95], the LLM generates natural language justifications for each item shown to the user, increasing the transparency of the system.
PNG
media_image5.png
540
601
media_image5.png
Greyscale
[using the trained dialogue classifier to determine a respective dialogue score for each dialogue of a set of dialogues;]);
wherein each dialogue of the set of dialogues comprises a respective anchor content item and a respective engagement content item associated with the respective anchor content item;
selecting a respective engagement content item of a particular dialogue of the set of dialogues based on the respective dialogue score for the particular dialogue ((Friedman Figure
PNG
media_image6.png
544
583
media_image6.png
Greyscale
[selecting a respective engagement content item of a particular dialogue of the set of dialogues]
Page 6, Figure 6 gives a schematic for the LLM ranker. For each candidate item, the LLM jointly generates a score and a natural language explanation for the score1. These scores implicitly induce a ranking of the items. The first step is to create a text summarization of the item that fits into the context window of the LLM based on metadata associated with the item. In the case of a YouTube video recommender, this metadata consists of information such as the title, knowledge graph entities associated with the video, developer description of the video, transcript of the video, and user comments [based on the respective dialogue score for the particular dialogue].
PNG
media_image5.png
540
601
media_image5.png
Greyscale
[wherein each dialogue of the set of dialogues comprises a respective anchor content item and a respective engagement content item associated with the respective anchor content item;]);
Liu and Friedman are considered to be analogous to the claim invention because they are in the same field of recommending to users using large language models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Liu to incorporate the teachings of Friedman to train a large language model. Doing so to create conversational recommender system that can generate synthetic data and create a diverse LLM for recommendations (Friedman Abstract line 14-25, In particular, we propose new implementations for user preference understanding, flexible dialogue management and explainable recommendations as part of an integrated architecture powered by LLMs. For improved personalization, we describe how an LLM can consume interpretable natural language user profiles and use them to modulate session-level context. To overcome conversational data limitations in the absence of an existing production CRS, we propose techniques for building a controllable LLM-based user simulator to generate synthetic conversations. As a proof of concept we introduce RecLLM, a large-scale CRS for YouTube videos built on LaMDA, and demonstrate its fluency and diverse functionality through some illustrative example conversations).
Evens teaches and causing at least a portion of the respective engagement content item of the particular dialogue to be presented in a graphical user interface as a highlighted engagement content item (Evans Fig. 2
PNG
media_image7.png
472
553
media_image7.png
Greyscale
Col 8 line 38-51, FIG. 2 illustrates an example interface including a video and several associated highlights. In particular embodiments, the computing system may send, to a client system (e.g., of a user), information configured to render one or more highlights ( e.g., highlights corresponding to predicted and/or specified noteworthy portions). As an example and not by way of limitation, referencing FIG. 2, the computing system may send information configured to display several highlights (e.g., the highlight 235) within a highlight menu 230. In particular embodiments, these highlights may be sent automatically to the client system when a trigger event occurs. In particular embodiments, trigger event may be an event that indicates an interest in an associated video by the user of the client system [to be presented in a graphical user interface as a highlighted engagement content item].).
Liu and Evans are considered to be analogous to the claim invention because they are in the same field of machine learning. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Liu to incorporate the teachings of Evans to highlight content that would be interesting to a user. Doing so to provide users with content that might pique their interest (Evan Col 53- 63, The disclosure contemplates predicting portions of a video that are noteworthy and presenting these portions as highlights that may 55 be quickly shared by users, for example, on an online social network. This sharing may allow other users to discover the video or pique their interest in viewing related videos. It may also increase user engagement with the highlight or the video from which the highlight was extracted. For example, social connections of a user may comment on a post by the user that includes a highlight, or may otherwise engage with video as a result of the highlight being shared).
Regarding claim 2 and analogous claims 9 and 16, Liu in view of Friedman and Evans teach the method of claim 1 and analogous claims 8 and 15.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Liu further teaches , wherein: each large language model prompt of the set of large language model prompts instructs the large language model to generate a number of engagement content items for a respective anchor content item of the set of anchor content items;
each large language model prompt of the set of large language model prompts comprises a respective set of engagement content item specifications;
and each engagement content item specification of the respective set of engagement content item specifications of each large language model prompt of the set of large language model prompts comprises a respective set of features of a respective engagement content item to be generated by the large language model (Liu page 4,
PNG
media_image2.png
352
775
media_image2.png
Greyscale
[each large language model prompt of the set of large language model prompts instructs the large language model to generate a number of engagement content items for a respective anchor content item of the set of anchor content items]
Page 4, Fine-Tuned LLM as Recommender para 1 line 12-21 In the inference process, sequential and direct recommendation tasks typically necessitate an item list as the target output. In that case, for sequential recommendation, beam search algorithms are employed to generate a list of potential next items, which is then evaluated under the all-item setting. In direct recommendation, recommended items are predicted from a candidate set. In this context, beam search is also utilized to decode a list of potential target items with the highest scores, after which evaluations are conducted [each large language model prompt of the set of large language model prompts comprises a respective set of engagement content item specifications; ] (i.e. each prompt has a specific engagement content item).
Page 11 Appendix.1 Dataset Statistics and Splits para 1,
Dataset Splits. Following P5 (Geng et al. 2022), we adopt different data partitioning strategies for different tasks. Specifically, for rating, explanation, and review tasks, we divide the Amazon Beauty dataset into training, validation, and testing sets using a ratio of 8:1:1. We ensure that each user and item has at least one instance included in the training set. To obtain ground-truth explanations, Sentires toolkit 1 is first utilized to extract item feature words from the reviews and then sentences that comment on one or multiple item features are regarded as explanations of user preferences. For the sequential recommendation task, we employ a leave-one-out strategy to split the dataset: for each interaction sequence, the last item is treated as the test data, the item before the last one as the validation data, and the remaining data for training. In the direct recommendation task, the training set is consistent with the training split of the sequential recommendation task to avoid data leakage issues during pretraining [a respective set of features of a respective engagement content item to be generated by the large language model].).
Regarding claim 3 and analogous claims 10 and 17, Liu in view of Friedman and Evans teach the method of claim 2 and analogous claims 9 and 16.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Friedman teaches wherein, for each engagement content item specification of the respective set of engagement content item specifications of each large language model prompt of the set of large language model prompts, the respective set of features of the respective engagement content item to be generated by the large language model comprises one or more of:
a length of the respective engagement content item to be generated by the large language model,
a specification that the respective engagement content item to be generated by the large language model is to comment on a particular point made in the respective anchor content item,
a specification that the respective engagement content item to be generated by the large language model is to comment on an overall topic of the respective anchor content item,
a specification that the respective engagement content item to be generated by the large language model is to have a social interaction with an author of the respective anchor content item,
an engagement content item depth type of the respective engagement content item to be generated by the large language model, or
an engagement content item tone of the respective engagement content item to be generated by the large language model (Friedman page 3,
PNG
media_image8.png
295
999
media_image8.png
Greyscale
page 8, Generating Synthetic Training Data. To use a user simulator to generate data for supervised training of one of the CRS system modules an additional property is needed: ground truth labels that the system can learn from. As a toy example, suppose we are trying to learn a sentiment classifier as part of a traditional dialogue state tracking module. For this we need to generate a set of examples 𝑆𝑖 , 𝑙𝑖 , where 𝑆𝑖 is a session 𝑠1,𝑢1, 𝑠2,𝑢2, ...𝑠𝑛,𝑢𝑛 and 𝑙𝑖 is a ground truth label for the primary user sentiment within 𝑆𝑖 coming from a set of possible labels 𝐿, e.g {angry, satisfied, confused,. We can use controlled user simulation to solve this problem, by defining a session level variable 𝑣 over this set of labels 𝐿. First we sample a variable 𝑣 from 𝐿 (e.g. "angry") and then condition the simulator based on this label, for instance in a priming implementation by appending the message "You are an angry user" to the beginning of the input of the simulator. If we are able to solve this LLM control problem effectively then we can attach a label 𝑙𝑖 ="angry" to the session 𝑆𝑖 and trust that with high probability it will be accurate
[a specification that the respective engagement content item to be generated by the large language model is to have a social interaction with an author of the respective anchor content item,]).
Regarding claim 6 and analogous claims 13 and 20, Liu in view of Friedman and Evans teach the method of claim 1 and analogous claims 8 and 15.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Evans teaches wherein the graphical user interface comprises a feed item, a notifications item, or an electronic mail message item;
and wherein the engagement content item of the particular dialogue is presented as the highlighted engagement content item in the feed item, the notifications item, or the electronic mail message item (Evans FIG. 2,
PNG
media_image9.png
916
635
media_image9.png
Greyscale
[graphical user interface comprises a feed item]
Col 6 line 30-41,
FIG. 1 by, for example, the reaction 130 which may have been submitted by a user during the portion 125) at a time point that is within a time range encompassing exciting portions of a football video, such as the portion 125 during which a touchdown occurred. In this example, the computing system may predict that portions of the video where there are peaks in the number of "wow" reactions ( or some combination of other suitable reactions) are noteworthy. As another example and not by way of limitation, referencing FIG. 1, a large number of comments may be submitted during a noteworthy portion, such as the comment 140.
(including the text "touchdown!")
Col 8 line 38-47, FIG. 2 illustrates an example interface including a video and several associated highlights. In particular embodiments, the computing system may send, to a client system (e.g., of a user), information configured to render one or more highlights ( e.g., highlights corresponding to predicted and/or specified noteworthy portions). As an example and not by way of limitation, referencing FIG. 2, the computing system may send information configured to display several highlights (e.g., the highlight 235) within a highlight menu 230 [and wherein the engagement content item of the particular dialogue is presented as the highlighted engagement content item in the feed item]).
Regarding claim 7 and analogous claim 14, Liu in view of Friedman and Evans teach the method of claim 1 and analogous claims 8 and 15.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Friedman each dialog example, of the set of dialogue examples, corresponds to an anchor content item of the set of anchor content items;
each dialog example, of the set of dialogue examples, comprises an engagement content item, of the set of engagement content items, generated by the large language model for the anchor content item to which the dialog example corresponds (Friedman page 2, Figure 2,
PNG
media_image10.png
485
478
media_image10.png
Greyscale
Page 7 Figure 8,
PNG
media_image11.png
412
500
media_image11.png
Greyscale
[each dialog example, of the set of dialogue examples, corresponds to an anchor content item of the set of anchor content items;]
4.1 User Simulation,
According to the conversational recommender setup considered in this paper (see Section 2), a session consists of a sequence 𝑆 = {𝑠1,𝑢1, 𝑠2,𝑢2, ..., 𝑠𝑛,𝑢𝑛}, where each 𝑢𝑖 is a natural language utterance by the user and each 𝑠𝑖 is a combination of a natural language utterance and possibly a slate of recommendations by the CRS. Therefore, a user simulator is defined by a function 𝑓 (𝑆′) = 𝑈𝑖 , where 𝑆′ = {𝑠1,𝑢1, 𝑠2,𝑢2, ..., 𝑠𝑖 } is a partial session and 𝑈𝑖 is a distribution over possible user utterances 𝑢𝑖 continuing the session. Given a fixed CRS and such a user simulator 𝑓 , we can generate a new sample session by having the CRS and 𝑓 interact for a given number of turns (i.e. the CRS generates each 𝑠𝑖 and 𝑓 generates each 𝑢𝑖 ) [each dialog example, of the set of dialogue examples, comprises an engagement content item, of the set of engagement content items, generated by the large language model for the anchor content item to which the dialog example corresponds;]);
the set of dialog examples are associated with a set of labels;
each label, of the set of labels, labels a respective dialog example of the set of dialog examples;
and each label, of the set of labels, indicates whether the engagement content item of the respective dialog example is an insightful comment on the anchor content item to which the respective dialog example corresponds ((Friedman Figure 8,
PNG
media_image12.png
407
482
media_image12.png
Greyscale
(i.e. the system response to the users graduate and learns if the response was insightful ))
page 8, Generating Synthetic Training Data. To use a user simulator to generate data for supervised training of one of the CRS system modules an additional property is needed: ground truth labels that the system can learn from. As a toy example, suppose we are trying to learn a sentiment classifier as part of a traditional dialogue state tracking module. For this we need to generate a set of examples 𝑆𝑖 , 𝑙𝑖 , where 𝑆𝑖 is a session 𝑠1,𝑢1, 𝑠2,𝑢2, ...𝑠𝑛,𝑢𝑛 and 𝑙𝑖 is a ground truth label for the primary user sentiment within 𝑆𝑖 coming from a set of possible labels 𝐿, e.g {angry, satisfied, confused, ...} [the set of dialog examples are associated with a set of labels;]. We can use controlled user simulation to solve this problem, by defining a session level variable 𝑣 over this set of labels 𝐿. First we sample a variable 𝑣 from 𝐿 (e.g. "angry") and then condition the simulator based on this label, for instance in a priming implementation by appending the message "You are an angry user" to the beginning of the input of the simulator [each label, of the set of labels, labels a respective dialog example of the set of dialog examples;]. If we are able to solve this LLM control problem effectively then we can attach a label 𝑙𝑖 ="angry" to the session 𝑆𝑖 and trust that with high probability it will be accurate [and each label, of the set of labels, indicates whether the engagement content item of the respective dialog example is an insightful comment on the anchor content item to which the respective dialog example corresponds]).
Claim(s) 4, 11, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Liu in view of Friedman and Evans and further in view of Lu, Wenhao, Jian Jiao, and Ruofei Zhang. "Twinbert: Distilling knowledge to twin-structured compressed bert models for large-scale retrieval." Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2020 (“Lu”).
Regarding claim 4 and analogous claims 11 and 18, Liu in view of Friedman and Evans teach the method of claim 1 and analogous claims 8 and 15.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Liu does not teach the trained dialogue classifier comprises a first trained [bidirectional] encoder representations from transformers model, a second trained bidirectional encoder representations from transformers model, and a trained fully connected layer;
and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises inputting the engagement content item of the particular dialogue into the first trained bidirectional encoder representations from transformers model, and inputting the anchor content item of the particular dialogue into the second trained [bidirectional] encoder representations from transformers model.
Friedman teaches wherein:
the trained dialogue classifier comprises a first trained [bidirectional] encoder representations from transformers model, a second trained [bidirectional] encoder representations from transformers model, and a trained fully connected layer;
and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises inputting the engagement content item of the particular dialogue into the first trained [bidirectional] encoder representations from transformers model, and inputting the anchor content item of the particular dialogue into the second trained [bidirectional] encoder representations from transformers model (Friedman page 5 3.2.1 Retrieval para 1 line1-5, Generalized Dual Encoder Model. A popular solution to retrieval in traditional deep learning based recommenders is to use a dual encoder model consisting of two neural net towers, one to encode the context and one to encode the items (see e.g [102] and Figure10a).
page6, 3.2.2 Ranking/Explanations, para 1 line 1-5, After candidate items have been retrieved, a ranker decides which of them will be included in the recommendation slate and in what order. Unlike the retrieval module, the ranking module does not need to perform tractable search over a large corpus and is therefore less constrained in the types of computation that are possible. In a traditional recommender system, this usually manifests in the ranker crossing context and item features (instead of processing them in separate towers as is done in a dual encoder) and potentially using custom ranking losses during training that directly compare candidate items [10]. In the case of RecLLM, we take advantage of this extra room for computation to use an LLM that reasons sequentially about how well an item matches the context and generates a rationalization for its decision as a byproduct.
Figure 6,
PNG
media_image13.png
456
500
media_image13.png
Greyscale
[wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained bidirectional encoder representations from transformers model].
page 10 Figure 10 (a),
PNG
media_image14.png
272
344
media_image14.png
Greyscale
[: a first trained [bidirectional] encoder representations from transformers model, a second trained [bidirectional] encoder representations from transformers model, and a trained fully connected layer]).
Lu teaches [the trained dialogue classifier comprises a first trained] bidirectional [encoder representations from transformers model, a second trained] bidirectional [encoder representations from transformers model, and a trained fully connected layer;]
[and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises inputting the engagement content item of the particular dialogue into the first trained] bidirectional [encoder representations from transformers model, and inputting the anchor content item of the particular dialogue into the second trained] bidirectional encoder representations from transformers model] (Lu Page 2647 Figure 1,
PNG
media_image15.png
399
283
media_image15.png
Greyscale
Page 2647, 4.1 Model Architecture
As shown in Figure 1, the architecture of TwinBERT consists of two multi-layer transformer encoders and a crossing layer to combine the vector outputs of encoders and produce the final output. It is noteworthy that the parameters of the two encoders of query and keyword could be shared or different. The detailed comparison of the two styles is discussed in Section 5. Similar to BERT model architecture, at the bottom of each encoder is the embedding layer, where the query and keyword sentences are represented separately as embeddings and then fed into corresponding encoders (i.e. a first and second bidirectional encoders as TwinBERT).).
Liu, Friedman and Lu are considered to be analogous to the claim invention because they are in the same field of use of Large Language Models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Liu in view of Friedman to incorporate the teachings of Lu to use a bidirectional encoders. Doing to use a twin bidirectional encoders to process input separately and achieve close on on-par performance to BERT-Based models (Lu Abstract line 5-21, To address the problem, we present Twin-BERT model, which has two improvements: 1) represent query and document separately using twin-structured encoders and 2) each encoder is a highly compressed BERT-like model with less than one third of the parameters. The former allows document embeddings to be pre-computed offline and cached in memory, which is different from BERT, where the two input sentences are concatenated and encoded together. The change saves large amount of computation time, however, it is still not sufficient for real-time retrieval considering the complexity of BERT model itself. To further reduce computational cost, a compressed multi-layer transformer encoder is proposed with special training strategies as a substitution of the original complex BERT encoder. Lastly, two versions of TwinBERT are developed to combine the query and keyword embeddings for retrieval and relevance tasks correspondingly. Both of them have met the real-time latency requirement and achieve close or on-par performance to BERT-Base model.).
Claim(s) 5, 12, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Liu in view of Friedman and Evans and further in view of C. Channarong, C. Paosirikul, S. Maneeroj and A. Takasu, "HybridBERT4Rec: A Hybrid (Content-Based Filtering and Collaborative Filtering) Recommender System Based on BERT," in IEEE Access, vol. 10, pp. 56193-56206, 2022, (“Channarong”).
Regarding claim 5 and analogous claims 12 and 19, Liu in view of Friedman and Evans teach the method of claim 1 and analogous claims 8 and 15.
Liu, Friedman and Evans are combine in the same rational as set forth above with respect to claim 1 and analogous claims 8 and 15.
Liu does not teach wherein: the trained dialogue classifier comprises a trained bidirectional encoder representations from transformers model and a trained fully connected layer; and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained bidirectional encoder representations from transformers model.
Friedman teaches the trained dialogue classifier [comprises a trained bidirectional encoder representations from transformers model and a trained fully connected layer]; and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained [bidirectional] encoder [representations from transformers model].
(page6, 3.2.2 Ranking / Explanations, para 1, After candidate items have been retrieved, a ranker decides which of them will be included in the recommendation slate and in what order. Unlike the retrieval module, the ranking module does not need to perform tractable search over a large corpus and is therefore less constrained in the types of computation that are possible. In a traditional recommender system, this usually manifests in the ranker crossing context and item features (instead of processing them in separate towers as is done in a dual encoder) and potentially using custom ranking losses during training that directly compare candidate items [10]. In the case of RecLLM, we take advantage of this extra room for computation to use an LLM that reasons sequentially about how well an item matches the context and generates a rationalization for its decision as a byproduct.
Figure 6,
PNG
media_image13.png
456
500
media_image13.png
Greyscale
[wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained [bidirectional] encoder [representations from transformers model].
page 10 Figure 10 (c),
PNG
media_image16.png
301
228
media_image16.png
Greyscale
(i.e. the structure of the LLM) [wherein: the trained dialogue classifier;].
Channarong teaches [trained dialogue classifier] comprises a trained bidirectional encoder representations from transformers model and a trained fully connected layer; [and wherein using the trained dialogue classifier to determine the respective dialogue score for the particular dialogue comprises separately inputting both the engagement content item of the particular dialogue and the anchor content item of the particular dialogue into the trained] bidirectional [encoder] representations from transformers model (Channarong page 56197
PNG
media_image17.png
813
1108
media_image17.png
Greyscale
[comprises a trained bidirectional encoder representations from transformers model and a trained fully connected layer]
Page 56198, 2) INJECTING POSITIONAL EMBEDDINGS INTO THE MASKED RATERS' SEQUENCE
After randomly masking the input in the user sequence, we extract the user representation in the masked user sequence (Sm v ) by feeding it into the embedding layer. Because the encoder of the transformer in our model involves neither recurrence nor convolution, it cannot consider the order of the input sequence. To address this issue, we must incorporate the position of users in the input sequence. Therefore, the embedding layer of our model is the summation of the user embeddings and the positional embeddings (see Fig. 2). We call this user sequence embedding of the target item v and denote it as S’v [representations from transformers model].).
Liu, Friedman and Channarong are considered to be analogous to the claim invention because they are in the same field of Large Language Models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filling date of the claimed invention to have modified Liu in view of Friedman to incorporate the teachings of Channarong to include the use of bidirectional encoder. Doing so to consider user interactions and increase the model accuracy (Channarong Abstract line 7-15, We believe that if BERT were to consider other users' interactions in its analysis, it would increase the model accuracy. Therefore, we propose a new method called HybridBERT4Rec, which applies BERT to both CBF and collaborative _ltering (CF). For CBF, we want to extract the characteristics of the target user's interactions with purchased items. (We implement this in the same way as in BERT4Rec, with our model generating a target user pro_le.) For CF, we want to _nd neighboring users who are similar to the target user. Here, we extract the target item's characteristics using all other users who rated the target item as a second input to BERT. This generates a target item pro_le. After obtaining both pro_les, we use them to predict a rating score. We experimented with three datasets, _nding that our model was more accurate than the original BERT4Rec.).
Pertinent Prior Art
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Aberle (US20240062019A1) - teaches a method for generating text using AI language model based on specification submitted to the AI model.
Moon et al. (US11442992B1) - teaches an assistant system to assist the user in engaging with an only social network.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ALFREDO CAMPOS whose telephone number is (571)272-4504. The examiner can normally be reached 7:00 - 4:00 pm M - F.
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, Michael J. Huntley can be reached at (303) 297-4307. 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.
/ALFREDO CAMPOS/Examiner, Art Unit 2129
/SCHYLER S SANKS/Primary Examiner, Art Unit 2129
Prosecution Insights