Granted Patent B2

US 12,488,224 · App. 18/484,733 · Granted Dec 2, 2025

Inventor: Ibrahim Badr (New York, NY)

Assignee: GOOGLE LLC

G06N3/0475G06N3/08

Abstract

Claims (61)

1 . A computing system for soft prompt tuning for proactive content generation, the system comprising:

one or more processors; and

one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:

obtaining input data, wherein the input data is descriptive of a particular user accessing a user interface;

obtaining, in response to obtaining the input data, a soft prompt associated with the particular user from a profile database, wherein the profile database stores a plurality of user-specific sets of parameters associated with a plurality of different users, wherein each user-specific set of parameters of the plurality of user-specific sets of parameters was tuned for user-specific content generation for a different respective user, wherein the soft prompt comprises a set of parameters, wherein the set of parameters comprise machine-learned weights;

obtaining search history data associated with the particular user, wherein the search history data is descriptive of a plurality of previous search queries associated with the particular user;

determining the search history data is associated with a particular topic;

generating a prompt input based on the particular topic;

processing the soft prompt and the prompt input with a machine-learned content generation model to generate a model-generated content item, wherein the model-generated content item is generated based on the set of parameters associated with the particular user, wherein the model-generated content item comprises model-generated fiction comprising one or more style attributes determined based on the soft prompt associated with the particular user, and wherein the machine-learned content generation model comprises a pre-trained generative model, wherein a model-generated content item style for the model-generated content item is determined based on the set of parameters of the soft prompt associated with the particular user, and wherein a topic of the model-generated content item is conditioned based on the prompt input;

providing the model-generated content item to the particular user via the user interface;

generating feedback data based on data retrieved from a user computing system via the user interface, wherein the feedback data is associated with one or more interactions with the model-generated content item; and

adjusting a subset of the set of parameters of the soft prompt associated with the particular user based on the feedback data, wherein adjusting the subset of the set of parameters tunes at least a subset of the machine-learned weights of the soft prompt to adjust style attribute conditioning.

2 . The system of claim 1 , wherein the operations further comprise:

obtaining browsing history data associated with the particular user, wherein the browsing history data is descriptive of a plurality of web resources viewed previously by the particular user; and

wherein the model-generated content item is generated based on the browsing history data.

3 . The system of claim 2 , wherein the operations further comprise:

determining the browsing history data is associated with the particular topic; and

wherein data descriptive of the particular topic and the set of parameters are processed with the machine-learned content generation model to generate the model-generated content item.

4 . The system of claim 2 , wherein the operations further comprise:

determining the browsing history data is associated with a particular content type; and

wherein data descriptive of the particular content type and the set of parameters are processed with the machine-learned content generation model to generate the model-generated content item, wherein the model-generated content item is the particular content type.

5 . The system of claim 1 , wherein the operations further comprise:

providing a feedback interface in the user interface with the model-generated content item, wherein the feedback interface comprises a user interface element for receiving feedback from the particular user.

6 . The system of claim 5 , wherein the operations further comprise:

obtaining a feedback input from a user computing system via the feedback interface, wherein the feedback input is descriptive of a user satisfaction metric; and

wherein the feedback data is generated based on the feedback input.

7 . The system of claim 1 , wherein the operations further comprise:

determining a viewing time associated with the model-generated content item being displayed via the user interface; and

wherein the feedback data is generated based on the viewing time.

8 . The system of claim 1 , wherein a plurality of parameters of the machine-learned content generation model are fixed during adjusting of the subset of the set of parameters, and wherein the operations further comprise:

storing the soft prompt comprising the set of parameters in a user profile database.

9 . A computer-implemented method for proactive content generation, the method comprising:

obtaining, by a computing system comprising one or more processors, input data, wherein the input data is descriptive of a particular user accessing a user interface;

obtaining, by the computing system and in response to obtaining the input data, a set of parameters associated with the particular user, wherein the set of parameters were trained based on interaction data associated with the particular user from a profile database that stores a plurality of user-specific sets of parameters associated with a plurality of different users, wherein each user-specific set of parameters of the plurality of user-specific sets of parameters was tuned for user-specific content generation for a different respective user, wherein the interaction data is descriptive of previous interactions by the particular user with previously generated content items, wherein the set of parameters comprise machine-learned weights, wherein the set of parameters including the machine-learned weights were fine-tuned to adjust style attribute conditioning based on user feedback associated with the particular user;

obtaining search history data associated with the particular user, wherein the search history data is descriptive of a plurality of previous search queries associated with the particular user;

determining the search history data is associated with a particular topic;

generating a prompt input based on the particular topic;

processing, by the computing system, the set of parameters and the prompt input with a machine-learned content generation model to generate a model-generated content item, wherein the model-generated content item is generated based on the set of parameters associated with the particular user, wherein the model-generated content item comprises model-generated fiction comprising one or more style attributes determined based on the set of parameters associated with the particular user, and wherein the machine-learned content generation model comprises a pre-trained generative model, wherein a model-generated content item style for the model-generated content item is determined based on the set of parameters associated with the particular user, and wherein a topic of the model-generated content item is conditioned based on the prompt input; and

providing, by the computing system, the model-generated content item to the particular user via the user interface.

10 . The method of claim 9 , further comprising:

obtaining, by the computing system, a manual prompt input, wherein the manual prompt input is descriptive of a request for a particular content type; and

wherein the manual prompt input, the prompt input, and the set of parameters are processed with the machine-learned content generation model to generate the model-generated content item of the particular content type.

11 . The method of claim 9 , wherein the particular content type comprises a poem.

12 . The method of claim 9 , wherein the particular content type comprises a joke.

13 . The method of claim 9 , wherein the particular topic is associated with a plot of a story.

14 . One or more non-transitory computer-readable media that collectively store instructions that, when executed by one or more computing devices, cause the one or more computing devices to perform operations, the operations comprising:

obtaining input data, wherein the input data is descriptive of a particular user accessing a user interface;

obtaining, in response to obtaining the input data, a soft prompt associated with the particular user from a profile database, wherein the profile database stores a plurality of user-specific sets of parameters associated with a plurality of different users, wherein each user-specific set of parameters of the plurality of user-specific sets of parameters was tuned for user-specific content generation for a different respective user, wherein the soft prompt comprises a set of parameters, wherein the set of parameters comprise machine-learned weights;

obtaining search history data associated with the particular user, wherein the search history data is descriptive of a plurality of previous search queries associated with the particular user;

determining the search history data is associated with a particular topic;

generating a prompt input based on the particular topic;

processing the set of parameters of the soft prompt and the prompt input with a machine-learned content generation model to generate a model-generated content item, wherein the model-generated content item is generated based on the set of parameters associated with the particular user, wherein the model-generated content item comprises model-generated fiction comprising one or more style attributes determined based on the soft prompt associated with the particular user, and wherein the machine-learned content generation model comprises a pre-trained generative model, wherein a model-generated content item style for the model-generated content item is determined based on the set of parameters associated with the particular user, and wherein a topic of the model-generated content item is conditioned based on the prompt input;

providing the model-generated content item to the particular user via the user interface;

generating feedback data based on data retrieved from a user computing system via the user interface, wherein the feedback data is associated with one or more interactions with the model-generated content item; and

adjusting a subset of the set of parameters of the soft prompt associated with the particular user based on the feedback data, wherein adjusting the subset of the set of parameters tunes at least a subset of the machine-learned weights of the soft prompt to adjust style attribute conditioning.

15 . The one or more non-transitory computer-readable media of claim 14 , wherein the machine-learned content generation model comprises a generative model trained to generate literary fiction.

16 . The one or more non-transitory computer-readable media of claim 14 , wherein the machine-learned content generation model comprises a language model trained on a plurality of different downstream tasks.

17 . The one or more non-transitory computer-readable media of claim 16 , wherein the operations further comprise:

obtaining preference data descriptive of a plurality of preferences associated with the particular user;

processing the preference data to generate a user-specific prompt; and

wherein the user-specific prompt and the set of parameters are processed with the machine-learned content generation model to generate the model-generated content item.

Assignments (1)

Continuity (1)

References Cited (11)