# US 2026/0289611 A1 · System

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0289611 A1
- Document ID: US20260289611A1
- Status: Published application
- Kind: A1
- Application number: 19/560,282
- Filed: 2026-03-09
- Published: 2026-09-24
- Earliest claimed: 2025-03-19
- Assignees: SoftBank Group Corp.
- Applicants: SoftBank Group Corp.
- Inventors: Kaichi NAKANO
- Web page: https://patentgrep.com/patent/US20260289611A1

## CPC

- G06Q 30/0276: {Advertisement creation}
- G06F 40/186: Templates
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06F 40/30: Semantic analysis
- G06F 40/40: Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30)
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0475: Generative networks
- G06N 3/092: Reinforcement learning
- G06F 40/126: Character encoding
- G06F 40/163: Handling of whitespace
- G06F 40/232: Orthographic correction, e.g. spell checking or vowelisation
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A system includes a processor that is configured to provide an interface for inputting information, receive information input through the interface and store the received information in a database, generate, based on the stored information, a prompt sentence for instructing a generative AI model to generate a promotional text, input the generated prompt sentence into the generative AI model and automatically generate the promotional text by using the generative AI model, output the generated promotional text via a user interface, collect and analyze user reaction data in order to measure predetermined evaluation indicators, and apply a reinforcement learning algorithm based on measurement results of the evaluation indicators and adjust parameters of the generative AI model so as to improve quality of the promotional text.

## Claims

### Claim 1

A system comprising:
circuitry configured to:
receive, via a packet-switched network from a terminal device, structured input data comprising a plurality of parameter fields;
store the structured input data as indexed records in a storage area of a non-transitory storage medium;
retrieve the indexed records from the storage area and construct, according to a configurable composition rule set, a prompt data sequence embedding the plurality of parameter fields into a natural-language instruction format;
record the prompt data sequence in a prompt storage area in association with the indexed records;
transmit, via a communication interface coupled to the packet-switched network, a generation request comprising the prompt data sequence and a set of generation condition parameters to a generative neural network model;
receive, from the generative neural network model, output text data generated in response to the prompt data sequence;
store the output text data in a result storage area in association with the prompt data sequence and the indexed records;
transmit the output text data to the terminal device for rendering on a display;
collect, from at least one external analysis apparatus coupled to the packet-switched network, interaction event data indicating actions performed in response to delivery of the output text data;
compute, by performing aggregation processing on the interaction event data in association with a performance metric specification contained in the structured input data, a performance index value;
generate a reward signal based on at least the prompt data sequence, the output text data, and the performance index value; and
execute a reinforcement learning algorithm that receives the reward signal and updates at least one of the set of generation condition parameters and the configurable composition rule set, such that subsequent generation requests yield output text data associated with a higher performance index value.

### Claim 2 (depends on 1)

The system according to claim 1, wherein the generative neural network model comprises a transformer architecture including a plurality of self-attention layers, a plurality of feed-forward sublayers, and learned token embeddings, and wherein the generation condition parameters comprise at least a temperature scalar controlling a sampling probability distribution over a vocabulary of tokens.

### Claim 3 (depends on 2)

The system according to claim 2, wherein the circuitry is further configured to set, as part of the generation condition parameters, at least one of a top-k sampling threshold, a nucleus sampling probability threshold, and a maximum output token count, and wherein the generative neural network model iteratively selects output tokens conditioned on the prompt data sequence until an end-of-sequence condition or the maximum output token count is reached.

### Claim 4 (depends on 3)

The system according to claim 3, wherein the reinforcement learning algorithm comprises a policy network realized as a neural network, wherein inputs to the policy network comprise feature vectors derived from the structured input data and historical performance statistics, and wherein outputs of the policy network comprise at least selection probabilities for composition patterns of the prompt data sequence and suggested values for the generation condition parameters.

### Claim 5 (depends on 4)

The system according to claim 4, wherein the reward signal is computed by a reward function that receives at least a normalized ratio derived from the interaction event data and explicit evaluation information received from the terminal device, and wherein the output text data comprises promotional text data generated for distribution via at least one of an electronic message format and a digital advertisement format.

### Claim 6 (depends on 1)

The system according to claim 1, wherein the circuitry is further configured to convert the output text data into distribution data formatted for transmission via an electronic communication medium, embed tracking identification information and at least one of a tracking resource identifier and a tracking image into the distribution data, and transmit the distribution data to at least one of a message delivery apparatus and an advertisement delivery apparatus coupled to the packet-switched network.

### Claim 7 (depends on 6)

The system according to claim 6, wherein the interaction event data comprises event records each including at least an event type, a timestamp, and the tracking identification information, and wherein the event type indicates at least one of an opening operation, a selection operation, a transition operation, and an outcome occurrence.

### Claim 8 (depends on 7)

The system according to claim 7, wherein the circuitry computes the performance index value by dividing a count of event records having a first event type by a count of event records having a second event type, and wherein the performance metric specification identifies at least one of a click-through rate, a conversion rate, and an open rate.

### Claim 9 (depends on 8)

The system according to claim 8, wherein the circuitry stores the performance index value in the non-transitory storage medium in association with the prompt data sequence, the generation condition parameters, and the indexed records, and wherein the reinforcement learning algorithm retrieves a plurality of past prompt-text-metric tuples from the non-transitory storage medium as training data for gradient-based optimization of the policy network.

### Claim 10 (depends on 1)

The system according to claim 1, wherein the circuitry is further configured to perform normalization processing on the structured input data comprising at least one of character encoding unification, whitespace trimming, punctuation standardization, and numerical format canonicalization, and to store a normalized representation as the indexed records.

### Claim 11 (depends on 10)

The system according to claim 10, wherein the configurable composition rule set comprises a plurality of instruction templates each defining placeholder positions and fixed instruction text, and wherein constructing the prompt data sequence comprises selecting an instruction template from the plurality of instruction templates and inserting normalized parameter fields into the placeholder positions.

### Claim 12 (depends on 11)

The system according to claim 11, wherein the circuitry records, in the prompt storage area, a template identifier and a version identifier representing a state of the configurable composition rule set at a time of constructing the prompt data sequence, such that the circuitry can reproduce historical generation conditions for evaluation.

### Claim 13 (depends on 1)

The system according to claim 1, wherein the circuitry is further configured to perform post-processing on the output text data comprising at least one of deletion of extraneous prefix text, truncation at a token boundary based on a character count limit, correction of notation inconsistencies, and format adjustment for a target distribution channel.

### Claim 14 (depends on 13)

The system according to claim 13, wherein the circuitry receives a plurality of candidate output text items from the generative neural network model, performs evaluation processing on each candidate based on at least one of length, phrase composition, presence or absence of prohibited terms, and a quality evaluation score, and selects at least one candidate as the output text data based on a result of the evaluation processing.

### Claim 15 (depends on 1)

The system according to claim 1, wherein the circuitry is further configured to perform natural language processing and classification processing on at least a portion of the structured input data to compute emotion type information indicating an emotional state and confidence information associated with the emotion type information, and to store the emotion type information in the non-transitory storage medium in association with the indexed records.

### Claim 16 (depends on 15)

The system according to claim 15, wherein constructing the prompt data sequence further comprises embedding the emotion type information as an additional parameter field into the natural-language instruction format, such that the generative neural network model generates the output text data conditioned on the emotional state.

### Claim 17 (depends on 1)

The system according to claim 1, wherein the circuitry is further configured to receive, from the terminal device, an arbitrary prompt data sequence composed by a user, augment the arbitrary prompt data sequence with historical performance statistics retrieved from the non-transitory storage medium, provide the augmented prompt data sequence to the generative neural network model, and transmit a resulting improvement proposal to the terminal device for rendering on the display.

### Claim 18

A system comprising:
circuitry configured to:
receive, via a packet-switched network from a terminal device, structured input data comprising a plurality of parameter fields;
perform normalization processing on the structured input data comprising character encoding unification and numerical format canonicalization, and store normalized indexed records in a non-transitory storage medium;
select an instruction template from a plurality of stored instruction templates based on the normalized indexed records, and construct a prompt data sequence by inserting parameter fields from the normalized indexed records into placeholder positions of the selected instruction template;
transmit, via a communication interface, a generation request comprising the prompt data sequence and generation condition parameters including a temperature scalar and a maximum output token count to a generative neural network model comprising a transformer architecture with a plurality of self-attention layers;
receive a plurality of candidate output text items from the generative neural network model;
perform evaluation processing on each candidate output text item based on at least one of length, phrase composition, and a quality evaluation score, and select at least one candidate as output text data;
convert the output text data into distribution data, embed tracking identification information into the distribution data, and transmit the distribution data via the packet-switched network;
collect interaction event data from at least one external analysis apparatus by decoding the tracking identification information, and aggregate the interaction event data to compute a performance index value;
generate a reward signal from at least the performance index value and execute a reinforcement learning algorithm comprising a policy network to update at least one of a composition rule for the instruction template and the generation condition parameters; and
construct, using updated parameters from the policy network, a subsequent prompt data sequence predicted to yield a higher performance index value.

### Claim 19 (depends on 18)

The system according to claim 18, wherein the circuitry is further configured to receive, from the terminal device, explicit evaluation information comprising at least one of ratings and comparative selection results among the plurality of candidate output text items, and to incorporate the explicit evaluation information into the reward signal used by the reinforcement learning algorithm.

### Claim 20

A method comprising:
receiving, by circuitry via a packet-switched network from a terminal device, structured input data comprising a plurality of parameter fields;
storing the structured input data as indexed records in a storage area of a non-transitory storage medium;
constructing, according to a configurable composition rule set, a prompt data sequence embedding the plurality of parameter fields into a natural-language instruction format;
transmitting, via a communication interface coupled to the packet-switched network, a generation request comprising the prompt data sequence and a set of generation condition parameters to a generative neural network model;
receiving, from the generative neural network model, output text data generated in response to the prompt data sequence;
transmitting the output text data to the terminal device for rendering on a display;
collecting, from at least one external analysis apparatus coupled to the packet-switched network, interaction event data indicating actions performed in response to delivery of the output text data;
computing, by performing aggregation processing on the interaction event data, a performance index value;
generating a reward signal based on at least the prompt data sequence, the output text data, and the performance index value; and
executing a reinforcement learning algorithm that receives the reward signal and updates at least one of the set of generation condition parameters and the configurable composition rule set.
