Personalized responses to chatbot prompt based on embedding spaces between user and society

Patent metadata

Field	Value
Publication number	US-2025238630-A1
Application number	US-202418418871-A
Country	US
Kind code	A1
Filing date	Jan 22, 2024
Priority date	Jan 22, 2024
Publication date	Jul 24, 2025
Grant date	—

Abstract

Systems described herein may provide responses to chatbot prompts that correspond to both a user's preferences and accepted views of society. A chat recommendation server may receive a prompt from a user device. The chat recommendation server may determine a general Overton window and a user-specific Overton window associated with the prompt. The chat recommendation server may generate a plurality of candidate response using the first machine learning model, input the prompt and the plurality of candidate responses to a second machine learning model, and receive, as output from the second machine learning model, a polarization score for each of the plurality of candidate responses. Based on the polarization scores, a recommended response may be selected which minimizes a distance between the user-specific Overton window and the general Overton window. Accordingly, the recommended response may be displayed on the user device.

First claim

What is claimed is:

1 . A computer-implemented method comprising:
receiving, by a computing device and from a user device associated with a user, a prompt; determining, based on the prompt and using a first machine learning model, a general Overton window associated with the prompt; determining, based on the prompt and using the first machine learning model, a user-specific Overton window associated with the prompt; generating, using the first machine learning model and based on the user-specific Overton window, a plurality of candidate responses; inputting the prompt and the plurality of candidate responses to a second machine learning model; receiving, as output from the second machine learning model, a polarization score for each of the plurality of candidate responses; selecting, from the plurality of candidate responses and based on the polarization score associated with each of the plurality of candidate responses, a recommended response that minimizes a distance between the user-specific Overton window and the general Overton window; and providing, by the computing device and to the user device, the recommended response to the prompt.

2 . The computer-implemented method of claim 1 , wherein the general Overton window corresponds to a first cluster of responses in a first embedding space of the first machine learning model, wherein the first cluster of responses indicates acceptable views of society.

3 . The computer-implemented method of claim 1 , wherein the user-specific Overton window corresponds to a second cluster of responses in a second embedding space of the first machine learning model, wherein the second cluster of responses indicates preferences of the user.

4 . The computer-implemented method of claim 1 , wherein the first machine learning model comprises at least one or a foundation model or a large language model.

5 . The computer-implemented method of claim 1 , wherein the second machine learning model comprises a transformer model.

6 . The computer-implemented method of claim 1 , wherein selecting the recommended response comprises:
determining that the general Overton window and the user-specific Overton window overlap with each other; and selecting, from the plurality of candidate responses that are located within an overlapping portion of the general Overton window and the user-specific Overton window, the recommended response.

7 . The computer-implemented method of claim 1 , wherein selecting the recommended response comprises:
determining that the general Overton window and the user-specific Overton window do not overlap with each other; and selecting, from the plurality of candidate responses falling within the user-specific Overton window that minimizes the distance between the user-specific Overton window and the general Overton window, the recommended response.

8 . The computer-implemented method of claim 1 , further comprising:
training, based on historical chat information that represents one or more acceptable views of society, the first machine learning model to determine training general Overton windows, wherein the historical chat information comprises a plurality of prompt-response pairs from a plurality of users.

9 . The computer-implemented method of claim 1 , further comprising:
training, based on a chat history of the user, the first machine learning model to determine training user-specific Overton windows, wherein the chat history of the user comprises a plurality of prompt-response pairs from the user.

10 . The computer-implemented method of claim 1 , further comprising:
re-training, based on the polarization score generated by the second machine learning model, the first machine learning model.

11 . A system comprising:
one or more processors; and memory storing instructions that, when executed by the one or more processors, cause the system to:
receive, from a user device associated with a user, a prompt;
train, based on historical chat information from a plurality of users representative of acceptable views of society, a first machine learning model to determine training general Overton windows, wherein the historical chat information comprising a plurality of prompt-response pairs from the plurality of users;
determine, using the first machine learning model and based on the prompt, a general Overton window associated with the prompt;
determining, based on the prompt and using the first machine learning model, a user-specific Overton window associated with the prompt;
generate, using the first machine learning model and based on the user-specific Overton window, a plurality of candidate responses;
input the prompt and the plurality of candidate responses to a second machine learning model;
receive, as output from the second machine learning model, a polarization score for each of the plurality of candidate responses;
re-training, based on the polarization score generated by the second machine learning model, the first machine learning model;
select, from the plurality of candidate responses and based on the polarization score associated with each of the plurality of candidate responses, a recommended response that minimizes a distance between the user-specific Overton window and the general Overton window; and
cause display of the recommended response to the prompt on the user device.

12 . The system of claim 11 , wherein the general Overton window corresponds to a first cluster of responses in a first embedding space of the first machine learning model, wherein the first cluster of responses indicates acceptable views of society.

13 . The system of claim 11 , wherein the user-specific Overton window corresponds to a second cluster of responses in a second embedding space of the first machine learning model, wherein the second cluster of responses indicates preferences of the user.

14 . The system of claim 11 , wherein the instructions, when executed by the one or more processors, cause the system to selecting the recommended response by:
determining that the general Overton window and the user-specific Overton window overlap with each other; and selecting, from the plurality of candidate responses that are located within an overlapping portion of the general Overton window and the user-specific Overton window, the recommended response.

15 . The system of claim 11 , wherein the instructions, when executed by the one or more processors, cause the system to selecting the recommended response by:
determining that the general Overton window and the user-specific Overton window do not overlap with each other; and selecting, from the plurality of candidate responses falling within the user-specific Overton window that minimizes the distance between the user-specific Overton window and the general Overton window, the recommended response.

16 . The system of claim 11 , wherein the first machine learning model comprises at least one or a foundation model or a large language model, and the second machine learning model comprises a transformer model.

17 . One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause a computing device to:
receive from a user device of a user, a prompt; determine, based on the prompt and using a first machine learning model, a general Overton window associated with the prompt; determining, based on the prompt and using the first machine learning model, a user-specific Overton window associated with the prompt; generate, using the first machine learning model and b

Assignees

• Capital One Services Llc

Inventors

• Turner Taylor
• Rafferty Galen
• Sharpe Samuel
• Barr Brian
• Goodsitt Jeremy
• Reinert Owen
• Farnan Tyler

Classifications

• G06F40/35

  Discourse or dialogue representation · CPC title

• G06F40/30

  Semantic analysis · CPC title

• G06F40/40 Primary

  Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30) · CPC title